Federal Court of Australia

Epic Games, Inc v Google LLC [2025] FCA 901

ORDERS

THE COURT ORDERS THAT:

1. The further hearing of the proceeding be stood over to a date to be fixed.

2. Save for the oral summary of these reasons given by Beach J at the time of delivery of this judgment and any republication in any form of that summary, pursuant to s 37AF(1)  of the Federal Court of Australia Act 1976 (Cth) and on the ground set out in s 37AG(1)(a), subject to further order, disclosure and publication of these written reasons and the content thereof shall only be made to the parties’ external legal advisors in this proceeding and in proceedings NSD 1236 of 2020, VID 341 of 2022 and VID 342 of 2022, and to no other person.

3. Liberty to apply.

4. Costs reserved.

Note:    Entry of orders is dealt with in Rule 39.32 of the Federal Court Rules 2011.

REASONS FOR JUDGMENT

BEACH J:

1 There are four proceedings before me, the joint trial of which took place over a period of four months on questions of liability.

2 In the first proceeding, the Epic parties have made claims against the Apple parties principally concerning alleged contraventions of s 46 and other provisions of the Competition and Consumer Act 2010 (Cth) (CCA).  I have set out the relief that they seek elsewhere.

3 The Epic parties are involved in developing entertainment software for personal computers, smart mobile devices and gaming consoles, with their most popular game being Fortnite. They allege that the Apple parties’ contravening conduct has forced app developers including Epic to only use Apple’s App Store to distribute their software applications (apps) to the broad base of iOS device users, and to only use Apple’s payment system for processing the purchases of their in-app digital content by iOS device users.  Apple removed Fortnite from the App Store on 13 August 2020.  That step has provoked the present litigation against the Apple parties.

4 I have delivered a separate set of reasons dealing with the first proceeding (Epic Games, Inc v Apple Inc [2025] FCA 900).

5 In the second proceeding, which these present reasons deal with, the Epic parties have made claims against the Google parties principally also concerning analogous alleged contraventions of s 46 and other provisions of the CCA.  The relief that they seek is directed at gaining more access to Google’s digital ecosystem which does allow some access but is akin to a nature reserve protected by a large barbed-wire fence.

6 They allege that the Google parties’ contravening conduct has hindered the ability of app developers including Epic to distribute apps to Android devices in a realistic and practical way other than through Google’s Play Store.  It is said that Google has achieved this by imposing various contractual and technical restrictions, which stifle or block Android device users’ ability to download app stores and apps directly from developers’ websites and prevent or hinder competition in the distribution of apps to Android devices.  It is also said that Google has inappropriately imposed on app developers Google’s payment system for processing the purchases of in-app digital content by Android device users.  I will explain the detail of this in a moment.  Google removed Fortnite from the Play Store on 13 August 2020.  That step has also provoked the present litigation against the Google parties.

7 The third and fourth proceedings are representative proceedings commenced by the representative applicants under Part IVA of the Federal Court of Australia Act 1976 (Cth) where similar allegations have been made by them on behalf of group members against the Apple parties and the Google parties concerning the same contraventions of s 46 of the CCA as alleged in the first proceeding and the second proceeding.  One of the class actions is against the Apple parties concerning inter-alia the overcharging of commissions on the App Store.  The other class action is against the Google parties concerning inter-alia the overcharging of commissions on the Play Store.  In each case it is said that the contraventions caused app developers to pay materially higher commissions on paid app downloads and in-app purchases of digital content than they would otherwise have had to pay had the contravening conduct not occurred.

8 I have delivered a separate set of reasons dealing with both the third and fourth proceedings (Anthony v Apple Inc; McDonald v Google LLC ([2025] FCA 902).

9 I will say something more about the joint trial and how these reasons should be read a little later.  But for the moment let me introduce the case against the Google parties in some more detail.

10 Epic’s case is that Google has used its control of the Android ecosystem to stifle competition in app distribution and in payment solutions for Android apps.  It is said that as a result of Google’s conduct, Google’s Play Store is the default and predominant distributor of Android apps, and is not meaningfully subject to the forces of competition.  And it is said that in exchange for access to the Play Store, app developers are compelled to use Google’s payment solution, Google Play Billing, as their sole supplier of services for facilitating, including accepting, processing, collecting, and remitting, payments for in-app purchases of digital content.

11 Further, it is said that in order to secure, and to prevent any challenger from threatening, Google’s control of app distribution and in-app payment solutions within the Android ecosystem, Google imposes and enforces a set of contractual and related arrangements on app developers and on the original equipment manufacturers (OEMs) of smartphones and tablets.  

12 Epic’s case is that every OEM that wishes to sell smart mobile devices needs to pre-install an operating system (OS) for those devices.  Now the Android OS, which is owned by Google, is an operating system for mobile devices being smartphones and tablets, based on the open source Linux OS.

13 Google makes Android OS available for free under an open source licence being the Apache 2.0 licence via the Android Open Source Project (AOSP). So, Google makes the source code for that software publicly available for use and modification. At the time of trial there had been 14 iterations of the Android OS, with Android 15 in beta phase.  But apparently Android 15 was released in late 2024.  The Android OS is available to be modified by OEMs.

14 The Android OS is the foundation of what has been described as the Android ecosystem, which comprises the Android OS, the products and services that support the Android OS, as well as the participants in that ecosystem, being users, OEMs, developers, carriers and Google.

15 Now the development and success of Android as an OS depends upon its ability to attract OEMs to develop Android-compatible devices, its ability to attract developers to develop apps for Android and its ability to attract users to acquire Android devices and apps, each of which is related to the others.

16 The notion that Android is an ecosystem reflects the fact that there is interdependence between these participants.  This dynamic of interdependence is relevant to understanding the commercial realities of any competition or any competitive constraints faced by Google, Android and the Play Store.

17 Android connects OEMs who make Android devices, telecommunications carriers who supply mobile services on Android devices, app developers who develop apps for Android devices, and consumers who use Android devices and Android apps.  As I have said in my reasons in the Apple proceeding, multi-sided platforms involve indirect network effects which can create feedback loops across the platform.  These indirect network effects manifest strongly with respect to Android.  The more consumers that want to acquire Android devices, the greater the incentive for OEMs to make Android devices and for developers to develop apps for the Android OS. The better the range and quality of Android devices and Android apps, the more attractive Android devices become for consumers.

18 These network effects create economic challenges.  In particular, the continued success of Android depends upon it overcoming coordination problems and agency problems.  Now a coordination problem arises when the value that one participant derives from a platform is linked to the participation of other users on the platform.  Further, an agency problem arises in situations where participants are incentivised to act in ways that may harm other participants or the platform overall, such as some OEMs or developers engaging in malicious privacy or security practices that reduce consumers’ willingness to purchase Android devices.

19 Coordination and agency problems manifest in multi-sided platforms due to indirect network effects.

20 In the case of Android, the interests of users, OEMs, developers and carriers must be balanced so that each group is sufficiently incentivised to participate. In the absence of such balance, the Android ecosystem may be unable to attract and sustain sufficient engagement and investment from participants on all sides of the platform, being OEMs, developers, users and carriers.  Were that to occur, Android may cease to remain competitive.

21 In the case of Android, risks are mitigated, in part, by the contracts Google enters into with OEMs.  These contracts overcome coordination problems, and enhance the attractiveness of Android to both users and developers.  By comparison, Apple manages risks through vertical integration.

22 Now as I have said, Google owns the Android OS and makes its source code available to third parties free of charge under an open source licence.  I will refer to this version of Android as open source Android.  Open source Android does not include a suite of software apps, software development kits (SDKs) and application programming interfaces (APIs) that are owned by Google and known as Google Mobile Services (GMS).

23 A smart mobile device which does not have GMS installed is functionally limited, because most non-iOS apps are built using GMS APIs or SDKs supplied by Google and those apps will only run as intended on devices that have GMS installed. Further, GMS includes six core applications developed by Google that are among the most used apps in the world, being the Play Store, Google Search, Chrome, Google Maps, Gmail and YouTube.

24 In theory, an OEM can do without GMS. It can license an Android fork, which is a modified version of open source Android which does not comply with compatibility standards set by Google. It can also create its own mobile OS.  But few OEMs do either of these things and those that do end up with market shares that are miniscule.  Instead, the vast majority of non-iOS smart mobile devices distributed outside of China with a licensable mobile OS are pre-installed with a combination of open source Android and GMS.  For convenience, I will refer to that combination as GMS Android or Google Android, although I accept Google’s point that this is not supplied by it as a unitary product.

25 Now to obtain the right to distribute devices with GMS Android pre-installed, OEMs must enter into a mobile application distribution agreement (MADA) with Google.

26 The MADA grants OEMs a license to distribute their devices with GMS pre-installed, but on conditions that, first, certain apps nominated by Google, including the Play Store, must be pre-installed on those devices, second, those Google apps must be placed in specified prominent locations on the OEM’s device, third, the OEM must comply, and their devices must be tested for compliance, with numerous technical requirements set by Google, including restrictions on which apps are allowed to install other apps and, fourth, the OEM must ensure ongoing compliance with the Android Compatibility Commitment (ACC) as part of the Android compatibility program.

27 An ACC is a nominally separate contract between the OEM and Google. In general terms, it precludes OEMs from distributing devices that run on an Android fork. It operates together with a MADA so that, for so long as an OEM wants to distribute devices with GMS Android pre-installed, it cannot distribute any devices that run on an Android fork. This explains the dearth of non-Apple smart mobile devices distributed outside China that run on something other than GMS Android.

28 The effect of the anti-fragmentation agreements (AFAs) and the ACCs is that some OEMs have committed to certain core functionality and baseline standards, which ensures that developers and users have a consistent experience in developing for and using Android devices. That is because departures from or modifications to the Android OS source code may result in so-called “forked” devices, being devices on which apps developed for Android either do not function, or do not function correctly.

29 This is the problem of device fragmentation, which arises when there is incompatibility between devices notwithstanding that they notionally operate on the same operating system.

30 As their names suggest, the AFAs and ACCs address the risk of device fragmentation by promoting compatibility between Android devices. Devices that comply with the compatibility standards agreed in the AFAs and ACCs are often referred to as compatible Android devices and are so approved under what has been described as the Android compatibility program.  Google did not, and does not, charge OEMs any fee under the AFAs or ACCs.

31 Further, some OEMs have entered into MADAs with Google under which, as discussed above, Google licenses the OEMs to use GMS. That package of software includes familiar Google apps such as Google Maps, Chrome and YouTube. These are high quality apps which certain OEMs want to have on their devices. Google also licenses OEMs to use the “Android” and “Google” trademarks under the MADAs in relation to compatible Android devices. Outside of Europe and Turkey, Google does not charge OEMs a licensing or other fee under the MADAs.

32 Of course, many OEMs do want to develop, manufacture and sell Android devices that satisfy the AOSP compatibility standards, which use GMS and which can be marketed and sold using the Android trademark. Such devices may be referred to as Android devices although they might more accurately be described as GMS devices.

33 That OEMs who wish to develop, manufacture and sell Android devices must enter into a MADA with Google is unremarkable in circumstances where that is the means by which Google licenses OEMs to use Google’s proprietary software, being GMS, and intellectual property.

34 According to Epic, to ensure this situation never changes, Google has entered into incentive agreements with most OEMs, whereby they receive a portion of Google’s advertising revenues earned from the OEM’s devices. Such incentives are conditional upon OEMs complying with their MADAs and ACCs. Further, certain OEMs have entered into incentive arrangements which entitle them to a portion of Google’s Play Store revenues, provided the OEMs do not pre-install or promote a competing app store.  I will address what have been described as the Premier-tier revenue sharing agreements (RSA3s) and mobile incentive agreements (MIAs) later.

35 Let me now say something about the distribution of Android apps.

36 It is possible to distribute Android apps to Android device users via pre-installation on the device prior to its sale, or else via download to the device after sale, from the Play Store, a third-party app store that may come pre-installed on the device such as the Samsung Galaxy Store, or be directly downloaded onto the device, or via a website. Acquiring an app from a website is referred to as direct downloading or, in some cases, sideloading.

37 Google does not permit third-party app stores or any app that facilitates the distribution of another app to be distributed through the Play Store.

38 Pre-installation requires the agreement of the device’s OEM. Often the agreement of mobile telecommunications carriers is also necessary. Carriers supply devices to users subject to mobile plans, being so-called locked devices, and have a say in what apps are pre-installed on locked devices. Different carriers operate in different territories such that securing pre-installation is a complex endeavour.

39 Although direct downloading of apps and app stores is possible on Android devices, less than 3% of new apps downloaded to Android devices in Australia during 2020 were directly downloaded or downloaded from an app store that was itself directly downloaded. Globally, the relevant proportion was under 12%.  Three major challenges attend this distribution method.

40 The first challenge is one of discoverability, that is, of being found or noticed by users. If the app or app store is made available in the vast cyberspace of the internet, rather than within a dedicated and searchable app store that is already installed on users’ devices, it is far less likely to be found by users.

41 The second challenge is that several additional steps and warnings are injected into the sideloading process. These frictions stem from restrictions with which Google requires OEMs to comply as a condition of distributing devices with GMS installed. I will refer to these collectively as the technical restrictions.

42 The third challenge is that directly downloaded apps are functionally deficient, in that they generally do not update automatically. Instead, they must be updated manually, which requires the user to reinstall the app from the original source to obtain the updated version. Until Android 12 was released in October 2021, the same was generally true for apps sourced from a directly downloaded app store. It is only since October 2021 that Google has permitted directly downloaded app stores to update apps automatically.

43 Now it is said by Epic that within this realm, developers like Epic, who wish to distribute apps to the users of smart mobile devices with GMS installed, which for convenience I will refer to as Android devices or GMS devices, have been forced to operate.  And it is a realm in which the Play Store is pre-installed in a prominent location on almost every Android device, and in which about 97% in Australia and about 91% globally, excluding China, of new app downloads through Android app stores are undertaken via the Play Store. Its closest rival in Australia, being Samsung’s Galaxy Store, has a share of only about 1.5% of app downloads and its closest rival globally, being Oppo Store, has a share of only about 2.5%.

44 Now it is possible for developers to distribute their apps to Android device users by methods other than app stores. But Epic says that none of those methods are a substitute for distribution through an Android app store. And the only Android app store with a more than negligible market share is the Play Store.

45 Now developers do not need Google’s consent or permission to develop or distribute apps for Android. Developers have a range of alternatives for distributing apps on Android. These include reaching pre-installation or pre-loading agreements with OEMs or carriers, installing via a pre-installed third-party app store, downloading or sideloading apps or app stores from a website, via streaming or in the form of a web app. A web app is an application accessed through a web browser.  And developers can distribute through any of these means without entering into any agreement with Google.

46 But developers cannot distribute their apps through the Play Store unless they enter into a developer distribution agreement (DDA) with Google, the terms of which are essentially non-negotiable.

47 Under the DDA, developers are prohibited from distributing an app store or any app that distributes another app through the Play Store. Google claims this prohibition is necessary for security reasons.  The result is that rival Android app stores, and indeed any Android app that distributes another app, are denied access to the Play Store. They cannot be distributed via the most effective means of Android app distribution.

48 The DDA also provides that app developers that monetise their Play Store apps through in-app purchases of digital content must use Google’s own payment solution, Google Play Billing, to facilitate such payments. The parties have described this as the payments tie. When a user makes an in-app purchase of digital content using Google Play Billing, Google charges the developer a service fee, which is ordinarily 30% or 15% of the in-app purchase price.

49 Google requires that app developers use Google Play Billing as their payment solution for in-app purchases of digital content within apps downloaded from the Play Store.  There are some limited exceptions to this. Google does not require the use of Google Play Billing for in-app purchases of physical goods and services.  Other payment solutions including those provided by third-parties such as PayPal, Stripe, Adyen and Braintree are used to purchase physical goods or services within apps distributed through the Play Store.

50 Additionally, the DDA includes an anti-steering rule, which prohibits Play Store apps from leading users to a payment method besides Google Play Billing.  So, developers cannot include within their apps a link to a website where payment can be made.

51 The final set of agreements which are of present concern are known as the Games Velocity Program (GVP) agreements and Apps Velocity Program (AVP) agreements, whose genesis was Project Hug which I will describe in detail later.  These were made with certain large developers, which were described by Google as “beacons of the ecosystem” and many of whom Google considered to be at risk of distributing their Android apps outside the Play Store.

52 Now it is Google’s conduct in entering into and enforcing these various arrangements which is the subject of these proceedings.

53 Epic says that this conduct occurred in three relevant markets being, first, a market for the supply of mobile OS licenses to OEMs, second, a market for the supply of services for the distribution of Android apps and, third, a market for the supply of services for facilitating payments for the purchase of digital content within Android apps.

54 Now the posited mobile OS licensing market is a market for the supply of licenses for mobile operating systems, where the geographic dimension of the market is a global market excluding China, of which Australia forms part.

55 The posited Android mobile app distribution market is a market for the supply of services for the distribution of Android apps to Android smart mobile device users, with the following elements or dimensions.  It is said that the services consist of the provision of services to app developers enabling or facilitating app developers to reach, offer and provide Android smart mobile devices users with Android apps and associated updates and/or the provision of services to Android smart mobile devices users enabling or facilitating Android smart mobile devices users to be presented with or to find, obtain and utilise Android apps and associated updates.  And it is said that the geographic dimension of the market is Australia or a global market excluding China, of which Australia forms part.

56 The posited Android in-app payment solutions market is a market for the supply of services to app developers for accepting and processing payments for the purchase of digital content, including by way of subscriptions, within an Android app, with the following elements or dimensions.  It is said that the services consist of the provision of services to app developers enabling and/or facilitating app developers to accept and process payments for the purchase of digital content, including by way of subscriptions, within an Android app.  And it is said that the geographic dimension of the market is Australia or a global market, excluding China, of which Australia forms part.

57 Now there is an alternative posited to the Android mobile app distribution market.  This is the mobile app distribution market, which is said to be a market for the supply of services for the distribution of Android apps and native apps written for iOS to smart mobile device users, with the following elements or dimensions.  It is said that the services consist of the provision of services to app developers enabling or facilitating app developers to reach, offer and provide smart mobile device users with native apps and associated updates and/or the provision of services to smart mobile device users enabling or facilitating smart mobile device users to be presented with or to find, obtain and utilise native apps and associated updates.  And it is said that the geographic dimension of the market is Australia or a global market excluding China, of which Australia forms part.  I will put this alternative to one side for the moment.

58 Now it is said that Google has a substantial degree of power in each of these markets.

59 It is said that the purpose and effect of Google’s conduct in both the mobile OS licensing market and the Android mobile app distribution market has been to ensure that the Play Store remains the predominant supplier of Android app distribution services and to prevent or hinder the emergence and expansion of rival Android app stores. It is said that its conduct has substantially lessened competition in the Android mobile app distribution market.  It is said that in the absence of Google’s conduct, rival app stores would enter that market and compete with the Play Store, including by offering developers lower fees. And it is said that existing Android app stores such as the Amazon App Store and the Galaxy Store could compete more effectively.

60 Similarly, it is said that the purpose and effect of Google’s conduct in the Android in-app payment solutions market has been to ensure that Google remains the sole supplier of services for the facilitation of payments for the purchase of digital content within Play Store apps. It is said that its conduct has substantially lessened competition in that market.  It is said that in the absence of Google’s conduct, rival payment solutions providers would enter the Android in-app payment solutions market and compete with Google Play Billing.

61 In summary, Epic says that Google has contravened ss 46 and 47, alternatively, s 45 of the CCA on and from 6 November 2017 (the relevant period).

62 It is said that Google has contravened s 46 because it has a substantial degree of power in all three relevant markets and its conduct had the purpose, effect, or likely effect of substantially lessening competition in two of those markets.

63 It is said that Google has contravened s 47 by supplying Android app distribution services to developers via the Play Store on the condition that they not obtain services for the facilitation of in-app purchases of digital content in Play Store apps from anyone besides Google.  Alternatively, it is said that Google has contravened s 45 because it has made and given effect to contractual provisions that have the purpose, effect, or likely effect of substantially lessening competition.

64 Further, it is said that in all the circumstances, Google has engaged in unconscionable conduct contrary to s 21 of the Australian Consumer Law.

Summary of findings

65 In summary, I have made the following key findings.

66 First, I have found in favour of Epic concerning its three posited markets being, first, the mobile OS licensing market, second, the Android mobile app distribution market and, third, the Android in app payment solutions market.

67 And relevant to the question of substitution concerning the second posited market, in my view other Android app stores do provide a competitive constraint on the Play Store such as to be a substitute.  Further, there is some substitutability between sideloading of apps and Android app distribution services, but it is not a close constraint, and users and developers do not consider it to be a close substitute.  Further, the pre-installation of apps is not a close substitute for Android app distribution services.  Further, there is some competition with PC and console operators and mobile app stores, but there is not close substitution.  Further, out of app payments say little about the question of substitution for distribution services and in any event would not constrain the hypothetical monopolist of Android app distribution services.

68 Now I should note that the fact that I have found that various other products and services are not close constraints and therefore are not part of the relevant product dimension of market definition does not entail that the availability of such products and services do not provide some form of competitive constraint.  In my view they do provide a competitive constraint to some extent although not such as to deny the market definition that I have found or Google’s substantial market power in such a market.

69 Second, I have found in favour of Epic to the effect that at all relevant times Google has and had a substantial degree of power in each market.

70 Third, I have found that contrary to s 46 of the CCA, Google engaged in conduct in the latter two of the three posited markets that has had or is likely to have the effect of substantially lessening competition in such markets being conduct that prevents or prohibits developers and users from using alternative payment methods to Google Play Billing for purchasing digital in app content.

71 In terms of the Google Play Billing restrictions being the payments tie provision and the anti-steering provision, I have found that they were imposed or have been maintained at the start of the relevant period with the effect or likely effect of substantially lessening competition in the payment solutions market.

72 Fourth, I am satisfied that Google engaged in conduct concerning Project Hug, the GVP agreements, the AVP agreements and the RSA3s that had the purpose of substantially lessening competition in the Android mobile app distribution market and so contravened s 46.  Further, in relation to the conduct constituted by Project Hug and the GVP agreements, I am satisfied that such conduct had a likely effect of substantially lessening competition in that market; I am using “likely” here in terms of a real chance.

73 In all other respects I have not accepted Epic’s case against Google concerning the other alleged contraventions of Part IV of the CCA or its unconscionable conduct case.  And to be clear, I have not accepted Epic’s case concerning the technical restrictions or its case concerning clause 4.5 of the DDA.  Let me briefly make some observations on these last two matters.

74 As to the first matter in terms of the technical restrictions relating to install flows and frictions, I have concluded the following. First, such technical restrictions caused or imposed by Google are disproportionate to the security risks that Google has sought to protect against. Second and notwithstanding the first point, in my view such technical restrictions were not imposed by Google or sought to be maintained by Google for the purpose of substantially lessening competition. Third, such restrictions do not separately have any actual or likely effect of substantially lessening competition.  And fourth, I have also considered these restrictions in combination with other restrictive conduct that Epic has alleged, but in my view no anti-competitive purpose or effect has been shown by the combination.

75 Now as to the second matter in terms of clause 4.5 of the DDA and the ban on, in effect, an app store within an app store, I am not prepared to find that Google imposed this with the purpose or effect or likely effect of substantially lessening competition in the Android app distribution services market.

76 Now, it must be recalled that Google has permitted direct downloading or sideloading of other competing Android app stores.  So such competition was not substantially hindered by Google.

77 Further, the clause 4.5 ban is targeted if at all against competitors rather than competition.

78 Further, Google seems to have had various purposes for the clause 4.5 ban.  One purpose was to address security risk as Google perceived it.  Another purpose was to prevent free riders.  And an associated purpose was to protect the asset that it had invested in and built up. Clearly there was no anti-competitive purpose.  Moreover, given the other avenues for the direct downloading or sideloading of alternative Android app stores, in my view no actual or likely anti-competitive effect has been shown in the relevant markets.

79 Now I should say here that the position is a little more complicated concerning Apple as it precluded direct downloading or sideloading of alternative iOS app stores.

The joint trial and these reasons – structural aspects

80 Now as I have said, the joint trial before me concerned the concurrent hearing of four proceedings being Epic’s proceeding against Google, Epic’s proceeding against Apple and two class actions, the detail of which I have set out elsewhere.

81 Epic’s two proceedings had been issued out of the NSW Registry of this Court.  The two class actions had been subsequently issued out of the Victoria Registry of this Court.  The two class actions in terms of competition law liability questions were derivative of or sought to leverage off Epic’s cases and evidence on liability against Apple and Google, save for the question of the overcharging of commissions relating to paid app downloads or in-app digital purchases on the Play Store or Apple’s App Store.

82 Now in terms of the concurrent hearing of all four matters, I made orders that the evidence in one case should be treated as evidence in all other cases unless I indicated to the contrary, say, in respect of an admission made by one party that was not cross-admissible against other parties.

83 As I have said earlier, I have produced three sets of reasons being the present reasons, another set of reasons in the Epic v Apple proceeding and a third set of reasons dealing with and combining my analysis and findings in the two class actions.  Now a number of points.

84 First, the sequence in which the reasons should be read for ease of comprehension are my reasons in the Epic v Apple proceeding, the present reasons and then the reasons in the class actions.

85 Second and relatedly, it should also be understood that findings of fact or law in my reasons in the Epic v Apple proceeding form part of the foundation for my present reasons and then my reasons in the class actions, unless I indicate to the contrary.  Indeed, all sets of reasons should be read as a composite whole.  So, although I have indicated the sequence in which the reasons should be read, which in one sense can be seen as an incremental vertical structure, I also intend for my findings in this proceeding to be treated as findings in the Epic v Apple proceeding as well as findings in the class actions, and for the findings in the class actions to the extent relevant to be findings in this proceeding and the Epic v Apple proceeding.

86 Third, occasionally, my reasons in Epic v Apple and Epic v Google may overlap in relation to a particular topic.  This is partly a function of how the cases have been presented.  Nevertheless, my conclusions on such a topic in each proceeding have considered and been based upon all the evidence in all proceedings.

87 Fourth, some aspects of my discussion in the reasons in the class actions concerning the relevant counterfactual analysis should be taken as being incorporated by reference in these reasons. But I do accept that to some extent the dimensions to such an analysis are different.

88 In terms of assessing whether the relevant contraventions occurred, one is looking at and comparing the future from the time of the contravention with the relevant impugned conduct against the future from the time of the contravention without such conduct.  But of course in looking at the future with and without the relevant conduct you do not just simplistically hold all else equal.  But in terms of the class actions where one is considering causation questions involving what would have occurred if the relevant impugned conduct had not been engaged in, a different calculus and standard of proof is required as I have explained in my reasons in the class actions.  A different retrospective hypothetical scenario is involved.  Nevertheless the two types of analysis overlap and accordingly my reasons here also include aspects of my reasons in the class actions.

89 Fifth, I have dealt with the survey evidence in my reasons in Epic v Apple.

90 Sixth, I have set out all relevant legal principles concerning market definition, s 46 and other provisions of Part IV of the CCA and aspects of the ACL in my reasons in Epic v Apple.

91 Seventh, I have also set out some applicable economic principles in my reasons in Epic v Apple, including on the question of tied products.

92 Eighth, annexed to my reasons in this proceeding and the Epic v Apple proceeding is both a schedule of abbreviations and acronyms and also a dramatis personae which also includes the detail of lay and expert witnesses in terms of their positions at the start of the trial.

93 Now in terms of the venue for and mode of the joint trial, I have made various observations in my reasons in the Epic v Apple proceeding which I will not repeat save as to two matters.

94 As I have said in my reasons in the Epic v Apple proceeding, the trial required the use under my control of the largest electronically fitted out trial court room of its type in Australia with the necessary technical and other assistance from numerous court staff, all of whom were directly answerable to my requirements.  What was also necessary was to provide the parties with work spaces, conference rooms and storage facilities on the same level as the court room.  All of this was arranged and provided months in advance of the trial starting.  If I may say so I am most grateful to the administrative and court staff for making this all possible.  I should also express my appreciation to the in court electronic operator for her first class work.

95 Further, let me express my appreciation for the considerable effort put in by Mr Neil Young KC and his team for Epic, Mr Cameron Moore SC and his team for Google, Mr Matthew Darke SC and his team for Apple and the cameo performance of Mr Anthony Bannon SC and his team for the class applicants.  The timeliness and quality of the work of counsel and their instructing solicitors was excellent, and the skill of counsel in their notably efficient presentations including the handling of witnesses resulted in this complex and lengthy trial finishing precisely on time.

96 Let me now proceed to the heart of the matter.  For convenience I have divided my reasons into the following sections:

(a) The Google entities and background matters – [97] to [148].

(b) Epic, Google and the Play Store – [149] to [191].

(c) The OEM restrictive conduct and the OEM agreements – [192] to [367].

(d) App developer restrictive conduct and the developer distribution agreements – [368] to [545].

(e) Mobile OS licensing and Android mobile app distribution markets – s 46 – [546] to [594].

(f) In app payment solutions restrictive conduct – [595] to [642].

(g) The Android in app payment solutions market – s 46 – [643] to [687].

(h) Google’s lay evidence – [688] to [766].

(i) Market definition concerning Android mobile app distribution – the parties’ cases – [767] to [861].

(j) Some relevant concepts debated by the economic experts – [862] to [1095].

(k) Epic’s posited Android mobile app distribution market – what is the correct focal product? – [1096] to [1280].

(l) The experts’ approaches to HMT quantitative and qualitative analysis – [1281] to [1444].

(m) Fortnite removal event and diversion ratios – [1445] to [1816].

(n) Logit demand model, transition probability analysis and counting of diversion estimates – [1817] to [2017].

(o) Critical loss analysis including the Lerner index – [2018] to [2402].

(p) What are the close substitutes, if any? – [2403] to [2650].

(q) Is there any competitive tension between Google and Apple? – [2651] to [2789].

(r) Market power – Android mobile app distribution – [2790] to [3162].

(s) Market definition and market power concerning the mobile OS licensing market – [3163] to [3376].

(t) The MADAs including questions of anti-competitive purpose and anti-competitive effect – [3377] to [3584].

(u) The revenue share agreements including the RSA3s and MIAs – anti-competitive purpose and anti-competitive effect of the RSA3s – [3585] to [3918].

(v) Project Banyan and Project Hug – purpose and effect of Project Hug and the GVPs – [3919] to [4280].

(w) Clause 4.5 of the DDA – prohibiting the distribution of rival app stores – [4281] to [4406].

(x) Android ecosystem security including Google Play Protect – [4407] to [4596].

(y) Behavioural economics evidence – [4597] to [4716].

(z) The technical restrictions – [4717] to [4815].

(aa) Real world impact of installation frictions – [4816] to [5178].

(ab) The technical restrictions – purpose and effect of SLC – [5179] to [5346].

(ac) Epic’s s 46 case – Android in app payment solutions, contracts and mechanisms – [5347] to [5567].

(ad) In-app payments – some technical evidence – [5568] to [5687].

(ae) Android in-app payment solutions – market definition, market power, purpose and effect of SLC – [5688] to [6057].

(af) Overall purpose and effect of Google’s conduct – Android mobile app distribution – [6058] to [6128].

(ag) The effect of Google’s conduct on Epic and Fortnite – [6129] to [6221].

(ah) Epic’s claims for breach of ss 45, 46 and 47 and for unconscionable conduct – [6222] to [6345].

(ai) Relief and conclusion – [6346] to [6370].

(aj) Schedule of abbreviations and acronyms, dramatis personae of witnesses and other individuals, and other annexures.

The Google entities and background matters

97 The following matters in this section are largely uncontentious, unless I indicate otherwise.

98 Google LLC is incorporated in Delaware, USA and its ultimate holding company is Alphabet Inc., which is also incorporated in Delaware.  Google LLC is headquartered in California, USA.

99 Google LLC is the largest of the entities owned by Alphabet. For public financial reporting purposes, Google LLC is divided into two segments, being Google Services and Google Cloud.  The core products and platforms which comprise Google Services include what are known as or what have been described as Android, Chrome, Hardware, Gmail, Google Drive, Google Maps, Google Photos, the Play Store, Search and YouTube.

100 Google LLC owns and continues to develop the Android OS. Google LLC also owns and develops GMS and licenses GMS to OEMs. But as Google points out, it does not  supply or license a unitary product of both the Android OS and GMS.  They are separately supplied, although of course GMS has some relevant OS aspects.

101 Let me at this point say something about GMS Android, which I have used as a convenient composite label, given that both in Australia and globally excluding China, over 99.9% of smart mobile devices which utilise a licensed OS are GMS Android devices, that is, they have the Android OS and GMS.  From 2018 to 2021, approximately 4 million smartphones were sold in Australia each year with GMS Android pre-installed. Globally excluding China, approximately 800 million smartphones were sold each year with GMS Android pre-installed.

102 As I have said earlier, GMS is Google’s proprietary software that runs on top of open source Android and is necessary for a number of key OS features not available in open source Android alone.

103 GMS includes key modules being APIs and SDKs collectively known as Google Play Services, which enable core functionality for the device and for many apps and services. Android devices which do not also have GMS installed cannot access Google Play Services and are significantly functionally limited as a result. One limitation is that if app developers build apps using Google Play Services APIs or SDKs, those apps become dependent on GMS and will only run as intended on devices that have GMS installed.

104 So, by using Google Play Services, app developers can design their app to make use of Google Maps functionality, such as features that respond contextually to a user’s location, via specific Google Play Services APIs.  But if such an app is installed on a device that lacks GMS, the app will not have that functionality. Many popular apps, such as Facebook, YouTube, WhatsApp, Gmail, Twitter, Netflix, Uber, Pokemon Go, Slack, and CommBank cannot be run on an Android device that does not have GMS installed.

105 GMS also includes apps developed by Google. Among them are the core applications, being the Play Store, Search, Chrome, Google Maps, Gmail and YouTube. The remaining five core applications were among the top six apps by usage in the Play Store as of April 2023. GMS also includes what are described as flexible applications.  These differ between territories and, in Australia, consist of Drive, YouTube Music, Play Movies, Duo and Photos.

106 Now as I have indicated, in Australia and globally excluding China, the vast majority of smart mobile devices which utilise a licensed OS are pre-installed with a combination of open source Android and GMS, which on occasion I will refer to as GMS Android or Google Android.

107 As I have said, Google’s GMS apps are pre-installed under the terms of the MADAs on virtually all Android devices sold outside of China.

108 As I have said, Google licenses OEMs to distribute GMS on their Android devices in accordance with the terms of a MADA.  In doing so, more users are introduced to the GMS apps, which include Google Search and the Play Store, and therefore more advertising revenues are derived by Google. This has led Google to make most of its money selling advertising on Google Search and a considerable amount from the Play Store.

109 Such a licence includes the following software and apps. First, a bundle of closed-source, proprietary Google apps including the Play Store, Google Search, Google Chrome, Google Maps, Gmail and YouTube. Second, a Google proprietary software layer that provides background services, SDKs and APIs for the integration of apps with Google Play Services.

110 Google Play Services is an important component of the Android ecosystem and is relied on by many app developers.  Its main components are the Google Play Services Android Application Package (APK) and the Google Play Services client library.  And without Google Play Services, many Android apps could not function properly.

111 The elements of GMS perform functions that were previously performed by open-source Android OS source code.

112 The GMS apps and Google Play Services are listed in the Google Product Geo Availability Chart, as modified or updated by Google from time to time.

113 Let me at this point say something about the Play Store and Google Play Billing.

114 The Play Store is an app store and is itself a native app.  It is owned and was developed by Google. It was originally known as Android Market, which launched in 2008 in the United States and in 2009 in Australia, before being rebadged as the Play Store in 2012.

115 The Play Store distributes Android native apps to Android devices.  It is the principal distribution point for Android native apps.  Through the Play Store, users can search or browse for apps developed by Google or third-parties, and then download and install such apps onto their devices. There are approximately 3 million apps available to users in Australia for download on the Play Store. Users can discover apps within the Play Store using the search function, and recommended apps are presented to users based on their search and download history. The Play Store also facilitates app updates.

116 From December 2017 to September 2022, the Play Store was pre-installed on between 97.7 and 99.6% of active Android devices in Australia, and between 97.2 and 99.2% of active Android devices worldwide excluding China. Further, in 2020, some 97.9% of new app downloads to Android devices in Australia that occurred through an app store were undertaken via the Play Store, while globally excluding China some 91.4% of new app downloads from an app store occurred via the Play Store.

117 If an app developer wishes to distribute its app via the Play Store, the developer must enter into a DDA with Google.

118 With limited exceptions, app developers who wish to distribute their apps via the Play Store are required to use Google Play Billing to facilitate in-app purchases of digital goods.   Google Play Billing enables the facilitation of payments from users in many countries, via a range of payment methods. Google charges developers a fee of up to 30% of the price of in-app purchases made through Google Play Billing.

119 Google Play Billing can only be used on native Android apps that are downloaded from the Play Store and installed on Android devices. Hundreds of thousands of developers monetise their apps through the use of Google Play Billing.

120 Google Play Billing has a function which facilitates the transfer of money from the user to the app developer.  Google Play Billing can be considered to be either a payment solution or it can be considered to have a function which includes a payment solution.

121 Let me at this point say something about Google Asia Pacific Pte Ltd and Google Payment Australia Pty Ltd.

122 Google Asia Pacific Pte Ltd is incorporated in Singapore.  Its ultimate holding company is Alphabet Inc. Google Asia Pacific Pte Ltd is a party to a payment processing service agreement with Google Payment Australia Pty Ltd.

123 Google Asia Pacific Pte Ltd is also the contracting Google entity and marketplace service provider under the DDA with app developers who select to distribute apps through the Play Store to Android users in Australia.

124 The DDA provides that it is a contract between the app developer and “the applicable Google entity based on where [the developer has] selected to distribute” its apps. The “applicable Google entity” is Google Asia Pacific Pte Ltd in relation to apps distributed through the Play Store in Australia. And the applicable entity is Google LLC in respect of US distribution.  Consequently, a developer that distributes apps through the Play Store in Australia and in the USA enters into the DDA with both Google Asia Pacific Pte Ltd and Google LLC.

125 Google Payment Australia Pty Ltd is an Australian corporation and subsidiary of Google LLC. Google Payment Australia Pty Ltd holds an Australian financial services licence which authorises it to offer the Google Payments Service as that term is used in the Google Payment Australia Pty Ltd product disclosure statement.  Google Payment Australia Pty Ltd enters into agreements with users of the Google Payments Service who register for that service online. The terms of those agreements are set out in the PDS issued by Google Payment Australia Pty Ltd.

126 By a payment processing services agreement with Google Asia Pacific Pte Ltd, Google Payment Australia Pty Ltd has undertaken to perform payment processing functions and services as directed by Google Asia Pacific Pte Ltd.

127 Now both Google LLC and Google Asia Pacific Pte Ltd carry on business within Australia for the purposes of s 5(1)(g) of the CCA. They do so directly and through Google Payment Australia Pty Ltd.  As such, the relevant provisions of the CCA apply to their conduct outside Australia.  Now by the end of the trial these matters ceased to be contentious.  But for completeness it is worth stating the following.

128 The fact that Google LLC carries on business within Australia is evident from the following facts.

129 First, Google LLC supplies services in Australia, including but not limited to the service known as Google Play to Australian users of the Play Store.  The Play Store terms of service expressly reflect such a position.  Further, the Google terms of service stipulate that “Google services are provided by, and you’re contracting with, Google LLC”. Those “Google services” referred to within the Google terms of service are not confined to the Play Store. They also include Android OS, Chrome, Gmail, Google Maps and Google Search. That such Google services are supplied in Australia is evident from the fact that Australian users of Android OS, Chrome, Gmail, Google Maps and Google Search all use those products on devices physically located in Australia.

130 Second, in addition to the supply of the Google services in Australia, Google LLC also enters into the contracts under which those services are supplied in Australia.  In each instance where the Google terms of service and the Play Store terms of service are accepted by users within Australia when they first use the Google services and the Play Store on their Android device, Google LLC can be taken to have entered into that contractual relationship in Australia.

131 As at 30 September 2022, there were more than 1.9 million Android device users in Australia, which is indicative of the number of Android device users who had agreed to the relevant terms of service in Australia.

132 Google LLC’s supply of services and entry into contracts in Australia amounts to business activities carried on by Google LLC in the jurisdiction on a continuous and repetitive basis.

133 Third, and in addition to the services that Google LLC supplies to Australian consumers, Google LLC also supplies, in Australia, various services to Google Payment Australia Pty Ltd. Appendix IV of Google Payment Australia Pty Ltd’s Compliance Manual records that those services include “[s]oftware maintenance and development”, “[o]perations and support” and “hosting” for the “[Google] Wallet account management platform”, “IT support and services for the office environment”, among other services relating to risk management, human resources, auditing and legal services.

134 Google Asia Pacific Pte Ltd likewise carries on business within Australia.

135 First, like Google LLC, Google Asia Pacific Pte Ltd also supplies services and enters into contracts in Australia.

136 For a developer to distribute apps via the Play Store, the developer must enter into a DDA. Where the developer wishes to have its apps distributed in Australia, the relevant Google counterparty to the DDA is Google Asia Pacific Pte Ltd. So, developers located in Australia who wish to distribute apps in Australia via the Play Store must enter into the DDA with Google Asia Pacific Pte Ltd.  Those developers’ entry into the DDA occurs by the developer accepting the terms of that agreement by accessing a website from their device located in Australia, such that Google Asia Pacific Pte Ltd will be taken to have entered into those contractual relationships within Australia.

137 Further, under the DDA, Google Asia Pacific Pte Ltd is the relevant entity authorised to make apps available in Australia via the Play Store, such that Google Asia Pacific Pte Ltd supplies app distribution services in Australia.

138 Accordingly, Google Asia Pacific Pte Ltd’s supply of services and entry into contracts in Australia amounts to business activities carried on by it in the jurisdiction on a continuous and repetitive basis.

139 Second, Google Asia Pacific Pte Ltd has contracted with Google Payment Australia Pty Ltd for the provision of payment processing services in Australia in exchange for a service fee, which Google Asia Pacific Pte Ltd pays to Google Payment Australia Pty Ltd under that agreement. This ongoing commercial relationship with an Australian entity, the subject matter of which is services supplied by that entity in Australia in consideration for a fee, similarly amounts to the carrying on of business in Australia by Google Asia Pacific Pte Ltd.

140 Further, both Google LLC and Google Asia Pacific Pte Ltd carry on business within Australia by and through Google Payment Australia Pty Ltd.

141 Google Payment Australia Pty Ltd conducts a confined business as one of several Google payment entities used to process payments made on the Play Store.  It performs that role pursuant to the terms of the payment processing services agreement between it and Google Asia Pacific Pte Ltd.

142 The payment processing services agreement provides that Google Payment Australia Pty Ltd shall provide the defined “Services” in consideration for a set fee from Google Asia Pacific Pte Ltd. Those “Services” are defined to include “all services, consulting, advice and assistance required by [Google Asia Pacific Pte Ltd] in connection with performing payment processing functions and services”. The agreement stipulates that those services shall be provided “as directed by” Google Asia Pacific Pte Ltd. The agreement further stipulates that Google Payment Australia Pty Ltd “shall provide other services in relation to the Services as directed by Company”.

143 The fee for which Google Payment Australia Pty Ltd performs these services at the direction of Google Asia Pacific Pte Ltd is tied to the amount of operating expenses incurred by Google Payment Australia Pty Ltd in the performance of those services, plus an additional 8% of those expenses.

144 Google Payment Australia Pty Ltd adopts the organisational systems and compliance measures, processes and procedures and the employment and personnel policies of Google LLC, and it considers that its employees are bound by the code of conduct published by Google LLC.

145 Indeed, many of the functions required to deliver its services are provided either in whole or in part by Google affiliate companies. And Google Payment Australia Pty Ltd’s compliance manual lists a multitude of functions that are outsourced to Google LLC and Google Asia Pacific Pte Ltd.

146 Further, Google LLC has given Google Payment Australia Pty Ltd a financial undertaking so as to enable it to satisfy the requirements of its Australian financial services licence.  Similarly, Google Asia Pacific Pte Ltd has indemnified Google Payment Australia Pty Ltd in respect of any financial risk, expenses, civil and criminal liability, or any kind of financial liability whatsoever that may be incurred as a result of any third-party litigation in which both entities are named as co-defendants.

147 In summary, both Google LLC and Google Asia Pacific Pte Ltd carry on business within Australia.  And because they carry on business within Australia, Part IV of the CCA and the ACL (other than Part 5-3) apply to conduct engaged in by Google LLC and Google Asia Pacific Pte Ltd outside Australia.

148 It is convenient in dealing with Epic’s main competition case to refer to Google LLC and Google Asia Pacific Pte Ltd together as Google unless I indicate otherwise; I will not generally put Google Payment Australia Pty Ltd under that label given its more limited function.

Epic, Google and the Play Store

149 I have summarised the nature and evolution of Epic’s business up to the launch of Fortnite on iOS in my reasons in the Apple proceeding.  It is unnecessary to repeat that detail here.  Let me at this point set out a narrative of some of the relevant dealings between Epic and Google and Epic and various OEMs, the primary facts of which were not contentious unless I state otherwise.

150 Let me begin by saying something about the launch of the Fortnite installer on Android.

151 In August 2018, Epic launched the Fortnite installer on Android. The Fortnite installer was an app from which users could download, and which would thereafter be used to update, another native app, that is, Fortnite.

152 Although initially Fortnite was the only mobile app available for download through the Fortnite installer, Epic’s intention was that the Fortnite installer would be expanded in the future to become the Epic Games App, through which users could download other mobile apps developed by Epic, and ultimately mobile apps developed by third parties.

153 Since the primary function of the Fortnite installer was to distribute and update another app, the DDA prohibited Epic from distributing the Fortnite installer through the Play Store.

154 But it is possible for developers to distribute native apps to Android device users outside the Play Store by reaching arrangements with OEMs to pre-install the app on devices manufactured by that OEM, via third-party Android app stores, or via direct downloads from a website. And given these possibilities, Mr Sweeney proposed launching the Fortnite installer on Android outside the Play Store via a combination of direct downloading from Epic’s website and pre-installation deals negotiated with Android OEMs.

155 Epic ultimately chose not to distribute Fortnite on the Play Store because it believed that Google’s contractual terms would require Epic to use Google’s payment solution, Google Play Billing, for in-app purchases, require Epic to pay Google a 30% commission on those purchases, and prevent Epic from distributing the Fortnite installer as distinct from Fortnite, because the former facilitated the distribution of another app contrary to Google’s contractual terms.

156 Epic saw the Fortnite installer as an initial step towards having its own mobile app store. This was important to Epic, including because Epic was intent on developing its own ecosystem, being a set of developers and users with whom Epic had an ongoing relationship.

157 Let me at this point say something about Epic’s collaboration with Samsung.  The following chronology is not contentious in terms of the base facts.

158 From about April to August 2018, Epic negotiated with Samsung and sought its agreement to pre-install the Fortnite installer on Samsung devices. Epic’s original proposal disclosed its intention to later grow the Fortnite installer into a mobile game distribution service for third-party gaming apps. Samsung liked Epic’s intention to disrupt the Play Store but advised that it had to get approvals from Google and various carriers which may not be forthcoming.

159 Whilst these negotiations were occurring, Epic informed Google of its intention to make Fortnite available on Android outside the Play Store. Soon afterwards, Google, which was concerned with a potential revenue loss of up to USD 3.6 billion if other developers followed Epic in not launching on the Play Store, offered Epic various incentives worth over USD 200 million to reverse that decision.

160 When Epic declined this offer, Mr Lockheimer, the senior vice president, platforms and ecosystems at Google, said to Mr Sweeney that Google viewed any launch of software outside the Play Store as a threat to Google and that Google would have to make changes to Android to protect users.

161 In July 2018, an internal Google presentation titled Project Elektra was prepared. It noted Epic’s plan to launch Fortnite directly to the public, thereby bypassing the Play Store, and observed that a large strategic investment into Epic could help advance broader discussions about using the Play Store for Fortnite.

162 On 19 July 2018, an off-cycle meeting of the Google business council was convened to discuss the proposed investment in Epic. The business council is an internal Google committee which is responsible for reviewing significant deals or transactions being proposed by Google’s various business divisions.

163 In a presentation titled EPIC/Fortnite BC Deal Review of the same date, the objectives of the deal were said to include that “[n]ot having Fortnite on Play creates significant risks for Android ecosystem & Play business, and threatens our business in the long term”. Among the risks identified in the presentation were the “Contagion effect” of other top developers following Epic, “increased rev[enue] share pressure” and “Fortnite may legitimize “Samsung” store & 3rd party stores; fragmenting distribution on Android”.

164 On 21 July 2018, the business council approved the proposal. The proposed investment in Epic did not eventuate.

165 In August 2018, the negotiations with Samsung culminated in Epic entering into the Samsung collaboration agreement. The Samsung collaboration agreement provided that Samsung would distribute the Fortnite installer or the Epic Games App through the Galaxy Store. It also licensed Samsung to pre-install the stub version of the Fortnite installer and, if Epic transitioned to the Epic Games App, the stub version of the Epic Games App on Samsung Galaxy Note 9 and Galaxy Tab S4 devices.

166 These stubs were icons which when clicked took the user to the Galaxy Store, where the user could then initiate a download of the Fortnite installer in the same way that any other app could be downloaded from within a pre-installed Galaxy Store, which had a privileged install permission and so did not confront the technical restrictions.

167 But the Samsung collaboration agreement was limited and did not overcome the discoverability challenges faced by Android apps distributed outside the Play Store.

168 Samsung only agreed to use reasonable efforts to pre-install the stub on two new device models, and to make reasonable efforts to perform an update which would install the app on existing Samsung devices.  So, Epic could not compel pre-installation on locked devices, even of the two new models, without the consent of carriers.  Further, the Fortnite installer stub was located within Samsung’s Game Launcher, which was located in the app tray and not on the home screen of the relevant Samsung devices. Further, if a user searched for an app via the search bar of their Samsung device, they were redirected to Google Search, the results of which prioritised apps available on the Play Store.

169 Now pursuant to the Samsung collaboration agreement, Fortnite apps and any apps downloaded through the Epic Games App were required to use Samsung’s in-app payment solution. Use of Samsung IAP was not mandatory where Epic’s apps were directly downloaded onto a Samsung device.

170 The agreed revenue shares were as follows. In the scenario of in-app purchases within Fortnite apps, the revenue share was 85% to Epic and a commission of 15% to Samsung for the first 18 months, and an 88/12 percentage split thereafter.  And in the scenario of in-app purchases within other apps downloaded through the Epic Games app, the revenue share was REDACTED] to Epic and a commission of [REDACTED] to Samsung for the first REDACTED] months, and a REDACTED] percentage split thereafter.

171 On 10 August 2018, Epic announced the launch of Fortnite on Samsung devices at Samsung’s Galaxy Unpacked event. The Fortnite installer was also made available for Samsung users to download via the Galaxy Store and Epic made an Android version of the Fortnite installer available for direct download from its website.

172 Further, in the period from 2018 to 2020, Epic negotiated with various OEMs besides Samsung, with a view to having the Fortnite installer, and later the Epic Games App, pre-installed on as many Android devices as possible. Epic also negotiated with carriers to pre-install its apps on “locked” Android devices. Epic’s goal in these negotiations was to improve both the discoverability and the install flow of Epic’s apps on old and new, locked and unlocked Android devices.

173 At about the time of these negotiations, Epic launched the Epic Games Store on PC in December 2018 and, in October 2019, Epic began replacing the Fortnite installer with the Epic Games App for Android devices. At that time, the Epic Games App enabled users to download, install, and update another Epic app besides Fortnite.

174 Let me now turn to the question of the launch and subsequent removal of Fortnite on the Play Store.

175 In December 2019, Mr Sweeney contacted Google and advised that Epic would submit a version of Fortnite to the Play Store that used Epic’s own payment solution being Epic Direct Payment (EDP) to facilitate in-app purchases instead of Google Play Billing. The app was duly submitted with EDP installed but Google refused to allow it onto the Play Store and suspended Epic’s publishing status.

176 Following further unsuccessful negotiations, Mr Sweeney concluded that all avenues for negotiating terms with Google to distribute Fortnite on the Play Store using EDP were exhausted. He concluded that Epic’s attempt to connect with a large proportion of Android users outside the Play Store, via direct download, pre-installation, and the Galaxy Store, had been unsuccessful.

177 Mr Sweeney decided that Epic should adopt the same approach toward both Apple and Google, namely, to distribute Fortnite on the Play Store and on the App Store, and then publicly challenge both Apple and Google to alter their policies.

178 On 21 April 2020, Epic launched Fortnite on the Play Store, using Google Play Billing. It continued to make the Epic Games App available on Android through other means, including the Galaxy Store and Epic’s website. The contractual terms which applied to Epic’s distribution of Fortnite through the Play Store were set out in a DDA with Google that was entered into in November 2016.

179 After Fortnite launched on the Play Store, Epic experienced a large increase in new Fortnite users on Android. In the first month after that launch, 3.14 million Fortnite users downloaded the app from the Play Store, and 1.7 million of them were new to Fortnite. Fortnite’s gross revenue from the Play Store in that month was approximately $1.3 million, or about $50,000 per day.

180 In June and July 2020, Mr Sweeney sought Google’s agreement to allow Epic to offer its own payment solution within Fortnite and other Epic apps distributed via the Play Store, and to allow Epic to launch a mobile Epic Games Store app on Android, either via the Play Store or by a direct download process which did not have to confront the technical restrictions. Google declined those requests.

181 On 3 August 2020, Epic submitted Fortnite version 13.40 to Google. This version contained code that would allow Epic to offer EDP as an alternative payment option within Fortnite on Android devices when what has been described as a hotfix was activated.  Epic did not disclose the hotfix to Google, because it believed that Google would then not approve the update.

182 On 5 August 2020, Mr Sweeney requested from Mr Lockheimer and Mr Samat that Google support Epic using competing payment options in time for a planned Fortnite price reduction. Lockheimer’s response on 11 August 2020 indicated that Google was not willing to do so.

183 On 13 August 2020, Epic activated the hotfix and thereby made EDP available within Fortnite apps downloaded from the Play Store and the App Store.  As a result, Play Store users were presented with a choice when purchasing V-Bucks within the Fortnite app: they could make their purchase using Google Play Billing or using EDP. At the same time, Epic introduced the “Fortnite mega drop”.

184 Within a few minutes of Epic activating the hotfix and implementing the mega drop, Mr Sweeney emailed Mr Lockheimer and other senior executives at Google, informing them that Epic had introduced EDP in the version of Fortnite distributed through the Play Store.

185 Later on 13 August 2020, Google removed Fortnite from the Play Store.  Epic later commenced proceedings against Google in the US District Court for the Northern District of California, alleging, among other things, violations of the Sherman Act and California’s Unfair Competition Law.

186 Fortnite is still excluded from the Play Store and users have been unable to download Fortnite from the Play Store since the removal. Further, users who previously downloaded Fortnite from the Play Store have been unable to update that version of the app since 13 August 2020. The Epic Games App from which users can download Fortnite is available to Android users in Australia on the Galaxy Store and by direct download from Epic’s website.

187 Shortly after the removal, the average monthly active users of Fortnite on Android devices declined to below 3 million, from 5.9 million immediately prior to the removal in July 2020.

188 Now I should note here that Google says that it removed Fortnite from the Play Store because Epic had breached the DDA and the payments policy by introducing EDP via a hotfix.  Google says that Epic’s move was part of a pre-planned scheme intended to provoke a response from Google. It says that Google’s conduct only amounted to enforcing the provisions of the DDA and the payments policy.

189 I will say something more about the relationship between Epic and Google much later in my reasons and addressed to the question of whether Google’s conduct or policies impeded Epic or Fortnite.

190 Further, as to the relevant Apple entities and background matters relevant to such entities, I refer to what I said in my reasons in the Apple proceeding.

191 Let me now turn to Epic’s case and begin by more precisely identifying what Epic says to be the OEM restrictive conduct.  Before dealing with the question of markets, it is important to be clear about what precisely is said to be the impugned conduct.

The OEM restrictive conduct and the OEM agreements

192 Generally speaking, Epic says that Google has engaged in conduct by which it prevents or hinders competition for the supply of services for the distribution of Android apps and for facilitating in-app payments made within Android apps.  Specifically, it is said that Google has spun a web of contractual terms and policies that traps OEMs and app developers, so as to preclude robust competition for the Play Store and Google Play Billing.

193 Now for convenience and in this context, Epic’s case concerning Google’s anti-competitive conduct can be divided into three categories.

194 First, there are the alleged contractual restrictions which Google imposes on and enforces against OEMs that license GMS.  I will refer to this as the OEM restrictive conduct.

195 Second, there are the alleged contractual restrictions which Google imposes on and enforces against app developers that distribute Android apps via the Play Store and additional arrangements with large developers by which Google has sought to prevent or neutralise competition in app distribution. I will refer to this as the app developer restrictive conduct.

196 Third, there are the alleged contractual restrictions which Google imposes on and enforces against app developers with respect to in-app payment solutions and related matters. I will refer to this as the in-app payment solutions restrictive conduct.

197 Let me in this section address the OEM restrictive conduct only.  Now the alleged OEM restrictive conduct includes the following aspects.

198 The first aspect is that it is said that Google’s conduct required OEMs that wish to distribute devices with GMS to enter into a MADA.

199 Epic has given the following description, the terms of which are largely accurate subject to any qualification that I have made.

200 The terms of the MADA require OEMs, who wish to distribute an Android smart mobile device with GMS, to pre-install the Play Store and to place the icon giving access to the Play Store on the device’s default home screen (DHS) or, in the case of certain Samsung devices, in the device hotseat.  OEMs must place the icons giving access to Google Search, the Play Store and a folder providing access to a collection of icons for the other GMS apps on the DHS of their Android smart mobile device or, in the case of certain Samsung devices, in the device hotseat.

201 And OEMs may only distribute devices with GMS if their devices comply with the following GMS requirements. Devices must pre-install all core applications on the device, including the Play Store. The Google Search widget, the Play Store app icon and a folder labelled “Google” containing the core applications must feature on the DHS.  Devices must pre-install the GooglePackageInstaller app, as a privileged app, in place of the default package installer available with the Android OS.  The Play Store must be placed in the apps menu and should be in the top level of the apps menu and may not reside within any folder of the apps menu. Devices running Android must replace the global “Unknown Sources” setting with the “Allow apps installs” setting, which grants the “REQUEST_INSTALL_PACKAGES” permission on a per app basis, allowing the granted app to provide other apps for installation. Devices must not implicitly grant this permission to any pre-installed or downloaded third-party apps that do not have a privileged installation permission (unknown sources). And devices must comply with the CDD and pass applicable testing suites, including the CTS and GMS testing suite.

202 OEMs are required to comply with terms of the MADA or documents incorporated into the MADA which have the effect that if an Android smart mobile device user seeks to download an Android app from an unknown source, including an internet browser or an alternative app store that does not have a privileged install permission, the user is confronted with various warnings including that their device is more vulnerable to attack by “unknown apps” and must alter their device’s settings to allow the download to occur.  Further, those warnings and the requirement for an Android smart mobile device user to alter their device settings will not apply to downloads from the Play Store.

203 OEMs are also subject to the following requirements under the MADA.

204 OEMs may only distribute GMS on an Android compatible device, being a device that complies with the Android Compatibility Definition Document (CDD), being part of the Android compatibility program.  I will return to the detail of the CDD later.

205 Further, OEMs may not distribute an Android compatible device unless the GMS apps, including the Play Store, and Google Play Services, have been preinstalled on that device, except as otherwise approved by Google in writing.

206 Further, OEMs may only distribute a device with one or more GMS apps if they make all of the Google mobile services apps, including the Play Store, available on that device.

207 Further, in the case of MADAs entered into before 2017, OEMs must not take any actions that may cause or result in the fragmentation of the Android OS.

208 Further, OEMs may not distribute an Android compatible device in any territory without Google’s written approval prior to the launch of each device model, and OEMs must send the final software build of their devices, as well as test reports confirming the device passes the Android Compatibility Test Suite (CTS), to Google for final approval.

209 The second aspect is that it is said that Google’s conduct required OEMs that wish to distribute devices with GMS to enter into an AFA containing the terms referred to below. So, OEMs that wanted to distribute devices with GMS had to enter into an AFA and ACC, being part of the Android compatibility program.

210 OEMs are prevented from distributing, in the case of OEMs subject to an AFA, or from manufacturing, distributing or marketing, in the case of OEMs subject to an ACC, devices based on the Android OS that are not Android compatible devices. A device is an Android compatible device if it complies with the CDD.  I will say something more about the CDD shortly.

211 Further, OEMs are prevented from distributing, in the case of OEMs subject to an AFA, or from developing, distributing or marketing, in the case of OEMs subject to an ACC, any software based on the Android OS that is not designed to run on Android compatible devices.

212 Further, OEMs subject to an AFA are prevented from taking any actions that may cause or result in the fragmentation of the Android OS.

213 Further, OEMs are prevented from, in the case of OEMs subject to an AFA, distributing an SDK derived from the Android OS or derived from Android compatible devices and participating in the creation of, or promotion in any way of, any third-party SDK derived from Android, or derived from Android compatible devices, or in the case of OEMs subject to an ACC, distributing or marketing an SDK based on the Android OS to third parties, or participating in the development of such an SDK, save for their own internal use.

214 Further, OEMs that wanted to distribute devices with GMS had to comply with various requirements of the CDD, which I will detail more specifically in a moment.

215 The third aspect concerns Google’s conduct in entering into what have been described as the Premier revenue sharing agreements, which I have previously referred to as RSA3s, or mobile incentive agreements, which I have previously referred to as MIAs, with certain OEMs.

216 Those terms promise financial incentives to OEMs in exchange for the OEMs configuring devices so that the Play Store is set as the default app store, no alternative app store or other app substantially similar to the Play Store is pre-installed on the device, only apps that are available on the Play Store are pre-installed on the device, and the Play Store icon is placed on the device’s DHS.

217 The RSA3s or MIAs with certain OEMs contained various terms.

218 The incentives offered to OEMs in the RSA3s, conditional on the OEM entering into a MADA and AFA, are only payable in respect of revenue generated from devices configured by the OEMs as follows. Devices configured in the appropriate way are referred to as “Premier” devices.

219 A Premier device must not include any pre-installed app that is substantially similar to the Play Store, nor any launcher, over the air prompt or functionality that has the primary purpose of providing access to any service substantially similar to the Play Store.

220 Further, a Premier device must set the Play Store as the “default” app store or marketplace for apps and all other digital content including subscriptions.

221 Further, a Premier device must only pre-install an app that is available on the Play Store, and does not contain privileged install permissions, which facilitate the downloading of other apps, although in some instances, there is an exception for an app store owned by the OEM or certain carriers.

222 Further, a Premier device must not be configured so as to allow an OEM or any third-party to promote, introduce or suggest alternative app stores to an Android smart mobile device user.

223 Further, a Premier device must have the Play Store icon placed on the DHS.

224 The incentives offered to OEMs in the MIAs, conditional on the OEM entering into a MADA and AFA, are only payable in respect of revenue generated from devices configured by the OEMs with, generally speaking, the same requirements as set out.

225 The fourth aspect is said to be Google’s conduct in giving effect to and/or enforcing the terms referred to above including in relation to Epic by engaging in the relevant conduct.

226 I will refer to the MADAs, AFAs and ACCs, and RSAs and MIAs collectively as the OEM agreements.  Let me delve deeper into some of these agreements and arrangements.

Anti-fragmentation agreements and Android compatibility commitments

227 Now as I have mentioned earlier, if an OEM wishes to distribute Android devices with relevant APIs known as GMS pre-installed, Google requires the OEM to enter into either an AFA or an ACC as part of the Android compatibility program, in addition to a MADA.  Of course an AFA or an ACC can be entered into without entering into a MADA.  Now the AFAs and ACCs are relevant to Epic’s case in two ways, as Epic has explained in the following terms.

228 First, they mandate compliance with compatibility standards that include the technical restrictions.

229 Second, they buttress Google’s power in the mobile OS licensing market and constitute a barrier to entry and expansion in that market.

230 Now Epic does not say that all of the compatibility standards imposed on OEMs via the AFAs and ACCs are anti-competitive and unlawful. It was not Epic’s case that Google cannot set common compatibility standards for Android devices or that Android OS must be allowed to fragment. Instead, Epic impugns a single aspect of the compatibility standards, namely, that aspect which forms part of the technical restrictions.

231 Epic otherwise relies on the AFAs and ACCs as elements of Google’s power in the mobile OS licensing market.

232 Now as I have indicated, Google made an AFA and/or an ACC with every OEM that distributed Android devices with GMS pre-installed during the relevant period. After all, the GMS distribution licence granted under the MADA is expressly subject to the OEM being in compliance with a valid and effective AFA or ACC.

233 In any event, the evidence includes copies of agreements covering the relevant period which Google made with ten OEMs, being Samsung, Huawei, OPPO, Xiaomi, Sony, Motorola, Nokia, Vivo, OnePlus, and LG; I will refer to these as the selected OEMs.

234 The selected OEMs accounted for between about 74 and 85% of all Android devices and Android fork devices shipped in Australia from 2018 to 2022, and over 70% of all Android devices and Android fork devices shipped globally excluding China during the same period.

235 And as I have indicated, both AFAs and ACCs use a defined term that is also used in the MADA, namely, Android compatible device(s), which term is defined to mean an Android device that complies with the CDD and successfully passes the CTS.  The CDD contains numerous baseline requirements for the configuration, behaviour, and security of Android devices.  The CTS is an automated software suite developed by Google that scans an OEM’s installed software and tests for compliance with the CDD. Devices that run on an Android fork do not comply with the CDD and will not pass the CTS, whereas all Android compatible devices do comply with the CDD and have passed the CTS.

236 The AFAs relevantly provide that OEMs may only distribute devices developed to run on the Android OS if those devices are Android compatible devices. The AFAs further provide that OEMs may only distribute software developed for use on devices that run on the Android OS if those devices are Android compatible devices.

237 Finally, the AFAs prohibit OEMs from taking any actions that may cause or result in the fragmentation of Android. They also prohibit OEMs from distributing, creating or promoting an SDK that is derived from Android or from Android compatible devices.

238 In 2017, Google replaced most AFAs with ACCs. The terms of the ACCs are similar to those of the AFAs. The ACCs provide that all devices based on Android that the OEM manufactures, distributes or markets will be Android compatible devices. They further provide that all Android based software that the OEM develops will be designed to run on Android compatible devices. Further, the ACCs prohibit the OEM from distributing or marketing an SDK based on Android to third parties.

239 The ACCs contain exceptions whereby OEMs are permitted to manufacture devices and components for devices that are not Android compatible devices, if they do so for a third-party brand and do not market such devices. OEMs are also permitted to develop, distribute, or promote apps for any platform.

240 ACCs entered into from 2020 onward also contain an exception for devices supplied into the European Economic Area (EEA). For those devices only, the OEM may develop, manufacture, distribute or market, first, Android devices that are not Android compatible devices, second, software derived from the Android OS that is not designed for Android compatible devices and, third, SDKs associated with those devices, provided that such devices, software and/or SDKs comply with the Android and Google brand guidelines. This exception largely abrogates the ACC within the EEA.

241 Google nonetheless insists that ACCs must be entered into and enforced throughout the rest of the world.

242 The AFAs and ACCs have the effect of committing OEMs to only distribute Android compatible devices and to refrain from manufacturing, distributing, or marketing, except as a contractor for third-party brands, any smart mobile devices that operate on an Android fork or are otherwise non-compliant with the CDD.

243 ACCs but not AFAs may be terminated by the OEM on 60 days’ notice. But termination of an ACC would cause the OEM to lose its distribution licence under the MADA.

244 Now the ACC is an agreement entered into voluntarily by device manufacturers to promote compatibility and ensure compliance with the CDD.  But if an OEM wants to comply with the CDD, it can do so without signing an ACC.  And it follows that a voluntary desire to promote compatibility and ensure compliance with the CDD cannot explain why OEMs sign binding ACCs with Google, although Google did not accept this.

245 If an OEM wants the right to pre-install GMS on its devices, it has no choice but to enter into, and comply with, an ACC. And I agree with Epic that the primary reason why an OEM would sign an ACC is because it wants the right to distribute GMS under a MADA.

246 Google’s own documents state that almost all OEMs which sign an ACC also sign a MADA. Further, Google has not led evidence of any OEM that has not signed both.

247 The primary reason why OEM’s sign an ACC is that they want the right to distribute GMS under a MADA. If they want that right, OEMs have no choice but to enter into, and comply with, an ACC, or previously an AFA.

248 So, although formally separate documents, the ACCs and AFAs are in substance part of the bargain embodied in the MADAs.

249 In this way, one of the choices supposedly available to OEMs in the Android ecosystem is foreclosed. If an OEM wants to distribute any Android devices, it must sign an ACC. And for so long as it is a party to an ACC, the OEM cannot distribute any devices that run on an Android fork.

250 Google executives who prepared a 12 October 2010 presentation titled Android OC Quarterly Review – Q4 2010 said of the similarly framed AFA that it “[s]tops our partners and competitors from forking Android and going alone”. That was evidently an important objective which Google sought to achieve via the AFAs and ACCs. It was far more advantageous for Google to make sure that there was just one operating system, which it controlled, rather than multiple forked operating systems, which it did not.

251 Now it has been said that an OEM can always choose to distribute devices that run on an Android fork, or develop their own operating system, instead of distributing Android devices. But the evidence demonstrates that this is not really an option for OEMs if they wish to make a profit.

252 Apps developed for use on Android compatible devices are not likely to function correctly on forked devices, let alone on a newly developed operating system. Google’s proprietary apps, including the Play Store, are unlikely to function properly on forked devices. That is because those apps, along with hundreds of thousands of third-party apps, use one or more of the Google APIs that are collectively known as Google Play Services.

253 In 2016, some 826 of the top 1,000 Android apps used one or more Google Play Services APIs. If a device does not have Google Play Services installed, none of those apps will work as intended, if at all, on that device.

254 Google Play Services are necessary for several key operating system features not available in the AOSP alone and a number of very popular apps, including Facebook, YouTube, WhatsApp, Netflix and Uber, cannot be run on non-GMS-certified versions of Android.

255 It is also not possible to utilise the Google Play Services APIs without installing the Play Store. Those APIs require and use Play’s infrastructure.

256 As far back as 2010, Google’s executives were planning to train developers on Google’s APIs. They recognised that if developers used Google’s APIs, their apps would not function properly outside of the Android ecosystem which Google controlled.

257 The end result is that throughout the relevant period, an OEM which chose to distribute devices that ran on an Android fork would have to persuade users to buy devices on which most Android apps would not function properly, if at all.

258 The OEM would also have to persuade developers to develop apps or versions of apps that did function properly on its Android fork.

259 The only way for an OEM to ensure that most Android apps do function properly on its devices is to pre-install GMS which includes Google Play Services. And to do that, the OEM must sign both an ACC and a MADA.

260 So, it is unsurprising that few OEMs develop their own operating systems or choose to pre-install Android forks.

261 From 2017 to 2022, less than 0.2% of the smart mobile devices shipped in Australia and globally excluding China each year were smartphones with an operating system other than GMS Android or iOS pre-installed.

262 If one focuses on operating systems that are available for license by OEMs, then from 2017 to 2022 less than 0.1% of smart mobile devices shipped in Australia and globally excluding China each year pre-installed a licensable operating system other than GMS Android. The remaining 99.9% were Android devices.

263 So, excepting China where GMS is not available, and Apple whose proprietary operating system is not licensed to any other OEM, almost no OEMs believe that it is in their commercial interests to distribute smart mobile devices that use an operating system other than GMS Android.

264 The commercial reality is that, besides Apple, OEMs operating outside China have no practical choice but to sign an ACC and a MADA.

The mobile application distribution agreements

265 Let me now delve more deeply into the MADAs.

266 As I have already said, if an OEM wishes to include GMS on its smart mobile devices or to use the Android trademark, it must enter into a MADA with Google which contains the relevant licence terms for GMS.  Each of the selected OEMs was party to a MADA throughout the relevant period, and the same is no doubt true of all OEMs who distributed Android devices.

267 The Android trademark signals to users that the device is an Android-compatible device which participates in the Android app ecosystem.  The MADA relevantly grants, “subject to Google’s prior written approval for each use”, a licence to display Google owned trademarks, which include the Android trademark, “solely for the purposes of this Agreement”.

268 The “purposes of this Agreement” are to grant the OEM a licence to distribute devices with GMS installed and to prescribe the conditions of that grant.  So, the licence to use the Android trademark which Google confers upon OEMs under the MADA permits use of that trademark in connection with the distribution of Android devices with GMS.  It does not permit use of the Android trademark in connection with the distribution of devices that do not have GMS installed.

269 So, OEMs only have a choice to distribute non-GMS Android devices insofar as it is commercially sensible for them to sell smart mobile devices without being able to use the Android trademark when doing so.

270 Apart from the trademark licence, under the MADAs, Google grants OEMs a non-exclusive licence to distribute GMS on, and only on, Android compatible devices and to reproduce GMS to the extent necessary to exercise that licence.

271 GMS principally consists of specific Google owned apps, as well as the collection of SDKs and APIs that make up Google Play Services.

272 The GMS apps include the core applications, being, the Play Store, Search, Chrome, Google Maps, Gmail and YouTube, and flexible applications.

273 Google Play Services include SDKs and APIs that enable core functionality for many apps and services such as Google Sign-In and Google Maps.  Since many Android apps are developed using Google Play Services SDKs and APIs, those apps will not function properly on a device that does not have Google Play Services installed.

274 The OEMs’ licence to distribute devices pursuant to the MADA is subject to them complying with a valid and effective ACC.

275 Further, and importantly, under the MADAs OEMs are only permitted to distribute a device containing GMS if all core applications including the Play Store and flexible applications have been pre-installed on that device, save as approved by Google in writing. It is not permissible for OEMs to choose which of those GMS apps they wish to pre-install and which they do not.

276 The MADAs also prescribe placement requirements pertaining to the GMS apps.

277 Specifically, OEMs must place on the device’s DHS the icons which give access to the Google Search app, the Play Store and a folder providing direct access to icons for the core and flexible applications.

278 The phrase “default home screen”, which I have referred to as DHS, is defined at clause 1.14 of the MADA as the default display of a device, including the lock screen and/or the notification tray, prior to any changes made by end users that appears without scrolling after initial boot-up,  without scrolling after each subsequent power-up or without scrolling after an end user selects the “home” icon.

279 OEMs must also ensure that any GMS app that is not a core or a flexible application is placed no more than one level below the DHS.

280 Samsung’s MADA imposes placement requirements to the same effect, save that, since 8 January 2018, those requirements have been varied in respect of three categories of device only being Galaxy S9, Galaxy Note9 and Galaxy Fold.  On those devices, Samsung may remove all app icons from the DHS.  But where Samsung does so, the Play Store must be placed in the device hotseat, which is the dock along the bottom row of every screen, including the DHS.  Further, a folder of pre-installed GMS apps must be placed in the first row of the application tray, which is accessed by swiping up from the bottom of the DHS.

281 These three categories of Samsung devices accounted for 13.9% of all Android device shipments in Australia in 2018, and that number rapidly diminished, such that it was only 1.2% in 2019 and zero by 2020.  Globally excluding China, these devices only represented 2.1% of all Android device shipments in 2018, 0.5% in 2019, and zero in 2020.

282 The MADAs also provide that OEMs may not distribute an Android compatible device without Google’s written approval prior to the launch of each device model, and OEMs must send the final software build of their Android compatible devices, as well as test reports confirming they pass the CTS, to Google for final approval. Google may provide approval in its sole discretion.

283 Let me deal with one other matter at this point.

284 To comply with decisions of the European Commission and the Turkish Competition Authority, different versions of the MADA apply in the EEA and Turkey.  Under these MADAs, Google Search and the Chrome app are not included in the GMS licence granted to OEMs for devices that are distributed in the EEA and Turkey. Instead, Google Search and Chrome are separately licensed to OEMs at no cost. Further, OEMs must pay a licensing fee to Google to license the remaining GMS, including the Play Store. Otherwise, these MADAs are in substantially similar form to the other MADAs.

The Google Mobile Services requirements

285 The GMS distribution licence granted under a MADA is expressed to be “subject to” the “GMS requirements”. These requirements are published by Google, which may change them at any time in its sole discretion. The GMS requirements include the following which I have mentioned above.

286 First, the requirement to pre-install all core applications, including the Play Store.

287 Second, the requirement to feature on the DHS the Google Search widget, the Play Store icon, and a folder labelled “Google” containing the core applications, although this does not apply to certain Samsung devices as explained above

288 Third, the requirement that every device preloads a particular Google-signed package installer in place of the default package installer available with open source Android.

289 Fourth, the requirement that the Play Store be placed in the top level of the apps menu and not in any folder.

290 Fifth, the requirement that devices must replace the global “Unknown Sources” setting with the “Allow app installs” setting, which grants the “REQUEST_INSTALL_PACKAGES” permission on a per app basis, allowing the granted app to provide other apps for installation. Devices must not implicitly grant install permission to any pre-installed or downloaded third-party apps that do not have a privileged installation permission.

291 Sixth, the requirement that devices comply with the CDD and pass applicable testing suites, including the CTS and the GMS testing suite, being a suite of tests to ensure compliance with the GMS requirements. Relevantly, a device will not comply with the GMS requirements or pass the GMS testing suite if a pre-installed app has been granted a privileged installation permission but has not been approved by Google for inclusion on an “allowlist”, that is, for the grant of a privileged installation permission.

The Android CDD and technical restrictions

292 The MADA incorporates the requirements of the CDD. It does so in two ways. First, it makes the GMS distribution licence subject to compliance with a valid and effective ACC. Second, it confines the GMS distribution licence to Android compatible devices.

293 The CDD contains requirements for the configuration, behaviour, and security of Android devices. Throughout the relevant period, among other things, the CDD included requirements as to when, and how, a device may install apps from “unknown sources”.

294 By requiring OEMs to assume and perform the contractual obligations referred to above and/or by reason of decisions made by Google in the course of determining whether the OEM’s devices have complied with the Android CDD and the GMS requirements, and have passed the CTS and the GTS to its satisfaction, Google has imposed on all OEMs who wish to distribute Android smart mobile devices with GMS a requirement that the OEM must configure their Android smart mobile device so that, if an Android smart mobile device user attempts to download an app from an unknown source including via their internet browser, the user is confronted with all or at least some of the following aspects or consequences.

295 First, a warning with words to the effect that the APK can harm their device and words that ask whether the user wishes to keep the APK anyway.

296 Second, if the user indicates that they do wish to keep the app, the download commences and, after the APK is downloaded, a statement that for their own security their device is not allowed to install unknown apps on their Android smart mobile device.

297 Third, an option to cancel the installation or to alter their device settings.

298 Fourth, if the user wants to proceed with the installation, a statement and/or functionality that they must go to the settings for their Android smart mobile device and alter their settings to allow the installation of “unknown apps”.

299 Fifth, on navigating to the settings menu, another warning that their Android smart mobile device is more vulnerable to attack by “unknown apps” and that by installing apps from this source the user agrees they are responsible for any damage to their device or loss of data that may result from the app.

300 Sixth, a step such that to proceed with the installation of the app, the user must then toggle the setting to indicate that they wish to make the change to the default setting.

301 Seventh, a step being that the user is then presented with an installation confirmation prompt that asks the user whether they want to install the app, and once the user acknowledges the prompt and presses “install” the installation proceeds.

302 The precise text the user is confronted with as a result of each of the matters referred to above can vary between OEMs, versions of the Android OS and, in instances of direct downloading, the browser through which the app is downloaded.

303 Further, any device had to be configured such that prior to the release of the Android 12 version of Android OS on 4 October 2021, it was not possible to automatically update apps downloaded from unknown sources; and to update an app downloaded from an unknown source the user had to repeat all or at least some of the steps outlined above.

304 Further, any device had to be configured such that since the release of the Android 12 version of Android OS on 4 October 2021, apps downloaded from an unknown source can be automatically updated, but subject to conditions which are not imposed on Android apps distributed via the Play Store.

305 As a result of all of these matters, OEMs are prevented from performing or permitting the following actions.

306 They are prevented from distributing Android smart mobile devices without configuring those devices as referred to above, such that all attempts to directly download an Android app, or to download an app from an alternative app store that is not pre-installed with a privileged install permission, will confront the restrictions referred to, but the restrictions referred to will not apply to downloads from the Play Store.

307 Further, they are prevented from allowing direct downloading of apps or downloading of apps from an alternative app store that has not been approved by Google for pre-installation with a privileged install permission on Android smart mobile devices without displaying warnings to users and requiring users to take steps to allow such downloading of apps outside the Play Store.

308 I have referred to the above matters collectively as the technical restrictions.

309 Now Epic says that by reason of the technical restrictions, OEMs that want to distribute devices with GMS installed must configure those devices so that, if a user attempts to download an app from a source that does not have a privileged install permission, the user cannot download and install that app without manually navigating through a series of warnings and additional steps or clicks.

310 The technical restrictions ensure that only Android apps and app stores which have a privileged install permission can install other apps via a one click process.

311 Where GMS Android is installed on a device, the Play Store is granted a privileged install permission and Google has ultimate control over which other apps or app stores (if any) are granted a privileged install permission. Google exercises this control to disallow the grant of privileged install permissions to apps and app stores of which it does not approve.

312 If a user wants to download and install an app from a source, be it a web browser or an app store, that has not been granted a privileged install permission, being sources Google refers to as “unknown sources”, the user must explicitly grant permission to install apps from that source by changing a setting on their Android device.

313 Epic says that the technical restrictions are designed to apply frictions based on the source of the app being installed, that is, the web browser or the app store, not based on the app being installed or the developer of the app being installed.  Epic says that this leads to absurd results.

314 So, if an Android device user seeks to install the Google Wallet app using the Google Chrome web browser, the technical restrictions apply and frictions are applied on the basis that Google Chrome is an “unknown source”.

315 But once the user changes their settings to allow Google Chrome to install the Google Wallet app, the technical restrictions will not apply for any app, however malicious, which the user subsequently finds on the web using the Google Chrome browser and seeks to install on their device.

316 There is nothing secure about this system, and evidently the technical restrictions are not aimed at protecting Android device users from potentially harmful apps. They are aimed at potential sources of app distribution, not at potentially harmful apps or at potentially nefarious app developers.

317 Until Android 8 was released in 2017, this required users to adjust their device’s settings once, following which the device would allow all so-called “unknown sources” to install apps without causing warnings to appear. With the launch of Android 8, this was changed to a “per source” permission, such that users have been required to adjust their device’s settings each time they wish to allow a new “unknown source” to install apps.

318 In practical terms, this means that although it is possible to download and install apps otherwise than from a privileged app store, users who attempt to do so encounter frictions in the form of explicit warnings and additional steps or clicks.  The number, form and sequence of these install frictions is described later.

319 So, to directly download the Amazon App Store to various Android devices, the user encounters three separate warnings.

320 In addition, on devices that post-date Android 8, the first time a user tries to download an app from the Amazon App Store which has been directly downloaded onto the device, they must undertake another three or four steps and encounter a fourth warning.

321 Further, because directly downloaded apps do not update automatically, and updated versions must be reinstalled from the original source, users have to undertake three of the same steps and re-encounter one of the same warnings each time they directly download an updated version of the app. Until October 2021, the same experience confronted users when seeking to update an app which they had obtained from a directly downloaded app store.

322 The number, form and sequence of the install frictions that result from the technical restrictions is described in the evidence. Taken together, they not only create a disincentive for users to directly download apps, by making the process more difficult and time consuming than downloading apps through the Play Store, but they are apt to cause users to think that directly downloading apps is harmful and to dissuade them from completing the process.

323 Epic says that the technical restrictions are also designed in such a way that apps being directly downloaded from well-known and safe sources are treated in the same way as truly “unknown sources”. In contrast, on Windows PCs, the degree of install friction varies depending upon the extent to which the developer of the app being installed has been verified as trustworthy by means of a certification process.

GMS non-technical requirements

324 Some of the GMS requirements form part of the technical restrictions which I have identified elsewhere and will not repeat here. As Epic has identified, the remaining relevant GMS requirements do not form part of the technical restrictions and are as follows.

325 Section 2.2 provides that new product launches must pre-install “all Core services and apps in the latest GMS bundle no later than 60 days after the latest bundle is released by Google.”

326 Section 3.1 provides that the DHS must feature the Google Search widget,  the Play Store icon and a folder labelled “Google” containing the core applications.

327 Section 3.3 provides that all GMS apps must be placed in the apps menu and that the Play Store should be placed in the top level of the apps menu and not in any folder.

328 In summary, OEMs may only distribute devices with GMS, including GMS apps and Google Play Services, if their devices comply with the GMS requirements, which include the following requirements.

329 First, to pre-install all core applications on the device, including the Play Store.

330 Second, to feature on the DHS of the device the Google Search widget, the Play Store app icon and a folder labelled “Google” containing the core applications.

331 Third, to pre-install the GooglePackageInstaller app, as a privileged app, in place of the default package installer available with the Android OS. A package installer is a system application that is responsible for handling the installation and uninstallation of apps on the device; the package installer confirms that an app that is attempting to install other apps has permission to do so, or has otherwise sought the consent of the user to install apps.

332 Fourth, to place the Play Store in the apps menu and, specifically, within the top level of the apps menu and not any folder on the apps menu.

333 Fifth, to replace the global “Unknown Sources” setting with the “Allow app installs” setting on Android devices, which grants the REQUEST_INSTALL_PACKAGES permission on a per app basis, allowing the granted app to provide other apps for installation. A device must not implicitly grant this permission to any pre-installed or downloaded third-party apps that do not have a privileged installation permission, that is, unknown sources.

Revenue sharing agreements and mobile incentive agreements

334 Let me at this point say something further concerning the RSAs and the MIAs.

335 Under the RSAs Google promises to pay OEMs a share of certain Google revenues, which are essentially advertising revenues generated from Google Search results displayed on the OEMs’ devices, in exchange for those OEMs configuring their devices in particular ways.

336 Prior to mid-2019, each of the selected OEMs had entered into RSAs.

337 RSAs also require OEMs to comply with their MADAs. Where an OEM breaches this obligation, Google is not required to make payments to the OEM under its RSA.

338 In this way, the RSAs serve to ensure that OEMs will prefer distributing Android devices over possible alternatives, and will comply with the restrictive terms imposed under the MADA, including the requirement to pre-install the Play Store on the DHS and the technical restrictions, and the ACCs.

339 Unlike ACCs and MADAs, it is not mandatory for an OEM to enter into an RSA. But most OEMs are incentivised to enter into a RSA in order to receive the lucrative revenue shares.

340 In November 2021, about 97% of Android devices that were active worldwide excluding China were manufactured by an OEM that had an RSA with Google.

341 Let me say something more about the RSA3s and Premier devices, which excited some attention during the running of the trial.

342 Until mid-2019, RSAs do not appear to have contained specific requirements relating to the Play Store or to provide for any sharing of revenue generated through the Play Store.

343 But between January and June 2019 Google developed its Google Distribution on Android Framework, which was approved by Google’s business council in June 2019 and which I have described elsewhere as simply GDAF.

344 RSAs were modified in accordance with this approved framework and the modified versions are referred to within Google as RSA3s.  At least six of the selected OEMs have entered into an RSA3.

345 Pursuant to the approved framework, there are three tiers of device configurations which the OEM may choose to adopt, with each tier being more prescriptive than the last, and with OEMs entitled to receive a greater share of Google revenues in respect of devices configured as Tier 3 or Premier devices, than for Tier 2 or Tier 1 devices.

346 Unsurprisingly, OEMs that have an RSA3 tend to configure most of their devices as Tier 3 or Premier devices.  As at October 2020, some 65% of RSA3 activations were Premier tier.

347 Significantly, under the RSA3s, Google promised OEMs who configured their devices as Premier devices a share of the revenue generated from the Play Store on those devices. In exchange, OEMs were required not to pre-install any alternative app store on their devices.

348 For a device to be classified as a Premier device and for OEMs to be entitled to a share of Play Store revenues generated from that device, OEMs must configure that device as follows.

349 First, the device must not include any pre-installed app that is substantially similar to the Play Store, as determined in the sole discretion of Google, nor any launcher, over the air prompt or functionality that has the primary purpose of providing access to any service substantially similar to the Play Store.

350 Second, the device must set the Play Store as the “default” app store or marketplace for apps and all other digital content including subscriptions.

351 Third, the icon for the Play Store must be placed on the DHS.

352 Fourth, the device must not be configured so as to allow the OEM or any third-party to promote, introduce or suggest alternative app stores to users.

353 Fifth, the OEM must comply with the Premier device program requirements document, which may be updated by Google. Effective 31 July 2020, the Premier device program requirements document provided that a Premier device must only pre-install an app that is available on the Play Store, that is offered via Play auto-install unless separately agreed to in writing by Google, that does not interfere with the Premier device program requirements, that does not overlap with the functionality or features of the Play Store and other specified GMS apps, and that does not contain privileged install permissions which facilitate the downloading of other apps.

354 Let me say something about payments under RSA3s.  RSA3s entitle an OEM to a proportion of one or two types of revenue.

355 The first type of revenue is net ad revenue, which is the revenue recognised by Google from ads displayed on a device’s results page and clicked by the user, minus REDACTED] of those revenues. These ads are displayed in response to a valid Search query from a series of agreed-upon Search access points such as Google.com or Google Assistant.

356 The second type of revenue is net play transaction revenue, which is the gross amount spent by device users on Play Store transactions, minus applicable taxes, adjustments, refunds, chargebacks and reversals, the portion paid to app developers, and 20% of the remainder. Play Store transactions are purchases of third-party apps and in-app purchases by a user within the Play Store on a Premier device.

357 The portion of these revenues payable to the OEM varies depending on whether a given device is configured in accordance with the requirements that apply to a Tier 1, Tier 2 or Tier 3 (Premier) device.

358 Now a share of Play Store revenues and the highest portion of advertising revenues are only payable under the relevant RSA3s in respect of Premier devices, that is, devices which do not include any pre-installed app that is substantially similar to the Play Store and do not promote or suggest any alternative app store.

359 Let me say something more about the MIAs and the Samsung incentive arrangements.

360 The MIAs are another form of incentive agreement, which have very similar features to the RSAs. Three selected OEMs have entered into MIAs rather than RSA3s, being Motorola, LG, and Samsung.

361 Under the Motorola and LG MIAs, very similar requirements to those specified in the RSA3s must be met for a device to qualify as a Premier device.

362 Relevantly, Motorola’s and LG’s MIAs entitle those OEMs to a monthly incentive payment if a minimum percentage of their devices activated that month are Premier devices. Further, the amount of that payment increases if the percentage exceeds higher thresholds. Further, the greater the proportion of activated devices that are Premier devices, the greater the amount payable by Google.

363 Motorola’s and LG’s entitlement to these incentives is conditional upon ongoing compliance with their MADAs.

364 As for Samsung, it has been party to an RSA but not an RSA3 with Google throughout the relevant period. The first was entered into in September 2017, and entitled Samsung to receive REDACTED] of net ad revenue earned from its devices, provided, among other things, that Samsung complied with the MADA and the AFA. An internal Google presentation indicates that, in 2019, Google was paying Samsung around USD 1.7 billion per annum under this RSA.

365 On the same day in November 2020, Google entered into both a replacement RSA and an MIA with Samsung. The second RSA was on relevantly similar terms to the first although with varying revenue shares. The MIA entitled Samsung to additional incentive payments, of between [REDACTED] of certain kinds of [REDACTED][REDACTED] earned from Samsung devices.  The MIA also entitled Samsung to a bonus payment of USD 150 million.

366 Samsung’s MIA is in a different form to that of Motorola and LG. It does not provide for different tiers of devices and nor does it impose restrictions on Samsung relating to the Play Store besides those prescribed in the MADA. But it does make Samsung’s revenue share entitlement conditional upon Samsung complying with its MADA.

367 Let me turn to what Epic has alleged to be the app developer restrictive conduct.

App developer restrictive conduct and the developer distribution agreements

368 The app developer restrictive conduct alleged by Epic includes the following aspects.

369 The first alleged aspect is Google’s conduct in requiring app developers that wish to distribute apps through the Play Store to enter into a DDA, and comply with the associated developer program policies, containing the terms referred to below.

370 Those terms preclude app developers from distributing an alternative app store via the Play Store. Further, they preclude app developers from using user information obtained via the Play Store to distribute apps outside the Play Store. Further, they require app developers offering Play Store app in-app purchases to use Google Play Billing for those transactions except in limited circumstances. Further, they require app developers to pay a fee of 30% on Play Store app in-app purchases subject to limited exceptions.  Further, they preclude apps distributed via the Play Store from leading users to a payment method other than Google Play Billing subject to limited exceptions. And further, they empower Google to terminate the DDA on 30 days’ notice (or less) if the above terms are breached.

371 So, Google required app developers that wished to distribute Android apps to Android smart mobile device users via the Play Store to agree to the following terms of the DDA and developer program policies.

372 Google Asia Pacific Pte Ltd is appointed as the app developer’s marketplace service provider in respect of apps distributed through the Play Store in Australia, and Google LLC is appointed as the app developer’s agent in respect of apps distributed through the Play Store in the US and various other countries.  This is for the purpose of making the developer’s apps available to Android smart mobile device users.

373 Further, app developers must not use the Play Store to distribute or make available any product that has a purpose that facilitates the distribution of software applications for use on Android smart mobile devices outside of the Play Store.

374 Further, app developers are prohibited from using user information obtained via the Play Store to sell or distribute products outside the Play Store.

375 Further, Google may, in its discretion, determine that an app or a portion of an app violates the DDA or an applicable policy, or may have an adverse impact on Google including economic impact, in which case Google may reject, remove, suspend, or limit the visibility of the relevant app on the Play Store, and/or suspend and/or bar the app or app developer from the Play Store or Android smart mobile devices.

376 Further, Google may terminate the DDA for any reason on 30 days’ written notice, or immediately upon written notice, or with 30 days’ prior written notice if required by law, where the app developer breached the DDA or other agreement relating to the Play Store or the Android platform and/or the app developer or its app posed a potential risk of economic, reputational or security-related harm to Google.

377 Further, Google is authorised to give users refunds in accordance with the Play Store refund policies and may deduct the amount of those refunds from payments to the app developer.

378 Further, app developers wishing to offer Play Store app in-app purchases must have a valid payments profile under a separate agreement with a payment processor and be approved by a payment processor for a payments profile, with payment profile being a financial service account or profile provided by a payment processor to an app developer that enables the payment processor to collect and remit payments on the app developer’s behalf for “Products” distributed by the Play Store, payment processor being a Google-affiliated entity providing services that enable the app developer to be paid for “Products” distributed via the Play Store, and “Products” being software, content, digital materials, and other items and services as made available by app developers via the Google Play Console.

379 Further, app developers must pay Google a fee of 30% on Play Store app in-app purchases with certain limited exceptions.

380 Further, all claims arising out of or relating to the DDA or the app developer’s relationship with Google under the DDA are expressed to be governed by the laws of the State of California, excluding California’s conflict of laws provisions. App developers agree to submit to the exclusive jurisdiction of the federal or state courts of the county of Santa Clara, California to resolve any legal matter arising from or relating to the DDA or the app developer’s relationship with Google under the DDA.

381 Further, app developers offering Play Store app in-app purchases must use Google Play Billing for those transactions.

382 Further, in the period from at least January 2021 in respect of new apps, and 30 September 2021 in respect of pre-existing apps, which could be extended to 31 March 2022 on request, an app distributed via the Play Store must not lead users to a payment method other than Google Play Billing subject to certain exceptions.

383 Further, an app distributed via the Play Store must not modify, replace or update itself using any method other than the Play Store’s update mechanism, download executable code from a source other than the Play Store or allow users to install another app.

384 Further and as I have already indicated, where Google considers that an app developer or app has not complied with the terms of the DDA, including the developer program policies, Google may reject, remove, suspend, or limit the visibility of the app from the Play Store, suspend and/or bar the app and/or the app developer from the Play Store or Android smart mobile devices, and/or terminate the DDA.

385 Further, app developers who wish to offer Play Store app in-app purchases are required to enter into an agreement with Google Payment Australia Pty Ltd in respect of the use of Google Play Billing as the method of payment in Australia.

386 The second alleged aspect is Google’s conduct in entering into the GVP agreements and the AVP agreements with certain app developers.

387 Google entered into agreements with approximately 20 games app developers which required those app developers to do the following.

388 They had to release, or in the case of some app developers, use reasonable endeavours to release, Android apps on the Play Store at the same time as, or before, the release of the same app, that is, an app of the same title, on any other mobile distribution channel.

389 Further, for certain of those app developers, they had to ensure that the Play Store was not disadvantaged in terms of the quality and promotion of the developer’s apps.

390 Further, they had to ensure that the core content, features and functionality, or, in the case of some app developers, that the core game content and quality, of their Android apps released on the Play Store were the same on other mobile app stores or, in the case of some app developers, other comparable platforms.

391 Further, for certain of those app developers, they had to not remove any of their Android apps from the Play Store unless required by law or in other limited circumstances specified in the agreements.

392 Further, for certain of those app developers, they had to make good faith efforts or, in the case of one app developer, reasonable efforts, to include their apps in pre-registrations, open betas, and other early access programs on the Play Store.

393 Further, Google entered into AVP agreements with at least 4 app developers which required those app developers to act or refrain from acting in the following fashion.

394 First, they were not to remove specified Android apps from distribution on the Play Store unless required by Google or by law.

395 Second, they had to ensure that the specified Android apps had the same core content and features as are made available in the same app or service on platforms other than the Play Store.

396 Third, they had to promote its apps on the Play Store in the same or equivalent manner as it promotes the same app or service on platforms other than the Play Store.

397 Fourth, they could not make exclusive content available on the same app or service as those specified Android apps on platforms other than the Play Store, unless it is not technically feasible to make such content available on the Android OS.

398 Fifth, they were to make available to Play Store users the same SKUs or service tiers in each specified Android app as are made available to users on other platforms.

399 Sixth, with the exception of one developer, they had to ensure that the subscription and/or purchase price applicable to each specified Android app on the Play Store is at least as favourable as the subscription and/or purchase price of the same app or service on other platforms.

400 Seventh, they had to use Google Play Billing as the sole payment solution in the specified Android apps across all devices in all markets in which those apps are available on the Play Store.

401 The third alleged aspect is Google’s conduct in giving effect to and/or enforcing the terms referred to above including in relation to Epic, for example by suspending Epic’s publishing status on the Play Store, removing Fortnite from the Play Store, refusing to allow Epic to offer EDP for any Play Store app in-app purchases and failing to agree to Epic’s request for in-principle approval to make the Epic Games Store available on the Play Store.

402 Let me now turn to saying something more about the relevant agreements concerning app developers.

The developer distribution agreements

403 In addition to its agreements with OEMs, Google has imposed various contractual restrictions on app developers, primarily through the DDA.   I have mentioned some of this earlier, but let me set out the detail.

404 Any app developer that wishes to distribute Android apps through the Play Store must enter into a DDA. Its terms are non-negotiable, and clause 15.1 empowers Google to vary them at any time.

405 Now the relevant terms of the DDA may be divided into three categories.

406 First, one has terms governing the relationship between Google and the developer, as well as the service supplied under the DDA.

407 Second, one has a term prohibiting the distribution of rival app stores such as clause 4.5.

408 Third, one has terms relating to Google’s service fee and payments policy.

409 The DDA is a contract between three separate Google entities and the developer. Those entities are Google LLC, Google Commerce Limited, and Google Asia Pacific Pte Limited.

410 Pursuant to clauses 2.1 and 3.1 of the DDA, the developer makes the following appointments.

411 First, the developer appoints Google LLC as its “agent” to make its products available in the Play Store to users in various countries and territories predominantly located in North and South America; there is a definition of “Products” that I will identify in a moment.

412 Second, the developer appoints Google Commerce Ltd as its “agent” to make its products available in the Play Store to users in various countries and territories predominantly located in Europe, the Middle East and Africa.

413 Third, the developer appoints Google Asia Pacific Pte Ltd as its “marketplace service provider” to make its products available in the Play Store to users in various countries and territories predominantly located in Asia and the Pacific, including Australia.

414 Each of those appointments expressly states that “the purchase and sale of [the developer’s] Product(s) is governed by a contract of sale directly between [the developer] and users”, save where Google Commerce Ltd is the merchant of record for sales in specified countries in Europe, the Middle East and Africa. So, neither Google LLC nor Google Asia Pacific Pte Ltd effects any contracts of sale on behalf of the developer under the DDA.

415 Let me then turn to some of the specific provisions.

416 “Products” are defined by clause 1 as “Software, content, digital materials, and other items and services made available by Developers via the Play [Store]”. They include apps and digital content purchased within the app.

417 Clause 2.1 refers to the developer’s “use of Google Play to distribute Products” and provides that Google will “display and make [the developer’s] Products available for viewing, download, and purchase by users”.

418 Clause 3.1 similarly provides that the developer appoints Google as its agent or marketplace service provider (as the case may be) “to make [the developer’s] Products available in Google Play”.

419 “Google Play” is defined as the “software and services, including the Play Console, which allow Developers to distribute Products to users of Devices”. Various other provisions refer to products being “distributed” or “made available” via the Play Store.

420 Clause 3.2 stipulates that the DDA “covers both Products that users can access for free and Products that users pay a fee to access”. It then provides that “[i]n order for [the developer] to charge a fee for [the developer’s] Products and to be paid for Products distributed via Google Play, [the developer] must have a valid Payments Profile under a separate agreement with a Payment Processor”.

421 A “Payments Profile” is defined as a “financial service account or profile provided by a Payment Processor to a Developer that enables a Payment Processor to collect and remit payments to the Developer on the Developer’s behalf for Products sold via Google Play”.

422 As for a “Payment Processor”, it is defined as a “Google affiliated entity providing services that enable Developers to receive payments for Products sold via Google Play.”

423 So, the terms of the DDA make it clear that the “services that enable Developers to receive payments for Products sold via Google Play” are provided by payment processors “under a separate agreement”. Those services are not provided by Google LLC, Google Commerce Ltd or Google Asia Pacific Pte Ltd.  Nor are they provided under the DDA.

424 But what is provided under the DDA is the distribution service constituted by the Play Store. Google itself defines the Play Store as the “services … which allow Developers to distribute Products to users of Devices”.

425 Pursuant to clauses 2.1 and 3.1 of the DDA, the developer appoints Google as its agent or “marketplace service provider” to make the developer’s products available for viewing, download and purchase by users in the Play Store. “Products” are defined by clause 1 as “Software, content, digital materials, and other items and services made available by Developers via the Play [Store]”. They include apps and digital content purchased within the app.

426 Where Google determines in its sole discretion that an app developer’s product violates the DDA or the developer program policies, or may have an adverse impact on Google, Google is entitled under clause 8.3 of the DDA to “reject, remove, suspend, limit the visibility of a Product” on the Play Store and/or suspend and/or bar any product and/or app developer from the Play Store or from Android smart mobile devices.

427 Clause 4.1 of the DDA requires app developers and their products to adhere to the developer program policies. Two of those policies are presently relevant, being the device and network abuse policy and the payments policy. The latter compels developers to use Google Play Billing for in-app purchases of digital content.

428 A breach of either policy entitles Google to, among other things, remove the developer’s app from the Play Store and terminate the DDA immediately or on 30 days’ notice.

429 Clause 4.5 of the DDA prohibits app developers from using the Play Store to distribute or make available any product that has a purpose that facilitates the distribution of other Android apps for use on Android devices outside of the Play Store.

430 In other words, clause 4.5 prohibits an app store or any app that facilitates the distribution of another app from being distributed through the Play Store.

431 This prohibition has existed in this form since 25 September 2014. Prior to that date, clause 4.5 only prohibited products whose primary purpose was to facilitate the distribution of Android apps.  Prior to this change there were apps on the Play Store that distributed other apps. Google altered the terms of clause 4.5 which had the effect of stifling emerging competition for the Play Store in app distribution.

432 Clause 4.9 of the DDA prohibits app developers from using user information obtained via the Play Store to sell or distribute products outside the Play Store. This prohibition has been present in all versions of the DDA since at least June 2014. It precludes app developers from using information about their own customers. It is problematic given that Google purports to be the app developer’s agent or marketplace service provider.

433 Clause 3.8 of the DDA authorises Google to give users refunds in accordance with the Play Store’s refund policies, and to deduct the amount of any refunds from payments due to the developer. Google’s refund policies generally provide that Google will administer refunds within the first 48 hours and thereafter users should contact the developer.

434 By the combination of clauses 2.1, 3.1, 3.8 and 4.9 of the DDA, Google interposes itself between the developer and its customers, with the result that the control of important interactions is assumed by Google. The outcome of refund applications, for example, will often depend on Google’s refund policy and its application thereof, not the developer’s.

435 More broadly, these terms limit the developer’s ability to form a direct commercial relationship with its customers, and to expand that relationship beyond the Play Store.

436 For app developers, this direct relationship with users of their apps is important. Differing levels of control over the customer relationship is one of the ways in which alternative providers of app distribution services and in-app payment solutions for Android apps would be likely to compete with the Play Store and Google Play Billing, if Google’s conduct did not foreclose or impede such competition.

437 Clause 3.4 of the DDA requires app developers to pay Google a service fee, being a percentage of the price of Android apps purchased through the Play Store and of in-app purchases of digital content. The service fees payable by app developers under the DDA have varied over time, as set out in the following table:

438 As the table shows, the service fee varies depending on the app developer’s revenue and the type of digital content offered for sale. Generally, app developers now pay a service fee of 15% on the first USD 1 million of revenue earned by the developer each year, 30% of revenue in excess of USD 1 million of revenue each year, and 15% of revenue for automatically renewing subscriptions.

439 Developers that are eligible for special programs pay a 15% or lower service fee. These programs generally target media companies, or developers offering video, audio or music subscriptions, or digital books.

440 Further, Google charges different service fees that are 4 percentage points lower where app developers use an alternative payment solution in South Korea or India, 3 percentage points lower where app developers of non-gaming apps use an alternative payment solution in the EEA, and 4 percentage points lower where app developers of non-gaming apps use an alternative payment solution as part of Google’s pilot program for user choice billing.

441 Now the service fee is only payable on apps and in-app products sold through the Play Store’s billing system or through an alternative billing system, the latter being a system permitted under Google’s user choice billing (UCB) or developer only billing (DOB) programs.

442 The service fee payable by app developers under the DDA has varied over time to time, as summarised in the table that I have set out earlier.

443 Generally, developers now pay a service fee of 15% on the first USD 1 million of revenue earned by the developer each year; a service fee of 30% of revenue in excess of USD 1 million of revenue each year; and a service fee of 15% of revenue for automatically renewing subscriptions regardless of the revenue earned by the developer.

444 Developers eligible for special programs pay a 15% or lower service fee. These programs generally target media companies, or developers offering video, audio or music subscriptions, or digital books.

445 In addition, Google charges a varied service fee in circumstances where UCB or DOB apply.

446 Outside of special programs for eligible developers, UCB, and DOB, Google has made only limited changes to the service fee that it charges for in-app purchases of digital content over time. When in-app purchases were first introduced to the Android Market in 2011, the service fee for such purchases was 30%.  Some 13 years later, the service fee is still 30%, save for the first USD 1 million in earnings each year and subscriptions, both of which are charged at 15%.

447 Google’s service fees are similar to those generally charged by Apple, although there are two differences.

448 First, Apple charges a 15% fee to developers whose earnings were under USD 1 million in the prior year, as opposed to charging all developers 15% on their first USD 1 million.

449 Second, Apple’s fee for subscriptions is 30% in the first year and 15% thereafter.

450 Let me set out a little history concerning the changes.

The subscription pricing change in 2018

451 The first change to Google’s subscription pricing occurred with effect from 1 January 2018, when Google reduced its service fee for subscriptions that had been active for more than one year from 30 to 15%.

452 Google witnesses said that this change was driven by a need to remain price competitive with Apple, which had reduced its fee on subscriptions active for more than one year from 30 to 15% some 18 months earlier, with effect from 13 June 2016.

453 But the document in which this change was proposed, a document that Mr Feng co-presented, gave five reasons for the change, only one of which related to Apple.  The other four reasons for the change listed in the document included encouraging subscription developers to adopt Google Play Billing and easing the path to a potential Play Store payment policy change, being the removal of the external consumption exemption. As to Apple, the document merely said, “Bring Play’s pricing for subscription SKUs in line with market – Apple offers 30% for 12 months, 15% thereafter”.

454 Google’s behaviour did not amount to competing with Apple on price, or on any other metric. That is a fortiori when the price adjustment was made some 18 months after that of Apple. And it is especially true in this case, given the other reasons for the change. Moreover, bringing one’s pricing into line with the market is not competing on price.

The service fee reduction for small developers and a further subscription pricing change in 2021

455 On 18 November 2020, Apple announced a small business program, whereby its service fee was reduced to 15% for those developers who earned less than USD 1 million in the prior year.

456 With effect from 1 July 2021, Google reduced its service fee to 15% for the first USD 1 million earned each year by all developers. In October 2021, Google reduced the service fee for subscriptions to 15% in the first year.

457 Although the reasons for the change are contentious and a topic to which I will return later, it is convenient to note here that in my view one of the motivations for making the first of those changes was the pressure Google was facing from regulators.

458 In an internal document dated 16 November 2020 titled Project Runway: Proposal for changes to Play business model, Google set out its hypotheses, including that “[w]e have an a la carte menu of additional levers to pull in a set of forecasted PR/regulatory scenarios for further discussion”, and “[w]e should proactively make moderate changes in 2021 to forestall more drastic required changes”.

459 Under the heading “Scenario Planning”, the following scenarios were contemplated.

460 The first stated:

* Should we be REQUIRED to adjust our service fee, we would need to re-imagine the store promotional strategy and GPB strategy.

* If government mandated requirement to make play billing optional, we respond with aggressive store changes and perhaps lower service fee …

461 The second stated “[i]f government mandated price for play billing (say 20% or cost + 5%) (WIP, needs more thinking)…”.

462 The third stated:

* We c/should VOLUNTARILY adjust our service fee now to improve optics, align value delivered with fees charged, and reduce agitation --

…

* We recommend the first magical bridge option (30% for the first year after user purchase, 15% thereafter) …

…

* Match Apple’s discount for small developers and perhaps be more aggressive by tiering it lower in emerging markets.

…

463 Further, the document makes no reference to a supposed competitive dynamic with Apple as somehow necessitating or justifying its proposals. Rather, the focus is on “forestall[ing] more drastic required changes” by way of regulation and “optics”.

464 On 21 November 2020, the day of Apple’s small business announcement, Mr Samat prepared a document which commenced with a summary of Apple’s small business program announcement, including reactions from the press and developers. The document then provides an overview of the reaction from regulators, which includes that Apple’s announcement had been “informally described … as too little too late and will not help Apple’s position in the Digital Markets Act debate in Europe” and that Google was “expecting to receive” a request from South Korean legislators to “follow Apple’s lead”.

465 The document then set out Google’s “take”, as follows:

Waiting for Apple to make the first move has been our plan. Our strategy has been to wait to make any public revenue share changes until Apple moved first. We did not want to move downward on rev share while Apple stayed at 30%, nor did we want to give them the opportunity to “one-up” any announcement we made and cause us to need to react again.

We have a very similar program scoped out…

We will take the next several weeks to gain more feedback on how Apple’s changes landed. The benefit of going second here is we have the opportunity to be more thoughtful and take into account the reaction. …

466 So, Google wanted Apple to lower its prices and was not at all concerned that this would cause Google to lose customers.

467 Google’s motives are also confirmed by a 2 December 2020 document titled Runway January 2021 Announcement Proposal and Status, which records that Google had “decided to announce in January 2021 an adjustment to our fee schedule, focused on providing relief to small developers and defusing some of the regulatory pressure experienced in several countries”.  Defusing regulatory pressure was the real reason for this change.

468 The following “goals” are set out in the document:

Defuse immediate regulatory pressure

Be seen as helping people build their business…

Reduce the new perceived gap between Apple and Play’s billing policies

Continue to allow Apple to make first moves on service fee structure changes

469 The penultimate “goal” was not about price competition but about the regulatory pressure that had arisen due to the “gap between Apple and Play’s billing policies”. That it was not about price competition is confirmed by the last goal.

470 The document then outlines the “Desired PR and Regulatory Landings” and says “our message will be evaluated based on how many devs pay less. Press will line us up vs Apple”. Again, it was all about “PR and regulatory landings”, not winning customers from Apple.

471 Further, a drafting comment records:

… we need a headline PR/reg message – esp given that the message is the primary goal in this effort … Arguably the most believable, as well as most helpful reason is competitive pressure/response…Helpful to our case – if we’re trying to create the meme intense competitive pressure.

472 In other words, competition was a “message” or “meme” Google was “trying to create”. If competition were the reality, this would not have been necessary.

473 I will return to the evidence of some of the Google witnesses on these questions much later in my reasons.

Google Play Billing — the DDA and the payments policy

474 Google Play Billing is the payment solution that Google requires app developers and consumers to use, and is the only payment solution that Google permits app developers and consumers to use, for accepting and processing payments for Play Store in-app purchases.

475 Google Play Billing, as a payment solution, provides for and facilitates the acceptance and processing of payments from Android smart mobile device users for apps or in-app purchases.

476 It does this by, among other things, providing for the acceptance and collection of payments from Android smart mobile device users, all necessary engagements with credit providers, payment processors and financial institutions, the deduction and payment of any relevant fees or commissions, and the payment of the remaining balance to the app developer.

477 Google Play Billing is charged by Google Payment Australia Pty Ltd, or alternatively by Google LLC, Google Asia Pacific Pte Ltd and/or Google Payment Australia Pty Ltd, at a 30% commission to app developers for Play Store in-app purchases or, in some limited circumstances, 15%.

478 Clause 4.1 of the DDA provides that the developer must adhere to the developer program policies, which include Google’s payments policy.

479 The payments policy provides that developers charging for app downloads from the Play Store must use Google Play Billing.

480 The payments policy provides that Play Store-distributed apps requiring or accepting payment for access to in-app features or services, including any app functionality, digital content or goods, must use Google Play Billing unless section 3 or section 8 of the payments policy applies.

481 Section 3 stipulates that Google Play Billing must not be used where the payment deals with any of the following items or matters.

482 First, it must not be used where the payment is primarily for the purchase or rental of physical goods, for the purchase of physical services or a remittance in respect of a credit card bill or utility bill.

483 Second, it must not be used where the payment includes peer-to-peer payments, online auctions or tax exempt donations.

484 Third, it must not be used where the payment is for content or services that facilitate online gambling.

485 Fourth, it must not be used where the payment is in respect of any product category deemed unacceptable under Google’s payments centre content policies.

486 Section 8 was first added to the payments policy in late 2021. The provision was initially limited to in-app purchases effected by users in South Korea, but its scope was enlarged with the subsequent introduction of inter-alia the UCB pilot and the DOB program.

487 These programs permit developers to offer users an alternative billing system in countries where UCB or DOB are available, but only if they comply with requirements prescribed by Google.

488 In summary then, except where developers are participating in UCB or DOB, the payments policy relevantly requires developers to exclusively use Google Play Billing for in-app purchases of digital content within apps distributed via the Play Store. The parties before me referred to this as the payments tie.

489 And as I have said, the payments policy does not require the use of Google Play Billing for in-app purchases of physical goods or services. In fact it prohibits such use.

490 And since the service fee is only payable on apps and in-app products sold through Google Play Billing or an alternative billing system, no service fee is payable on in-app purchases of physical goods or services. And nor is a service fee payable in respect of apps which are distributed without charge and without charging for in-app purchases, even if the developer monetises the app in a different way, for example, with paid advertising.

491 There is also an exemption from the requirement to use Google Play Billing in around 15 countries where Google Play Billing is not offered. Developers can use their own billing systems for purchases made in those countries.

492 Now it is convenient at this point to say something about the external consumption exception.

493 Between August 2013 and 20 January 2021, the payments policy contained another exception, whereby developers offering additional content, services or functionality within a non-gaming app downloaded from the Play Store were not required to use Google Play Billing where the payment was for digital content or goods that may be consumed outside of the application itself. The parties have referred to this as the external consumption exception.

494 With effect from 20 January 2021, Google removed the external consumption exception from the payments policy.  Google described this as a “clarification”, but this was an unsatisfactory description.  The change was a unilateral change to the contractual obligations of developers which required them to use Google Play Billing and to pay a service fee on in-app purchases of digital content that, until then, had been exempt from these requirements.  Existing apps that used an alternative payment solution were given until 30 September 2021 to comply with the amended payments policy.  I have discussed this so-called Orwellian-like “clarification” later.

495 Google announced that from 1 June 2022, any app developer that was non-compliant with the amended payments policy would be removed from the Play Store.

496 Before Google removed the external consumption exemption from its payments policy in January 2021, around 5,000 developers had chosen to use a payment solution other than Google Play Billing to facilitate in-app purchases of digital content within Play Store apps. Google expected more to follow.

497 Google expected contagion or a domino effect, with an increasing number of developers having recourse to alternative payment solutions.

498 I will return to the topic of the external consumption exemption much later in my reasons.

499 But I would say here that in my view the major reason Google removed the external consumption exemption was to raise more revenue by requiring more developers to use Google Play Billing rather than an alternative payment solution.

500 Now the requirement to use Google Play Billing is not imposed for security reasons. If it were, that would surely have been foremost among the considerations in favour of removing the external consumption exception and it would not have taken several years for Google to alter its policy.

501 Moreover, Google positively requires alternative payment solutions to be used for in-app purchases of physical goods and services, and permits them where UCB and DOB apply. That conduct is not consistent with a genuine concern that such alternative solutions are inadequate from a security perspective.

502 Now many developers affected by the removal of the external consumption exception preferred to use an alternative payment solution and did not want to use Google Play Billing.

503 Google has suggested that this is explained by a desire to avoid paying the service fee. But many developers considered Google Play Billing to be a deficient product that was harmful to their business.

504 The evidence shows that many developers wanted to use an alternative payment solution because the relevant alternative better suited their business and payment solution needs.

505 Developers wanted better forms of payment (FOP) coverage, better involuntary churn rates, and the ability to maintain a direct relationship with their customers, including the ability to respond to questions and complaints. Developers wanted to use the same payment solution for in-app purchases of both digital and physical products and to use the same payment solution across more than one distribution channel. Developers wanted better access to data about their users and the ability to bill customers at the time they chose and for a price they chose. And developers wanted the ability to adjust the payment solution features that they acquired from one country to the next, based on local market conditions.

506 The desire amongst some large developers to use an alternative payment solution was so great that they were willing to use an alternative payment solution even if there was no reduction in Google’s service fee.

The anti-steering  prohibition

507 On 20 January 2021, Google amended the payments policy to prohibit apps that were required to use Google Play Billing from leading users to a payment method other than the Play Store’s billing system.

508 This prohibition expressly extends to leading users to other payment methods via an app’s listing in the Play Store, in-app promotions related to purchasable content, in-app webviews, buttons, links, messaging, advertisements or other calls to action, and in-app user interface flows, including account creation or sign-up flows.

509 The parties before me referred to this prohibition as the “anti-steering rule”.

510 Google has enforced the anti-steering rule.

Device and network abuse policy

511 Clause 4.1 of the DDA requires app developers and their products to comply with the developer program policies. One such policy is the device and network abuse policy, which stipulates that an app distributed via the Play Store must not “modify, replace, or update itself using any method other than Google Play’s update mechanism”, and must not “download executable code … from a source other than Google Play”, or allow users to install another app.

512 This restriction plays a similar role to that of clause 4.5 of the DDA, insofar as it precludes the distribution of apps that distribute other apps. But the device and network abuse policy does more than this.

513 Developers tend to update their apps regularly such as to effect improvements, to make new features available to users, and to resolve software glitches. Yet the device and network abuse policy precludes developers from updating their own apps, except through the Play Store, even after a user has selected and downloaded the app from the Play Store onto their own device. The result is that an app distributed through the Play Store is tied to the Play Store.

514 Google, a so-called agent of app developers, has inserted itself, permanently, between those developers and users of their apps, and the developers remain subject to Google’s control.  A decision by Google to remove their app from the Play Store or to terminate the DDA will make it impossible for the developer to update apps previously downloaded from the Play Store, and will preclude the developer from continuing to serve its users.

515 Further, clause 4.9 of the DDA prohibits app developers from using information obtained via the Play Store to contact those users and suggest that they might want to reacquire the app outside the Play Store.

516 I have kept these realities in mind when evaluating Google’s arguments as to the alternative choices available to app developers. How many of them can truly afford to do other than bend the knee and pay whatever fee, or accept whatever condition, Google is minded to impose, when the alternative is to abandon every user the developer has ever acquired via the Play Store and hope to reacquire them somewhere else within the Android ecosystem that Google controls?

Other provisions of the DDAs

517 Let me address some other provisions of the DDA.

518 First, clause 15.1 empowers Google to vary the terms of the DDA at any time.  This is a power which Google has exercised in the past, including to make material changes.

519 One example is Google’s alteration of its payments policy, which Google changed to require developers, for the first time, to pay a service fee on purchases of digital content that could be consumed outside the app.

520 Another example is Google's amendment to clause 4.5, by which Google expanded the class of apps with distribution capabilities that were prohibited from being distributed through the Play Store.

521 Second, clause 15.3 provides that if developers do not agree with Google’s changes to the DDA, their “sole and exclusive remedy” is to terminate their use of the Play Store. In other words, developers must take it or leave it. And if they leave it, developers will lose the ability to distribute or update their apps previously distributed through the Play Store. Current users of their apps will not be able to make any in-app purchases within those apps.

522 Third, clause 10.3 permits Google to terminate the DDA “for any reason” on 30 days’ notice where “allowed under applicable law”. The same clause permits Google to terminate the DDA “immediately upon written notice or with thirty (30) days prior written notice if required under applicable law” if the developer has “breached any provision” of the DDA or the developer or its product “pose a potential risk for economic, reputational, or security-related harm to Google, users, or other third-party partners.”

523 Termination of the DDA will cause the developer to lose the ability to distribute or update their apps previously distributed through the Play Store, and current users of their apps will not be able to make any in-app purchases within those apps.

524 Any user that downloaded the app from the Play Store will no longer have an app that updates, or that facilitates in-app purchases. Before they can obtain an up-to-date version of the app, or make an in-app purchase, users will have to find the app elsewhere and download it again.

525 Fourth, clause 8.3 provides that Google may “suspend and/or bar any Product and/or Developer from Google Play” if inter-alia Google “determines in its sole discretion that a Product … violates this Agreement, applicable policies, or other terms of service, as may be updated by Google from time to time”. A suspension results in the developer’s app being removed from the Play Store, such that it is no longer available for download by users, it cannot be updated, and current users are unable to make any in-app purchases within the app.

526 Fifth, clause 4.9 prohibits app developers from using user information obtained via the Play Store to sell or distribute products outside the Play Store. This precludes developers from using information about their own customers. It is incongruous, given that Google purports to be the developer’s agent or marketplace service provider.

527 Sixth, clause 3.8 authorises Google to give users refunds in accordance with the Play Store’s refund policies, and to deduct the amount of any refunds from payments due to the developer. Google’s refund policies generally provide that Google will administer refunds within the first 48 hours and thereafter users should contact the developer.

528 An exercise of Google’s power to terminate the DDA or to suspend apps from the Play Store is likely to have a material adverse impact on most developers that have been distributing their apps through the Play Store for any length of time. Affected developers would lose their Play Store customers and their in-app purchase revenues from those users.

529 Those powers, which include a right to terminate for any reason on 30 days’ notice, confer considerable leverage or bargaining power on Google vis-à-vis developers.

530 Likewise, Google’s right to alter the terms of the DDA confers considerable leverage or bargaining power on Google vis-à-vis developers, whose only “remedy” is to terminate the DDA and lose all their Play Store users and associated revenues.

The Games Velocity Program agreements

531 The origins of the GVP agreements are sourced to Project Hug which I will describe later.

532 The GVP agreements were structured as addenda to existing DDAs between Google and the relevant app developers. Under the GVP agreements, in exchange for benefits provided by Google, which included marketing support, advertising credits, and a share of Google’s Play Store revenue from the developer’s apps, to be provided in the form of Google Cloud credits, the developers undertook to do various things.

533 First, they undertook to simultaneously launch new titles on the Play Store and other mobile distribution channels.  This has been described by the parties and I will describe it as sim-ship.  Google described this as the key criteria of the GVP.

534 Second, they undertook to ensure that Play Store users were not materially disadvantaged compared to users on other mobile distribution channels, with regard to core game content, features and functionality. In other words, there had to be content parity.

535 Third, they undertook to ensure that titles were not removed from the Play Store unless the developer was required to do so by applicable law, was required to do so by Google, or there was agreement between Google and the app developer to do so.

536 Fourth, they undertook to maintain game quality, with a prescribed minimum user rating.

537 Fifth, they undertook to make best efforts to include titles in pre-registrations, open betas and other early access programs in the Play Store, and to participate in strategic programs.

538 In addition, the Tencent GVP agreement required Tencent to ensure that prices for its Android apps on the Play Store, including the purchase price, any subscription prices and the price for any in-app purchases, were at least as favourable as the prices on any other platform.

539 Further, Tencent was required to make available all product and service offerings applicable to its Android apps within the Play Store that are available on any other platform, unless a product or service offering is not compatible with the Play Store, or it is not technically feasible to make a product or service available within the Play Store.

The Apps Velocity Program agreements

540 Let me say something more about the AVP agreements although I have already touched on them.

541 In early 2020, Google decided to extend Project Hug, which I will detail later, to 20 top non-game app developers. The extended Project Hug proposal, which became the AVP, was presented to the business council in February 2020. It was approved on 6 March 2020.  Google thereafter entered into AVP agreements with four developers.

542 The AVP agreements were structured as addenda to the existing DDAs between Google and the relevant developers. Under the AVP agreements, Google agreed to provide certain benefits to app developers such as marketing support and credit for purchasing advertisements through Google, in exchange for the app developers agreeing to various restrictions.

543 First, the app developers’ designated apps were required to use Google Play Billing as the sole payment solution across devices in all markets in which the designated apps were available on the Play Store.

544 Second, the app developers’ designated apps were required to have the same core content and features as any comparable product, being an application or service that is the same as the applicable title or any in-app purchase product or subscription product that can be used in titles, except that it is made available by the app developer on platforms other than the Play Store, and to promote apps in the same or equivalent manner.

545 Third, the app developers’ designated apps were required to make available the same “SKUs or service tiers” on the Play Store as other platforms. Further, some of the developers were required to ensure that the subscription or purchase price applicable to each designated app on the Play Store was at least as favourable as the subscription or purchase price of comparable products.

The mobile OS licensing and Android mobile app distribution markets — section 46

546 I have laid out what is said to be the impugned conduct and the relevant contractual matrix.  Let me now turn more generally to Epic’s s 46 claim against Google and begin by again identifying Epic’s posited markets.

547 Epic says that there is a mobile OS licensing market, which is a market for the supply of licenses for mobile operating systems, where the geographic dimension of the market is a global market excluding China, of which Australia forms part.

548 Further, Epic says that there is an Android mobile app distribution market, which is a market for the supply of services for the distribution of Android apps to Android smart mobile device users, with the following elements or dimensions.

549 First, the services consist of the provision of services to app developers enabling or facilitating app developers to reach, offer and provide Android smart mobile devices users with Android apps and associated updates and/or the provision of services to Android smart mobile devices users enabling or facilitating Android smart mobile devices users to be presented with or to find, obtain and utilise Android apps and associated updates.

550 Second, the geographic dimension of the market is Australia or a global market excluding China, of which Australia forms part.

551 Further, Epic says that there is an Android in-app payment solutions market, which is a market for the supply of services to app developers for accepting and processing payments for the purchase of digital content, including by way of subscriptions, within an Android app, with the following elements or dimensions.

552 First, the services consist of the provision of services to app developers enabling and/or facilitating app developers to accept and process payments for the purchase of digital content, including by way of subscriptions, within an Android app.

553 Second, the geographic dimension of the market is Australia or a global market, excluding China, of which Australia forms part.

554 Further, Epic says that in the alternative to an Android mobile app distribution market, there is a mobile app distribution market, which is a market for the supply of services for the distribution of Android apps and native apps written for iOS to smart mobile device users, with the following elements or dimensions.

555 First, the services consist of the provision of services to app developers enabling or facilitating app developers to reach, offer and provide smart mobile device users with native apps and associated updates and/or the provision of services to smart mobile device users enabling or facilitating smart mobile device users to be presented with or to find, obtain and utilise native apps and associated updates.

556 Second, the geographic dimension of the market is Australia or a global market excluding China, of which Australia forms part.

557 Now Epic’s case is that Google has contravened s 46(1) of the CCA by engaging in conduct that had the purpose, and the effect or likely effect, of substantially lessening competition in the Android mobile app distribution market.

558 Now there are two markets which are relevant to this allegation.

559 First, there is the mobile OS licensing market, which is a market for the supply of mobile OS licenses to OEMs.

560 Second, there is the Android mobile app distribution market, which is a market for the supply of services for the distribution of Android apps.

561 Epic’s case is that Google has a substantial degree of power in each of these markets and has engaged in conduct for the purpose, and which had the effect or likely effect, of substantially lessening competition in the Android mobile app distribution market.

The purpose of Google’s conduct in the Android mobile app distribution market

562 Epic says that a substantial purpose of Google in engaging in the OEM restrictive conduct and the app developer restrictive conduct, which in this context includes the payments tie and the anti-steering rule, was to substantially lessen competition in the Android mobile app distribution market.

563 It is said that Google’s anti-competitive purpose ought to be inferred from the OEM restrictive conduct and the app developer restrictive conduct themselves.

564 Further, it is said that internal business records of Google demonstrate that a substantial purpose of the impugned conduct was to prevent or hinder competition for the Play Store.

The effect of Google’s conduct on the Android mobile app distribution market

565 Epic says that Google’s OEM restrictive conduct including the terms and enforcement of the ACCs, the MADAs and the RSA3s, and Google’s app developer restrictive conduct including the terms and enforcement of the DDAs, the GVP agreements and the AVP agreements have created and sustained many of the barriers to entry in the Android mobile app distribution market and operate, individually and cumulatively, to prevent or hinder competition in that market.

566 Epic says that the effect of the OEM restrictive conduct and the app developer restrictive conduct has been and is to prevent or hinder alternative sources of supply in the Android mobile app distribution market, to create a virtual monopoly, and to increase barriers to entry such that there is almost no independent rivalry or substitution in the market. There is almost no market pressure on Google to reduce its costs of production and prices or improve the quality of its Android app distribution services.

567 Epic says that Google has closed off every effective path by which an alternative Android app store might achieve a significant market share, and it has severely hindered app developers’ capacity to distribute Android apps outside the Play Store.

568 Epic says that the ways in which the OEM and app developer restrictive conduct have prevented or hindered competition in the Android mobile app distribution market include the following seven points or matters.

569 First, the OEM restrictive conduct ensures that the Play Store is pre-installed in the most prominent locations on virtually every Android device in the world. This guarantees the Play Store can distribute apps to virtually every Android device user, makes it virtually essential for app developers wishing to reach those users to distribute their apps via the Play Store, enhances network effects that benefit the Play Store, and increases market concentration at the expense of its rivals.

570 In a world without the OEM restrictive conduct, rival Android app stores could and would negotiate with different OEMs for pre-installation on all or some of that OEM’s devices, including exclusively, or else in a more prominent location than the Play Store, to improve the discoverability and reach of the rival app store relative to that of the Play Store.

571 Second, the OEM restrictive conduct embodied in the RSA3s and the Motorola and LG MIAs hinders competition because it creates a disincentive for the relevant OEMs to pre-install or promote any rival Android app store. These terms, together with the technical restrictions requiring Google’s approval for privileged install permissions, hindered and ultimately prevented Epic’s attempt to pre-install the Epic Games App on Sony, LG, and OnePlus devices.

572 In a world without these terms, the relevant disincentives would not exist.

573 Third, the OEM restrictive conduct comprising the technical restrictions hinders competition from rival app stores in several ways.

574 Since Google has a discretion to approve (or not) the inclusion of a rival app store on the “allowlist” before Android devices may be distributed under a MADA, Google ultimately controls whether, and if so which, alternative app stores may be granted a privileged install permission on Android devices.

575 Further, those restrictions hinder the installation of rival app stores on Android devices by way of direct downloading, including by deterring users from installing those rival stores.

576 Further, the restrictions also hinder the installation of the first app a user seeks to download from a directly downloaded app store.

577 Further, the restrictions disadvantage a rival app store that lacks a privileged install permission when seeking to compete with an app store that has such permission.

578 Overall, the technical restrictions mean that any rival app store which is wholly or partly reliant upon direct downloading to be installed on Android devices must overcome a poor initial user experience that is apt to dissuade a large proportion of users from completing the installation process or ever using the rival app store.

579 In a world without these restrictions, OEMs could choose which app stores ought to be granted a privileged install permission on their devices without needing to obtain Google’s approval, and rival app stores could be distributed more successfully by way of direct downloading. Both eventualities would enable rival app stores to find their way onto more Android devices and to compete more effectively against the Play Store.

580 Fourth, by the app developer restrictive conduct, being specifically clause 4.5 of the DDA, as well as clause 4.1 and the device and network abuse policy, Google denies rival Android app stores access to the most effective means of distribution to Android device users, that is, via the Play Store.

581 The effect of these terms is that there is no truly effective means by which rival Android app stores can be distributed to the vast majority of Android devices. Pre-installation and direct downloading each involve the obstacles outlined already. And no other Android app store besides the Play Store has a share of app store downloads above about 2.5%.  There is next to no prospect of a rival emerging or expanding to impose material competitive constraints on the Play Store with these terms on foot.

582 Conversely, in a world without these terms, rival app stores could be distributed to almost every Android device in the world, through the most popular Android app distribution channel, without users having to confront any of the technical restrictions.

583 Fifth, by clause 4.1 of the DDA and the device and network abuse policy, Google precludes developers from updating their own apps, except through the Play Store. This means developers cannot switch from distributing their apps through the Play Store to doing so only through a rival app store, without losing the ability to update apps previously downloaded from the Play Store. As a result, developers are unlikely to cease distributing apps through the Play Store and rival app stores will be unable to secure those apps as exclusive content.

584 Sixth, by the app developer restrictive conduct comprised in the GVP and AVP agreements, Google has prevented or hindered potential competitors of the Play Store from establishing themselves and offering differentiated content from some of the world’s most popular app developers, including “beacons” of the mobile gaming industry.

585 Without this conduct there is a real chance that rival app stores would be established, including by Riot Games and Tencent, and would offer exclusive or differentiated content likely to attract large numbers of Android device users.

586 Seventh, reference was made at one stage in the material to Google preventing app developers who did not list their apps on the Play Store from accessing the App Campaigns program.  The parties hardly addressed this and I do not propose to say anything further about this in these reasons; if the parties want more reasons from me, they can request it.

587 The effect is to prevent or impede potential competitors of the Play Store from differentiating themselves from the Play Store through exclusive content.

588 Epic says that there is at least a real chance that the OEM and app developer restrictive conduct have the effect of substantially lessening competition in the market for the supply of Android app distribution services.

589 Epic says that absent that conduct, a number of entities, including Epic, are likely to establish rival Android app stores and compete with the Play Store. Further, existing Android app stores would be able to compete on a more equal footing. If that occurred, there would be increased competition for the supply of Android app distribution services. That competition is likely to manifest through alternative app stores offering different terms and conditions, including lower prices.

Cross-effect of conduct in the mobile OS licensing and Android mobile app distribution markets

590 Further, Epic says that there is an important nexus between Google’s OEM restrictive conduct and its app developer restrictive conduct, and competition in the Android in-app payment solutions market.   It has made the following points.

591 In short, quite apart from the in-app payment solutions restrictive conduct, Google’s conduct with respect to OEMs and app developers ensures that app developers who wish to reach most Android device users have no real choice but to distribute their apps through the Play Store and accept the DDA terms.

592 That conduct facilitates and supports the anti-competitive conduct in which Google engages in the Android in-app payment solutions market, because it ensures that developers are, in practical terms, required to accept the payments tie and the anti-steering rule.

593 Indeed, both the payments tie and the anti-steering rule form part of the DDA terms, and as such, are an integral part of Google’s conduct vis-à-vis app developers.

594 So, Epic says that the effect and likely effect of the conduct referred to is not limited to the Android mobile app distribution market. It also substantially lessens competition in the Android in-app payment solutions market.

In-app payment solutions restrictive conduct

595 The payment solutions restrictive conduct alleged by Epic includes the following aspects.

596 The first alleged aspect is Google’s conduct in requiring app developers that wish to distribute apps through the Play Store to enter into a DDA, and comply with associated developer program policies, containing the terms referred to below.

597 Those terms require app developers offering Play Store app in-app purchases to use Google Play Billing for those transactions save in limited circumstances. They require app developers to pay a fee of 30% on Play Store app in-app purchases subject to limited exceptions. And they preclude apps distributed via the Play Store from leading users to a payment method other than Google Play Billing subject to limited exceptions.

598 So, Google required app developers who wished to offer Play Store app in-app purchases to agree to the terms of the DDA and developer program policies relating to Google Play Billing.

599 The second alleged aspect is Google’s conduct in entering into the AVP Agreements with certain app developers, and requiring those app developers to use Google Play Billing as the sole payment solution in all markets in which the relevant apps are available on the Play Store.

600 The third alleged aspect is Google’s conduct in giving effect to and/or enforcing the terms referred to above relating to Google Play Billing, including against Epic, by removing Fortnite from the Play Store, refusing to allow Epic to offer EDP for any Play Store app in-app purchases and failing to agree to Epic’s request for in-principle approval to make the Epic Games Store available on the Play Store.

601 Let me say something about the payments tie conduct.

602 As alleged by Epic, the payments tie conduct describes the conduct of Google, by the terms of the DDA and developer program policies relating to Google Play Billing, in making its supply, or offers to supply, app developers of services for the distribution of Android apps to Android smart mobile device users via the Play Store conditional upon app developers not acquiring payment solutions for accepting and processing payments for Play Store app in-app purchases from any person other than Google.

603 The second developer program policy which is relevant and made binding on developers by clause 4.1 of the DDA is the payments policy.

604 Between at least 31 July 2012 and 20 January 2021, the payments policy required app developers distributing apps through the Play Store to use Google Play Billing for in-app purchases, except where the payment was primarily for physical goods or services or for digital content or goods that might be consumed outside of the app itself.

605 In August 2013, the payments policy was amended to require all gaming apps selling digital in-game currencies or other digital goods to use Google Play Billing, even if the in-game currency or digital good could be used outside of the app. As a result, the second exemption referred to above no longer applied to gaming apps.

606 On 20 January 2021, what remained of the exemption for non-gaming digital content or goods that could be consumed outside of the app was removed entirely.

607 Google described this as a “clarification”, rather than what it actually was, being a unilateral amendment to a longstanding policy with contractual effect.

608 Existing apps that used an alternative payment solution were given until 30 September 2021 to comply with the amended payments policy. Google subsequently announced that, from 1 June 2022, any app developer that was non-compliant with the amended payments policy would be removed from the Play Store.

609 So, since January 2021, clause 2 of the payments policy has provided that “Play distributed apps requiring or accepting payment for access to in-app features or services, including any app functionality, digital content or goods … must use Google Play’s billing system for those transactions …”. This obligation is subject to certain exceptions, currently set out in clauses 3 and 8 of the payments policy.

610 Section 3 stipulates that Google Play Billing must not be used where the payment is of one of the following types. First, it is primarily for the purchase or rental of physical goods, for the purchase of physical services, or is a remittance in respect of a credit card bill or utility bill. Second, it concerns peer-to-peer payments, online auctions, and tax exempt donations. Third, it is for content or services that facilitate online gambling. Fourth, it is in respect of any product category deemed unacceptable under Google’s payments centre content policies.

611 Section 8 was added to the payments policy in response to legislative or regulatory changes in South Korea and India. It permits developers that require or accept payment from users in India and South Korea to offer users an alternative billing system alongside the Play Store’s billing system if they agree to certain prescribed requirements.

612 So, subject to limited exceptions, throughout the relevant period the payments policy has required app developers to exclusively use Google Play Billing for in-app purchases of digital content within apps distributed via the Play Store.  The parties before me referred to this as the payments tie.

613 Now many developers besides Epic expressed dissatisfaction with the payments tie throughout the relevant period, particularly following the “clarification” of the payments policy in January 2021.

614 From at least 2016 until April 2022, YouTube, which is a subsidiary of Google, used a Google-owned payments solution that was not Google Play Billing for in-app purchases of digital products on Android devices. YouTube resisted replacing its own payment solution with Google Play Billing.

615 YouTube was concerned about the features and functionality of Google Play Billing, that the integration of Google Play Billing would involve significant migration costs, including time and engineering effort, and that switching to Google Play Billing would result in the loss of YouTube’s autonomy and ability to innovate and compete with their main competitors, Spotify and Netflix.

616 Ultimately, following the September 2020 update to the payments policy, and an extended period of negotiation between YouTube and Google, YouTube reluctantly complied with the payments tie in April 2022.

617 Other developers have also expressed dissatisfaction with the payments tie during the relevant period such as Spotify, Netflix, Match, Calm, Bumble and Parship Group. It would seem that these developers considered Google Play Billing to be inferior to other payment solutions. Google nonetheless enforced the payments policy over the developers’ objections.

618 I have also mentioned above the anti-steering rule which is relevant in this regard.

619 Let me turn to three other related topics.

User choice billing in South Korea and India

620 Since 18 December 2021 in South Korea, and since 26 April 2023 in India, Google has permitted app developers to offer users an alternative in-app payment solution in addition to Google Play Billing under Google’s user choice billing (UCB) program.

621 If an app developer wishes to offer UCB, it must comply with several requirements imposed by Google.

622 First, it must follow Google’s interim user experience requirements, including displaying an information screen and a separate billing screen which presents Google Play Billing and the alternative payment solution to the user in an equal manner.

623 Second, it must manually report to Google the amount of all paid transactions from users in South Korea / India each month.

624 Google has stated that when a user selects an alternative payment solution under UCB, Google charges a service fee that is 4 percentage points below the fee it would charge the developer for that transaction had the user selected Google Play Billing.

625 In other words, where UCB is deployed and the user selects an alternative payment solution, the service fee is said to be typically 26% or 11%, rather than 30% or 15%.

626 So, unless a developer can find an alternative payment solution that charges a service fee of less 4%, the developer is financially worse off under UCB.

Developer only billing in the EEA

627 Since 19 July 2022, in the EEA only, Google has allowed non-gaming app developers only to use an alternative in-app payment solution without also having to integrate Google Play Billing alongside it. This program is the developer only billing (DOB) program. It was introduced by reason of the passage of the Digital Markets Act.

628 Under DOB, if a non-gaming app developer wishes to integrate an alternative in-app payment solution, it must comply with requirements imposed by Google, including following Google’s user experience guidelines, and accounting for and reporting to Google the amount of all paid transactions from the alternative payment solution for invoicing.

629 Google has stated that, under DOB, when an in-app purchase of digital content is made using an alternative payment solution, it charges a service fee that is 3 percentage points lower than the fee that Google would charge had the user selected Google Play Billing.

630 So, unless a developer can source a payment solution that charges a service fee of less than 3%, the developer will be worse off financially under DOB.

631 Google’s own assessment is that payment processing costs vary between about 3.5 and 6%.

User choice billing pilot in other countries

632 Since 1 September 2022, Google has also allowed eligible non-gaming app developers only to join its UCB pilot program in certain geographies outside India and South Korea, including Australia, India, EEA countries, Indonesia and Japan.

633 On 10 November 2022, the program was extended to the United States, Brazil and South Africa.

634 If a non-gaming app developer wishes to use an alternative in-app payment solution alongside Google Play Billing pursuant to the UCB pilot program, it must comply with the same requirements imposed by Google in South Korea and India outlined above.

635 Google has stated that, pursuant to this arrangement, when a user selects an alternative payment solution, it charges a service fee that is below what it would otherwise charge for that transaction had the user selected Google Play Billing by 4 percentage points.

636 So, unless a developer can source a payment solution that charges less than 4%, the developer will be worse off financially under UCB. Again, Google’s assessment is that the payment processing costs vary between about 3.5 and 6%.

637 App developers entitled to access the UCB program remain subject to the payments tie, albeit in a varied form. They must still integrate Google Play Billing as a payment solution for in-app purchases of digital content. They have the option of offering an alternative payment solution in addition to Google Play Billing but are not permitted to only offer that alternative payment solution.

638 Further, Google precludes price competition under UCB. Everywhere other than South Korea, app developers using UCB must offer users the same price, irrespective of the payment solution selected by the user. App developers therefore cannot incentivise users financially to choose a third-party payment solution in lieu of Google Play Billing.

639 There are also limitations on the alternative payment solutions that app developers can offer under UCB and DOB.

640 The terms of the UCB pilot prevent app developers from using a payment solution provider that operates as a “merchant of record”.

641 Now the evidence establishes that acting as a merchant of record is a way to take administrative and operational burdens such as handling taxes, currencies and financial compliance away from the app developer, whereby the payment solution provider operates as a reseller of the app developer’s software for each action.  So, payment solution providers that act as a merchant of record, such as Paddle, cannot offer payment solutions under UCB.

642 Again, the choice being presented to developers by Google is not commercially real. App developers remain constrained by the payments tie, even when they have access to UCB.

The Android in-app payment solutions market — section 46

643 Epic’s case is that there is an Android in-app payment solutions market, which is a market for the supply of services to app developers for the facilitation of payments for the purchase of digital content within Android apps.

644 Alternatively, it is a narrower market, for the supply of services to app developers for the facilitation of payments for the purchase of digital content within Play Store apps only.

645 But nothing turns on the distinction between the two alternative markets relied on by Epic. The Android in-app payment solutions market is only marginally wider than the alternative market, and this difference does not alter the analysis of market power or the likely competitive effects of Google’s conduct.

646 The points below refer only to the Android in-app payment solutions market, although they should also be taken to also include the narrower alternative market.

647 Epic has made the following points.

648 Through clause 4.1 of the DDA and the payments policy, Google imposes the payments tie, namely, it requires app developers that sell digital content in apps downloaded from the Play Store to use Google Play Billing as their payment solution for all in-app purchases of digital content within those apps.

649 But Google’s payments tie does not dictate that a service for facilitating in-app payments is not a separate product from app distribution services. From an economic perspective, two products are distinct if it is technically feasible for them to be supplied separately, and in the absence of tying, there would be distinct demands for each of them, as customers are prepared to acquire each product without the other attached.

650 Google Play Billing can be, and is, provided separately from Android app distribution services. The following may be noted.

651 Google Play Billing is not available in 15 countries where the Play Store operates. In those countries, app developers can and do distribute apps through the Play Store but use a payment solution other than Google Play Billing.

652 Until about 2021, when Google amended its payments policy to prevent it, at least 90 different developers distributed apps through the Play Store using a payment solution other than Google Play Billing for in-app purchases of digital content.

653 Under the UCB and DOB programs, developers are permitted to use a payment solution for in-app purchases other than Google Play Billing, and various developers have sought to do so.

654 Between November 2021 and October 2022, approximately 103 app developers in South Korea sought to integrate an alternative payment solution into their apps.

655 Between September 2022 and October 2022, approximately 19 developers of non-gaming apps sought to integrate an alternative payment solution into their apps, pursuant to the UCB pilot program.

656 And between 19 July 2022 and October 2022, 29 app developers applied to Google to use a payment solution other than Google Play Billing, pursuant to the DOB program.

657 Various other developers of Android apps have sought to use, or expressed interest in using, a payment solution for in-app purchases of digital content other than Google Play Billing, including Epic, Parship Group and customers of Paddle.

658 App developers who sell non-digital content within Play Store apps do not use Google Play Billing. Instead, they acquire payment solutions from third parties separately from the distribution services supplied by Google.

659 For these reasons, Epic says that it is appropriate to treat services for the facilitation of in-app payments as a distinct product.

660 The next question that arises is whether there are close substitutes to Google Play Billing for the supply of services to app developers for facilitating in-app payments for digital content within Android apps.

661 Out-of-app payment solutions connote at least two requirements, namely, a requirement for the developer to build or acquire a solution that facilitates out-of-app payments, and a further requirement for the user to leave the app to effect a purchase elsewhere, for example, on a website using an out-of-app solution which the developer has built or acquired.

662 Epic says that out-of-app payment solutions are not a viable substitute for Google Play Billing.

663 First, as I have said above, Google’s anti-steering rule prohibits apps from “leading users to other payment methods”. The result is that developers cannot create links to out-of-app payment solutions within their Play Store apps or even tell their users about the out-of-app solution via in-app promotions or user-interface flows. Rather, out-of-app payment solutions can only be implemented separately from the app. Very few developers have taken this option.

664 Second, developers would not substitute in-app payment solutions with out-of-app solutions because the latter are a worse experience for users, who are more likely to complete the transaction within the app than if they must exit the app. Leaving the app can be disruptive to users, takes more time, and is often frustrating. App developers will prefer the payment solution that leads to more, not fewer, completed purchases.

665 Further, Epic’s case is that the geographic dimension of the Android in-app payment solutions market includes at least Australia and is likely global, excluding China and possibly other territories where competitive conditions began to differ towards the end of the relevant period.

Google’s power in the Android in-app payment solutions market

666 Epic says that Google has a substantial degree of power in the Android in-app payment solutions market.  It has made the following points.

667 If the market is confined to payment solutions for in-app purchases within Play Store apps, then Google is a virtual monopolist.

668 The in-app payment solutions restrictive conduct means that Google Play Billing is the only payment solution permitted for facilitating payments for in-app purchases of digital content within Play Store apps, save for apps distributed in the 15 countries where Google Play Billing is not offered, and apps that take advantage of the UCB or DOB programs. The take up of those programs has been relatively low.

669 The in-app payment solutions restrictive conduct also prevents new entry by competing payment solution providers. The barriers to entry with the conduct are insurmountable outside of the limited exceptions and app developers have no effective countervailing power.

670 Even within the exceptions, the scope for competitive constraint is very limited.

671 Under the UCB program, developers must still offer Google Play Billing alongside an alternative, and, save in South Korea, developers are not allowed to quote a lower price for using the alternative solution.

672 Further, the DOB and the UCB pilot programs are only available to “non-gaming” app developers, and the UCB pilot program precludes app developers from using a payment solution provider that operates as a “merchant of record”.

673 Alternative providers are thereby precluded from competing with Google Play Billing on equal terms, given that Google does act as merchant of record.

674 If the market also includes payment solutions for in-app purchases of digital content within Android apps more generally, the conclusion is the same.

675 The Play Store accounts for 98% of new app downloads from Android app stores in Australia and 91% of such downloads globally, and Google Play Billing is the payment solution for almost all in-app purchases of digital content that occur in those apps.

The purpose of Google’s conduct in the Australian Android in-app payment solutions market

676 Epic says that a substantial purpose of the in-app payment solutions restrictive conduct has been to substantially lessen competition in that market.

677 It is said that Google’s anti-competitive purpose ought to be inferred from the in-app payment solutions restrictive conduct itself.

678 Epic says that, notably, even when Google has been forced to allow developers to offer alternatives to Google Play Billing, it has done so in ways that force developers to continue offering Google Play Billing whether they want to or not, that disincentivise the use of any alternative solution, or that prohibit developers charging their customers a lower price for the alternative.

679 Epic says that none of this is aimed at protecting user security or privacy. It is aimed at hindering competition and maximising Google’s profits.

680 Further, it says that Google’s internal business records demonstrate that a substantial purpose of the in-app payment solutions restrictive conduct was to prevent or hinder competition.

The effect of Google’s conduct in the Android in-app payment solutions market

681 As to the state of competition in this market with Google’s conduct, Epic says that there is next to no competition at all.  Epic makes the following points.

682 Epic says that in the counterfactual, without the in-app payment solutions restrictive conduct, Google Play Billing would be likely to face significant competition.

683 There would be demand from app developers for Android payment solutions other than Google Play Billing, just as there was in the past and is currently.

684 There would also be suppliers with the incentive and ability to meet that demand. Several providers already supply payment solutions for in-app purchases of physical goods within Android apps. It would not be difficult for these providers to supply an in-app payment solution for purchases of digital content within Play Store apps. As a result, one would expect entry by those suppliers into the market if Google’s in-app payment solutions restrictive conduct ceased.

685 At least one payment solution provider, namely, Paddle, has already expressed a desire to enter the market and compete with Google Play Billing.

686 For these reasons, Epic says that the in-app payment solutions restrictive conduct has the effect, or likely effect, of substantially lessening competition in the Android in-app payment solutions market.

687 Now that I have set out the claims that Epic has made, let me say something about the witnesses.

Google’s lay evidence

688 I have addressed Epic’s lay witnesses in my reasons in the Apple proceeding.  Essentially, Epic called the same lay witnesses in both the Apple and Google proceedings.  I adopt that commentary here.

689 It is convenient at this point to say something about Google’s lay witnesses.

690 Mr Lee Rawles, the managing director of Global Games, Play Partnerships at Google, gave written evidence in chief. He was cross-examined by Mr Young KC for Epic.

691 Generally speaking he gave reliable evidence, although at times he was a little evasive and occasionally was prone to overly elaborate responses. As to his evidence which was relevant to the question of market definition and competition, I have addressed that subject matter elsewhere.

692 I should also note that Mr Rawles reported to another witness in the case, Ms Purnima Kochikar.

693 Ms Kochikar, the vice president of Play Partnerships at Google, gave written evidence in chief. She was cross-examined by Mr Rich SC for Epic and Mr De Young KC for the class applicants.

694 Clearly, Ms Kochikar is commercially highly skilled particularly in her role in dealing with many facets of Google’s dealings with developers including their demands, needs and preferences, and also questions dealing with billing and Google’s monetisation of its dealings with developers.

695 Ms Kochikar generally speaking gave reliable evidence, but I would note three cautionary aspects.

696 First, I thought that she over pushed and overly injected the corporate theme of Google vigorously competing with iOS/Apple, and on occasion in a non-responsive way to Mr Rich SC’s precise and pithy questions. The competition for app developers and users was more directly within the Android ecosystem between the Play Store and non-Play Store platforms or other mechanisms, although of course the broader context included competition with iOS/Apple.  Further, I accept that the broader competition concerning Android mobile devices and iOS mobile devices is not irrelevant to the case.  But such competition has little to do with market definition in my view.  But it is relevant to any broader context considering competitive constraints.

697 Second, her evidence concerning Project Hug had some problematic aspects to it, which I have specifically addressed elsewhere.

698 Third, like some other witnesses, Ms Kochikar also sought to push the line of a “clarification” to policy in an Orwellian fashion.

699 Mr Richard Byers, a principal software developer at Google, gave written evidence in chief. He was cross-examined by Mr Young KC for Epic. Generally he gave reliable evidence, although in my view he overstated the virtues of web apps. I have dealt with this topic in more detail elsewhere.

700 Mr Paul Gennai, a product management vice president at Google, gave written evidence in chief. He was cross-examined by Mr Garry Rich SC for Epic and Mr Bannon SC for the class applicants.

701 I found his evidence at times to be evasive. Moreover on occasion he made what I consider to be problematic distinctions in some of his answers. Moreover, some of his evidence concerning Project Hug and the RSA3s was not reliable. I have discussed specific matters concerning his evidence elsewhere.

702 Mr Paul Feng, the vice president of product management at Google, gave written evidence in chief. He was cross-examined by Mr Young KC for Epic and Mr Bannon SC for the class applicants. His principal topics of evidence concerned questions of monetisation, service fees, the Play Store’s billing system and the like.

703 Generally he gave reliable evidence on most topics. But there were aspects which were problematic such as his attempt to explain a policy change which eliminated an exception as a mere “clarification”. Several of Google’s witnesses persisted with such an Orwellian expression, which did them no credit. Further, at times his evidence was evasive or incomplete, such as what he said and sought to justify at [152] of his first affidavit.  I have dealt with specific topics of his evidence in more detail elsewhere.

704 Finally, there were two other lay witnesses called by Google on security and technology questions, namely, Mr Porst and Mr Cunningham.  I will say something more about them later.

Jones v Dunkel inferences

705 Now a key question in this case is Google’s subjective purpose in imposing and enforcing relevant terms in the AFAs, ACCs, MADAs, RSAs, DDAs and payments policy and undertaking Project Banyan and Project Hug.

706 Epic says that Google’s contemporaneous business records support an inference that a substantial purpose of Google’s conduct was to substantially lessen competition in the pleaded markets.

707 But Google has both denied that it held such an anti-competitive purpose, and positively asserted that its subjective purposes were either procompetitive or otherwise justified.

708 But Google has neither called the senior executives who bear executive responsibility for the impugned conduct and contractual terms nor explained why those senior executives were not called to give evidence.

709 Now the rule in Jones v Dunkel (1959) 101 CLR 298 is that an unexplained failure by a party to call a witness may in appropriate circumstances support an inference that the uncalled evidence would not have assisted the party’s case.  The failure to call a witness may also permit a court to draw, with greater confidence, any inference unfavourable to the party that failed to call the witness, if that uncalled witness appears to be in a position to cast light on whether the inference should be drawn.

710 The inference permitted to be drawn is not that evidence not called by a party would have been adverse to the party, but that it would not have assisted the party. In other words, the rule does not enable me to infer that the evidence of the absent witness would have been positively adverse to that party.

711 It follows that the rule does not preclude me from drawing inferences in favour of a party that has not called a witness if such a favourable inference is otherwise open on the documents.

712 The rule in Jones v Dunkel operates in circumstances in which a party is required to explain or contradict something. A Jones v Dunkel inference will not always be drawn and whether it is drawn depends on weighing all of the evidence.  The basis on which an inference may be appropriate depends on all of the circumstances of the specific case.

713 Now there are limitations to drawing an inference under the principle in Jones v Dunkel. The inference cannot be used to rectify any deficiencies or gaps in the evidence of the party seeking to rely on the adverse inference. Similarly, conjecture or speculation cannot be converted into an inference.  The inference must be available from direct and proven evidence. In this regard, disputed questions of fact must be decided according to the evidence that the parties adduce, not according to some speculation about what other evidence might possibly have been led.

714 Epic has made the following points.

715 Mr Sundar Pichai was the CEO of Alphabet Inc, which was Google’s parent company, during the relevant period.  Mr Pichai previously held positions as chief business officer at Google, and “Head of Android”.

716 As Google’s most senior executive during the relevant period, Mr Pichai was a person who could speak with authority to Google’s subjective purpose.  Mr Pichai was directly involved in driving decisions regarding the Play Store placement requirements under the MADA. Mr Pichai was also involved in the amendment to clause 4.5 of the DDA. Further, as the CEO, Mr Pichai had ultimate executive responsibility for Google’s proposed “special deal” with Epic, Project Banyan, Project Hug and Google’s longstanding approach to imposing and enforcing the impugned contractual provisions and policies.

717 Mr Pichai was not called to give evidence in this proceeding and Google provided no explanation for the failure to call him.  He is still employed in the role of CEO.

718 Sitting directly under, and reporting to Mr Pichai during the relevant period, was Mr Lockheimer. Mr Lockheimer was the most senior executive, other than Mr Pichai, with responsibility for the Play Store.

719 Mr Gennai, whose own boss Mr Rosenberg reported to Mr Lockheimer, gave evidence that either Mr Lockheimer or Mr Rosenberg would have been responsible for approving the changes to clause 4.5 of the DDA.

720 It is also apparent from the documentary record that Mr Lockheimer was extensively involved in the issues and decisions the subject of this proceeding, including Google’s strategy for Android since 2010.

721 Mr Gennai, by contrast, described his role in the 2010 process as being a scribe.  Mr Lockheimer was involved in decision-making for Project Banyan, and Project Hug.

722 Further, Mr Lockheimer, along with Mr Rosenberg and others, was listed as a deal representative for the Google business council approval of the framework that included the RSA3s; I have discussed this later.  Mr Gennai was not a representative, nor was he present at the relevant meeting.

723 Mr Gennai’s suggestion that he was travelling only serves to emphasise the oddity of calling him rather than Mr Lockheimer or Mr Rosenberg to shed actual light on the decision. Indeed, the travelling suggestion was offered by Mr Gennai on various occasions to explain his failure to have attended important meetings.

724 Mr Lockheimer was not called to give evidence in this proceeding and Google did not explain its failure to call Mr Lockheimer, a very senior executive whom Google still employs.

725 During the relevant period, Mr Rosenberg and Mr Samat sat below Mr Lockheimer and held various senior responsibilities relating to the Play Store and Android. The evidence demonstrates that they were key decision-makers concerning the attempted deal with Epic in 2018, Project Hug, Project Banyan and the RSA3s.

726 Both Mr Rosenberg and Mr Samat remain employed by Google. Neither was called to give evidence. Despite the documentary record indicating their close involvement in the relevant conduct, Google provided no explanation for its failure to call these senior executives. Instead, it called Mr Gennai, who reported to Mr Rosenberg.

727 Not only did Google not call Mr Lockheimer, but it did not even call his direct reports, Mr Rosenberg and Mr Samat. Instead, it called executives another rung down the ladder: Mr Gennai, Ms Kochikar and Mr Feng. Mr Gennai reported to Mr Rosenberg from 2010 until about August 2020. Ms Kochikar had also previously reported to Mr Rosenberg and Mr Samat. Mr Feng had previously reported to Mr Samat.

728 That forensic choice might have made sense if Mr Gennai, Mr Feng, or Ms Kochikar had been  decision-makers within Google.  But the evidence indicates that they were not. Instead, Mr Gennai’s involvement in decisions was in the nature of him being an advisor rather than as part of the decision-making team, Ms Kochikar was downstream of the decisions, and Mr Feng was not responsible for any final sign-off on policies concerning the Play Store or Google Play Billing.

729 In these circumstances, Epic says that it is appropriate to infer that Google did not call Mr Pichai, Mr Lockheimer, Mr Rosenberg, or Mr Samat because it was afraid that their evidence would expose facts unfavourable to Google.

730 Epic says that I should infer that the evidence of each of those senior executives would not have assisted Google’s case and more readily draw inferences adverse to Google which are available on the evidence.

731 But in my view and largely for the reasons that Google has advanced, no Jones v Dunkel inference should be drawn against Google.

732 Google called evidence from very senior executives being Mr Gennai, Ms Kochikar and Mr Feng.  Those witnesses were intimately involved in Google’s decision-making in connection with the conduct at issue, and in a position to give evidence of Google’s purposes in connection with that conduct.

733 Mr Gennai is a product management vice president at Google LLC. Between 2010 and 2020, he was involved in the launch of the Play Store, managed the Play Store’s weekly product council, managed the strategy and analytics function for Android and managed and grew the Play Store analytics team, which is responsible for collecting, processing and analysing data for the Play Store business.

734 Mr Gennai was closely involved in developing Google’s strategy for the Play Store and Android since the Play Store’s launch in March 2012 until he changed roles in August 2020. He was in a position to speak to that strategy and to why Google made and implemented the decisions it made during the period of critical relevance to Epic’s claim.

735 Ms Kochikar is the vice president of Play Partnerships at Google and has been in that role since November 2020. Ms Kochikar was closely involved in the development of Project Hug and the GVP. She was involved in the internal process of proposing and progressing Project Hug for approval by the Play Store’s business council, and attended the business council meetings where that project was discussed. Ms Kochikar was in a position to give evidence as to Google’s subjective purpose in connection with Project Hug and the GVP.

736 Mr Feng is the vice president of product management at Google LLC. Between 2016 and 2022, he led the Play Store’s monetisation initiatives. He led the team that was responsible for the management and functions of the Google Play billing system, including the Play Store’s payments policies. As the leader of the team, he was responsible for product strategy, and for defining and implementing products to assist developers to build and monetise their digital content businesses on Android and the Play Store.  Mr Feng was involved in Google’s consideration of its monetisation strategy and payments policy. He was in a position to give evidence as to Google’s rationale for that policy and its subjective purpose.

737 Each of Mr Gennai, Ms Kochikar and Mr Feng appears on numerous internal Google documents that are at the core of Epic’s case against Google. That demonstrates their central involvement in, and knowledge of, Google’s decision-making in connection with the conduct that Epic impugns. They were appropriate witnesses to address those matters.

738 Now Google is a large organisation and it is customary for senior executives to attend meetings and be copied on documents and emails.  And Google is not required to call each and every witness who can speak to a matter.

739 Further, I agree with Google that Epic has made no attempt to identify, with the requisite precision, what inference ought to be drawn in each case.  Even if I was inclined to draw any generalised adverse inference, that would not take any particular forensic issue far in any event.

Chat deletions

740 Epic says that an aspect of the evidentiary landscape which should affect my assessment of Google’s purpose is Google’s destruction of its internal Google Chat records. Google Chat is an instant messaging service which is used regularly by Google employees to communicate with other Google employees, including about business matters. Records of those chats, if they still existed, would constitute a repository of contemporaneous business communications.

741 Epic says that those records to a significant extent were deliberately destroyed and consequently have not been discovered in this proceeding. Moreover, it is said that that occurred in circumstances where there is evidence that Google employees, knowing that Google’s Chats would be destroyed, channelled sensitive business communications into Google Chat.

742 Now in September 2008, Google issued an internal company-wide announcement of a new company policy in the light of Google continuing to be in the midst of several significant legal and regulatory matters and its recognition that Google was going to keep facing these kinds of challenges.  The policy announcement was sent by Mr Bill Coughran then the senior vice president of engineering, and Mr Kent Walker a senior lawyer within Google and apparently now the chief legal officer and president of global affairs for Alphabet.

743 The new policy, which was said to help avoid the “inadvertent retention” of instant messages, was a decision to make “off the record” the Google corporate default setting for “Google Talk”. In this context, “off the record”, which is also known as “history off”, was a setting whereby chats were permanently deleted within 24 hours.

744 So, Google instituted a default setting that resulted in the systematic destruction of Google Chat records every 24 hours, for the purpose of ensuring that none of those records would have to be produced to third-parties in the several significant legal and regulatory matters that were ongoing at the time, or in the context of future legal and regulatory challenges that Google was going to keep facing.  Google maintained this policy from September 2008 until February 2023, when the default setting was changed.

745 Both Ms Kochikar and Mr Gennai confirmed that during that period they used Google Chat regularly to communicate with other Google employees, including concerning Google business matters.

746 Now Google’s policy was not influenced by any technical constraints regarding its document retention capabilities. Google possessed the technical ability to preserve its Google Chat records by setting the chat history to “on”.

747 Google’s efforts to avoid the so-called “inadvertent retention” of documents also extended to the training it provided to its employees.

748 In a training program titled You Said What?! 10 Things To Ensure You Are Communicating With Care, employees were advised that “[a]t Google, [w]e are constantly in the public eye … and the courthouse. We often have to produce employee communications as evidence … Our communications can hurt or embarrass us as a company …”.   With respect to the advisability of communicating by “Chat ‘off the record’ via Hangouts” instead of by email, employees were instructed that this option is “[b]etter than sending the email, but not without risk” because “[w]hile “off the record” Hangout chats … are not retained by Google as emails are, any chat participant may save the conversation by simply copying and pasting it into a doc or email …”.

749 Google’s policy also set out instructions to those employees subject to a “litigation hold”, directing them that if “you must chat regarding matters covered by that hold, please make sure that those chats are ‘on the record’”.

750 Relevantly, both Mr Gennai and Ms Kochikar were subject to legal hold notices since around September 2020 in relation to the lawsuit brought by Epic against Google in the United States. But despite being subject to litigation holds, until the policy change in February 2023, the default setting for their Google Chats remained history-off.  Google left it to the discretion of employees to determine whether their communications were “covered by” the hold.

751 Now Epic says that determining whether a communication such as a Google Chat record is relevant to the issues in dispute in legal proceedings and necessary to retain is a question which cannot reasonably be delegated to the lay participants in those communications, particularly in a case which involves the many and complex issues which arise in this proceeding.

752 Epic says that Google’s conduct gives rise to the real likelihood that a significant quantity of otherwise relevant material was not discovered.  And it says that as to such chats being relevant, that likelihood is increased given that Ms Kochikar directed her co-workers to communicate with her via Google Chat, rather than by email, in respect of sensitive business matters.

753 So, Epic says that in an email chain concerning Google’s assessment of a policy violation by Epic, Ms Kochikar advised the participants that “[t]his developer is not acting in good faith. Please be very careful about anything you say and do not put anything in writing. Please call or ping me if you have any questions”. By “ping”, Ms Kochikar meant communicate using Google Chat.

754 Further, it says that another instance occurred in an email chain regarding a meeting with Mr Don Harrison of Google about a proposed amendment to Google’s payment policy, in respect of which Google was attempting to obtain Spotify’s agreement. Ms Kochikar responded that she:

… caught up with Don briefly after the Spotify call - his concerns are less about the commercial terms and more about implications of the product path we are picking. Let me ping you in a bit ... It’s best to discuss this live.

755 In another email dealing with Google’s response to the Fortnite hotfix and removal situation, Ms Kochikar directed her staff not to discuss the topic using a form of Google Chat whose records are retained.

756 In the present case, Epic says that I should take an approach analogous to the rule in Jones v Dunkel.  So, if the evidence as a whole is consistent with a particular adverse inference relating to Google’s conduct or purpose, Epic says that that inference may be drawn more readily by reason of the absence of contemporaneous Google Chat records that might have shed light on the true position.

757 Epic says that here, the evidence establishes that the conduct which resulted in the destruction of Google Chat records was not accidental but a deliberate company policy that was conceived and implemented with a view to preventing the retention of records that might have to be produced in legal and regulatory proceedings.

758 It is said that the effect of Google’s conduct in destroying its Chat records is that I do not have the benefit of knowing what was contained in those Chat records and the way in which they may have borne upon the issues in dispute.

759 But in my view no adverse inference should be drawn against Google concerning the Chats.

760 Chats are not core business documents. They are ephemeral communications and the digital equivalent of idle chatter or vacuous prattle around a water cooler.

761 Epic has not demonstrated that Chats were commonly used by Google’s executives to discuss any of the matters at issue in these proceedings. Nor is there a basis to infer that any significant documents have been lost by reason of the fact that Chats were not historically retained.

762 Further, Epic has been able to identify documents which it asserts record or constitute Google’s decision-making and actions in connection with the conduct at issue.  Further, this is not a case where I have any basis to infer that Epic has been deprived of the opportunity to cross-examine witnesses by reference to contemporaneous emails.  And of course, Chats are nothing like emails.

763 On the evidence that I have seen, Google makes decisions collaboratively, over time and in an elaborate documented fashion.

764 Indeed, when one has regard to the over 3 million documents that Google has discovered in this proceeding, the idea that any meaningful business communication occurred in Chats, without reflection in some email, presentation or other document, is quite unlikely, as Google points out.

765 For these reasons, I will not draw the inference that Epic seeks.

766 Let me now turn to the question of market definition and begin by saying something about the parties’ cases.

Market definition concerning Android mobile app distribution – the parties’ cases

767 Let me begin by setting out again Epic’s posited markets.

768 Epic says that there is a mobile OS licensing market, which is a market for the supply of licenses for mobile operating systems, where the geographic dimension of the market is a global market excluding China, of which Australia forms part.  I will put that to one side for the moment.

769 Further, Epic says that there is an Android mobile app distribution market, which is a market for the supply of services for the distribution of Android apps to Android smart mobile device users, with the following elements or dimensions.

770 First, the services consist of the provision of services to app developers enabling or facilitating app developers to reach, offer and provide Android smart mobile devices users with Android apps and associated updates and/or the provision of services to Android smart mobile devices users enabling or facilitating Android smart mobile devices users to be presented with or to find, obtain and utilise Android apps and associated updates.

771 Second, the geographic dimension of the market is Australia or a global market excluding China, of which Australia forms part.

772 In this section of my reasons I will concentrate on this market.

773 Further, Epic says that in the alternative to an Android mobile app distribution market, there is a mobile app distribution market, which is a market for the supply of services for the distribution of Android apps and native apps written for iOS to smart mobile device users, with the following elements or dimensions.

774 First, the services consist of the provision of services to app developers enabling or facilitating app developers to reach, offer and provide smart mobile device users with native apps and associated updates and/or the provision of services to smart mobile device users enabling or facilitating smart mobile device users to be presented with or to find, obtain and utilise native apps and associated updates.

775 Second, the geographic dimension of the market is Australia or a global market excluding China, of which Australia forms part.

776 Further, Epic says that there is an Android in-app payment solutions market, which is a market for the supply of services to app developers for accepting and processing payments for the purchase of digital content, including by way of subscriptions, within an Android app, with the following elements or dimensions.  I will put that to one side for the moment.

777 In this part of my reasons I will just address the posited Android mobile app distribution market and the alternative market as outlined.

Android mobile app distribution market

778 Before proceeding further I should say a little more about Epic’s articulation of this market.

779 In summary, Epic says that the Android mobile app distribution market is a market for the supply of services for the distribution of Android apps to Android device users.

780 It is said that those services are supplied to app developers and enable them to reach, offer, and provide their Android apps to users, as well as to update those apps. In addition, those services are supplied to Android device users and enable them to find and obtain Android apps, as well as updates to those apps.

781 Now as Epic correctly contends, it is appropriate to start the market definition process by identifying the product which Google supplies and asking whether there are any close substitutes.

782 Google supplies services for the distribution of Android apps through the Play Store. The services which Google supplies to developers via the Play Store are described in the DDA.

783 In particular, clause 2.1 recognises that the developer is making use of the Play Store “to distribute Products”, that is, apps, and provides that “Google will … display and make [the developer’s] Products available for viewing, download, and purchase by users”.

784 Clause 3.1 further provides that developers appoint Google as their agent or “marketplace service provider” in order “to make [the developer’s] Products available in Google Play”.

785 Other pertinent aspects of the DDA include the express recognition in clause 3.2 that the developer’s products are “distributed via Google Play” and the statement in clause 3.3 that “Products are displayed to users” on the Play Store.

786 Further, in clause 3.4 there is reference to products being sold or made available to users “via Google Play”. Further, there is authorisation granted to Google under clause 5.1 to market products on the Play Store and to provide “hosting services … to allow for the storage of and user access to the Products”.

787 Further evidence that Google supplies app distribution services to developers and users via the Play Store can be found in the text of the developer program policies, which are incorporated into the DDA by clause 4.1. Those policies include statements such as “[t]hese Developer Program Policies, along with the [DDA], ensure that together we continue to deliver the world’s most innovative and trusted apps to over a billion people via Google Play” and “[p]eople from all over the world use Google Play to access apps and games every day”.

788 Further, Google describes the Play Store as an app distribution system and a distribution platform.  And as Epic points out, Ms Kochikar accepted that the services that the Play Store provides directly to developers include facilitating discovery of apps by users and app distribution.  Further, Mr Gennai explained that using the Play Store, users can search for or browse for apps developed by third-party developers and download and install such apps onto their devices. He also said that distribution on the Play Store is valuable to developers in that it provides them with the ability to reach billions of Android device users around the world via a single platform, through which they can distribute their apps.

789 Further, Epic’s case is that only other Android app stores are sufficiently close substitutes to warrant inclusion in the same market as the distribution services which Google supplies to developers and users through the Play Store.

790 Epic also makes the following ten points.

791 First, a small but significant change in the price or quality of those distribution services would not provoke sufficient switching by developers and/or users to other possible substitutes as to justify their inclusion in the same market.

792 Second, the possibility of developers entering into pre-installation arrangements with OEMs to distribute their Android apps is not a close substitute to distribution via an Android app store for various reasons.

793 Third, the distribution of Android apps by means of direct downloads from a website on an Android device is also not a close substitute to distribution through an Android app store.

794 Fourth, accessing web apps, which run through the device’s web browser rather than natively on the device itself, is not a close substitute to the distribution of Android apps through an app store.

795 Fifth, the absence of close substitutability between services for the distribution of apps through Android app stores and web apps is demonstrated by users’ clear preference for native apps over web apps.

796 Sixth, the methods of distributing apps on non-Android smart mobile devices are not substitutes for distribution through an Android app store.

797 Seventh, the vast majority of users do not own both an Android and a non-Android mobile device, and do not readily switch between Android and non-Android devices. The intent to use another OS is low, potentially driven by high switching costs and/or loyalty to their existing OS. Switching costs include both direct financial costs and loss of the features of a particular eco-system, such as Google Drive, headsets, smart watches, paid apps or app related data.

798 Eighth, services that enable the distribution of apps on PCs are not a close substitute for distribution through an Android app store, including because there are functional differences between smart mobile devices and PCs which mean that they tend to be used in different situations, rather than as substitutes for one another.

799 Ninth, most of the top 50 revenue-generating apps on the Play Store are not available in PC app stores, reinforcing the conclusion that users are unlikely to switch to PC app stores in response to changes in the price or quality of Android app stores.

800 Tenth, distribution of apps on consoles is not a substitute for the distribution through Android app stores. The smart mobile device user base is about 3.7 billion, more than twelve times larger than the console user base of about 300 million. Given that the large majority of smart mobile devices distributed outside China consists of Android devices, it is clear that most Android device users do not own a console, and could not switch to using apps distributed to consoles without incurring the additional cost of acquiring one.  In addition, gaming consoles are functionally different from smart mobile devices and the most widely used such as Xbox and PlayStation are designed for use at a fixed location.

801 In summary, Epic says that the market in which Google supplies app distribution services through the Play Store is one for the supply of services for the distribution of Android apps to Android device users, and the participants in the market are the Play Store and other Android app stores. I have referred to this as the Android mobile app distribution market.

802 The geographic dimension of the Android mobile app distribution market is at least Australia-wide, and likely global excluding China. The market excludes China because the Play Store, being the primary supplier of services, is not available in China. The primary supplier of services in the market is Google. Further, there is no price or other differentiation across geographic areas indicative of a national or regional market.

The mobile app distribution market

803 Now contrary to the above, if services for the distribution of Android apps to Android devices are supplied in the same market as services for the distribution of non-Android apps on non-Android smart mobile devices, then Epic maintains an alternative case based on the existence of a mobile app distribution market.

804 That market is relevantly the same as the Android mobile app distribution market, except that it also incorporates services for the distribution of non-Android apps to non-Android devices, including iOS devices.

805 In other words, this market would include distribution services relating to both Android app stores including the Play Store and Apple’s App Store.

806 Epic says that including the App Store in the market effectively makes the mobile app distribution market a duopoly market.

807 In 2020, the Play Store was responsible for about 38% of new app downloads from all app stores in Australia and 67% globally. The Apple App Store was responsible for about 61% of new app downloads in Australia and 26% globally.

808 Consequently, Apple and Google are responsible for about 99% of downloads in Australia and 94% of downloads globally. The market thus remains a highly concentrated one.

809 Epic says that Google and Apple do not compete closely so as to impose a real competitive constraint on one another in the field of app distribution. Their market shares are relatively stable, and neither has a history of altering prices or other terms to undercut the other.

810 But even if the appropriate market for analysis was the mobile app distribution market, Epic says that Google would still have a substantial degree of market power.  And Epic’s case concerning the purpose and effect of Google’s conduct would continue to be as I have set out elsewhere.

Google’s criticisms – markets

811 Now Google says that I cannot be satisfied that any of these markets is an appropriate market for assessing whether Google has substantial market power, or whether its alleged conduct has the proscribed purpose or effect.

812 Google has referred to various legal and economic principles of relevance to market definition.  I have discussed the relevant principles in my reasons in the Apple proceeding and, of course, have applied such principles to the Google proceeding.

813 Google has asserted various general deficiencies in Epic’s case on market definition. It is convenient to identify some of these asserted deficiencies at this point.

814 Google’s first main criticism concerns the asserted failure by Epic to capture relevant incentives and rivalry affecting the Play Store.

815 Google says that Epic’s approach to market definition, and its expert economic evidence in support of that approach, focuses on the use that can be made of Android OS and app stores by ecosystem participants, that is, OEMs, carriers, app developers and device users, rather than the transactions that involve spending and the facilitation of those transactions.

816 Google says that a deficiency arises because Google’s services are unique in that, at several points, there is a decoupling of revenue-generating transactions from use of the service features.

817 It is said that the central problem with Epic’s approach is that it eschews consideration of the very revenue generating transactions that create the relevant incentives to compete and in respect of which the rivalry between market participants principally occurs.

818 This is notwithstanding that Ms Edwards-Warren conceded that monetisation is relevant to the market definition inquiry, that monetisation of the product is what matters to the supplier, that if rivals can divert monetisation of the product, this could constrain the supplier, and that the focus of the analysis is on what constrains the Play Store, which impacts Google’s profits.

819 So, by excluding the relevant transactions or facilitation of transactions that generate revenue for Google, Google says that the pleaded markets cannot provide the necessary lens for assessing the purpose and effect of the alleged conduct on the competitive process.

820 This is because competitive constraints derive from incentives, which in turn are driven by profits that ultimately must be sourced from monetary transactions.

821 Take Epic’s pleaded Australian Android mobile app distribution market, which comprises a focal product of app distribution services but not facilitation of sales of digital content, which it says is inconsistent with the contractual position.

822 Google says that by its very definition, this market excludes virtually all the transactions from which Android app stores earn revenue and over which they compete, because it focuses exclusively on distribution of apps and app updates rather than facilitation of transactions for app-based digital content.

823 This can be seen from the Play Store’s revenue composition, being REDACTED] from paid ads on the Play Store, which is not in any of Epic’s markets, and REDACTED] from service fees on transactions between developers and users for digital content.

824 In 2022, the service fee revenue was REDACTED] from in-app purchases and REDACTED] from subscriptions, which are both not in Epic’s market, while less than 1% came from paid app downloads, which are in Epic’s market.  In other words, Epic’s market includes only 1% of the transactions which earn revenue for the Play Store.

825 Google says that the difficulty with a market for app distribution services that excludes facilitation of in-app transactions for digital content is that it cannot assist in assessing whether the relevant conduct is likely to have a meaningful impact on the way in which the Play Store and other Android app stores compete for developers and device users.

826 The relevant incentives to compete on the part of the Play Store and other app stores are driven by the choices which app developers and device users make with respect to paid transactions, because it is those choices which directly affect app store profits.

827 Google says that by excluding consideration of how the Play Store and other app stores monetise their services, the proposed market fails to capture the incentives of relevant firms to compete and the competitive constraints that they face. In short, it does not focus upon what emerges as the clearest picture of the relevant competitive process in the light of commercial reality.

828 Now Google says that Epic attempts to answer this problem by pointing to indirect network effects and the fact that usage could be a proxy for spend. The suggestion is that, because services for distributing apps and app updates, including free apps and apps which provide for transactions for physical goods, improve the attractiveness of the Play Store, this ultimately leads to paid transactions for digital content.

829 It is also suggested that there is a connection between usage and spend because it contributes to app revenue generation by developers and Google, whether via paid app downloads, in-app purchases or apps that display ads.

830 In that way, usage is said to be a proxy for spend, and therefore sufficient to analyse the relevant competitive incentives.

831 But Google says that it is neither necessary nor appropriate in this case to focus on usage as a proxy for spend. This is for several reasons as Google puts its case.

832 Spending data is the most informative data point concerning competitive incentives, so it would be erroneous to focus the market definition exercise on usage to the exclusion of the transactions which result in actual spend. This is particularly problematic in a case such as this where detailed spending data is available.

833 Further, the HMT focuses on how a hypothetical monopolist’s profits would change in response to a price increase or quality reduction. This necessarily requires a focus on how spending translates into revenue for the hypothetical monopolist if one is to apply the test quantitatively.

834 In addition, it is not clear in any event that usage and app spending on digital content are closely correlated.

835 Professor Asker’s analysis demonstrated a lack of correlation between app downloads and service fee revenue, as well as a lack of correlation between Android Fortnite usage and spending. And no expert has analysed the extent to which app store or app usage is correlated with other categories of app store revenue such as spending on advertising.

836 Google says that this further points against the utility of a market definition exercise focused on usage to the exclusion of spend, when there can be no confidence that usage data would accurately capture the relevant competitive incentives or rivalry with respect to the Play Store’s services.

837 The second main criticism by Google is that it says that Epic’s markets are said to be inconsistent with the evidence of choices and constraints.

838 Google says that a deficiency with Epic’s pleaded markets is that they are inconsistent with the commercial realities that Google faces with respect to threats to its revenue, in particular, the evidence concerning the choices available to app developers and device users, and the fact that the Play Store’s revenue is fragile given it is concentrated amongst a small number of high value users and large app developers.

839 Google says that when these choices are understood in this broader context, it becomes clear that the sources of alternatives available to users and developers impose a substantial competitive constraint on the Play Store and Android OS and should be included within the relevant markets.

840 Google says that there are choices available to Android device users with respect to how they use their apps and how they purchase app-based digital content.  It has advanced the following points.

841 A user first makes a choice when purchasing their mobile device, that is, choosing whether to purchase an Android or an iOS device. Next, a user must select which device to use from among those they already own, that is, multi-homing. Finally, a user can access and spend on an app from a multitude of locations, including through the Play Store, or an alternative app store, a sideloaded app, a pre-installed app, a web app, a streaming service or an out-of-app purchase on a developer’s website.

842 Out of all of these choices, the Play Store is only able to generate revenue if the user decides to transact via the Play Store.

843 Likewise, there are a range of choices available to app developers with respect to how and where they monetise their apps, and as to the extent to which they invest in one platform versus another.

844 These choices include the platforms for which they develop an app, that is, Android, iOS, PC, console, or web based. There is the choice of the extent to which they invest in apps for certain platforms, that is, a developer could invest more in apps for certain platforms relative to other platforms. There is the choice of how the app is distributed, that is, the Play Store versus an alternative app store, sideloading, pre-installation, web app or streaming service. There is the choice of the monetisation strategy, that is, in-app purchases, paid app download, subscription, in-app advertising decisions or a combination. And there is the choice of the transaction venues used for monetisation, that is, within the app, cross-wallet functionality, or out of app purchase on a website.

845 Further, the sources of these various user and developer choices in fact impose close competitive constraints on Android OS and the Play Store, and they are close constraints both individually and in aggregate.

846 The principal sources of this constraint are Apple’s iOS, the option of Android app stores or sideloaded or pre-installed apps in the Android ecosystem, gaming consoles and PC app stores, and web apps, websites and streaming services.

847 Google says that the evidence regarding these choices and competitive constraints is both qualitative and quantitative.

848 The qualitative evidence includes numerous Google business records that consistently describe its perception of serious competitive threats from iOS, rival Android app stores, and other platforms, and its response or proposed response to these threats.

849 Similar perceptions of competitive threats with respect to monetisation of in-app transactions are contained in Apple’s internal documents, communications between Epic and gaming console makers in the context of negotiations over cross-play and cross-wallet functionality in Epic’s games, and a communication between Epic and Valve.

850 Google also says that the quantitative evidence is provided by Professor Asker and is corroborated in material respects by the quantitative analysis contained in Professor Goldfarb’s second report.

851 It says that this econometric evidence based on Fortnite spend data shows not only that non-Android transaction venues including iOS are closer substitutes for user spend on the Play Store than alternative Android app stores, but that Epic’s pleaded Android app store market fails the hypothetical monopolist test.

852 Further, Google says that Professor Asker’s critical loss analysis demonstrates that even when adopting the inputs advanced by Professor Goldfarb and Professor Davila, Epic’s Android mobile app distribution market does not have robust empirical support for a 10% SSNIP. It fails under a range of reasonable assumptions and specifications.

853 Google says that the effect of the quantitative evidence is that the relevant market in which the Play Store competes must include, at a minimum, game consoles and Apple’s App Store.

854 Google’s third main criticism concerns the position of the OEMs.

855 Google says that Epic has adduced little evidence to understand the true position of OEMs or their commercial incentives to participate in the Android ecosystem.

856 It has adduced no evidence from any OEM to suggest that they consider the OEM restrictive terms to be restrictive or contrary to their interests. Nor has it adduced evidence as to how OEMs might operate if the OEM restrictive terms were removed.

857 And it has adduced no evidence that would permit an understanding of the financial position of OEMs, the nature and extent of their commercial incentives to participate in the Android ecosystem or how those incentives may be affected by the relief which Epic seeks.

858 Google says that these omissions have consequences for the analysis of the conduct at issue.

859 First, they impact the market definition inquiry and the evaluation of Epic’s proposed mobile OS licensing market.  One cannot select what emerges as the clearest picture of the relevant competitive process without first obtaining a detailed understanding and proper characterisation of the commerce in question.

860 Second, they impact the question of market power.  One cannot understand whether Google has market power in the putative mobile OS licensing market without understanding the countervailing power of OEMs.

861 Third, they impact the competitive effects inquiry.  Any question of commercially realistic likelihoods requires a consideration of whether, and how, OEMs would act differently in a future without the impugned conduct.

Some relevant concepts debated by the economic experts

862 It is convenient at this point to refer to some economic concepts that were the subject of discussion and debate before me in the Google proceeding.

863 The first topic discusses some general themes concerning the HMT.  The second topic concerns elasticities, diversion ratios, multi-homing and switching.  The third topic concerns whether one should look at user spending or actual usage.  The fourth topic concerns the distinction between a profit-maximising step and a profitable step.  The fifth topic concerns the application of quantitative testing to two-sided markets. And the sixth topic concerns bundled products.

864 Now many of these topics were the subject of greater debate between the experts in the Google proceeding than in the Apple proceeding.  This is a reflection of the fact that greater quantitative analysis was carried out by the experts in the Google proceeding.

865 But it should be noted that my discussion in this proceeding takes as its foundation the detailed legal and economic discussion concerning general principles relating to market definition contained in my reasons in Epic v Apple.

866 Now before proceeding further I should say something about the economic experts.

The experts

867 Professor John Asker, an expert economist, was called as a witness by Google. He is a professor in the department of economics at the University of California, Los Angeles. His research interests are industrial organisation, competition policy and anti-trust economics.

868 He gave one report in Epic’s proceeding against Google and another report in the class action against Google.

869 He also participated in a joint expert report with Ms Edwards-Warren and Professor Goldfarb and in a separate joint expert report with Mr Holt.

870 He gave evidence in various concurrent evidence sessions.  He was principally cross-examined for Epic by Mr Young KC, although Mr Prince was given a cameo role on some technical quantitative questions.  Professor Asker was also cross-examined by Mr Bannon SC for the class applicants.

871 There is no doubting Professor Asker’s expertise, but I did think at times that he over-played his position on some of the quantitative analysis.  Clearly he had a predilection to mathematise anything that he could.  But the relevant formulae and other mathematical analysis that he used could only be one tool.  Further, some relevant data inputs were lacking. Moreover, simplifying assumptions had to be made. But nevertheless, his detailed treatment was impressive and I have seen fit to set out some of his mathematical derivations for formulae that have been derived for the specific purpose of this proceeding.

872 But when his quantitative treatment came up against Professor Goldfarb, who seemed to have more specialist econometric expertise, I thought that Professor Goldfarb’s views were preferable. And there was no occasion where Professor Goldfarb over-played his hand.

873 Further, in terms of the inter-play between Professor Asker, who was more quantitative, and Epic’s expert Ms Kirsten Edwards-Warren, who was more qualitative, the latter’s views sat more easily and commercially with me, as did Professor Wright’s views in the Apple proceeding as I have said elsewhere.

874 Further, I thought that Mr Holt and Professor Asker were equally matched in the areas of subject-matter overlap.

875 Now rather than comment further at this point, it is better that I address the specific aspects of Professor Asker’s evidence under the particular topics later.

876 Now Ms Edwards-Warren is an expert economist. She is an executive vice president at Compass Lexecon, a global economic consulting firm, and the deputy head of its European practice. She had been the director of economics at both the UK Office of Fair Trading and the UK Competition Commission before they were combined.

877 Her first report dealt with an economic analysis of the relevant markets, market power and effects and was qualitative in its emphasis.  Her second report provided a response to Professor Asker’s report, and indeed was longer than her first report given that she responded to Professor Asker’s detailed quantitative analysis.

878 Ms Edwards-Warren participated in a joint expert report with Professor Asker and Professor Goldfarb and also gave evidence in concurrent evidence session. She was cross-examined by Mr Moore SC for Google.

879 Now it is fair to say that I engaged more with Professor Asker than I did with Ms Edwards-Warren during the concurrent evidence session, but I needed to understand more of the quantitative analysis in Professor Asker’s evidence, and indeed query some aspects to improve my understanding.  Contrastingly, her style of anti-trust analysis was more familiar to me and sat better with me commercially, as too did the approach of Professor Wright in the Apple proceeding.

880 There is no doubting Ms Edwards-Warren’s considerable expertise and reliability. It is appropriate that I leave further commentary to the more detailed discussion of particular topics.

881 I should also note that Professor Goldfarb and Mr Holt also gave relevant evidence on some of the economic topics. I repeat what I said in my reasons in the Apple proceeding concerning their background and expertise.

882 Let me then turn to some of the themes debated between the experts.

Hypothetical monopolist test — general themes

883 In what follows I will assume that the reader has read what I have said in my reasons in Epic v Apple concerning the HMT.

884 Ms Edwards-Warren usefully identified the conceptual steps in applying the HMT. It is convenient to adopt her framework with which for the most part I substantially agree. Some of this is a repetition of what I have said in my reasons in Epic v Apple.

885 The test begins by identifying a narrow candidate market that incorporates the product of interest.

886 The next step is to consider demand-side (and potentially supply-side) responses to a SSNIP. As I have said elsewhere, SSNIP is short for small but significant non-transitory increase in price. Small but significant is typically interpreted as a 5 to 10% price increase, holding prices of other products constant.

887 Now a SSNIP test is inter-alia an organising framework for assessing markets. It raises the question, if customers were faced with a small increase in price or, alternatively, a small worsening in contractual terms or quality, how would they likely respond? Answering this question involves identifying the potential substitute products and/or geographies and assessing based on available evidence whether customers would switch to these potential substitutes.

888 As part of the consideration of demand-side responses, it is important to consider who the customers are, and how they impact each other’s behaviour.

889 A market is two-sided if firms in the market are platforms which intermediate between two sets of customers, and the size of one side of the market has an impact on the size of the other side of the market.

890 In a two-sided market, reactions on one side of the platform, for example, by Android smart mobile device users, can influence reactions on the other side, for example, by app developers.

891 It may also be relevant to consider supply-side responses to a small hypothetical price increase, say, where a supplier of a closely related product may start to supply the product in question within a relatively short time frame and without incurring significant sunk costs.

892 Supply-side substitution could, in theory, render a hypothetical monopolist’s price increase (SSNIP) unprofitable.  But in practice the most immediate constraint typically comes from the demand-side.

893 Further, the time period within which switching has to take place to constrain the hypothetical monopolist might vary depending on the context. If switching is unlikely to happen sufficiently quickly, a SSNIP can be considered profitable.

894 The next step is to assess whether a SSNIP would likely be profitable.

895 If demand-side or, if relevant, supply-side responses are insufficient to prevent a small price increase by the hypothetical monopolist, that is, the small price increase is profitable, then the candidate market being assessed is a viable antitrust market and the test is complete.

896 If the hypothetical price increase is unprofitable because of switching to a substitute product, then the market should be widened to include that product. One then repeats the thought experiment starting with a wider set of products or geographies. The closest substitutes should be considered before more distant substitutes, to the extent this is known.

897 Now products that are not within the relevant market based on the SSNIP test may still exert some competitive constraint on products within the market. It is appropriate to consider such competitive constraints from potential rivals as part of the market power assessment.

898 Now as I have indicated, in practice it is frequently the case that one cannot perform an empirical or quantitative SSNIP test due to limitations on data availability. Instead, it is appropriate to evaluate other types of evidence such as functional substitutability, evidence on customer purchase drivers and evidence on past switching or threatened switching to inform a qualitative application of the HMT.

899 Let me say something about multi-sided platforms and network effects. Now I have discussed this topic in detail in my reasons in Epic v Apple, but let me make a few points here.

900 Digital platforms connect different groups of customers and may generate a network effect. A network effect arises when one group of customers on the platform benefit when there are more participants on the other side of the platform and/or vice versa. The platform can be described as multi-sided.

901 The Play Store is an example of a digital platform which supplies services to two distinct customer groups. It intermediates between app developers and Android smart mobile device users, enabling the distribution of apps. App developers benefit when there are more users accessing apps on the Play Store, and Android smart mobile device users benefit when there are more app developers distributing via the Play Store.

902 The prices charged by a multi-sided platform to each user group will depend on several different factors, including which user group benefits most from there being more participants on the other side of the platform, whether charges are levied per transaction or on a lump sum basis, and whether users use only one platform or multiple different platforms, that is, multi-homing.

903 If the HMT is applied without taking into account network effects, then the results could be misleading and markets could be defined incorrectly, typically too narrowly.

904 For example, if the test is applied only considering the behaviour of app developers, it could overlook how a reduction in the number of app developers would reduce the number of Android smart mobile device users, and thus reduce the profitability of a price increase.

905 Where relevant, in analysing different candidate market definitions, one should take into account the potential two-sidedness or multi-sidedness of the putative market and the resulting likely feedback and network effects between different groups of customers in, for example, the effects that the choices of users and app developers have on each other.

906 I will say something more about two-sided markets and applying a SSNIP test later.

907 Now before proceeding further, let me say something about the application of the HMT.

908 When assessing whether it would be profitable for a hypothetical monopolist to raise prices by a small amount, or reduce quality or worsen contractual terms more generally, it is not necessary that all customers switch away or that customers switch away all of their purchases for the small price increase to be unprofitable.

909 If the sales generated by marginal customers, that is, by customers who are most likely to switch away, whether fully or by reducing their purchases, are sufficiently large, a SSNIP may be unprofitable, even if a large number of infra-marginal customers remain with the hypothetical monopolist.

910 The following types of evidence are relevant to assessing how much switching would be expected in response to a SSNIP.

911 The first type of evidence relates to the comparability of potential substitutes from a functional perspective. So, Android smart mobile device users or app developers would be less likely to substitute towards products that are functionally very different in response to small changes in relative prices.

912 The second type of evidence relates to the cost of switching between different alternatives for OEMs, for app developers and for Android smart mobile device users or suppliers of similar products. The cost of switching may be financial or non-financial such as time or inconvenience. All else being equal, higher switching costs make it less likely that switching would occur in response to relatively small changes in prices.

913 The third type of evidence relates to actual switching, involving understanding the products to which customers or suppliers switch and their motivation for switching where known and the context for such switching.

914 The fourth type of evidence relates to what is revealed by Google’s internal documents to the extent that they reveal any competitive pressure or lack of pressure on Google Android, the Play Store or Google Play Billing.

915 Now if there is insufficient data to quantitatively assess actual losses from a SSNIP, one has to make qualitative judgements to decide whether a price increase by a hypothetical monopolist would likely be profitable.

Elasticities, diversion ratios, multi-homing and switching

916 A competitor constrains a firm when customers of the firm view a competitor’s product as a substitute. As Professor Asker explained, economists define substitutes in terms of elasticities.  An elasticity measures the percentage change of a variable in response to the percentage change of another variable.

917 As Professor Asker explained, when thinking about competition there are two types of elasticities to focus on.

918 First, own-price elasticity measures the percentage change in a firm’s quantity sold in response to a percentage change in its price. Own-price elasticities answer the question “if a firm raises prices by, for example, 1 percent, by what percentage is the quantity sold reduced?”

919 Second, cross-price elasticity measures the percentage change in a firm’s quantity sold in response to a percentage change in a competitor’s price. Cross-price elasticities answer the question “if a firm raises prices by, for example, 1 percent, by what percent does its rival’s quantity sold increase or decrease?”

920 Two firms’ products are substitutes when the cross-price elasticity is positive, that is, when the quantity sold by one of the firms increases when its rival’s price increases or, equivalently, when its rival’s quality decreases.

921 Contrastingly, two firms’ products are complements when the cross-price elasticity is negative, that is, when the quantity sold by one of the firms decreases when its rival’s price increases.

922 Analogous definitions for substitutes and complements can be defined with respect to changes in quality instead of changes in price.

923 Now as Professor Asker explained, a useful reformulation of price elasticities is that of diversion. Diversion refers to lost sales due to a price increase or product removal that are captured by a supplier of a substitute.

924 The diversion ratio between two products A and B is the proportion of sales that are captured by product B when product A increases its price by some amount. The diversion ratio tells you about substitutability between products and it is sometimes approximated by examining the effect of removing product A from the market entirely and seeing if the customers move across to buy product B. Diversion is typically measured by a diversion ratio, which refers to the proportion of total sales lost that is captured by a rival.

925 Diversion answers the question “if a firm increases its product’s price, of the lost quantity sold, what percent goes to a given rival?” Positive diversion implies that cross-price elasticities of demand are positive and that the two firms’ products are substitutes.

926 As Professor Asker explained, arithmetically the diversion ratio from good i to good j is equal to the ratio of market shares of good j to good i multiplied by the ratio of the cross-price elasticity of good j with respect to the price of good i to the absolute value of the own-price elasticity of good i.

927 Now the greater the degree of substitutability between two firms’ products, the closer they are as competitors and the higher the diversion ratio.

928 Now diversion ratios are invariant to measurement in terms of quantities or revenue units, as they are measured with reference to the firm that increases its product’s price. Revenue is price times quantity. Because diversion is expressed as a ratio, measuring diversion using revenue units is the same as measuring diversion using quantity because the price term is common to the numerator and denominator. So, price cancels out.

929 Now the key aspect for accurate measurement of diversion is that quantities (equivalently revenues) correspond to the product or service being purchased by consumers. Depending on the context, it can be easier from a data perspective to use revenues instead of quantities.

930 Further, as Professor Asker explained, two other related concepts to substitution are those of multi-homing and switching.

931 Multi-homing occurs when customers of a given product purchase that product from multiple firms, for example, a developer publishing to multiple app stores. The term multi-homing is more commonly applied with respect to multi-sided platforms.

932 Switching occurs when customers of a given product completely stop using one firm’s offering and start using another’s, for example, a consumer replacing an Android smartphone with an iOS smartphone.

933 Multi-homing and switching are distinct from, but informative about, substitution. They are distinct because neither concept measures changes in behaviour that can be solely attributed as responses to changes in price (or quality). Nonetheless as Professor Asker explained, they can be indicative about price-based and quality-based substitution when combined with other information.

934 So, high levels of multi-homing may indicate that it would be relatively easy for customers to substitute in response to any one firm raising prices because they already transact with rivals. For example, when the price of a game on a given hardware gaming platform such as Xbox or PlayStation 2 increases, customers who own multiple hardware gaming platforms may substitute to a platform on which the game is cheaper for purchase compared to customers who only own one platform. An important reason is that multi-homing customers have already purchased and transacted on other hardware platforms, and therefore do not need to incur additional adoption costs that are unavoidable for single-homing customers. This means that the costs to the customer of finding an alternate supplier, and any uncertainty associated with the quality of that supplier, are lower than if customers were not multi-homing.

935 And high levels of switching may indicate that it would be easy for some customers to substitute in response to a price increase to a rival’s product. When consumers change their choice of product, say, because consumers switch their preferences for characteristics that vary across differentiated products, absent a price change, then price changes will incentivise more consumers to switch products. This is because marginal customers already switch between the products in response to their experience with the product, indicating that it is feasible and likely relatively easy to do so.

User spending or actual usage? What is the relevant metric?

936 Professor Asker’s analysis focused on user spending to understand the relevant constraints facing the Play Store.

937 Epic’s experts instead focused on usage, relying on metrics such as the number of downloads and the amount of time spent on an app. They said that usage is informative because it is a proxy for spending.

938 But according to Professor Asker, actual spending is the best proxy for spending and the correct metric to use. It is unnecessary to interpose a different proxy in circumstances where there is no economic, statistical or practical reason not to base an analysis on the rich data available concerning the primary and most informative metric, spending.

939 First, the HMT is about understanding what constrains a hypothetical monopolist’s profits. This means that substitution that affects revenue, that is, where consumers substitute their spending, is central to the HMT because it requires the economist to consider things in terms of dollars.

940 Second, Professor Asker said that Professor Goldfarb’s focus on Fortnite usage was divorced from the commercial reality within which the Play Store operates.  The Play Store primarily earns revenues from the service fee it charges on consumer spending on digital content, not the use of apps downloaded via the Play Store.  This means that the relevant substitution to understand the market within which the Play Store operates is substitution in spending on paid digital content transactions. While there can be contexts where usage is relevant for market definition, for example, if the firm does not monetise with a price, they do not apply here.

941 Third, economists do not use usage as a proxy for spending when spending data is available. Usage may be a good proxy for spending in some contexts, but spending is always the best proxy for spending.

942 According to Professor Asker, in this matter rich data exists to measure spending, including data on Fortnite users’ spending, spending via the Play Store, spending via the iOS App Store and spending on digital game content. In any event, usage is not a good proxy for spending in the relevant circumstances. Usage of app content and spending on app content by consumers are separate and divergent inquiries.  Further, he said that the fact that usage and spending can be and often is decoupled increases the importance of focusing on the relevant metric, namely, spending.

943 Fourth, he said that spending is a sufficient statistic to evaluate competitive constraints facing the Play Store.  Now Epic’s experts pointed to indirect network effects in which greater usage leads to greater spending, although they did not quantify the magnitude of these effects and instead relied on hypothetical examples.  But Professor Asker said that the fact that such network effects exist does not mean that usage should be given primacy over spending. A decrease in usage may affect spending, whereas a decrease in spending will decrease spending. In this sense, focusing on spending is conservative. Further, if the network effects are significant, the analysis of spending captures them anyway.

944 But Professor Goldfarb gave evidence which I found persuasive that usage data and analysis is appropriate for several reasons.

945 First, usage is directly tied to the benefit that users receive from an app. Most users derive value without ever spending.

946 Second, the value of usage is evident regardless of the specific app monetisation strategy employed by an app developer.

947 Third, use by one consumer increases value for other consumers through direct network effects.

948 Let me develop these points further based upon Professor Goldfarb’s evidence which in my view was conceptually coherent and commercially sound.

949 First, I agree with Professor Goldfarb that usage is important for understanding consumer substitution behaviour as it quantifies the benefit a consumer derives from an app. The focus on actual consumer use of apps therefore measures how apps generate value to the consumer. As consumption is directly tied to how a consumer values an app, more use is a sign of more value to the consumer.

950 Nevertheless, Professor Asker said that use is distinct from spending. He supported this claim through several examples of goods like computer screens or grocery stores where usage appears to be unrelated to spending.

951 But these examples do not correspond to the issue at hand as apps are durable digital goods. For durable goods, the benefit of the product depends on usage itself. All else being equal, those who use a durable product derive more benefit from it through continued use.

952 For computer screens, the amount a consumer uses a screen provides a greater indication of the value they derive, net of the purchase price, than the purchase price alone. This aspect is heightened for digital goods, as much of the consumer surplus from a product is derived from consumption itself, that is, scrolling on an app, often regardless of payment. Consumption and thus the consumer benefit is directly related to time use on an app.

953 Second, I also agree with Professor Goldfarb that usage creates economic value for developers across all app monetisation strategies.

954 Consumer usage is also relevant to understanding business decisions that developers make.  Usage contributes to app revenue generation across multiple app monetisation strategies, that is, the developer’s choice of how to generate app revenue. So, usage is a key metric to understanding consumer behaviour and firm revenues. Professor Asker himself said that usage as a metric can be helpful in situations where firms monetise based on usage-based ads.  In an ad-based app monetisation strategy, usage serves as a proxy for consumer behaviour that drives a firm’s revenues. But this app business model is not the only monetisation strategy where usage is a key metric to understand consumer behaviour and firm revenues.

955 The nature of the connection between usage and spending depends upon the app monetisation strategy. Available monetisation strategies through the Play Store include pay-to-download apps, free-to-download apps with in-app purchases including recurring subscriptions, or free-to-download apps that may rely on ad-supported models. For each of these monetisation strategies, there is a link between use and revenues.

956 For pay-to-download apps, use is linked to spending because users download the app for the purpose of using the app. On the developer side, while a user must first pay before they engage with the app, developers seek to maximise user engagement with the app to attract other users to make paid downloads.

957 For apps utilising in-app purchases, use and spending are directly linked. A user engages in in-app purchases to enhance their engagement with the app, for example through a game or through purchasing specialised material in an education app.  For recurring subscription models, developers similarly value user engagement so that users are incentivised to continue making a recurring purchase and so that other users are motivated to initiate a subscription. So, use is closely related to continued subscription transactions.

958 For ad-supported apps, it is use that matters and not spending because developers seek to maximise individual user engagement and number of users overall to command higher advertising revenues. Professor Asker highlighted that because customers do not spend money on ad-supported apps, customer spending is not the appropriate metric to understand constraints on the relevant firm. He said that in these situations examining substitution in usage in response to increased ads or worsening quality can be informative.

959 App developers may also have one or multiple monetisation strategies, for instance, the “freemium model” where developers may offer a free tier and a paid tier of an app.  Free tiers often increase the number of sales of paid versions of apps available on Android, along with apps available on iOS.

960 Third, I also agree with Professor Goldfarb that direct network effects create value for apps from use alone.

961 Now Professor Asker further dismissed usage as uninformative because the Play Store does not earn revenues based on the time users spend within an app and users can scale this time up or down or shift this time to a competitor of the Play Store and the Play Store’s revenue is unchanged.  These statements ignore network effects.

962 Usage is linked to revenues because of the importance of network effects for many app types. By network effects, what is meant is that the value of an app increases with the number of users engaged.

963 For most gaming apps in particular, the number of users and the time they spend on an app directly impacts the experience of other users, particularly for multiplayer games, regardless of whether these players engage in in-app purchases or not.

964 Ad-supported social media apps also increase in value with more (non-transacting) users engaging across the platform. More users streaming content on a given app helps developers decide what content to invest in to meet demand, an indirect network effect, which in turn affects the decisions of users to pay for subscriptions.

965 For apps that facilitate the exchange of physical goods or services, for example, ride hailing or home rental, more drivers or property owners increase consumer surplus through more choice and improved quality for consumers. More users impact driver or property owner willingness to participate on the platform. Use therefore increases revenues.

966 Professor Asker further said that network effects in usage on apps do not translate to network effects in spending, particularly in the context of Fortnite.  But if network effects in usage do not translate into network effects in spending, then developers or platforms would not be motivated to develop or include free apps or freemium apps. In reality, freemium models are acknowledged to be popular in the video game industry and in other industries. Free versions of apps lead to more purchases of paid app versions on Android. This contradicts Professor Asker’s statements.

967 Professor Asker provided a measure of the correlation between total weekly spending through the Play Store in Australia and worldwide weekly usage across all channels as evidence that usage and spending are not related.

968 This approach is uninformative. The correlation figure presented by Professor Asker (-0.42) is a statistically insignificant figure that compares weekly spend on one channel in one country, being weekly spend on the Play Store in Australia, with the global weekly number of users across all channels. There are many factors that would explain this non-statistically significant negative correlation. For example, factors such as different holiday schedules or seasons across the Northern and Southern Hemispheres may cause differences between expenditure and usage patterns for Australia as compared to the rest of the world.

969 Professor Goldfarb considered the more appropriate comparison to be between the total weekly spending across all channels in Australia and the total weekly users across all channels in Australia. In estimating this figure, he found a statistically significant correlation of 0.52. The large figure and the statistical significance together indicate that these two variables move strongly together.

970 Fourth, I also agree with Professor Goldfarb that by focusing on only users that spend, Professor Asker excluded the majority of users and apps. This is because most apps do not offer paid transactions to users and the majority of users do not make paid app transactions.

971 The majority of users of Fortnite do not spend. Indeed Professor Asker estimated that approximately 55 to 70% of Fortnite users did not spend on Fortnite at all over the three and a half year period for which he had data.

972 On the developer side, few developers offer paid apps or in-app purchases.

973 So, considering only paid transactions omits the majority of users and the majority of apps from an analysis. This is despite the fact that usage contributes directly and indirectly to revenues, consumers gain consumer surplus through their usage, and developers gain value through network effects.

974 So, across each of his analyses, Professor Asker in effect excluded most users and a major source of consumer surplus. Let me turn to some general matters.

975 Now Professor Asker said that Epic’s economic experts ignored evidence of substitution in spending by focusing on substitution in usage.  But as Professor Goldfarb pointed out, this criticism is problematic given several relevant points about app products that Professor Asker disregarded, along with the relationship between spending and usage.

976 Usage is appropriate for understanding whether Android smart mobile device users treat web apps, iOS native apps, and native apps on other devices as interchangeable with apps on Android smart mobile devices. Usage quantifies the benefits that consumers derive from an app. As apps are digital durable goods,  the benefit of the product depends on usage itself, often regardless of a direct purchase on the app.

977 Consumer usage is also relevant to understanding the business decisions that developers make, including which monetisation strategy to pursue.

978 Usage is related to revenue generation and this is true across app monetisation strategies. A monetisation strategy is how the developer chooses to earn revenue, whether via subscription, in-app purchases, advertising or other means.

979 Professor Goldfarb explained the links between usage and revenue for pay-to-download apps, apps with in-app purchases, including recurring subscriptions, ad-supported apps, and mixed monetisation strategies like “freemium” models.

980 Further, usage itself generates value for users and developers. This includes network effects in usage translating to network effects in consumer spending. This runs counter to Professor Asker’s claim that usage and spending are not closely related.

981 Professor Asker presented a problematic, negative and statistically insignificant correlation between weekly spending on Fortnite through Play Store in Australia and global weekly usage across all channels as evidence for his claim.

982 Professor Goldfarb corrected this, and applied a more appropriate correlation metric between weekly spending and usage in Australia across all channels. Professor Goldfarb’s correlation metric is more appropriate because it is regionally specific and representative of the correlation between spending and usage regardless of channel.

983 He found a robust correlation of 0.52, which shows that usage and spending are highly correlated.

984 A further issue with Professor Asker’s spending analysis is that the majority of users do not spend, and so considering only spending to understand consumer substitution omits the majority of app consumers.

985 The same issue applies to the developer side. Most developers do not pursue monetisation strategies where users pay directly in or for the app. Alternatives include ad-supported apps, for example. So, considering only spending also fails to account for the majority of app developers.

986 Given the relevance of usage for understanding whether consumers treat different app types as substitutes, Professor Goldfarb did not consider spending to be the most useful metric through which to understand consumer substitution behaviour. Nevertheless, he explained that use is a strong proxy for spending for the users who make paid app transactions, for several reasons.  Whenever he referred to spending, he referred to paid app downloads and in-app purchases of digital content including subscriptions.

987 First, Professor Goldfarb demonstrated that use and spending are highly correlated for users who spend. This means that changes in revenue are associated with similar changes in usage, and vice versa. Professor Asker also presented correlation coefficients for use and spending, oftentimes nearing 0.50. A correlation coefficient runs on a scale of -1 to 1, with numbers closer to 1 indicating that two variables move closely together in the same direction, that is, when there is an increase in one coefficient there is also an increase in the other coefficient. He interpreted these coefficients to show minimal correlation.

988 This interpretation is incorrect given common practice is to consider both the magnitude and statistical significance from zero to interpret the strength of a correlation coefficient. It also ignores the established literature defining the assumptions required for one variable to be considered a good proxy variable for another.

989 Second, Professor Goldfarb found that usage is a good proxy for spending because users almost never play and spend across different channels.  Across all users prior to the removal of Fortnite from the Play Store, he found that a minority of users spent across more than one spending channel overall within a given day. He also assessed multi-homing across spending and play channels within a day and found that less than 2% of users unambiguously spend and play across different play channels within a day.

990 This is in direct contradiction to Professor Asker’s presented data on user substitution across spending channels. Professor Asker did not show evidence of substitution between spending channels, either on the same device or across devices. His evidence showed only that a consumer had ever used a given spending channel. This is not evidence of whether consumers would consider that spending channel as reasonably interchangeable with another.

991 Professor Goldfarb’s view is that usage is a critical outcome to assess when trying to understand consumer substitution behaviour.

992 Further, even if spending was the most appropriate metric for understanding consumer substitution, usage remains a strong proxy variable for spending.

993 Now Professor Asker said that usage is not a proxy variable for spending because usage and transactions are not correlated. He presented evidence first on the correlation between downloads and service fees revenues across apps in the Play Store.  He also presented evidence on the correlation between usage and revenues for Android or Play Store Fortnite users.

994 Professor Asker used a rank correlation as his main analysis for both measures but also presented correlations based on actual downloads, use, and spending as robustness checks. For measuring the correlation between downloads and revenues from service fees, Professor Asker found correlations of 0.27 to 0.31, depending on the exact methodology used.

995 Professor Goldfarb did not find these correlations to be informative in determining the extent to which usage is a proxy of spending because downloads do not measure usage.

996 For revenue versus use, he found a correlation of 0.39 to 0.46, depending on methodology. Again, each of these correlation coefficients are statistically significant.

997 This methodology likely understates the correlation between use and spending.

998 Usage and spending do not need to be perfectly correlated to be considered proxy variables.

999 In S Athey, R Chetty, G Imbens and H Kang, “The Surrogate Index: Combining Short-Term Proxies to Estimate Long-Term Treatment Effects More Rapidly and Precisely” (2019) National Bureau of Economic Research (working paper), the authors discuss surrogates, that is, proxies, as outcome variables that can be used to measure another outcome variable that may often be harder to measure or based on a longer time horizon, for example, school test scores as surrogates for long-run earnings.

1000 The authors note that the key assumption for one variable being a good surrogate, for example, usage, for a given outcome, for example, spending, is that the outcome must be independent of treatment conditional on the surrogate.

1001 Here, usage is a valid surrogate variable for spending if, after conditioning for usage, spending is independent of removal. This is to say that usage is a valid surrogate variable for spending if changes to spending due to removal can be explained by changes to usage alone. Based on this commonly used assumption, usage is a good proxy for spending.

1002 Professor Goldfarb performed his own analysis to understand the correlation between Fortnite usage and spending. He linked Fortnite game play and purchases data to study the relationship between match play time, being usage, and paid app transactions, being spending to obtain digital content, across all spending channels and usage channels and across a longer time frame.

1003 He found that only 26.1% of users ever spent on Fortnite through any spending channel during the time window from 26 February 2018 until 9 August 2020 (pre-removal).

The themes of “device first” and “clicks are costly”

1004 Professor Goldfarb explained that user behaviour in choosing devices such as a smartphone or PC is influenced by external factors such as use context, such as a free afternoon or a spare five minutes before class, and environment at home, work or, say, on the bus. He said that the choice of device itself influences choice of activity, since difference devices are conducive to difference usage experiences.

1005 This concept is referred to as “device first”.  These concepts are important in understanding consumer behaviour, particularly in relation to substitution patterns. Further, “device first” is a useful framework to understand why products that may appear to be functionally similar are not substitutes in reality for consumers.

1006 Further, the concept of “clicks are costly” refers to the phenomenon of users finding the requirement to execute additional steps to perform activities on electronic devices, that is, information gathering or transactions, to be burdensome enough to deter the user from completing the activity. “Clicks are costly” provides a framework to discuss the relative costs and benefits that consumers implicitly weigh in their decisions. This includes whether to move out-of-app to make a paid transaction or complete the transaction in-app.

1007 Professor Goldfarb disagreed with the claims of Professor Uri Gneezy and Professor Asker that the “clicks are costly” concept is uninformative for the matter at hand.

1008 Professor Asker mischaracterised the “clicks are costly” concept and provided unfounded critiques of Professor Goldfarb’s supporting empirical analysis.  Professor Asker also criticised his empirical study of the impact of switching to biometric-enabled devices on the number of transactions undertaken by the consumer.

1009 The criticisms of Professor Asker of Professor Goldfarb’s analysis of how switching to a biometric device impacts the number of transactions a consumer undertakes are incorrect. Furthermore, when one corrects for the errors in Professor Asker’s methodology to assess the effect of switching to a biometric-enabled device using a different event study design, the results confirm the conclusion that reducing the frictions associated with completing transactions increases the number of transactions consumers undertake.

1010 Further, Professor Gneezy refused to engage with Professor Goldfarb’s “clicks are costly” analysis.

1011 Professor Gneezy stated, in respect of Professor Goldfarb’s “clicks are costly” analysis, that the application of that concept to the analysis of Android smart mobile device users’ behaviour on electronic devices absent proper empirical evidence is inappropriate.

1012 Professor Goldfarb provided empirical analysis by way of his consideration of the introduction of biometric features on iOS devices. That analysis was contained within the “clicks are costly” section of Professor Goldfarb’s report to which Professor Gneezy was instructed to respond.

1013 The analysis was also referred to in Professor Goldfarb’s response to the same question in the joint report to which Professor Gneezy was responding.

1014 Despite those matters, it became clear from the cross-examination that Professor Gneezy had not considered Professor Goldfarb’s biometric analysis in preparing his written report.

1015 Professor Gneezy was willing to describe Professor Goldfarb’s analysis as inappropriate without paying any real attention to the material upon which it was based.

1016 In my view, his criticisms of Professor Goldfarb on this aspect lacked substance.

1017 I accept Professor Goldfarb’s themes of “device first” and “clicks are costly” which I have also discussed elsewhere.

1018 Let me turn to another topic that excited the interest of the parties and their experts.

Profit-maximising or just profitable?

1019 Professor Asker considered that the HMT asks whether the hypothetical profit-maximising monopolist would raise prices by a small, significant, non-transitory amount (say 5%). This is important, because if one does not require profit-maximisation, the test would be hypothesising a scenario that would be economically irrational for the hypothetical monopolist. The reason for conducting the critical loss analysis with a 10% price increase, which merely asks whether that price increase would be profitable, as opposed to profit-maximising, is because of a result in the economics literature that if a 10% price increase is profitable, the profit-maximising price increase is generally at least 5%, whereas if it is profitable at 5% but not 10%, the profit-maximising price increase is generally below 5%.

1020 Contrastingly, Ms Edwards-Warren considered that the HMT merely asked whether the hypothetical monopolist would find it profit increasing, in contrast to profit-maximising, to increase price by a small, significant, non-transitory amount (say 5%). But she conceded that on her version of the test it could mean that while a 5% price increase may be profitable, it may not be the level to which a hypothetical monopolist rationally would increase their price. She acknowledged that her version of the test really means “5 to 10 per cent being profit increasing, which one can also interpret as – broadly as two and a half per cent to 5 per cent profit maximising”.

1021 Google says that Ms Edwards-Warren’s version of the test should be rejected, because it is inconsistent with economic theory and competition regulatory practice for the following reasons.

1022 First, no regulatory guideline or authority has been put forward by Epic to corroborate Ms Edwards-Warren’s suggestion that a 2.5% to 5% profit-maximising price increase is the appropriate way to implement a SSNIP.

1023 Second, the evidence shows that the HMT is normally employed as a profit-maximising test.

1024 Third, the ACCC merger guidelines are consistent with a profit-maximising test and Professor Asker’s approach.  I will return to this in a moment.

1025 Fourth, P Davis and E Garcés, Quantitative Techniques for Competition and Antitrust Analysis (2009, Princeton University Press) cited by Professor Asker supports the profit-maximising application of the HMT, not Ms Edwards-Warren’s profit-increasing version.

1026 Now I note that under cross-examination, Professor Asker observed that “this is a frustrating text in this regard because economists are loose with words and precise with math”.  But Professor Asker said that when one focuses on the mathematical explanation of the HMT, it is clear that the authors posit an inequality where profits are always increasing, and require it to hold for all price increases between 0% and 5%.  In other words, the test requires that the point of profit-maximisation is not reached before 5%. So, what was required was a profit-maximising price increase of at least 5%.

1027 Fifth, the fact that many Australian judgments have described the HMT as whether the hypothetical monopolist could profitably engage a SSNIP does not gainsay that the correct approach requires profit-maximisation, which is a nuance that arises only explicitly when conducting the HMT quantitatively.

1028 This is because in almost all competition law cases where market definition is in issue, the HMT has been applied as a qualitative thought experiment. There is no need for precision if the data does not permit a quantitative test.  By contrast, mathematical precision is important if one is conducting the test quantitatively.  So far as the quantitative analysis is concerned, the more precise formulation should be followed being the requirement for profit-maximisation.

1029 So, it follows that the HMT requires a profit-maximising price increase of at least 5%.  This means that the critical loss analysis should be implemented with a profitable 10% price increase, to ensure that at least a 5% price increase would be profit-maximising.

1030 If so, Professor Asker’s approach of applying a 10% price increase is correct, and his analysis demonstrates that Epic’s Australian Android mobile app distribution market fails the HMT, whether using his preferred diversion estimates or Professor Goldfarb’s.

1031 So, it is too narrow and must be expanded to include non-Android platforms, particularly app stores on gaming consoles.

1032 Now Professor Asker’s HMT asked whether a hypothetical monopolist would find it profit-maximising to increase prices by at least 5% rather than asking if the price increase would be profitable.

1033 So, a given increase in price might be profitable, that is, profit-increasing, but not profit-maximising.  But a profit-maximising formulation is more likely to lead to a broader market than a profitable formulation. Indeed, Professor Asker considered that asking whether a price increase is profit-maximising is equivalent, under certain assumptions, to asking whether double that price increase is profitable. That is, a profit-maximisation formulation is equivalent to doubling the magnitude of the price increase that must be profitable.

1034 I reject Professor Asker’s profit-maximising formulation of the HMT.  It is not consistent with the preponderance of Australian and US case law or the literature.  Let me first set out some Australian law on the profitable versus profit-maximising formulation of the SSNIP.

1035 In Australian Competition and Consumer Commission v Pacific National Pty Ltd (No 2) [2019] FCA 669; (2019) ATPR ¶42-633 I said (at [91] and [500]):

The “hypothetical monopolist” test or what is sometimes known as the “SSNIP” test can be used to assess substitution possibilities on the demand side.  The question posed is whether a hypothetical monopolist supplier could profitably impose a small but significant non-transitory increase in price (so the SSNIP acronym) say between 5-10%, for the supply of the relevant product or service, but holding constant the terms of sale of all other products and services.  The SSNIP being considered is relative to the prices that would prevail but for the merger.  The thought experiment, whether built upon detailed data of the relevant variables or assessed more qualitatively, starts with the hypothetical monopolist supplier and the product or service in issue to define the product and geographic market; the test does not assist concerning defining the functional dimension.  The question is then posed as to whether a SSNIP could be profitably imposed.  If not, the next best substitute on the demand side is added to the market definition.  The new market is then tested with the SSNIP test.  If a SSNIP could be profitably imposed, then you have your market definition.  If not, you again add the next best substitute to your market definition and test again, and so on.  The idea is to find the smallest area (product and geographic) over which the hypothetical monopoly supplier can impose the SSNIP which then provides the boundaries of the market as Yates J discusses in Australian Competition and Consumer Commission v Metcash Trading Ltd (2011) 198 FCR 297 at [247] to [251].

…

Second, as the ACCC correctly points out, the first of PN’s proposed preconditions, that the hypothetical monopolist is able to identify the customers to whom price can be increased, is not to be confused with the hypothetical monopolist test.  As I have said, that test can be used in the analysis of demand-side substitution that is undertaken for the purpose of product market definition.  Contrastingly, PN’s first proposed precondition is a simple question as to the feasibility of identification of the relevant users.  But the hypothetical monopolist test asks whether a hypothetical monopolist who supplied the particular service at issue would find it profitable to impose a SSNIP in the order of 5 to 10% in relation to that service.  If, in response to the price rise, customers would switch to acquiring other services in sufficient numbers so as to render the price increase unprofitable, those other services ought to be included in the relevant market.  But this test assumes that the hypothetical monopolist could apply a SSNIP to the relevant users in relation to the relevant services, and asks whether that SSNIP would be profitable.  Moreover, satisfaction of the hypothetical monopolist test does not require me to be satisfied that PN could have in the past, or could at present, implement a SSNIP to the relevant users.  PN has not been in the past, and is not at present, a monopolist supplier of Rail Services in any of the relevant markets.  As such, its past and present experiences have no bearing on the hypothetical monopolist test, notwithstanding that they are relevant to the analysis of whether the relevant conduct would substantially lessen competition.

1036 On appeal, in Australian Competition and Consumer Commission v Pacific National Pty Ltd (2020) 277 FCR 49, Middleton and O'Bryan JJ, and Perram J writing separately did not take issue with that description.

1037 I note that I cited Australian Competition and Consumer Commission v Metcash Trading Ltd (2011) 198 FCR 297 at [247] per Yates J who did not use the profit-maximising phraseology for the relevant SSNIP.

1038 In Re Qantas Airways Limited (2004) ATPR 42–027 at [241] the Australian Competition Tribunal said:

…Where the available competitive alternatives vary systematically as between customer types, the area over which a hypothetical monopolist may profitably impose a small but significant and non-transitory increase in price (referred to as the "SSNIP test") will be determined by reference to customer type, provided that the hypothetical monopolist is able to engage in price discrimination between the various customer groups.

1039 In Re Duke Eastern Gas Pipeline Pty Ltd (2001) 162 FLR 1 at [106] the Australian Competition Tribunal said:

…The small but significant, non-transitory increase in price analysis tests substitutability by asking whether a firm would gain or lose by attempting to impose a small but significant, nontransitory increase in price on its own product or service.

1040 In Australian Competition and Consumer Commission v Air New Zealand Ltd (2014) 319 ALR 388 Perram J referred to the profit-maximising formulation (at [220]):

… Through the lens of the HMT the relevant product market was generally the smallest collection of products or services such that a hypothetical monopolist over those products would find it profit maximising to impose a small but significant and non-transitory increase in price (a SSNIP) on at least one of the products or services.

1041 In Australian Competition and Consumer Commission v P T Garuda Indonesia Ltd (2016) 244 FCR 190, Dowsett and Edelman JJ stated at [24]:

His Honour explained the SSNIP at [220]. His explanation of the concept was not in dispute. In 1977, Professor (now Justice) Breyer summarised an early version of what is now the SSNIP test as “the notion of drawing a circle around a set of sales. Within the circle one should find sales with the following characteristic: were all those sales in the hands of a single seller, that seller would have the power to raise price significantly above the competitive level [and maximise profits]”: Breyer S, “Five Questions About Australian Antitrust Law” (1977) 51 Australian Law Journal 28 at 34. The SSNIP approach has been traced to the United States Justice Department 1892 Merger Guidelines, based on the work of Mr Werden. It has not been without controversy: Stigler GJ and Sherwin RA, “The Extent of the Market” (1985) 28(3) Journal of Law and Economics 555; Harris RG and Jorde TM, “Market Definition in the Merger Guidelines: Implications for Antitrust Enforcement” (1983) 71(2) California Law Review 464 cited in Re Fortescue Metals Group Ltd (2010) 242 FLR 136; 271 ALR 256 at [1033] (Full Tribunal).

1042 Now as is clear from the first passage, the statement of the SSNIP was not in issue in the Full Court. Further, Dowsett and Edelman JJ referred to Re Fortescue Metals Group Ltd (2010) 271 ALR 256 at [24], which adopted the profitable formulation of the SSNIP. Dowsett and Edelman JJ did not indicate that they were departing from this decision or endorsing the profit-maximising approach. Yates J, who dissented on the outcome in the appeal, framed the test as the profitable formulation, although there is nothing to suggest that Yates J was seeking to disagree with the majority judgment or the primary judgment.

1043 Let me next set out some US intermediate appellate case law examples from various circuits on the profitable versus profit-maximising formulation of the SSNIP.

1044 In United States v. American Express Company 838 F.3d 179 (2016) it was said at 198 to 199:

To define the relevant market, this Court often applies a "hypothetical monopolist test" ("HMT") asking whether a hypothetical monopolist acting within the proposed market "would be 'substantially constrain[ed]' from increasing prices by the ability of customers to switch to other producers." Todd, 275 F.3d at 202 (alteration in original) (quoting AD/SAT, Div. of Skylight, Inc. v. Associated Press, 181 F.3d 216, 228 (2d Cir. 1999)). Under the HMT, "'[a] market is any grouping of sales whose sellers, if unified by a hypothetical cartel or merger, could profitably raise prices significantly above the competitive level. If the sales of other producers substantially constrain the price-increasing ability of the hypothetical cartel, these others are part of the market.'" AD/SAT, 181 F.3d at 228 ...

The Court implements the HMT by imagining that a hypothetical monopolist has imposed a small but significant nontransitory increase in price ("SSNIP") within the proposed market. If the hypothetical monopolist can impose this SSNIP without losing so many sales to other products as to render the SSNIP unprofitable, then the proposed market is the relevant market. By contrast, if consumers are able and inclined to switch away from the products in the proposed market in sufficiently high numbers to render the SSNIP unprofitable, then the proposed market definition is likely too narrow and should be expanded.

1045 In Federal Trade Commission v. Penn State Hershey Medical Centre 838 F.3d 327 (2016) it was said at 338 and 344:

A common method employed by courts and the FTC to determine the relevant geographic market is the hypothetical monopolist test. Under the Horizontal Merger Guidelines issued by the U.S. Department of Justice's Antitrust Division and the FTC, if a hypothetical monopolist could impose a small but significant non-transitory increase in price ("SSNIP") in the proposed market, the market is properly defined. Merger Guidelines, § 4, at 7–8. If, however, consumers would respond to a SSNIP by purchasing the product from outside the proposed market, thereby making the SSNIP unprofitable, the proposed market definition is too narrow. … Important for our purposes, both the Government and the Hospitals agree that this test should govern the instant appeal. See Gov't Br. 25; Hosps. Br. 17–20.

…

… In determining the relevant market, we "look[ ] not to the contractual restraints assumed by a particular plaintiff," id. but instead, we answer whether a hypothetical monopolist could profitably impose a SSNIP.

1046 In Federal Trade Commission v. Advocate Health Care Network 841 F.3d 460 (2016) it was said at 465:

The court also heard testimony from several economic experts, including Dr. Tenn. He used the "hypothetical monopolist test" to identify the geographic market relevant to the case. That test asks whether a single firm controlling all output of a product within a given region would be able to raise prices profitably a bit above competitive levels. Economists and antitrust cognoscenti refer to such a price increase as a "SSNIP," a "small but significant [usually five percent] and non-transitory increase in price." U.S. Dep't of Justice & Federal Trade Comm'n, Horizontal Merger Guidelines 9 (2010). …

1047 In Epic Games Inc v. Apple Inc 67 F.4th 946 (2023) it was said at 975 to 976:

A market comprises "any grouping of sales whose sellers, if unified by a monopolist or a hypothetical cartel" could profitably raise prices above a competitive level. Rebel Oil Co., Inc. v. Atl. Richfield Co., 51 F.3d 1421, 1434 (9th Cir. 1995). If the "sales of other producers [could] substantially constrain the price-increasing ability of the monopolist or hypothetical cartel, these other producers must be included in the market." …

Often, this inquiry involves empirical evidence in the form of a "SSNIP" analysis. That analysis echoes Rebel Oil and uses past consumer-demand data and/or consumer-survey responses to determine whether a hypothetical monopolist could profitably impose a Small, Significant, Non-transitory Increase in Price above a competitive level.

…

In Kodak, the Supreme Court considered the question of whether a lack of market power in the foremarket (photocopier machines, generally) categorically precludes a finding of market power in the aftermarket (replacement parts for and servicing of Kodak-brand photocopiers), which Kodak had allegedly achieved by contractually limiting customers to Kodak-provided parts and services. 504 U.S. at 455, 466, 112 S.Ct. 2072. ... The Supreme Court thus folded aftermarkets into the framework for assessing markets generally, evaluating cross-elasticity of demand to determine whether a hypothetical monopolist could profitably charge a supracompetitive price. See id. at 469, 112 S.Ct. 2072. …

1048 It was said in Re Southeastern Milk Antitrust Litigation 739 F.3d 262 (2014) at 277 to 278:

… Using the hypothetical monopolist as an entity that controls all the suppliers in a proposed market, a question is posed: could a monopolist profit if it imposed a "small but significant non-transitory increase in price" ("SSNIP")? Id. Typically, the increase that is posited is five percent. Ky. Speedway, 588 F.3d at 918. If buyers in a defined area would respond to a small, lasting increase in price—a SSNIP—by purchasing from another supplier, rendering the SSNIP unprofitable, the market has been too narrowly defined. See FTC v. Tenet Health Care Corp., 186 F.3d 1045, 1053, n. 11 (8th Cir.1999) (citing Merger Guidelines § 1.21). …

1049 In Federal Trade Commission v Tenet Health Care 186 F.3d 1045 (1999) it was said at 1053:

Tenet argues that the evidence relating to "marginal consumers" shows that the merged entity would be unable to raise prices without causing the "critical loss" of enough patients to make the increase unprofitable. This "critical loss" approach is in fact employed by the FTC in its own Horizontal Merger Guidelines which are used by the FTC to ascertain a relevant geographic market in exercising its prosecutorial discretion to challenge a merger. See Department of Justice, Federal Trade Commission, Antitrust Division, 1992 Horizontal Merger Guidelines, 57 Fed.Reg. 41552 § 1.21.

1050 Now as to the literature, Professor Asker relied on P Davis and E Garcés (2009) as a standard reference.  But it consistently employs the profitable formulation. Pages 201, 202 to 204, 206 to 208 and 227 are replete with such references or their analogues rather than using terminology such as profit-maximising.

1051 For example, on pp 202 and 203 it was said:

… To do this, the HMT assumes that all products within the proposed market definition are owned by one single producer which sets each of their prices in an attempt to maximize the total profits derived from them. If the hypothetical monopolist finds it profitable to increase prices, we will have found that constraints from goods outside the proposed market definition are not a sufficient constraint on producers within the market to render a price rise unprofitable. In other words, prices were kept down by the competition within the market. In practice, to operationalize this idea we must, among other things, be a little more precise about exactly what we mean by a "price rise." To that end most jurisdictions apply the "SSNIP" lest, which looks at whether a "small but significant nontransitory increase in prices" would be profitable for the hypothetical monopolist. Usually, "small but significant nontransitory" is assumed to mean 5-10% for a year.

…

… We then need to evaluate whether a monopolist of this product could profitably raise prices by 5-10% for a year. If so, that single product will then constitute our antitrust market. If not, we must include the "closest" substitute, that product which provides the best alternative to consumers facing the price increase. We then assume again a hypothetical monopolist, this time of each of the products in our newly expanded set of products in our candidate market and we repeat our question, will a 5-10% price increase for a year be profitable?" …

1052 Now Professor Asker said that the formulae below support the profit-maximising formulation. But this is not clear, as the formulae are not in and of themselves saying profit-maximising.

1053 The statement from the above extract that “[a] monopolist of product 1 will increase price as long as it raises their profits” and the immediately following formula is consistent with a profitable formulation.

1054 Now in fairness to Professor Asker, the part of the statement that he relied on is the later statement "[w]e will want to evaluate whether this inequality holds for all prices between p1Comp and p15% = 1.05p1Comp ".  Now admittedly the reference in that sentence to all prices is consistent with a maximising criterion. But the formulae on the immediately preceding and following pages clearly support a profit-increasing approach as the formulae focus on the difference between the profits of a hypothetical monopolist imposing a SSNIP and the profits of a hypothetical monopolist pricing at the competitive level and whether that difference is greater than zero.

1055 So on the immediately preceding page it was said (p 204):

1056 And on the following page it was said (p 206):

1057 It was then said (p 211):

1058 And it was then said (pp 217 and 218):

1059 Further, Ms Edwards-Warren considered that in Australia the hypothetical monopolist test is normally applied asking whether a hypothetical monopolist would find it profitable or profit-increasing to raise prices by 5 to 10%.  Moreover, Professor Wright, Professor Hitt and Mr Houston all agreed on the profitable formulation of the SSNIP test in the Apple proceeding.

1060 Further, the test as set out in the ACCC merger guidelines ([4.19] and [4.20]) also asks whether a price increase would be profitable and is expressed in the following terms:

The HMT determines the smallest area in product and geographic space within which a hypothetical current and future profit-maximising monopolist could effectively exercise market power. In general, the exercise of market power by the hypothetical monopolist is characterised by the imposition of a small but significant and nontransitory increase in price (SSNIP) above the price level that would prevail without the merger, assuming the terms of sale of all other products are held constant.

The process of applying the HMT starts with one of the products and geographic areas supplied by one or both of the merger parties. If a hypothetical monopolist supplier of this product cannot profitably institute a SSNIP because of customers switching to alternative products, the next closest demand substitute is added. If a hypothetical monopolist supplier of this extended group of products cannot profitably institute such a price increase because of customers switching to alternative products, the next best substitute is added. The collection of products is expanded until a hypothetical monopoly supplier of all those products could profitably institute a SSNIP.

1061 Now [4.19] merely states that the hypothetical monopolist is profit-maximising.  But [4.20] states that under the HMT, one tests whether that profit-maximising hypothetical monopolist could profitably implement a SSNIP. The paragraph does not require the price increase to be profit-maximising. The perfect Platonic form of a monopolist might only profit-maximise, but I am only concerned with the imperfect and base monopolist who in the worldly realm is hypothesised to be taking (or not) a profitable step.

1062 Further, the formulation of the Continental test is to ask if a 5% or 10% increase in price is profitable.

1063 Moreover, Professor Asker’s reasoning for why he preferred the profit-maximisation formulation was problematic in any event.

1064 Professor Asker said that it is economically irrational for a hypothetical monopolist to give up profits by choosing a profitable price if an intermediate price is in fact profit-maximising.  But that is not an argument for a profit-maximising formulation of the HMT as Epic points out.  The HMT is a conceptual framework used to define the market. The idea is to seek to identify whether there is a significant price constraint on a given set of products, which constraint comes not from the intra-candidate market but from the availability of other products that are outside the proposed market definition and that offer viable substitutes. It is not necessary to ask what price rise a hypothetical monopolist would in fact implement when assessing whether it is price constrained.

1065 Now why am I discussing any of this? Why does it matter?

1066 Well Professor Asker’s profit-maximising approach based on a SSNIP of 5 to 10% would likely lead to a requirement that a 10 to 20% price increase is profitable.  But a 20% price increase is not a small increase, and such an approach would likely lead to unrealistically defined markets.

1067 Indeed, Professor Asker accepted that asking whether a 5% price increase is profit-maximising is, under certain assumptions, the equivalent of asking whether a 10% price increase is profitable.

1068 But even a 10% price increase may be problematic in a case such as that before me being a misuse of market power case.  A 5% SSNIP may be more appropriate where there is a risk of the cellophane fallacy which I have discussed elsewhere.  It would seem that a small SSNIP should be preferred out of an abundance of caution when the cellophane fallacy might be in play, such as in a misuse of market power case, as Dr Rhonda Smith, University of Melbourne, has pellucidly observed.

Two-sided markets – applying a SSNIP test

1069 As was explained in L Filistrucchi et al, “Market Definition in Two-Sided Markets: Theory and Practice” (2014) 10(2) Journal of Competition Law & Economics 293, interesting issues arise when attempting to extend the SSNIP test to a two-sided market. First, given that in a two-sided market firms set two prices, one on each side of the market, the question is which price the hypothetical monopolist should be raising. Second, given that in a two-sided market there are indirect network effects between demands, and therefore profits, on the two sides, the issue is whether one should consider profits on one or both sides of the market.

1070 Now I discussed that article in detail in my reasons in Epic v Apple and I will not repeat that discussion.

1071 But in addition, in L Filistrucchi, “A SSNIP Test for Two-sided Markets: The Case of Media” (NET Institute working paper 08-34, 2008), the author highlights that the logic behind the traditional SSNIP test is to define a market as the smallest set of substitute products on which a monopolist would find it profitable in the EU (or profit-maximising in the United States) to increase prices by a small-but-significant amount. So, it is designed to ensure that the market is defined in a way that a monopolist has market power, which is a basic requirement of economic theory.

1072 To maintain the same rationale when dealing with two-sided markets, one should allow the monopolist to optimally adjust the price structure. The author also proposed some critical loss analysis formulas to perform the SSNIP test in a two-sided non-transaction market under the usual assumptions of linear or isoelastic demand and constant marginal costs.

1073 The author then argues that, while using the standard single-sided critical loss analysis formulas would lead to the definition of a relevant market that is too narrow, adopting the formulas proposed by others would lead to the definition of a market which is too large. Indeed, not allowing the price structure to be adjusted optimally would overestimate the loss in profits due to the increase in prices, because by definition the optimal adjustment by the hypothetical monopolist will tend to reduce such a loss.

Bundled products

1074 Let me say something here on the question of bundling.

1075 When products are sold only in a bundle, there can be a question as to whether they are separate products from the viewpoint of customers or whether they should be treated as components of a single product.

1076 Now as Ms Edwards-Warren pointed out, if customers are prepared to acquire each product without the other attached, and if this is technically possible, then from an economic perspective they should be treated as separate products with separate demand. If customers would only buy one of the products with the other attached, then one product is a component of the other and they should not be considered separately.

1077 Now this is a relevant consideration in this case in relation to the Play Store and Google Play Billing, given that they are effectively supplied as a bundle in Australia to app developers that monetise via Play Store in-app purchases. Google Play Billing is not available as a standalone payment solution and is not available outside the Play Store.

1078 So, when assessing the relevant markets for each of these two products, one therefore needs to consider both whether there is demand for the Play Store as a stand-alone product, that is, for app distribution services separate from payment solutions services, and whether there is demand for Google Play Billing as a stand-alone product, that is, for payment solutions services separate from app distribution services.

1079 Now Ms Edwards-Warren and Professor Asker disagreed as to whether the Play Store and Google Play Billing are correctly understood as a single product or distinct products.

1080 According to Professor Asker, the framework Ms Edwards-Warren used to make the determination that they are distinct products, namely, if customers are prepared to acquire each product without the other attached and if this is technically possible, is incomplete.

1081 He expressed the view that as a conceptual exercise, almost all products can be viewed as a bundle of their components. This is true even in cases where customers typically or always buy an integrated product (a pure bundle) rather than combine the components themselves.  So, a pair of shoes is a bundle of a left shoe and a right shoe. A newspaper is a bundle of the different sections in the paper. And a car is a bundle of a chassis, an engine, safety equipment, and many other components.

1082 Now to devise relevant criteria for determining when a bundle constitutes one or multiple products, Professor Asker explored these examples in more depth in the following fashion.

1083 A newspaper may be both a single product or multiple products depending on whether it is distributed physically or electronically or both. A newspaper’s readers can have separate demand for different sections of the newspaper. Some consumers only want to read the fashion section, whilst others only want to read the business section. But there are production and transaction cost efficiencies to offering all sections of a newspaper as a bundle, depending on how the newspaper is distributed.

1084 For physical distribution, there are transaction costs to selling each section of a newspaper separately as that increases the number of products a vendor needs to price, the time to complete selling an entire newspaper, and the complexity of delivery on a paper route.

1085 This is not true for electronic distribution where readers can simply click to transact for the sections they want. This difference is reflected in the fact that physical newspapers are largely sold as a bundle, that is, all sections together, whereas newspapers online can and in some instances do offer separate subscription products for specific sections.

1086 It is natural to think of a car as a single product, and that the components of a car are not distinct products. This is because for almost all consumers purchasing a car, the components of a car are sufficiently strong complements such that no firm in a perfectly competitive equilibrium would offer an “unbundled” car, that is, a pile of parts that consumers assemble on their own, or any subset of parts.

1087 Put differently, the value to the consumer of any component of the car requires that the consumer have the other components properly assembled to have a functional car. This is also because suppliers of cars have production efficiencies, that is, an integrated assembly line, to supply an entire car rather than subsets of a car.

1088 A pair of shoes is a single product. Left shoes and right shoes are not distinct products. This is because for nearly all consumers purchasing shoes, left and right shoes are sufficiently strong complements such that no firm in a perfectly competitive equilibrium would offer an “unbundled” pair of shoes that consumers mix-and-match. This is also because a firm supplying left shoes would have strong production efficiencies to also supplying right shoes, rather than another firm incurring the fixed costs of establishing a separate manufacturing facility for right shoes.

1089 So, Professor Asker developed and deployed the following economic criteria when making this determination.

1090 His first bundled products criterion posed the question: does separate demand exist for different components of the bundle? This question asks whether the components are close complements for consumers.

1091 His second bundled products criterion posed the question: is it more efficient, that is, output expanding, to supply the components of the bundle as a pure bundle absent the challenged conduct at but-for prices? This question asks whether production efficiencies exist because of, for example, reductions in transaction costs or production costs via bundling regardless of the price charged.

1092 He expressed the view that the economic content behind these criteria is to determine whether, starting from the position that the components of the bundle are distinct products, they would be sufficiently close complements for customers or would give rise to sufficient production and distribution efficiencies for suppliers, such that in market equilibrium the components would always or nearly always be supplied as a bundle. This is because when bundling increases the value that is created by engaging in trade, firms in a competitive market will nearly always supply the components as a bundle. If the components are both close complements and give rise to supply-side efficiencies, then a bundle is all the more likely to arise as the competitive market offering.

1093 Now Ms Edwards-Warren used only the first of the two criteria Professor Asker deployed to determine whether components of a bundle are single or multiple products, that is, whether customers have separate demand for different components of the bundle.

1094 But according to Professor Asker, this is economically incomplete. A physical newspaper would plausibly have separable customer demand for different sections of the newspaper but is nonetheless always sold as a single product given the production efficiencies of doing so.

1095 But for the moment let me say that I agree with Ms Edwards-Warren that this additional criterion is not an accepted economic principle for conducting a competition assessment.

Epic’s posited Android mobile app distribution market — what is the correct focal product?

1096 As I have indicated, Epic’s case is that there is and has been at all relevant times a market in Australia for the supply of services to developers and users for the distribution of Android apps to Android device users. I will refer to this as the Android mobile app distribution market.

1097 Epic says that this is the market in which Google engages in the app developer restrictive conduct and in which that conduct has effects upon competition. And it says that it is also a market in which the OEM restrictive conduct has effects upon competition.

1098 Moreover, Epic says that it is a market in which Google has a substantial degree of power, which it leverages, not only in engaging in the app developer restrictive conduct but also in imposing the payments tie and the anti-steering rule in the Android in-app payment solutions market.

1099 Now Google disputes the existence of the Android mobile app distribution market and contends instead for markets for facilitating sales of different types of digital content, but I agree with Epic that these markets do not meet the purposive character of the market definition exercise and the need to be consistent with commercial reality.

1100 As Epic says, they are artificial, including because they are driven by the manner in which Google happens to monetise its services from time to time.  Further, they do not take into account various aspects of the services in fact supplied by Google.  Further, they do not accord with commercial reality.

1101 Moreover, they are not an appropriate framework within which to assess Google’s impugned conduct.

1102 Let me begin by addressing the principal issue that divides the expert economists being the product by reference to which they commenced their competition analysis, referred to as the focal product including services. And it is convenient at this point to summarise my principal views.

1103 Now the Android mobile app distribution market is the appropriate lens through which to analyse the app developer restrictive conduct and the effects of the OEM restrictive conduct.

1104 Now Professor Asker has criticised such a market based in part upon two fundamental areas of dispute between Professor Asker and Ms Edwards-Warren. The first relates to the selection of an appropriate focal product as the starting point for the market definition exercise. The second relates to Professor Asker’s econometric analysis.  At the moment I will only address in detail the first matter.  I will leave the econometrics until later.

1105 Now as to the first area of dispute, it must be said that I have not seriously entertained Professor Asker’s selection of the facilitation of sales of digital content as the focal product. Such a focal product is inconsistent with the commercial reality of the services Google offers developers and users via the Play Store.

1106 First, it does not capture all of the services affected by the impugned conduct, and it is inconsistent with the contractual position.

1107 Second, it does not account for all of Google’s incentives to compete, even assuming those incentives have any relevance to the selection of the focal product.

1108 Third, it is insufficiently flexible to account for Google’s ability to change its monetisation strategy at any time.

1109 These difficulties are really a manifestation of the fact that Professor Asker’s inquiry into the focal product did not properly commence with the challenged conduct, but instead with how Google monetises the Play Store.

1110 I agree with Epic that Professor Asker’s starting point is unorthodox, and results in an artificial focal product that obscures rather than elucidates the competition inquiry.

1111 Contrastingly, Ms Edwards-Warren’s identification of app distribution services as the focal product reflects the commercial reality of the core services offered by Google via the Play Store.  And it is supported by the descriptions of Google’s services in the DDA and the developer program policies, as well as Google’s documents and the evidence of its witnesses. It is a sound starting point for the market definition inquiry.

1112 Now as to the second central area of dispute, Professor Asker’s primary criticism of Ms Edwards-Warren’s conclusions as to the boundaries of the Android mobile app distribution market was premised on an analysis involving the application of three quantitative methods to Fortnite revenue data to estimate diversion ratios from the Play Store in response to a change in price or quality. But I should say upfront that in my view those quantitative analyses are flawed and do not support Professor Asker’s contention that there is strong substitution away from the Play Store to console stores and PC stores.  I will address these analyses in much greater detail later.

1113 Now after such analyses one is then left to consider the respective experts’ analysis of potential substitutes for the app distribution services Google provides via the Play Store.  I will also address such matters in greater detail later.  But it is worth noting here that I have reached the view that the only close constraint on Google’s app distribution services are other Android app stores.

1114 Now as I have said, my key conclusion is that the relevant market for analysing the effects of the OEM restrictive conduct, the app developer restrictive conduct, and Google’s market power, is the Android mobile app distribution market. The participants in that market are the Play Store and other Android app stores.

1115 But I should say that even if app distribution services provided by Apple via the App Store are a sufficiently strong constraint on Google’s app distribution services provided via the Play Store so as to warrant inclusion in the market, the relevant market would then be one that encompasses distribution of both iOS and Android apps.

1116 But even the adoption of that market would not save Google.  Such adoption would not affect the conclusion that Google has substantial market power and that its conduct has substantially lessened competition.

1117 Let me now turn in detail to the identification of the focal product.

Google’s arguments

1118 There are three principal matters that Google has put against Epic’s identification of the focal product. Let me summarise its position on each aspect.

1119 Now the first matter that Google has raised is that it says that the relevant commercial agreements and the commercial structure of its operations do not support Epic’s focal product.

1120 Epic says that the fact that the focal product is app distribution services should not be controversial, because it follows from Google’s contracts with developers. But Google says that the central problem with Epic’s approach to the product which the Play Store supplies is that it is an incomplete concept of the product.

1121 The pleading refers to the product as services that enable app developers to reach, offer and provide Android device users with Android apps and updates, and that enable Android device users to be presented with and/or find, obtain and utilise Android apps and updates.

1122 But Google says that this focus only looks at part of what the Play Store supplies and ignores Google Play Billing and the fact that the Play Store does in fact facilitate transactions for purchasing digital content between users and developers.

1123 Google says that it supplies an integrated product that comprises a range of service features that include both providing a way for developers to make apps and app updates available and for users to download, install and update apps, and Google Play Billing for the benefit of users and developers.

1124 Further, the underlying DDA with developers does not support Epic’s conceptualisation of the product supplied by the Play Store. Google says that the suggestion that the Play Store merely supplies distribution services under the DDA that exclude Google Play Billing should be rejected for the following reasons.

1125 First, clause 1 defines “Google Play” as “[t]he software and services, including the Play Console, which allow Developers to distribute Products to users of Devices”.  Further, the evidence shows that “Play Console” includes Google Play Billing and that it is therefore part of the Play Store service supplied under the DDA.

1126 Second, clause 3.3 of the DDA makes it clear that Google Play Billing is part of the services supplied under the DDA. It states that:

Products are displayed to users at prices You establish in Your sole discretion. If You utilize Google Play's billing system or the Play Console to control settings related to any taxes (including taxes, fees, or surcharges that are outside the scope of this Agreement) on the sale of Products, or if Google believes that Taxes may be owed by You or Google on the sale of Products, You grant Google permission to charge, collect, and discharge any such taxes as required in accordance with applicable law. …

1127 Third, clause 3.4 of the DDA provides for the payment of a service fee by the developer on products for which the developer charges a fee:

… A “Service Fee”, as set forth here and as may be revised by Google from time to time with notice to You as described in Section 15, will be calculated and charged on the sales price. …

1128 The “here” in this clause is a hyperlink to a document which describes the service fee. That document states that apps and in-app products sold through the Play Store’s billing system or an alternative billing system are subject to a service fee.

1129 At the bottom of the service fee document is another hyperlink to a 2023 document titled “Understanding Google Play’s Service Fee”. This document states that the service fee covers the cost of Google Play Billing and a range of other services supplied by the Play Store.  This is clear from the following statement under the question “[w]hat does the service fee pay for?”:

The service fee supports our investments across Android and Google Play, reflects the value provided by Android and Google Play, enables us to deliver an affordable and innovative user experience, helps developers reach users and build sustainable businesses, and keeps the platform safe and secure.

1130 One of the stated investment areas under this statement is the billing system. So, the service fee is charged not only for processing payments for apps and in-app digital content sold by developers, but for the whole integrated package of features and services that the Play Store provides to developers.

1131 The fact that one service fee is charged with respect to the whole package of features supplied by the Play Store is consistent with the proposition that the Play Store supplies an integrated service that includes app discovery, app distribution and Google Play Billing, amongst many other features.

1132 So, Google says that as a matter of commercial reality and economic substance, what the Play Store supplies to developers under the DDA is an integrated product that includes features such as discovery and distribution of apps, and Google Play Billing.

1133 It says that the fact that some developers do not use all the features of the service the Play Store provides, such as Google Play Billing when developers offer free apps, does not detract from the fact that the Play Store supplies an integrated product that comprises multiple features.

1134 Fourth, Google says that whilst it is true that clause 3.2 of the DDA states that developers who wish to be paid for products distributed via the Play Store must have a valid “Payments Profile” under an agreement with a “Payment Processor”, this does not deny that the Play Store provides an integrated product that includes Google Play Billing under the DDA.

1135 Clause 1 of the DDA defines the payment processor as a “Google-affiliated entity providing services that enable Developers to receive payments for Products sold via Google Play”. In Australia, the payment processor is Google Payment Australia Pty Ltd.

1136 The agreement which users and developers must enter into with Google Payment Australia Pty Ltd is the product disclosure statement (PDS). Under the PDS, the service which Google Payment Australia Pty Ltd supplies is the “Google Payments Service”, which is an online payment processing service designed to facilitate processing of payments between purchasers, that is, users, and merchants, that is, developers.  But payment processing is only part of what Google Play Billing provides.

1137 So, the fact that a separate Google entity provides the payment processing part does not mean that Google Play Billing is not part of the services supplied under the DDA.

1138 Fifth, Google says that there is some distinction between the product that the Play Store supplies as a matter of commerce, and the legal contracts which Google enters into with developers and users that govern and facilitate the supply of that product.

1139 Now it was put to Mr Feng that the legal contracts are inconsistent with the view that the Play Store supplies an integrated product that includes Google Play Billing.  But Mr Feng’s view was that it is an integrated product, and that there is a distinction between the legal and the commercial dimensions.  And Mr Feng considered it artificial to separate out part of the Play Store product based on the different contractual entities.

1140 Further, it is said that Mr Feng’s evidence is also corroborated by an August 2019 presentation titled “Project Magical Bridge” that he gave to Mr Samat, which conceptualises the Play Store as comprising multiple features including user acquisition, monetisation, distribution and developer tools and services.  Google views the Play Store including Google Play Billing as an integrated product.

1141 Sixth, Google says that after being taken to the DDA and associated the Play Store console help pages, Ms Edwards-Warren agreed that Google charges the service fee for a range of services including Google Play Billing.

1142 It is said that she also agreed that Google supplies an integrated product for which it charges a single service fee providing both billing services and a range of other services, and does so pursuant to a contract between Google and the developer.

1143 Seventh, for the purposes of market definition, it is not in doubt that I should focus on the nature of the product supplied in a commercial sense, not some artificially constructed product based on legal form that does not accord with commercial reality or the economic substance of what is supplied.

1144 Google says that it is artificial and contrived in this case to characterise certain distribution features of what the Play Store provides to users and developers as separate and distinct from everything which Google Play Billing provides, when in a commercial sense it is one integrated product that has several features, not all of which will be used by every user or developer. And nor do they have to be for the Play Store to be characterised as a single integrated product.

1145 The second matter that Google has raised is that Google’s internal documents do not support Epic’s focal product.

1146 Google says that its documents do not describe the core service of the Play Store as making apps and app updates available to users on behalf of developers to the exclusion of Google Play Billing and every other feature that the Play Store provides.

1147 Now Epic has drawn attention to the fact that a slide deck which Mr Gennai prepared in August 2019 titled “Let’s talk about our business model” mentions that “Play makes $ (primarily IAP) through strength in its distribution”.  But Google says that all that sentence says is that a high number of users will lead to a high number of in-app purchases. It says nothing about what services or features are included in the core product supplied by the Play Store.

1148 The third matter that Google has raised concerns Professor Asker’s analysis.

1149 Now Epic made three criticisms of Professor Asker’s approach. First, it said that his approach excluded all services that do not involve the transfer of money to purchase digital content, and so frustrated the analysis of competitive effects. Second, it said that his approach erroneously focused on how the Play Store monetises at the moment, which means the product dimension of the market will change if Google alters its monetisation strategy. Third, it said that Professor Asker was unable during the concurrent evidence session to justify his approach in a comprehensible way.

1150 But Google says that these points have no substance.

1151 As to the suggestion that Professor Asker excluded services for which there is no payment of money and therefore frustrated the competitive effects analysis, Google says that that is wrong for several reasons.

1152 First, the focal product of facilitation of transactions for digital content does not exclude the Play Store features such as distribution or search.  The transactions facilitation service that the Play Store supplies to consumers and developers involves several features including search and discovery of content, distribution of content, intermediation of all transactions by assuming counterparty risk, and a secure billing system.

1153 Second, Professor Asker’s facilitating transactions for digital content products also does not exclude all the services with respect to free apps or apps with transactions for physical goods because of indirect network effects.

1154 In other words, because there are indirect network effects at work, the services supplied to free apps or apps with transactions for physical goods make the Play Store more attractive and support the facilitation of paid digital content transactions, and are therefore captured within the notion of facilitating transactions for digital content.

1155 Third, Professor Asker did take incentives into account, unlike Ms Edwards-Warren’s approach which excluded the incentives associated with all the Play Store revenue apart from paid app downloads, which corresponds to less than 1% of the Play Store revenue.

1156 Fourth, in any event, Google says that Professor Asker’s approach does not frustrate the approach with respect to market definition, because it ensures that there is a focus on the source of app stores’ incentives to compete, and how the conduct affects those incentives.

1157 Professor Asker explained this in his report in terms that the relevant question for evaluating market definition is to understand how a hypothetical firm’s profits would change in response to a price increase or quality reduction. He said that what disciplines the firm from doing so is the possibility that the firm would lose revenue because its customers substitute in response to the firm raising prices or lowering quality. He said that in most cases, firms charge a price or fee for the goods or services they sell, and so the relevant metric for understanding this substitution is customer spending.

1158 As to the suggestion that Professor Asker’s approach was erroneous because it focused on how the Play Store monetises at the moment, and the product dimension could change if Google were to alter its monetisation strategy, Google says that that is misconceived for various reasons.

1159 First, it says that Professor Asker gave evidence that it is normal economic practice to focus solely on the current monetisation model employed by the firm of interest.

1160 Second, it says that conducting the economic analysis based on how Google and most of its competitors makes money is a reasonable place to start because it reflects what is actually happening in the marketplace and is therefore informative of what would happen today and in realistic futures. It is a market-informed way to understand the sources of competitive pressure that exist, grounded in objective fact.

1161 Third, it says that it would be erroneous not to approach the economic analysis informed by how firms operate in the marketplace today, because as Professor Asker said, the hypothetical monopolist is still subject to the realities of the market. The fact that one firm has chosen a way to monetise is informative as to what all the real world commercial constraints are. To conclude otherwise would be to depart from commercial reality.

1162 Fourth, it says that there is nothing problematic about the fact that Professor Asker acknowledged that if the Play Store’s monetisation strategy were different, it is possible that the focal product could be different. His position seemed to be that it was not illogical that the product market could change with a change in monetisation strategy.

1163 This view reflects the fact that the question of market, especially the product market, does not necessarily admit of a unique answer. It involves value judgments and is purposive. The product market in one situation could differ from the product market in another situation, depending on the nature of the relevant competitive process and the conduct at the time.

1164 Finally, as to the suggestion that Professor Asker did not explain how he had taken into account developers who do not make sales of digital content, and was therefore evasive, Google says that Professor Asker explained how services supplied for free apps were taken into account through indirect network effects.  Further, Google says that he explained that anything which supports paid digital content transactions was captured within his product market.

Analysis

1165 Now the first step in the market definition process is to identify the focal product which is supplied by Google. And as Epic points out, that requires considering the conduct being analysed, and the commercial context in which it occurs.

1166 The app developer restrictive conduct that Epic has alleged against Google in summary consists of the following.

1167 First, it consists of Google’s conduct in imposing the DDA terms on all developers who wish to distribute apps via the Play Store, including clauses 4.5 and 4.1 of the DDA.

1168 Second, it consists of Google’s conduct in imposing the developer program policies on all developers who wish to distribute apps via the Play Store, including the device and network abuse policy and the payments policy.

1169 Third, it consists of Google’s conduct in entering into the GVP agreements and the AVP agreements with certain developers, which agreements formed an addendum to each developer’s DDA and placed obligations on those developers in connection with the launch time, content, and removal of apps distributed via the Play Store.

1170 Such conduct is concerned with the terms on which Google supplies app distribution services to developers and users via the Play Store. Each of the impugned agreements is concerned with Android app distribution.

1171 The OEM restrictive conduct which Epic says has anti-competitive effects in this market is also centrally concerned with Android app distribution.

1172 The relevant MADA terms that I have been required to focus on concern the pre-installation of and default home screen placement for the Play Store which it is said ensure that the Play Store is the default choice of users for Android app distribution services and hinder the ability of rival Android app stores to compete for the provision of those services.

1173 It is said that the RSA3s incentivised the exclusive pre-installation of the Play Store and hindered the ability of rival Android app stores to compete with the Play Store for the provision of distribution services by disincentivising their pre-installation.

1174 Additionally, it is said that the technical restrictions imposed on OEMs hinder the ability of rival Android app stores to reach and supply distribution services to Android device users in competition with the Play Store.

1175 So, I agree with Epic that a consideration of the impugned conduct directs attention to the app distribution services which Google supplies via the Play Store.

1176 Let me at this point say something more about the DDA.  The DDA contains several provisions which identify the nature of the services supplied by Google under the DDA.

1177 Clause 2.1 describes the DDA as an agreement in relation to the developer’s “use of Google Play to distribute Products”, that is, apps, and provides that “Google will … display and make [the developer’s] Products available for viewing, download, and purchase by users”.

1178 So, the DDA relates to the developer’s use of the Play Store’s distribution service, and obliges Google to provide distribution services, that is, it obliges Google to make apps available for viewing, download and purchase by users.

1179 Clause 3.1 of the DDA further provides that the developer appoints Google as its agent or “marketplace service provider” in order “to make [the developer’s] Products available in Google Play”.

1180 So, the relevant Google entity’s role under the DDA is to make apps available in the Play Store, that is, to distribute the developer’s apps on behalf of the developer.

1181 Further, clause 3.2 provides that the developer’s products are “distributed via Google Play”. Further, there is the statement in clause 3.3 that “Products are displayed to users” on the Play Store. Further, there is the reference in clause 3.4 to products being sold or made available to users “via Google Play”.

1182 Further, there is the authorisation granted to Google under clause 5.1 to market products on the Play Store and to provide “hosting services … to allow for the storage of and user access to the Products”.

1183 Accordingly, the terms of the DDA support the conclusion that Android app distribution is the core service provided by Google to developers and users through the Play Store.

1184 The developer program policies lead to the same conclusion.

1185 Those policies include the statement that “[t]hese Developer Program Policies, along with the [DDA], ensure that together we continue to deliver the world’s most innovative and trusted apps to over a billion people via Google Play”. There is also the statement that  “[p]eople from all over the world use Google Play to access apps and games every day”.

1186 Let me turn then to some broader evidence and propositions advanced by Epic that in my view were either not contentious or could hardly be disputed.

1187 The Play Store provides a single and accessible mechanism by which developers can distribute their apps to a global audience using a click through distribution agreement. It facilitates the distribution of apps to mobile devices.

1188 The central purpose of the Play Store, from the perspective of both users and developers, is the process of getting apps from developers to users’ devices.

1189 The Play Store provides a method for app developers to make their apps available to users and enables users to download, install and update Android apps on their device.

1190 Those distribution services are provided to developers on the terms of the same DDA and subject to the developer program policies, regardless of whether the developer’s app is a free app, a paid app, or an app with in-app purchases, whether of physical or digital content.

1191 Moreover, the Play Store supplies app distribution services to users, regardless of whether they pay for the app or make in-app purchases of digital or physical goods within the app.

1192 Google has described the Play Store as an “app distribution system” and a “distribution platform”. Google has also described the services provided by the Play Store as “discovery and distribution services”.

1193 A slide deck which Mr Gennai prepared in August 2019 titled “Let’s talk about our business model” includes the following pertinent descriptions of what the Play Store does: “Android reaches lots of people + Google distributes its apps”; “Play is the preeminent app store within Android”; “[g]ame and app developers BUILD for Android and choose PLAY to reach these users”; “Play makes $ (primarily IAP) through strength in its distribution”.

1194 In cross-examination, Mr Gennai denied that “through strength in its distribution” referred to the Play Store’s ability to distribute other people’s apps, but I agree with Epic that it plainly does refer to the strength of the Play Store’s distribution service.

1195 A 2015 document co-authored by Mr Gennai titled “Google Play | The Rise of New Storefront Platforms” described the “different levels of service that a mobile app store platform (Google, Apple, Amazon) provides for its developers”, namely, discovery, distribution, business services and development services. Mr Gennai agreed that the Play Store has provided each of those services to developers throughout the relevant period. Each is an aspect of the core service, being Android app distribution.

1196 Ms Kochikar described the services that the Play Store provides directly to developers as including facilitating discovery of apps by users and app distribution. Moreover, she described the value provided by Google to developers through the Play Store, and the broader Android ecosystem, as including the maintenance of a secure distribution platform.

1197 Mr Gennai explained that, using the Play Store, users can search for or browse for apps developed by third-party developers to download and install onto their devices. He also said that distribution on the Play Store is valuable to developers in that it provides them with the ability to reach billions of Android device users around the world via a single platform, through which they can distribute apps.

1198 Mr Feng described the Play Store and Apple’s App Store as two well-known app stores for the distribution of digital content on mobile devices, and the Play Store as an app store that facilitates the distribution of apps, games, movies, TV shows and books to users of Android devices.

1199 Clearly, as Epic correctly points out, Google’s witnesses conceptualise the Play Store as supplying app distribution services.

1200 On the commercial evidence, it seems clear to me that the focal product by reference to which the competition analysis should proceed is Android app distribution services.

1201 But let me turn to the views of an economist and address Professor Asker’s approach to the focal product.

1202 Now Professor Asker accepted that the Play Store provides a way for app developers to make their apps available to users, and a way for users to download, install and update such apps. That is, he accepted that the Play Store provides app distribution services.

1203 But Professor Asker then took an additional step, being to ask whether that service is a source of competitive pressure on the Play Store. He concluded that it was not.

1204 Rather, he found that the source of such pressure was the facilitation of sales of digital content. This led Professor Asker to identify the facilitation of sales of digital content as the focal product.

1205 Professor Asker defined “facilitating sales” to mean connecting a developer and a consumer and ensuring the transfer of funds and content in the transaction.

1206 So, Professor Asker’s focal product excluded all services that did not involve the transfer of funds to purchase digital content, including the distribution of free apps without in-app purchases as well as the distribution of apps which offer in-app purchases of physical goods and services.

1207 Professor Asker said that an inquiry into the source of competitive pressure on the Play Store was necessary because identifying the focal product for the market definition exercise requires an assessment of a firm’s incentives to compete. Because firms compete to earn profit, that directs attention to how the firm monetises.

1208 His conclusions about the focal product were therefore driven by his conclusion that the Play Store primarily monetises by facilitating sales of digital in-app content, subscription or otherwise, and not from downloads.

1209 But I reject Professor Asker’s analysis for the reasons advanced by Epic.

1210 First, his analysis excludes many services that the Play Store provides and which are affected by the impugned conduct.  Second, it is inconsistent with the contractual position.  Third, it does not account for all of Google’s incentives to compete. And fourth, his analysis identifies a focal product based on how Google chooses to monetise from time to time, such that the focal product would change if Google altered its monetisation strategy, despite no change in the underlying services being supplied.

1211 Professor Asker’s approach also combines the separate services of distribution and the facilitation of in-app purchases into a single product.

1212 Moreover, Professor Asker’s focal product does not capture all of the services which Google provides.

1213 I have rejected Professor Asker’s focal product. It is divorced from the impugned conduct, and does not grapple with the contractual or commercial evidence as to the services in reality supplied by Google to developers and users who interact with the Play Store.

1214 Now given that the exercise of market definition is purposive, the boundaries of the market should be drawn with a view to analysing the specific conduct in question, in its commercial context, to determine whether that conduct has the substantial anti-competitive effect(s) alleged.

1215 Google’s impugned conduct affects all apps distributed by all developers to users via the Play Store, irrespective of whether developers sell digital or physical content within their apps, whether developers pay the service fee, and whether users pay for their apps or make in-app purchases.

1216 Clause 4.5 of the DDA, for example, does not apply only to developers who sell digital content. It applies equally to developers who distribute their apps for free.

1217 In contrast, Professor Asker’s market definition excluded some 96% of apps distributed through the Play Store.

1218 In my view a market definition that excludes key elements of the services supplied by Google under the DDA, being a key contract whose provisions are impugned, frustrates an analysis of the competitive effects of that conduct.

1219 Instead, the product dimension of the market must capture the services which Google provides to app developers and users under the DDA, to which its impugned conduct extends.

1220 Now whilst Professor Asker seemed to accept that market definition is a purposive concept, and that the goal of market definition in the present case is to understand the competitive effects of the challenged conduct, his analysis did not commence with the impugned conduct but, rather, with how Google monetises the Play Store.

1221 His focus on monetisation seemed to arise out of the fact that when applying the HMT, one must consider flows of profit. That is apparent from the fact that his report commenced with a consideration of the HMT and then later turned to the question of the services provided by Google.

1222 But by focusing on monetisation at the initial step, Professor Asker moved away from the orthodox approach of identifying the focal product prior to considering substitution alternatives.

1223 Instead, Professor Asker’s approach used the HMT logic to define the focal product itself.

1224 A firm’s incentives to compete are relevant to the question of substitution, that is, the application of the HMT, not to the identification of the focal product by reference to which the substitution analysis is to occur. It is an approach that is inconsistent with his own concession that implementing the HMT presupposes that an economist has a candidate product in mind.

1225 Professor Asker was not able to justify that approach during the concurrent evidence. The following exchange with me identified the difficulty:

HIS HONOUR:     I’m just trying to work out – there are two ways to look at it. You look at the conduct and then work out what market definition best conceptually explains the conduct and its competitive nature, and if you start from the other end of trying to work out the competitive nature and retrofit that back to the conduct and try and fit the conduct back into that.

PROFESSOR ASKER:     So I think - - -

Q:     Are you saying the latter or the former?

A:    I’m saying somewhere between the two, in the sense that if I might suggest, if we were doing a merger between Ford and Toyota, we wouldn’t spend an enormous amount of time on exactly how, potentially, car dealers work. But if we were doing a case relating to the vertical arrangements between Ford and car dealers, that would be entirely sensible.

1226 The first methodology that I identified is the plain vanilla Australian approach.

1227 The second, retrofitted methodology is the approach which Professor Asker took. His inability to accept that, and his reliance upon an irrelevant example about car manufacturers, was indicative of his overall difficulty in justifying the approach that he took.

1228 The result of Professor Asker’s retrofitting approach is the identification of a focal product that does not extend to all of the subject conduct, which includes services provided by Google to developers who distribute apps via the Play Store, and users who download apps from the Play Store, where there is no sale of digital content at all.

1229 The fact that the impugned conduct extends to apps being provided for free was put to Professor Asker during the concurrent evidence session.  But as Epic rightly says, Professor Asker did not offer any coherent explanation as to how that aspect of Google’s conduct fits within his analysis.

1230 Now I also agree with Epic that Professor Asker’s approach is inconsistent with the contractual position.

1231 The exercise of defining a market miscarries when that exercise is used to construct, or deconstruct and reconstruct, the supply of a service in a manner divorced from the commercial context of the putative contravention which precipitates the analysis.  The analysis of market definition and market power must reflect the reality of the contractual arrangements under which services are provided.

1232 The “service” provided by Google under the DDA is expressly said under clause 3.1 to be “mak[ing] [a developer’s] Products available in Google Play”. As clause 3.2 states, that service is provided in respect of “Products that users can access for free and Products that users pay a fee to access”.

1233 It is inconsistent with the contractual regime to separate the services provided by Google under the DDA into separate and distinct segments, insofar as those same services facilitate the distribution of free apps, or in-app sales of physical goods and services, or sales of digital content versus monetised apps.  The DDA contemplates the same services being provided in respect of all “Products”.

1234 Further, Professor Asker’s focus on the facilitation of sales of digital content is in tension with the fact that under the contractual arrangements none of the Google entities responsible for distributing apps through the Play Store supplies the services which facilitate the payments made by users to developers, or is a party to the contract between the developer and user by which a developer’s digital content is sold.

1235 Professor Asker accepted that these are transactions between a developer and a user.

1236 Defining the focal product by reference to a transaction to which no Google entity is a party, and which is not part of the service supplied under the DDA, involves a further departure from the impugned contractual arrangements.

1237 Further, to describe the Play Store as supplying a single service of facilitating (or intermediating) sales of digital content between consumers and developers is also inconsistent with Google’s own descriptions of the service it provides via the Play Store in its internal documents, which focus upon app distribution.

1238 I agree with Epic that those documents are a better guide to the commercial reality of the service which Google provides than any theoretical analysis by Professor Asker.

1239 Let me now turn to the question of monetisation and Professor Asker’s obsessive focus on how Google monetises the Play Store.

1240 Now I agree with Epic that it is because Professor Asker’s analysis started with the HMT rather than with the product or service being supplied that Professor Asker came to identify the facilitation of sales of digital content as the focal product.

1241 The essential problem with that emerged in the concurrent evidence session, where it became clear that Professor Asker’s approach meant that the product dimension of the market would alter if Google changed its monetisation strategy, even though the underlying service supplied by the Play Store did not change and the impugned conduct did not change.

1242 Consider a scenario where Google reduced its commission rate from 30% to 20%, but increased the price it charged for search ads by 50%.  That scenario is consistent with changes Google considered making to its monetisation targets in 2019.

1243 In this scenario, applying Professor Asker’s “monetisation” criterion, the focal product with which Professor Asker’s analysis started would change.

1244 The focal product would be expanded to include the services in respect of search ads which Google provides to any app, whether including in-app purchases or not.

1245 And the focal product would continue not to capture any services provided by Google in respect of free apps if those developers do not purchase search ads.

1246 Consider now a scenario where Google reduced the commission charged to 10%, but required all developers to pay a flat fee of a specified amount when their app is made available for distribution. That hypothetical is also consistent with changes Google previously considered.

1247 In that scenario, the focal product with which Professor Asker’s analysis started would expand to include services which Google provides in respect of all apps, because the flat fee applies to all developers, and so Google monetises the services provided to all app developers.

1248 Professor Asker conceded that it is an empirical reality that Google could change its monetisation targets. Indeed, that possibility was central to his effects analysis. Yet that possibility is unaccounted for in his identification of the focal product.

1249 On Professor Asker’s logic, if Google were to adopt any of those alternative monetisation strategies, the focal product would change. Yet in circumstances where the services that Google supplies are unchanged, and the impugned conduct is unchanged, in each of the hypotheticals set out above the starting point for market definition analysis should not change.

1250 Professor Asker’s initial tentative acceptance of that proposition, followed by a later rejection of it demonstrated his fixed and mistaken approach.

1251 I also agree with Epic that another feature of Professor Asker’s analysis is that, following his logic, Professor Asker is led to identify multiple different markets comprising Google’s distribution services insofar as they facilitate sales of digital gaming content, Google’s distribution services insofar as they facilitate sales of physical goods and services, Google’s distribution services insofar as they facilitate other free apps, Google’s services insofar as they generate revenues through search ads, and additional markets for Google’s services insofar as they facilitate sales of different types of non-gaming digital content.

1252 This is so despite the fact that precisely the same conduct is the subject of analysis across each of Professor Asker’s markets, and Google provides the same services across each market. The artificiality of such a conclusion is immediately apparent.

1253 Further, Professor Asker’s focus on Google’s competitive incentives in identifying the focal product was misguided.

1254 But even if such an approach were justified, Professor Asker’s version was incomplete. His analysis was premised on an unduly narrow characterisation of the incentives driving Google’s behaviour, for two reasons.

1255 First, Professor Asker’s focus on the sales of digital content ignored the material value non-paying developers and users present to Google by reason of network effects.

1256 Non-paying developers provide app diversity and innovation, in turn attracting more users. More users mean more views for advertising, and also more opportunities for in-app purchases to be made, both of which earn revenue for Google.

1257 More users in turn mean more developers will distribute via the Play Store as those developers want access to users, many of whom will pay the service fee and thereby contribute to Google’s revenue.

1258 Further, every time network effects are increased in this way, this builds on Google’s market power because it heightens the barriers to entry for a new app store.

1259 So, Google gains significant benefits via network effects from attracting and retaining non-paying developers and users. That is demonstrated by Google’s investment in non-paying developers and users by distributing those developers’ apps to users without charge, and by offering features of the Play Store relating to search, review and security.

1260 Now the provision of those services involves costs. But Google would not incur those costs unless it considered attracting and retaining non-paying developers and users to be of commercial benefit.

1261 Further, Google itself recognises the importance of network effects in its internal documents.

1262 In an internal Google presentation titled “Amazon Competitor deep dive – Apr. 2017”, the speaking notes record that:

Play benefits from network effects.

Users come to Play because we have by far the most compelling catalogue of apps / games

Developers come to Play because that’s where the users are.

1263 An undated internal Google document titled “Strategic developments” states that:

2. App and game developers

The app and game ecosystem is a virtuous circle: better apps and games make Android more attractive to users, while more users increases the attractiveness to developers for investing in Android. To keep the virtuous circle, we need to continue supporting developers, with the goal of making Android the most profitable platform for their apps / games.

1264 Now Professor Asker accepted that the Play Store is a multi-sided platform that is characterised by network effects. He also recognised that attracting non-paying developers and users improves the competitiveness of the Play Store.

1265 He noted that the service fee supports the Play Store providing free services for many developers and apps that likely confer substantial value to consumers, and supports the competitiveness of the Play Store.

1266 And he noted that a platform that attracted no one would have no competition within the platform. Sellers on a platform need to have consumers to compete over.

1267 Moreover, Google competes with other Android app stores across the full range of services provided through the Play Store.

1268 Clearly, Google seeks to attract all developers and users, not just paying ones, and all such customers drive Google’s competitive incentives and constraints.

1269 But such incentives are ignored in Professor Asker’s identification of the relevant focal product for his analysis.

1270 Second, Professor Asker’s focus on the sales of digital content ignored the material proportion of revenue Google derives from the Play Store through routes other than in-app purchases of digital content.

1271 Some [REDACTED] of the revenue Google earns via the Play Store is through advertising on the Play Store itself. Advertising is therefore a significant means through which Google monetises the app distribution service it provides through the Play Store. Through such advertising, developers obtain greater discoverability and, therefore, greater prospect of their apps, whether free or paid, being downloaded.

1272 Considering the focal product through the paradigm of Google’s profit-incentives would require the focal product to extend to advertising. Professor Asker’s only justification for his exclusion of advertising was that Epic’s allegations in this case “focus on service fees and not on ad-revenues from search placement”.

1273 This demonstrates the difficulty with Professor Asker’s approach. Had he started with the conduct that is the subject of the proceeding, the drawing of such fine distinctions about which money-making opportunities he does and does not consider would be unnecessary.

1274 Further, I agree with Epic that Professor Asker was evasive when pressed on how he had grappled with the fact that Google’s impugned conduct extended beyond developers offering in-app purchases and therefore extended to services not encompassed within the focal product.

1275 Further, at no point did Professor Asker explain how he had considered developers who were subject to the DDA but who do not make sales of digital content within their apps and therefore do not fall into Professor Asker’s market definition.

1276 Further, Professor Asker sought to make reference to consumption-only apps when dealing with his focal product.

Conclusion

1277 For the above reasons, the facilitation of sales of digital content, be that gaming content or non-gaming content, cannot be adopted as the focal product for the purposes of defining a market to analyse the effects of the OEM restrictive conduct and the app developer restrictive conduct.

1278 Such a focal product is detached from commercial reality, fails to account for Google’s incentives to compete, and results in an unstable market definition.

1279 In my view the focal product is the supply of app distribution services to developers and Android device users.

1280 Let me now turn to the next topic concerning the HMT.

The experts’ approaches to HMT quantitative and qualitative analysis

1281 Detailed qualitative and quantitative evidence was led by the parties concerning the use and application of the HMT and related topics. Part of the quantitative evidence analysed data relevant to the Fortnite hotfix and removal event and diversion ratios.  Detailed analysis was also addressed to logit demand modelling, transition probability analysis and ultimately critical loss analysis including the Lerner index.

1282 Now a formal HMT requires estimating cross-price elasticities or a complete set of diversion ratios in response to a hypothetical price increase for each product in the relevant candidate market(s) as well as considering the economic margins for these products.

1283 But in the context before me, this required more data than was available.  Specifically, there was a lack of systematic information about how developers responded to changes in app store pricing and quality.

1284 But economists can still use the framework of the HMT to weigh candidate market(s) even when the necessary information to implement a formal HMT is absent.

1285 With information on alternatives available to, and used by, customers of the hypothetical monopolist, one can use the best available evidence to evaluate whether, in response to a price increase, customers would have an incentive to switch to rivals and one can determine whether evidence indicates that doing so is relatively easy.  And this evaluation of incentives and ease of switching can be used to draw inferences as to whether the hypothetical monopolist’s price increase would likely be defeated.

1286 Now the key evidence needed under the framework of the HMT is the degree of customer substitution.  And this can be measured empirically if one has the appropriate and available data.  Now in the context before me some data was available to assess customer substitution along certain dimensions, but such data was limited in other respects.  And to overcome these limitations Professor Asker performed other quantitative analyses.

1287 But before proceeding further, let me introduce some concepts that I will develop in much greater detail later.

Critical loss analysis

1288 In the absence of complete data on cross-price elasticities and margins required for a complete quantitative HMT, an empirical tool that economists can use to determine whether a candidate market is a well-defined market, but underpinned by the principles of HMT, is a critical loss analysis.

1289 A critical loss analysis first establishes a critical loss threshold that would make a price increase by the hypothetical monopolist unprofitable, and then determines whether actual losses from that price increase would be higher or lower than the critical loss threshold.

1290 In essence, critical loss answers the question: if the hypothetical monopolist raises the price of one product in the candidate market by X%, what percentage of the sales of that product would need to be lost, that is, not recaptured by the hypothetical monopolist’s other products, for the price increase to be unprofitable?

1291 In its standard formulation, critical loss is expressed as a threshold that is a function of the percent increase in price, X, and the firm’s margins, M.  The standard formulation of the critical loss threshold is expressed as a fraction where the numerator is X and the denominator is X plus M.  This means that the critical loss threshold increases with a higher price increase and decreases with a higher margin.  Actual loss answers the question: if the hypothetical monopolist raises the price of one product by X%, what percentage of sales of that product would be lost due to the price increase?

1292 Let me also couch this concept in profit maximising terms as Professor Asker has done.

1293 Critical loss analysis starts by asking the following question: would a hypothetical monopolist of the candidate market find it profit-maximising to increase the price of one of its products, say Product A, by at least X %? The profit-maximising price increase is at least X % if an increase of twice the size (2X %) is profitable. This occurs when the actual loss in sales, for a 2X % price increase, is less than what is called the critical loss threshold.

1294 To determine the actual loss in sales, one must balance two competing effects. On the one hand, by increasing the price, the firm will lose customers from Product A. On the other hand, the firm, as the hypothetical monopolist, can recapture some of those lost consumers through the other products it owns, that is, everything else in the candidate market except Product A. The amount of lost Product A sales can be calculated using consumer elasticity, where the elasticity can be imputed using the Lerner index. The amount of recapture can be calculated using an aggregate diversion ratio, that is, how much is diverted to all other products in the candidate market.

1295 Critical loss analysis is typically evaluated by estimating an aggregate diversion ratio.

1296 An aggregate diversion ratio answers the question: if the hypothetical monopolist raises the price of one product, what fraction of lost sales of that product does the hypothetical monopolist recapture by other products it owns?

1297 Calculating an aggregate diversion ratio for a hypothetical monopolist of Android app stores would answer the question: what fraction of sales that divert from the Play Store in response to an increase in the Play Store’s prices are recaptured by other Android app stores?

1298 Now the advantage of such an analysis is that it requires a parsimonious set of parameters to understand when a candidate market is a well-defined market being the aggregate diversion ratio from the product of interest to other products in the candidate market, the margin for the product of interest, and the hypothetical percent increase in the price of the product of interest.

1299 A critical loss analysis usually evaluates a 10% price increase by the hypothetical monopolist, as that corresponds to the hypothetical monopolist finding it profit-maximising to increase prices by at least 5%.

1300 Now a standard critical loss analysis that compares aggregate diversion ratios to the critical loss threshold is predicated on the Lerner index. The Lerner index relates a firm’s margins to the elasticity of customer demand it faces.  Obviously the more price-sensitive consumers are, the smaller margins a firm can sustain.

1301 Now to perform a critical loss analysis Professor Asker used the diversion ratios calculated based on the hotfix regression specifications, which is a method of conducting the HMT quantitatively.  But it only considers the hypothetical monopolist raising prices on one product, rather than all products in the market, as would be required under the HMT.

Diversion ratios

1302 Let me say something about diversion ratios, which were calculated and/or analysed by the experts based upon the Fortnite removal event as a result of the Fortnite hotfix.

1303 Professor Asker’s analysis involved three quantitative methods applied to Fortnite revenue data to estimate diversion ratios from the Play Store in response to a change in price or quality.

1304 First, he used a difference in differences analysis which purported to measure the extent to which Fortnite revenue was diverted from the Play Store to other channels following the removal of Fortnite from the Play Store.

1305 Second, he used a logit demand model which estimated diversion ratios based on data about consumer spending prior to the removal.

1306 Third, he used a transition probability analysis which estimated diversion ratios based on data about consumer spending prior to the removal.

1307 Fourth, Professor Asker also counted up the various estimates of diversion, in particular diversion from the Play Store to the Samsung Galaxy Store, and used the results in his critical loss analysis to support his conclusion that the Android mobile app distribution market proposed by Ms Edwards-Warren was too narrow.

1308 Now Google’s thesis is that the spending patterns in the Fortnite spend data before and after the hotfix illustrate that non-mobile platforms are substitutes for the Play Store when it comes to users executing transactions for app-based digital content.

1309 Professor Asker considered that this pattern of spending indicates that Android and the Play Store Fortnite users are able to substitute their spending to non-mobile transaction venues such as consoles and PCs, because the data shows that in practice this is what they in fact do.

1310 Google said that a significant percentage of the reduction in the Play Store spending was diverted to other platforms including non-mobile platforms.  It is said that this shows that non-mobile platforms are substitutes for the Play Store as digital content transaction venues.

Logit demand modelling

1311 Now the purpose of the logit demand model is to generate diversion ratios, and the purpose of diversion ratios is to inform substitution, which is the basis for defining relevant markets.

1312 The logit demand model translates observed consumer choices among a viable choice set into estimates of diversion and substitution.  The architecture of the logit demand model means that diversion is proportional to spending shares. If consumers choose one transaction venue more than another out of the available choice set, the logit model says that there will be more diversion to that transaction venue.

1313 Now Professor Asker used the logit demand model in the present context to derive a measurement of diversion to the iOS App Store which was not possible from available natural experiments like the hotfix.  Diversion estimates from Professor Asker’s logit model exercise were said by Google to corroborate diversion estimates from the analysis of the hotfix.

1314 The logistic model, typically referred to as the logit model, is a popular model in modern industrial organisation economics. It belongs to a group of economic models called discrete choice models which economists use to explain how agents choose one option from a set of alternatives, often referred to as the choice set. So, an economist can use the logit model to analyse purchase data from the cereal aisle of a grocery store, including the prices, number of boxes bought, and characteristics about each cereal available, to estimate how price sensitive consumers are and the substitution patterns between cereals.

1315 Choice sets must exhibit several properties for a discrete choice model to be appropriate for a given setting.  First, the alternatives must be mutually exclusive, choosing one option means that none of the other alternatives are being chosen. Second, the choice set must include all relevant alternatives to the decision maker. Last, there must be a finite number of alternatives.  In practice, economists include an alternative referred to as the “outside option” in the choice set. The outside option typically represents the option not to purchase any available alternative.  So, it can represent a consumer choosing to not to spend their money or purchase another option not in the choice set.

1316 With that general introduction to the concepts out of the way, let me return to Professor Asker’s quantitative framework and how he dealt with data limitations.

Professor Asker’s quantitative analysis

1317 Professor Asker accepted that in practice one cannot perform a full empirical or quantitative SSNIP test due to limitations on data availability.  Instead, one has to evaluate other types of evidence such as functional substitutability, evidence on customer purchase drivers and evidence on past switching or threats to switch to inform a hybrid qualitative and quantitative application of the HMT.

1318 Now whilst Professor Asker agreed with Ms Edwards-Warren that there were limitations on data availability that prevented performing a full quantitative SSNIP test, he disagreed with her that the available data was sufficiently lacking so as to require the use only of a wholly qualitative application of the HMT.  He said that detailed data did exist that allowed one to do the following.

1319 First, one could estimate switching rates, including net switching rates and gross churn rates, between iOS and Android smartphones so as to contextualise those rates in comparison to other products Australian consumers commonly interact with.

1320 Second, one could calculate diversion from Android smartphone models to iOS smartphones based on elasticity estimates from academic literature in combination with available data to understand the closest competitive constraints facing Android smartphones.

1321 Third, one could estimate diversion from the Play Store to determine which rivals were the Play Store’s closest competitors.

1322 Fourth, one could quantify the prevalence of multi-homing by developers and consumers across transaction venues that facilitated their ability to substitute across these venues.

1323 Fifth, one could conduct a variety of other quantitative analyses that informed whether and to what extent customers of a hypothetical monopolist would substitute in ways that would discipline the hypothetical monopolist.

1324 Now Professor Asker relied on three related quantitative analyses to support his findings of a market for the facilitation of digital sales of games apps.

1325 First, he relied on a critical loss analysis which he claimed disproved that there is a narrower market that only includes Android app stores.

1326 Second, he relied on an event study which analysed customer purchasing behaviour before and after the Fortnite hotfix on 13 August 2020, on which he concluded that console stores are closer competitors to the Play Store than rival Android app stores.

1327 Third, he relied on a logit demand model through which he claimed to estimate consumer diversion, which also showed that Play Store Fortnite users are substantially more likely to divert Play Store spending to console stores, PC stores and the Apple App Store than to rival Android app stores.

1328 Now Ms Edwards-Warren did not conduct these analyses but focused primarily on how comparable potential substitutes are from a functional perspective. She also focused on conceptual discussions of switching costs.

1329 Now although Professor Asker agreed with Ms Edwards-Warren that qualitative evidence including functional substitutability can in some circumstances be informative, he was of the view that placing too much weight on such evidence could lead to problematic conclusions of overly narrow markets.

1330 Professor Asker’s view was that in the present case, detailed quantitative data existed that was informative as to market definition.

1331 Let me say something about Professor Asker’s specific approach concerning multi-homing/switching and demand system modelling.  As he explained, economists typically use two quantitative approaches to empirically evaluate whether two products are substitutes. The first approach involves the estimation of multi-homing, switching or substitution, and the second approach involves demand system modelling.  Professor Asker used both approaches in the present context.

1332 Let me say something about estimates of multi-homing, switching, and substitution.

1333 Estimates of multi-homing and switching typically take the form of summary statistics.  These estimates do not require a triggering event, that is, a change in price or quality, to properly estimate.  One can contrast this with substitution where there is switching in response to a price or quality adjustment.

1334 Multi-homing and switching statistics are calculated from data on consumer spending on the products over time. So, with data on how consumers spend on Fortnite across transaction venues over time, it is possible to construct estimates of Fortnite user multi-homing and switching across these venues.

1335 Now estimates of substitution rely on variation and natural experiments. This is because estimates of substitution require a triggering event such as a change in price or availability. This triggering event is provided by the natural experiment.

1336 A natural experiment is when an exogenous event changes the environment in which individuals, families, firms, or cities operate.  This change occurs due to external factors, for example, institutional rule changes, policy changes and natural phenomena.  This change occurs in such a way that some people are exposed to the change being the treatment group, and some people are not exposed to the change being the control group.  These features of natural experiments permit one to identify causal relationships as opposed to correlations.

1337 Typically, when evaluating whether a candidate natural experiment would permit causal inference, economists evaluate whether the event is exogenous. By this, economists mean that the event itself is uncorrelated with any unobserved determinants of the outcome of interest.

1338 So, as an exogenous example, app stores may raise prices when they think customers would not meaningfully respond, for example, because a rival app store has already raised its price, and so analysing customer responses to such a price increase would underestimate the magnitude of substitution. A natural experiment would avoid this flaw and allow researchers to estimate the causal effect of the change on a given outcome of interest, here, consumer substitution.

1339 A standard natural experiment analysed by economists to understand substitutability concerns the entry or exit of a firm’s or a rival’s products.

1340 Under certain circumstances, the entry or exit of a firm’s or a rival’s products can safely be treated as exogenous.  So, for example, unexpected plant closures due to extreme weather conditions can be used as a natural experiment to analyse substitutability between the affected and unaffected products. Likewise, competition agencies have utilised the entry and exit of stores when defining relevant markets in the context of grocery retail level mergers.

1341 Now whilst such analyses examine the effect of removing or adding a product rather than a change in price, the diversion ratios estimated from these analyses are regarded as reasonable approximations to those estimated with price changes.

1342 Let me say something about demand system modelling.  Under certain modelling assumptions, observable market outcomes like shares of sales associated with each product in a candidate market can be informative as to the underlying consumer demand.

1343 A workhorse demand model used by economists is the multinomial logit, which is more simply known as the logit model.  As I have already indicated, a logit model captures consumer decision-making over discrete choice options. Each consumer is assumed to make a single choice over distinct products. The model structures consumer utility over these options as a function of the product’s price and other observable and unobservable characteristics, with consumers choosing the product that yields the most utility to them.

1344 By placing certain assumptions on the unobservable characteristics, the model derives expressions for elasticity ratios and diversion ratios that are ratios of market outcomes like shares of sales.

1345 Professor Asker used the logit model to contextualise consumer spending when determining relevant markets.

1346 Let me draw out some other themes of Professor Asker’s evidence.

Professor Asker’s qualitative analysis

1347 Now Professor Asker said that to supplement quantitative approaches in a context where available data is limited, economists also rely on qualitative evidence to determine the nature and degree of competition between rivals.  And economists use qualitative evidence to draw conclusions about relevant markets in several ways.

1348 First, qualitative evidence on technical or functional substitutes to the products offered by the firm engaging in the challenged conduct can inform reasonable substitutability, and so narrow the set of potential products to evaluate in a candidate market. Technical or functional substitutes refer to products that offer similar uses to customers. The products which are substitutes in use are sometimes known as the set of functional substitutes.

1349 Second, qualitative evidence like internal company strategy documents or public financial reports can be useful to determine which products and rivals a firm views as a competitive threat, and how they respond to this competition.  So, internal strategy documents can describe the competitive responses that the firm engaging in the impugned conduct took in response to competition from rivals.

1350 Third, competitive responses to rivals in terms of pricing and innovations are evidence of competition with those rivals.

1351 Fourth, one can also use survey evidence to draw inferences about market definition.

1352 Now usually the use of surveys for market definition is confined to retrospective buyer surveys that question a firm’s customers as to what products the buyers currently use or used in the past. Such retrospective surveys can be informative as to switching rates and multi-homing. When there have been past changes in price or quality, and those changes are plausibly exogenous, then retrospective surveys could also reveal changes in customer behaviour that is indicative of substitution.

1353 Contrastingly, surveys that focus on why customers made certain purchasing decisions, or what features of products customers found salient when they made a purchase, may be less informative as to how such customers would respond to price increases or reduced quality.

1354 Now Professor Dhar used such survey evidence to draw conclusions as to whether Android smartphone purchasers valued the quality of app stores and the availability and pricing of apps when purchasing their device. Ms Edwards-Warren relied upon Professor Dhar’s survey evidence to support her conclusion that the iOS App Store and Play Store are not in the same relevant market.

1355 Further, hypothetical questions in surveys are generally viewed as less reliable for understanding substitution when detailed evidence on actual purchasing behaviour exists.  Economists instead typically analyse revealed preference, that is, choices revealed in real-world data. Contrastingly, asking survey respondents how they would respond given a hypothetical price increase or quality reduction without real-world consequences or incentives may not elicit accurate responses.

1356 More generally, Professor Asker expressed the view that when appropriate data exists for quantitative empirical estimation of substitution, economists usually weigh such quantitative evidence more heavily than qualitative evidence.  Moreover, the importance of quantitative evidence is higher when addressing questions about the degree to which effects were substantial, which are more appropriately addressed through quantification. In such situations, qualitative evidence is usually used to supplement these data-driven approaches.

Ms Edwards-Warren’s approach

1357 Now Ms Edwards-Warren’s starting point for her market definition inquiry was the product, the customers who use the product and the conduct in question. This is because market definition exercises are purposive. The aim is to assess the effect of particular conduct in a market.

1358 The correct starting point for assessing market definition is the good or service supplied or acquired by the relevant firm and its close substitutes and to consider the functional dimension of the market.  I have already identified the focal product.

1359 Now Ms Edwards-Warren assessed six potential substitutes to Android mobile app stores that she identified being the pre-installation of individual apps on Android devices, the direct downloading of individual apps on Android devices, web apps including streamed web apps, app distribution services on non-Android devices, app distribution services on PCs and app distribution services on consoles.

1360 The first three are potential substitutes, for which users do not have to switch devices, whilst the next three are potential substitutes for which users would have to switch devices.  I have discussed elsewhere the six potential substitutes, and the extent and closeness of substitution (if any).

1361 With respect to each potential substitute, Ms Edwards-Warren assessed evidence on functional substitution, switching costs, both financial and non-financial, switching rates and quantification of the prevalence of multi-homing by developers and consumers. She considered multi-homing by app developers and consumers across the Play Store and the Apple App Store, developer multi-homing with PCs and consoles, and user access to PCs by smart mobile device users, and access to consoles.  I have discussed some of Ms Edwards-Warren’s evidence on switching in my reasons in Epic v Apple.

1362 On the basis of her evaluation, Ms Edwards-Warren concluded that the relevant market was for Android app stores, that is, Android mobile app distribution. She considered each substitute separately and concluded that developer and/or user switching to each substitute in response to a SSNIP would be very limited. On this basis she also considered that total switching to all six substitutes would be insufficient to enlarge the market.

1363 In summary, Ms Edwards-Warren evaluated qualitative and quantitative evidence to inform a qualitative application of the HMT.  Further, she also analysed Professor Asker’s analyses and gave her view as to where she saw the difficulties.

1364 Ms Edwards-Warren considered that there were a number of difficulties with Professor Asker’s quantitative analysis.

1365 First, she considered that the critical loss analysis suffered from methodological flaws and relied upon the flawed diversion ratios calculated by the hotfix analysis. But even if one accepted the methodology and inputs, the results were mixed and suggested that either there is an Android mobile app distribution market or that this market should be extended to include direct downloads but no wider. In her view the analysis did not support the market definitions put forward by Professor Asker under any scenario.

1366 Second, the Fortnite hotfix on 13 August 2020 was not a clean event that would have caused consumers to switch their purchases to another distribution channel. Any user that had downloaded Fortnite via the Play Store prior to 13 August 2020 could still both play and purchase V-Bucks in that version of Fortnite, at the same price as on other platforms. The only change was that they paid via EDP rather than via Google Play Billing. It was not until 27 August 2020 when a new season of Fortnite was available that users needed to switch to other platforms.

1367 Third, in her view the switching that was identified in Professor Asker’s analysis was generated by a flaw in the way the analysis was designed. She found that if one re-ran the same codes on other randomly selected dates, one would find substantively the same results.

1368 Fourth, in her view the results of the Fortnite hotfix analysis could not be generalised to all games apps. Professor Asker claimed that the fact that most of the top 25 games apps on the Play Store have cross-wallet and cross-progression functionalities meant that users have the same substitution options as Fortnite users. But of those top 25 games apps, only two are actually available on consoles.

1369 Fifth, as for Professor Asker’s logit demand model, it essentially estimated diversion ratios based on revenue market shares, in which the market had already been assumed to include the Apple App Store, PC stores and console stores. So, in her view the results of the logit model were driven by the starting assumptions.

1370 So, the fact that the highest switching is to consoles is by the design used by Professor Asker.  Consoles account for the highest share of revenues in this pre-assumed widely defined market and therefore switching to consoles is the highest.

1371 Now Professor Asker said that the hotfix event is what economists call a natural experiment and from the perspective of Fortnite users, the removal of Fortnite from these two app stores corresponds to the exit of a transaction venue.

1372 And it was on the basis of these three analyses that Professor Asker concluded that the relevant markets in which the Play Store compete is the markets for facilitating sales of digital gaming content.  Indeed, these three analyses were fundamental to his overall findings and for broadening the market for facilitating sales of digital gaming content to include PC, consoles and the Apple App Store.

1373 But in Ms Edwards-Warren’s opinion, these analyses were not determinative for market definition and/or the results were misinterpreted.

1374 First, and as I have said, she was of the view that the critical loss analysis of Professor Asker was flawed in relation to its design, implementation and interpretation.  But even if one accepted the methodology and inputs, the results were mixed and suggest that either there is an Android mobile app distribution market or that this market should be extended to include direct downloads, but no wider.  In her view, the analysis does not support the market definitions put forward by Professor Asker under any scenario.

1375 Second, in her view and as I have said, the Fortnite hotfix was not a clean natural experiment and should not be described as the exit of a transaction venue, as from the user perspective very little changed.  Smart mobile device Fortnite users were able to continue to play, and to make in-app purchases at the same price as via other transaction venues after 13 August 2020.  And in her view there were various other conceptual and analytical flaws with the analysis.  She gave evidence that if one re-runs Professor Asker’s analysis on three placebo dates prior to the hotfix, the analysis produces substantively the same results, thus demonstrating that it is not the hotfix event driving them.

1376 Third, in her view and as I have said, the logit model that was described by Professor Asker essentially assumed the answer to the question that it purported to answer.  In her view this analysis said nothing significant about market definition that has not been pre-assumed.

1377 Let me also identify some other differences between the experts.

Functional substitutability

1378 Professor Asker said that Ms Edwards-Warren’s approach to defining markets was incomplete given the primacy she gave to analysis of functional substitutability.  But this does not accurately reflect what she did. She identified a candidate market, and then assessed each of the potential substitutes that she identified from the Google witness statements.

1379 Her starting point for each assessment was functional substitutability. She then considered evidence of switching and other relevant evidence available to her, both quantitative and qualitative. She then reached her conclusions based on her assessment of all of this evidence.

1380 Now qualitative evaluation is usually the starting point of any market definition exercise.  And it is appropriate to start with the product characteristics and the intended use(s) of the product. Doing so allows one to define a broad and plausible set of possible demand substitutes. The products which are substitutes in use are the set of functional substitutes.

1381 So, Ms Edwards-Warren did not give primacy to analysis of functional substitutability, over and above other types of evidence. But she did consider it a necessary step of the market definition exercise to understand which other products fulfil the same functional purpose as the focal product, in line with the best practice described above.

1382 Now Professor Asker said that Ms Edwards-Warren ignored relevant empirical evidence on substitution. Professor Asker listed five categories of “rich data” in this case relevant for the application of the HMT and said that she did not conduct these analyses and that she ignored this available data.  But Ms Edwards-Warren disagreed with the accuracy of that assessment.

1383 As to switching rates, Ms Edwards-Warren analysed switching rates. But she did not analyse net switching rates and gross churn rates or benchmark them against other industries because she considered them to be uninformative for market definition.

1384 Further, as to diversion ratios from Android to iOS smartphones based on academic literature, she did not analyse this, but she explained why the calculations based on those datasets by Professor Asker were inaccurate, biased and did not support the conclusions drawn from them.

1385 Further, as to the diversion from the Play Store to other platforms using Fortnite hotfix data, she did not consider this informative for market definition.

1386 Further, as to the quantification of the prevalence of multi-homing by developers and consumers, she analysed multi-homing by app developers and consumers across the Play Store and the Apple App Store. She also analysed developer multi-homing with PCs and consoles, and user access to PCs by smart mobile device users and access to consoles.

1387 Now a variety of other quantitative analyses can inform whether and to what extent customers would substitute in ways that would discipline the hypothetical monopolist.

1388 In addition to quantification of switching rates and multi-homing, Ms Edwards-Warren relied on other quantitative evidence where there was data available.

1389 So, she analysed Google Android and iOS smart mobile device prices when evaluating substitution between licensable and non-licensable OSs, and the number and share of new app downloads by source when assessing the constraint direct downloading imposes on distribution via Android app stores.

The interpretation of rates of switching

1390 Another evidentiary issue on which Ms Edwards-Warren disagreed with Professor Asker was the interpretation of rates of switching that occur in the ordinary course of business, that is, customer churn.

1391 This is relevant in the assessment which they both conducted of switching by users between Android smartphones and iOS smartphones. The question of interest is the extent to which users would switch away from Android to iOS in response to a small increase in the price, or small decrease in quality, of the operating system, or in response to a small increase in the Play Store service fee, should it result in an increase in app prices.

1392 Since they could not directly observe this happening, they both analysed studies of user switching between phones with the Google Android OS and iPhones.

1393 She agreed that switching rates are distinct from, but informative about, substitution, because switching rates do not measure changes in behaviour that can be causally attributed as responses to changes in price or quality.

1394 But Ms Edwards-Warren disagreed with two aspects of the reasoning of Professor Asker.

1395 The first aspect was that she did not agree that the incremental switching in response to a price increase would be higher than the base level of switching. When conducting a HMT, one is interested only in the incremental switching driven by the price increase, not the base level of switching and the increment. The base switching rate is informative about whether consumers find switching easy in which case, all else being equal, customer churn will be high, or not. That is, the base switching rate is informative about the ability of consumers to switch. But if say 10% of customers switch from one product to another on an annual basis, this does not mean that the same proportion or a greater proportion of customers would switch in response to a small price increase.

1396 The base level switching is driven by many factors including changes in the product price or quality, consumer income, and dissatisfaction with service. So, Ms Edwards-Warren disagreed with Professor Asker’s conclusions.

1397 The second aspect was that she disagreed with the interpretation of switching levels in other industries as set out by Professor Asker. There is no benchmark level of churn above which products are in the same market and below which they are not.

Usage and spending data

1398 Let me at this point say something about usage and spending data.

1399 Now Professor Asker said that the relevant question for evaluating market definition is to understand how a hypothetical firm’s profits would change in response to a price increase or quality reduction. In most cases firms charge a price for the goods or services they sell, and so the relevant metric for understanding customer substitution is customer spending.

1400 Professor Asker said that economists consider usage in addition to or in lieu of spending when evaluating markets if there is lack of data on spending and usage is a proxy for spending, and when a firm does not monetise by charging a price.

1401 Ms Edwards-Warren agreed that when implementing a HMT quantitatively one is primarily interested in where customers switch their expenditures in response to a price increase. But a hypothetical monopolist will take into account where customers switch their volumes if it affects where the same customers or other customers switch their expenditures.

1402 If spend data is not available, then volumes or usage can be a proxy to assessing substitutability. This is the case if customers switch their volumes or usage, for example, play time of a game, to the same alternatives as they switch their expenditures in response to a price increase.

1403 Where revenue or spend data was available and robust, Ms Edwards-Warren used it.

1404 So, she used it when assessing spend on the Play Store by app and purchase types, supply shares of operating systems, the cross-availability of apps across the Play Store and the Apple App Store, and the impact of the Fortnite removal event on payment solutions.

1405 To the extent that she relied upon volume or usage data it was because she judged that the analysis could not be carried out robustly with value data, for example through lack of availability, or because it was otherwise relevant.

1406 The two analyses that Professor Asker criticised her for having relied upon usage data are the following.

1407 Professor Asker criticised her focus on how consumers and developers use the Play Store when discussing functional substitutability. Her analysis of whether a potential substitute performs similar functions to the focal product was informative about whether a developer or user is likely to switch both their usage and their expenditure to that potential substitute.

1408 Further, Professor Asker criticised her reliance on the analysis of Professor Goldfarb. Professor Goldfarb conducted analysis using usage data, which is relevant for understanding consumer behaviour and substitution patterns. Ms Edwards-Warren recognised that implementing a HMT quantitatively, which is something Professor Goldfarb was not instructed to do, requires assessing spend.

1409 With respect to market definition involving the Play Store, it is important to note that approximately 96% of developers do not pay a service fee for app distribution services.

1410 The choices of the non-fee paying developers will have an impact on the choices of the fee-paying developers, because they generate positive externalities from being on the same platform. I have referred to these elsewhere when discussing network effects.

1411 So, Professor Asker was incorrect when he said that spending is a sufficient statistic to evaluate competitive constraints facing the Play Store. He was also incorrect when he said that when data on spending is available, there is no reason to evaluate usage.

1412 Let me now turn to a more general matter concerning some aspects of principle.

Professor Asker’s five HMT principles

1413 Now Professor Asker said that Ms Edwards-Warren made several implementation errors when deploying the HMT that resulted in her defining improperly narrow markets. Professor Asker defined five implementation principles for the HMT which he said that he subscribed to, but Ms Edwards-Warren had not conformed to.

1414 Let me identify these principles and Professor Asker’s criticisms.  I might say that Ms Edwards-Warren agreed with most of these principles putting to one side for the moment an issue that arose between the experts concerning the difference between taking a profitable step as distinct from a profit maximising step.

1415 There is a difference between profit-maximising and profit-increasing, and I have addressed this in more detail elsewhere.

1416 Ms Edwards-Warren considered that a HMT involves assessing whether a small, typically 5% or 10%, price increase would be profitable, with care needed to ensure that one does not define over-wide markets due to the cellophane fallacy. This is how a quantitative SSNIP test is conducted, but that was not possible in this case.

1417 Absent a quantitative SSNIP analysis, one assesses the evidence on each substitute and considers whether there appears to be sufficient switching to them either individually or collectively.

1418 The first HMT principle was to the effect that the hypothetical monopolist of the candidate market is to be taken as not controlling any products outside that market. Now a failure to adhere to this principle may mean that substitution to products outside the candidate market that the hypothetical monopolist is “incorrectly” assumed to own may not discipline the hypothetical monopolist. This may result in an improperly defined market.

1419 Now Professor Asker said that for the purposes of analysing a price increase by a hypothetical monopolist, the terms of sale of products outside the candidate market are held constant.  But Professor Asker said that Ms Edwards-Warren made an error by assuming that a hypothetical monopolist over an Android app distribution market controls the payment solution which Ms Edwards-Warren viewed as a distinct product.

1420 But this is not a correct interpretation of what Ms Edwards-Warren did.

1421 Professor Asker claimed that by assuming a Play Store service fee of 30% Ms Edwards-Warren was including Google Play Billing. But Ms Edwards-Warren was hypothesising an increase in the app store service fee, and that the app store service fee is at most 30% of the value of in-app purchases.  This is because she did not know what the prices of the Play Store and Google Play Billing would be if they were unbundled, but she could assume that neither would exceed 30%. Her analysis was explicitly taking an upper bound for the app store service fee for illustrative purposes.

1422 But in any event, Professor Asker’s criticism is inconsequential.  If Ms Edwards-Warren had assumed that the app store service fee was 26% or 27%, as advanced by Professor Asker, her conclusion that a small increase in Android app store service fees would have a small impact on the average user expenditure relative to the cost of switching smart mobile devices would be unchanged.  This is because a 5% or 10% increase in Android app store service fees would be even smaller when applied to a lower service fee figure, for example, 26% or 27%, than to a 30% fee.

1423 The second HMT principle was to the effect that a proposed market fails the test if the hypothetical monopolist would not find it profit-maximising to raise prices (or worsen quality) by at least 5% given sufficient substitution to all possible rivals.  So, this principle states that even if no potential substitute sufficiently constrains the prices of the hypothetical monopolist on its own, one should consider whether collectively all possible substitutes in aggregate provide a constraint, and then consider which of those should be included in the market.

1424 This second principle enshrines the idea that iteratively applying the HMT to each individual rival or different groups of rivals and determining that they would not discipline the hypothetical monopolist on their own is not informative as to whether all the rivals together could discipline the hypothetical monopolist.

1425 Now a failure to adhere to this principle means that the HMT sought to be applied does not consider total substitution away from the hypothetical monopolist and so leads to inappropriately defined markets.

1426 Professor Asker said that Ms Edwards-Warren made this implementation error when defining all three of her proposed markets.  But Ms Edwards-Warren concluded that developer and/or user switching to each of six potential substitutes in response to a SSNIP would be very limited.  On that basis she also considered that the total switching to all six of them would also be insufficient to enlarge the market.

1427 The third HMT principle was to the effect that sufficient substitution from the hypothetical monopolist that would lead it to not find it profit-maximising to raise prices or worsen quality can arise from partial substitution by a subset of customers.  Now a failure to adhere to this principle means that the HMT ignores substitution that could discipline the hypothetical monopolist, that is, every dollar that could substitute away. This leads to inappropriately defined markets.

1428 Professor Asker said that Ms Edwards-Warren made this implementation error when defining all three of her proposed markets.  He said that she focused on complete substitution that requires customers to switch away all their purchases from the hypothetical monopolist but did not consider partial substitution by customers.  He said that Ms Edwards-Warren made this error despite accepting that when assessing whether it would be profitable for a hypothetical monopolist to raise prices by a small amount, or reduce quality or worsen contractual terms more generally, it is not necessary that all customers switch away or that customers switch away all of their purchases for the small price increase to be unprofitable.

1429 Now Ms Edwards-Warren agreed that partial substitution by a subset of customers can in principle constrain a hypothetical monopolist if the substitution is sufficient and if the hypothetical monopolist cannot price discriminate towards that subset of customers.

1430 But she did not take account of partial substitution in her assessment of market definition involving GMS Android because, as she saw it, OEMs could not substitute some of their devices away from GMS Android to Android forks.  But she did take into account partial substitution in her assessment of market definition involving the Play Store.

1431 The fourth HMT principle was to the effect that the relevant time frame for evaluating substitution from the hypothetical monopolist will vary based on the lifecycle of the products considered and can be several years in duration for durable goods. But for products with short lifecycles and high user churn, examining substitution over short time periods is appropriate. For products with longer lifecycles where users replace products every few years, examining substitution over longer time periods is appropriate.

1432 A failure to adhere to this principle means that the HMT would evaluate substitution over inappropriate time windows and lead to inappropriately defined markets.

1433 Professor Asker said that Ms Edwards-Warren made this implementation error when defining the proposed mobile OS licensing market. She considered narrow time frames when evaluating whether consumers would substitute to iOS devices and the iOS App Store.  He said that despite conceding that the average lifecycle for mobile devices is typically more than three years, she did not evaluate substitution over a period of at least that length. This understated substitution.

1434 Now Ms Edwards-Warren agreed that the relevant time frame depends on the type of product and on how quickly consumers can switch.  And in applying the HMT, she accepted that the time frame considered affects two elements.

1435 First, it affects the expected cost of switching to customers. She agreed that customers would likely take into account the expected lifetime cost of a product when switching. They would most likely value spending today more highly than spending in two years’ time, therefore applying some discount rate.

1436 Second, it affects the time frame over which customers are likely to switch away.

1437 Now as she said, it is not the case that one has to assume a time frame over which every single customer could switch away.  This is not appropriate, because a hypothetical monopolist, when considering whether to increase prices, would take into account whether some customers may or may not switch for a long period of time. In any event, she considered this criticism to be inconsequential, as using a longer or shorter time frame did not affect her overall conclusion.

1438 Now more specifically Professor Asker criticised Ms Edwards-Warren for overstating switching costs faced by consumers when switching from Android to iOS smartphones and misinterpreting evidence on switching because she had not assessed switching over an appropriate time frame of three to four years.  But Ms Edwards-Warren disagreed with this and said that a hypothetical monopolist when deciding whether to increase prices would take into account the fact that some customers will switch immediately and some would not switch immediately.

1439 Finally, the fifth HMT principle was to the effect that it is more appropriate to evaluate candidate markets that contain closer substitutes to the firm’s product than weaker ones.  Professor Asker said that a well-known challenge with the framework of the HMT is that the relevant markets defined using the framework can be sensitive to the order in which rivals’ products are evaluated.

1440 He said that consistently with the goal of market definition being to better evaluate the effects of the challenged conduct by narrowing the focus to the most relevant competitors and substitute products, one should ensure that the products of the closest competitors to the challenged firm are included in the candidate markets to evaluate using the HMT.

1441 Now Professor Asker said that Ms Edwards-Warren made this implementation error when defining the proposed mobile OS licensing market and Android mobile app distribution market.  Professor Asker said that this was evident in the fact that she included the licensable OS Sailfish in her proposed OS market whilst she excluded iOS, and in the fact that she included Android app stores like Aptoide in her proposed Android mobile app distribution market whilst she excluded webstores, console stores, PC stores, and the iOS App Store.

1442 Now Ms Edwards-Warren agreed with this fifth HMT principle, but she disagreed with the conclusions of Professor Asker about which were the closest substitutes to the Play Store and the Android OS.

1443 Now in my view Professor Asker’s criticisms of Ms Edwards-Warren are misplaced.  There are of course differences between them which I will come to in discussing the detail of some of the quantitative analyses. But Professor Asker’s so-called implementation or methodological errors as such which he has attributed to Ms Edwards-Warren are not made out.

1444 Let me now turn to the detail and begin with the topic of the Fortnite removal event and diversion ratios.

Fortnite removal event and diversion ratios

1445 Let me begin by saying something about diversion ratios, which were calculated and/or analysed by the experts based upon the Fortnite removal event as a result of Mr Sweeney’s hotfix strategy.

1446 Professor Asker used three quantitative methods applied to Fortnite revenue data to estimate diversion ratios from the Play Store in response to a change in price or quality.

1447 First, he used a difference in differences analysis which purported to measure the extent to which Fortnite revenue was diverted from the Play Store to other channels following the removal of Fortnite from the Play Store.  I will discuss the legitimacy or otherwise of this approach in this part of my reasons.

1448 Second, he used a logit demand model which estimated diversion ratios based on data about consumer spending prior to the Fortnite removal event.  I will discuss such a model in a separate section of my reasons.

1449 Third, he used a transition probability analysis which estimated diversion ratios based on data about consumer spending prior to the Fortnite removal event. Again I will discuss such an analysis in a separate section of my reasons.

1450 Fourth, Professor Asker also counted up the various estimates of diversion, in particular diversion from the Play Store to the Samsung Galaxy Store, and used the results in his critical loss analysis to support his conclusion that the definition of the Android mobile app distribution market was too narrow.  I will also discuss this critical loss analysis in a separate section of my reasons.  As I have already indicated, a critical loss analysis is underpinned by the principles of HMT and is the essence or central theme of Professor Asker’s analysis, with the diversion ratios used as a necessary input.

1451 Now Google’s thesis is that the spending patterns in the Fortnite spend data before and after the hotfix illustrate that non-mobile platforms are substitutes for the Play Store when it comes to users executing transactions for app-based digital content. This is said to be the case by Google for various reasons that were said to be justified by Professor Asker’s quantitative analysis.

1452 Professor Asker considered that the relevant pattern of spending indicated that Android and the Play Store Fortnite users were able to substitute their spending to non-mobile transaction venues such as consoles and PCs.  It was said that the data shows that in practice this is what they in fact do.

1453 Google said that a significant percentage of the reduction in the Play Store spending after the removal event was diverted to other platforms including non-mobile platforms.  It is said that this shows that non-mobile platforms are substitutes for the Play Store as digital content transaction venues.

1454 Now as Epic correctly identified, there are two interrelated questions that arise in relation to Professor Asker’s quantitative analyses.  First, is quantitative analysis in relation to a single app, Fortnite, informative for market definition more generally?  Second, assuming quantitative analysis concerning Fortnite is informative, what does that analysis show?

1455 Now Epic says that Professor Asker’s difference in differences regression analysis of diversion of Fortnite spending following the hotfix suffers from various flaws.  As a consequence it over-estimates total diversion.  It says that Professor Asker inappropriately used a levels difference in differences analysis rather than the percent approach deployed by Professor Goldfarb.

1456 Further, as to the relevant time frame(s) used by Professor Asker, Epic says that retaining the week of the hotfix (10 August 2020) and the two subsequent weeks (17 and 24 August 2020) in Professor Asker’s analysis was inappropriate.  Further, it says that adopting a 5-week post-period rather than the 13-, 26- and 52- week post-periods was not the preferable approach.

1457 Before proceeding further, I should say something more about the Fortnite hotfix and removal event.

1458 As I have discussed elsewhere, on 13 August 2020 Epic activated a hotfix on the Fortnite app available through both the Play Store and the App Store which presented users with a choice for purchasing V-Bucks either through Google’s or Apple's own payment solution or through EDP.

1459 At the same time, Epic introduced the “Fortnite Mega Drop”, being a permanent 20% discount for Fortnite in-app purchases on mobile, PC and console platforms. But users were not eligible for such discounts if they chose to use Google's or Apple's payment solution.

1460 On the same day, Fortnite was removed from the Play Store and the App Store,  and since 13 August 2020, Fortnite has not been available for download from, or updating via, the Play Store or the App Store.

1461 At or about the same time, users of Fortnite downloaded from the Play Store and the App Store were informed that they would not be able to play a new season of Fortnite when it was released on 27 August 2020.

1462 Now smart mobile device users who had already downloaded Fortnite through those app stores were able to continue playing the version of Fortnite they had downloaded, and to make in-app purchases at the same discounted price via EDP on their smart mobile device as on other platforms.  The main change was that smart mobile device Fortnite users could not re-download or update Fortnite through those app stores. Moreover, as I have indicated, users would not be able to update the Fortnite app downloaded from the Play Store and the App Store when a new season was released on 27 August 2020.

1463 Now the Fortnite removal event could be described as, or giving rise to, a natural experiment.  Professor Asker said that from the perspective of Fortnite users, the removal of Fortnite from the relevant two app stores corresponded to the exit of a transaction venue(s), although Ms Edwards-Warren took issue with this characterisation.

1464 Now Professor Asker focused on Fortnite users who spent on the Play Store during the four weeks before the hotfix. Professor Asker excluded Fortnite users who had spent on the App Store during the same time period from both the treatment and control groups.

1465 Professor Asker analysed where those users made in-app purchases during the five weeks after the hotfix.  He defined this five-week period after the hotfix as the period between 10 August 2020 and 13 September 2020. But this “after hotfix” period included three days prior to the hotfix, which occurred on 13 August 2020.

1466 Professor Asker considered that the diversion ratios estimated from switching that occurred in the five weeks after 13 August 2020 would be the same as those estimated from a small change in price or quality.

1467 Now to put the hotfix diversion analysis in context, I should note the following matters.

1468 First, Professor Asker and Ms Edwards-Warren agreed that analysis of substitution and diversion, including natural experiments and critical loss analysis, can be informative about market definition.

1469 Second, Professor Asker and Professor Goldfarb agreed that the removal of Fortnite from the Play Store is a natural experiment, they agreed that estimates of diversion in response to that event are informative and can be used to draw general conclusions as to substitution by consumers across devices and transaction venues, and they agreed that a difference in differences regression estimation strategy is a well accepted method, but that in some instances the suitability of a levels or percent approach may differ although in other instances it may not matter which approach is utilised.

1470 Third, Professor Asker and Professor Goldfarb agreed that diversion is central to market definition because it is closely related to substitution.  Substitution is defined as the percentage change in quantity purchased from a firm in response to the firm raising prices or worsening quality by a small amount.  Diversion measures the fraction of the sales a firm loses due to a small increase in price or decrease in quality that are recaptured by another firm’s product.  Of course the two concepts are related such that higher diversion to another firm’s product means that that product is a closer substitute.

1471 Fourth, Professor Asker and Professor Goldfarb agreed that to estimate diversion empirically usually requires a triggering event, being a change in price or quality.

1472 Now I should note here that both Professor Asker and Professor Goldfarb used similar study designs to measure diversion. They used a specification where the treatment group consisted of users that played or spent on Fortnite via the Play Store (the treatment group), and the control group included users that did not do so and played or spent via alternate transaction venues (the control group).

1473 Now in order to appreciate the materiality of the methodological debates between the experts and to identify the diversion estimates that are significant for market definition, it is necessary to note the way in which the diversion estimates are sought to be used.

1474 First, the diversion from the Play Store to other Android app stores, most notably the Galaxy Store, is an input into Professor Asker’s critical loss analysis, which as I have already indicated is a form of quantitative HMT that Professor Asker has used to evaluate Epic’s candidate Android mobile app distribution market.  As I say, I will discuss this critical loss analysis later.

1475 Second, the diversion from the Play Store to console stores relative to diversion to sideloading or other Android app stores is said to be informative as to whether console stores or other Android transaction venues impose closer competitive constraints on the Play Store, which indicates the next closest substitutes to be included in the market if the candidate market fails the HMT.

1476 Third, the diversion from the Play Store to all other Android transaction venues relative to all non-Android transaction venues is also said to be informative as to whether Android or non-Android transaction venues impose the greater competitive constraint on the Play Store, and indicates whether those venues provide meaningful competitive constraints for the Play Store.

1477 But I should note, more generally, that what matters is not the total amount of diversion from the Play Store, but the relative diversion to available choices.

1478 Now I should also note that Professor Asker and Professor Goldfarb presented numerous different sensitivities for estimating diversion from the hotfix.  I will return to these diversion estimates later.

1479 Let me now turn to some of the evidence of Ms Edwards-Warren and Professor Goldfarb.  In doing so and conveniently, I will pick up the salient aspects of Professor Asker’s evidence.

Ms Edwards-Warren’s evidence

1480 Ms Edwards-Warren disagreed with Professor Asker’s perspective on the Fortnite hotfix event. In her view, the hotfix should not be described as the exit of a transaction venue, as from the user perspective very little changed. Smart mobile device Fortnite users were able to continue to play, and to make in-app purchases at the same price as via other transaction venues.

1481 Ms Edwards-Warren considered that because from the perspective of smart mobile device Fortnite users little changed on 13 August 2020, the hotfix did not give these users an incentive to switch to other “transaction venues”.

1482 So, Ms Edwards-Warren considered Professor Asker to be erroneous in attempting to estimate substitution of user spend following the hotfix event and subsequent removal of Fortnite from the Play Store on 13 August 2020, in order to estimate how Fortnite users would substitute in response to a small change in relative prices.

1483 Now Professor Asker concluded that the Fortnite hotfix analysis demonstrated that there was more switching by Play Store users to consoles than to other Android app stores.  But Ms Edwards-Warren said that this result was due to an error in the analysis, being a poorly defined control group.

1484 To demonstrate this, she re-ran the same analysis around three different “placebo” dates, being 4 June, 18 June, and 9 July 2020, and she found substantively the same results, being that diversion to consoles was significantly higher than switching to Android app stores.  This demonstrates that the result was not driven by the Fortnite hotfix on 13 August 2020, but was due to errors in the analysis.  I should say that Professor Asker sought to explain this away, but I was not completely convinced.

1485 She said that another error in the model is that the “after period”, selected to be 10 August 2020 to 13 September 2020, contained an actual event that did give users an incentive to switch, and this event contaminated the results of the Fortnite hotfix analysis of Professor Asker.

1486 On 27 August 2020, a new season of Fortnite was launched. From this date onwards, users were affected by the removal of Fortnite from the Play Store, because they could not update their apps to get the new season. At this point in time switching might have occurred in response to essentially a large degradation in quality. This happened in the middle of the “after” period of Professor Asker, but was not controlled for in the analysis.

1487 Further, in her view the Fortnite hotfix analysis was conceptually incorrect given that the event on 13 August 2020 did not result in an option being removed, or indeed a price change, for users.  In her view the hotfix on 13 August 2020 did not reflect a forced diversion of users from the Play Store and the App Store to other “transaction venues” due to a “complete removal” as characterised by Professor Asker.

1488 Generally, Ms Edwards-Warren was of the view that Professor Asker was erroneous in attempting to estimate substitution of user spend following the hotfix event and subsequent removal of Fortnite from the Play Store in order to estimate how Fortnite users would substitute in response to a small change in relative prices.

1489 Let me elaborate further on some of her evidence.

1490 Now Ms Edwards-Warren identified what she said to be analytical errors in Professor Asker’s approach.

1491 Professor Asker’s analysis found that there was switching by users of their expenditure to other platforms after 13 August 2020. Ms Edwards-Warren considered that it is unclear why the hotfix itself should have driven user switching, so she investigated why Professor Asker’s analysis found that users switched. She considered it possible, for example, that Epic publicly challenging Google’s policies around the time of the hotfix might have affected user behaviour and therefore could have driven the findings of Professor Asker, although ultimately she concluded that the findings were driven by errors in his analysis.

1492 If the public challenge had led to user switching, this would not be a good approximation for developer and user response to a price increase, because she would not expect other developers to publicly challenge Google in the same way.

1493 In her opinion the most likely reason is that Professor Asker’s analysis relied on a poorly defined control group.

1494 Professor Asker conducted difference-in-differences regressions across a four-week period before and a five-week period after the hotfix event.

1495 This is a standard data analysis methodology, and the aim of such analysis is to compare how behaviour of a control group compares with a treatment group and to eliminate the effect of potentially confounding variables, for example, common trends that affect both groups.

1496 But Ms Edwards-Warren did not agree that Professor Asker’s analysis achieved this for the following reasons.

1497 Professor Asker defined the treatment and control groups as follows. He excluded Fortnite users who had spent on the App Store during the same time period from both the treatment and control groups.

1498 The treatment group was Fortnite users who spent on the Play Store during the four weeks before the hotfix. The treatment group included users who spent on the Play Store only, as well as users who spent on the Play Store and other transaction venues. It excluded Fortnite users that spent on Fortnite downloaded from the App Store in this period.

1499 The control group was defined as Fortnite users who only spent on Fortnite from non-Play Store and non-iOS sources during the same time period, being four weeks before the hotfix.

1500 For the analysis to be meaningful, the behaviour of the control group has to be a good approximation of the behaviour of the treatment group. This is known as the parallel trend assumption.

1501 Professor Asker presented no evidence that his chosen control and treatment groups experienced similar trends in user spend. And he presented no evidence that changes in spending by both user groups are similar.

1502 A common method of demonstrating if the parallel trend assumption holds, and therefore whether the chosen treatment group is appropriate for accounting for confounding effects, is to compare the differences between the chosen control and treatment groups before the event occurs that affects the treatment group. If such differences are stable over time, the chosen control group can be said to follow a similar trend to the treatment group.

1503 So, Ms Edwards-Warren compared the evolution of weekly average spend of users in the chosen control and treatment groups of Professor Asker. Her figure 11 below presents this comparison in Australia for user spend on the PlayStation, during the same time period chosen for the hotfix analysis by Professor Asker.

1504 The figure above shows that average weekly user spend of the treatment and control group of Professor Asker’s analysis evolved differently over time. During the “before” period (non-shaded area), average user spend in Australia on the PlayStation for the treatment group ranged between USD 1.16 and USD 1.80, but between USD 3.41 and USD 5.98 for the control group.

1505 To better assess if user spend of the treatment and control group users in Professor Asker’s analysis follow similar trends, Ms Edwards-Warren examined the difference between average weekly spend of both groups, that is, the difference between the light and dark blue lines in figure 11 above.

1506 If both sets of users’ expenditures follow similar trends over time, she would expect these differences to remain relatively similar each week before the hotfix.

1507 Ms Edwards-Warren’s figure 12 below presents the evolution of these differences in Australia for user spend on the PlayStation, during the same time period chosen for the hotfix analysis by Professor Asker.

1508 The figure above shows that the difference between average weekly spend, on the PlayStation, of treatment and control group users in Professor Asker’s analysis in Australia does not remain stable over time. During the “before” period (non-shaded area), the difference in average spend between both sets of users differed by as much as 86%, which is calculated as the difference between the minimum difference, being USD 2.25, and maximum difference, being USD 4.18, divided by the minimum difference.

1509 These comparisons suggest that the control group chosen by Professor Asker does not satisfy the parallel trends assumption required for a differences-in-differences analysis and cannot account for any confounding effect on user spend over time.  Now Professor Asker said that his analysis was robust as to this assumption, but again I was not convinced as to this.

1510 Ms Edwards-Warren’s view, which I accept, was that such analytical errors in the analysis will lead to unreliable and uninformative results.

1511 Ms Edwards-Warren had other criticisms.

1512 Putting aside the above errors of Professor Asker’s analysis, to further illustrate the point that Professor Asker’s hotfix analysis was not in fact estimating diversion, Ms Edwards-Warren re-ran the exact same analysis but centred around three different dates, being 4 June, 18 June, and 9 July 2020. These dates were randomly selected amongst possible dates where a similar analysis spanning nine weeks would still fit within the time period where Fortnite was available on the Play Store.

1513 Running Professor Asker’s analysis on these “placebo” dates, when there is no reason to expect Fortnite user diversion in spend, generates similar results to the hotfix analysis. That is, the results on these dates also show much higher switching to consoles than the Galaxy Store.  Ms Edwards-Warren’s table 13 below presents the diversion ratios estimated based on the regression analyses around the three placebo dates globally excluding China.

1514 The table above shows that in the three placebo analyses, diversion ratios indicate a substantial pattern of user spending diverting from the Play Store to console stores. In particular, diversion ratios from all three placebo analyses indicate higher switching to consoles than to the Galaxy Store.

1515 Because Ms Edwards-Warren generated substantively similar results to Professor Asker using placebo dates, which had no competitive significance, that is, where no change in user behaviour was expected on Fortnite downloaded from the Play Store, this provided further support for her conclusion that Professor Asker’s analysis was flawed and not capable of providing information about diversion of spend in response to a small or large change in prices or quality. So, it is uninformative for market definition.

1516 Now Professor Asker sought to explain away the placebo date conclusions, but I was not convinced.

1517 Further, even taking the results of Professor Asker’s analysis as given, Ms Edwards-Warren found a number of errors and inconsistencies in the interpretation of the results.

1518 First, diversion in response to an extreme event will not necessarily be informative in this case about diversion in response to a smaller event.

1519 Second, Professor Asker assumed that switching, as a proportion of the total lost user spend on the Play Store, following the removal of an option is equivalent to switching following a small change in price or quality. He said that while the hotfix event resulted in the complete removal of Fortnite from the Play Store and the App Store, and so does not correspond to a 5% change in the price or quality of the Play Store or the App Store for Fortnite users, it is nonetheless informative for analysing substitution for the purposes of market definition.

1520 But Professor Asker then contradicted this assumption when criticising Ms Edwards-Warren’s own analysis of the Fortnite data, which was limited to analysing whether users switched from Google Play Billing to EDP after the hotfix which led to Play Store in-app purchases made through EDP being 20% cheaper than when made through Google Play Billing. Professor Asker said that a major issue with Ms Edwards-Warren’s analysis was that consumer behaviour when faced with a 20% price reduction is uninformative about behaviour in response to a smaller price change.

1521 Ms Edwards-Warren disagreed that one can assume that diversion in response to an extreme event such as the Fortnite removal equates to diversion in response to a more limited change, being a small change in price or quality. And she did not make that assumption.

1522 This assumption requires that all users spending on Fortnite behave in the same way. But there is evidence that Fortnite spenders are heterogenous, which implies that this assumption may be incorrect.

1523 Third, the fact that Epic reduced prices across all platforms simultaneously does not, contrary to the claim made by Professor Asker,  suggest that Epic views them as substitutable transaction venues. On the contrary, if the other platforms were substitutable transaction venues, then there would have been limited value to Epic in either distributing via the Play Store in the first place, or in implementing the hotfix.

1524 Fourth, Professor Asker did not consider whether the limited switching to the Galaxy Store or direct downloading might itself be a function of the relevant Google conduct.

Other matters

1525 In Ms Edwards-Warren’s view, another error in Professor Asker’s model was that the “after period”, selected to be 10 August to 13 September 2020, contained an actual event that gave users an incentive to switch, and this event contaminated the results of Professor Asker’s Fortnite hotfix analysis.

1526 Now as I have said, on 27 August 2020, a new season of Fortnite was launched. Users that had previously downloaded Fortnite from the Play Store or the App Store had to either continue playing the previous season or switch to an alternative platform in order to download the new season of Fortnite.

1527 So, the new season launch on 27 August 2020 represented a decrease in the quality of Fortnite for players playing the old version downloaded from the Play Store relative to players who had downloaded the app from other sources. At this point in time it was feasible that those users, being the ones analysed by Professor Asker, switched both playtime/usage and in-app purchasing spend to another platform on which they could access the new season of Fortnite.

1528 So, the launch of the new season was an exogenous event or natural experiment that may have caused diversion of spend.

1529 But because this event occurred part way through the “after period” in Professor Asker’s analysis, being the five-week period after the hotfix between 10 August 2020 and 13 September 2020, it contaminated the “before” and “after” analysis being carried out.

1530 Professor Asker’s analysis defined the “after” period as the five-week period between 10 August 2020, being the start of the week containing 13 August 2020, and 13 September 2020. But only two of those weeks were after 27 August 2020.

1531 By focusing on the 13 August 2020 date, when it is likely that little changed for users who had already downloaded the Fortnite app from the Play Store, and considering only a narrow two-week period after 27 August 2020, Professor Asker was erroneous in attempting to estimate substitution of user spend following the hotfix in order to estimate how Fortnite users would substitute in response to a small change in relative prices.  I have touched on this above.

1532 From 27 August 2020 onwards, users were affected by the removal of Fortnite from the Play Store, because they could not update their apps to get the new season. At this point in time Ms Edwards-Warren would have expected switching to have occurred in response to essentially a large degradation in quality. This happened in the middle of the “after” period of Professor Asker, but was not controlled for in his analysis.

1533 So, in Ms Edwards-Warren’s view, the Fortnite hotfix analysis was conceptually incorrect given that the event on 13 August 2020 did not result in an option being removed, or indeed a price change, for users.

1534 In summary, Ms Edwards-Warren considered that Professor Asker’s analysis should not be attributed any weight for the purposes of defining markets.

1535 Let me deal with one final aspect of her evidence concerning the generalisation of the Fortnite hotfix analysis to other games apps.

1536 Even if one concluded that the Fortnite hotfix analysis was conceptually and analytically sound, Ms Edwards-Warren disagreed with the conclusion of Professor Asker that the results of this analysis can be generalised to all games.

1537 Now Professor Asker concluded that, on the basis that the top 25 games apps have cross-wallet and cross-progression functionalities, users of those apps have the same substitution options as users of Fortnite.  But Ms Edwards-Warren analysed those top 25 five games apps and found the following matters.

1538 First, Fortnite is and was at the time of the hotfix playable on consoles as well as PCs and mobile, whilst only two of the top 25 games apps on the Play Store are playable on consoles.

1539 Second, Fortnite is and was at the time of the hotfix available on six platforms. 23 out of the top 25 games apps on Android are available on only two or three platforms.

1540 Third, five of the top 25 games apps are available on mobile only. These five games account for 20% of the top 25 games apps and [REDACTED] of Australian consumer spend on Play Store in-app purchases across the 25 game apps in 2021.

1541 Fourth, Fortnite offers and offered at the time of the hotfix cross-wallet functionalities on consoles, except the Nintendo Switch, and on PC and mobile, while only two of the top 25 games apps offer cross-wallet functionality on consoles, and not all offer cross-wallet on PC.

1542 Fifth, Fortnite offers and offered at the time of the hotfix cross-progression functionalities on consoles, PC and mobile, while only two of the top 25 games apps offer cross-progression functionality on consoles, and not all offer cross-functionalities on PC.

1543 On the basis of this analysis, she concluded that the Fortnite hotfix analysis cannot be generalised to other games apps on Android.

Professor Goldfarb’s evidence

1544 Like Professor Asker, Professor Goldfarb also used the removal of Fortnite from the App Store and the Play Store as a natural experiment to estimate diversion to Play Store alternatives following what could be considered to be an infinite price increase on the Play Store.

1545 Now Professor Asker examined diversion of player spending, and so his analysis was restricted to only players that spend.  But Professor Goldfarb’s analysis assessed diversion in usage across all users.

1546 Now they both began with standard regression models to estimate the change in their respective outcome variables, being spending or usage, for Fortnite Play Store users based on a difference-in-differences estimation approach.  But Professor Goldfarb considered that Professor Asker’s primary diversion methodology led to problematic results.

1547 Professor Asker’s results differed from Professor Goldfarb’s primarily because he used a levels difference-in-differences approach. But according to Professor Goldfarb this is an inappropriate approach because the average spending amounts for Professor Asker’s treatment and control groups differ substantially prior to removal. This leads to inaccurate estimates of what spending would have been in the absence of the removal for the treatment group.

1548 Specifically, according to Professor Goldfarb, this drives the illogical result of negative average spending values for several channels, implying that Epic would pay users to spend through such channels but for the removal. When a more appropriate percentage-based approach to calculate diversion was used, Professor Goldfarb found diversion estimates that aligned with the results he found for usage.

1549 Professor Goldfarb also disagreed with Professor Asker’s choices regarding several regression approaches. These disagreements centred around his use of a short time period for assessment of the removal.

1550 Now beyond methodological differences, Professor Asker also emphasised that his disagreement with Professor Goldfarb’s removal analysis was that Fortnite player usage is the incorrect metric for understanding substitution.

1551 But I agree with Professor Goldfarb that usage is a direct measure of the consumer benefit from playing Fortnite. Focusing only on spending ignored that the majority of Fortnite users do not spend. Professor Asker’s analysis ignored that Fortnite still derives value from these non-spending users. A focus exclusively on spending therefore underestimates the true impact of removal.  I have discussed this debate as to the use of the appropriate metric elsewhere in my reasons.

1552 Now as I have touched on, Professor Asker calculated diversion to a given channel as the additional spending on that channel attributable to the removal, divided by the loss in spending on Play Store attributable to the removal. Professor Asker estimated any additional spending on Play Store alternatives using a levels differences-in-differences regression approach.

1553 Professor Asker said that he used a standard difference-in-differences methodology, which infers the treatment group’s counterfactual outcome by leveraging the level change observed in the control group. This means that he assumed, had the hotfix not occurred, the change in spending relative to the pre-period for the treatment group would have been by the same amount as the change observed for the control group. He said that this is a standard approach in the economic literature for estimating a difference-in-differences regression. He referred to this standard approach as a levels difference-in-differences.

1554 Professor Goldfarb explained that a key assumption in applying a levels-change methodology is that the treatment group, that is, Fortnite Play Store spenders, and control group, that is, Fortnite non-Play Store spenders, must be largely similar. When they are not, economists often take the log of the outcome variable so that proportional changes can instead be assessed. This is because the methodology uses the control group’s trajectory to construct a counterfactual trajectory for the treatment group. It infers that the observed level-change in spending for the control group from pre-removal to post-removal would have also occurred, at the same level, for the treatment group but for removal.

1555 Professor Goldfarb said that for some channels, during the pre-removal period the average spend for the control group was upwards of three times greater than the average spend for the treatment group.  This points to a substantial lack of comparability between the treatment and control groups in levels that must be addressed in the analysis design. But according to Professor Goldfarb, this approach is inappropriate and leads to illogical conclusions about what would have happened but for the removal.

1556 According to Professor Goldfarb, Professor Asker’s inappropriate levels differences-in-differences approach yields illogical results, for example, Fortnite spending absent the removal is estimated to be negative for PlayStation and Switch. In other words, but for the removal, Epic would have paid users of those channels to transact on Fortnite. Professor Asker estimated a negative counterfactual spending amount for some channels, which was unrealistic. Such results arose due to the substantial differences pre-removal between Professor Asker’s treatment and control groups for some channels.

1557 Professor Goldfarb said that one well-recognised solution is to assess the relative change rather than the absolute change in the treatment by the use of a logarithmic transformation of a given outcome variable. When a variable is so transformed, the estimated effect in a regression is then interpreted as a percentage increase in response to treatment.

1558 Professor Goldfarb updated Professor Asker’s analysis with this approach. He replicated Professor Asker’s results and corrected the approach by using a more appropriate proportional change in spending. He found substantially lower diversion rates to non-Android channels.

1559 Professor Goldfarb found that Professor Asker’s diversion results were skewed towards a diversion to PlayStation, Switch, and Xbox, which was on precisely these channels where the control group average spending is more than double that of the treatment group pre-removal. I will elaborate on this in a moment.

1560 Further, Professor Goldfarb said that a separate and important weakness in Professor Asker’s empirical analysis of the removal was his use of a narrow analysis period for assessing diversion. Professor Asker’s window of assessment was 4-weeks pre-removal and 5-weeks post-removal, including the week of removal.

1561 Professor Goldfarb used wider analysis periods, for example, 13-weeks pre-removal and 13- weeks, 26- weeks, 52-weeks post-removal, excluding the week of removal. And in his view, the use of a narrower analysis period reduces the reliability of Professor Asker’s results in several ways.

1562 First, consumers could still use the Play Store Fortnite app with full functionality during the week of removal and up to two weeks post-removal. Removal from the App Store and Play Store meant that users could no longer download the Fortnite app through these stores. Users who had already downloaded the app through these stores could continue to access the app, but would not be able to re-download, for example, if they factory reset their device. And they would not be able to update their Fortnite app.

1563 Now Professor Asker’s analysis was designed to compare a period where Fortnite was available on the Play Store to a period where Fortnite was not available on the Play Store. But consumers could still play with full functionality on the Fortnite Play Store app for over half of the period in which Professor Asker counted it as unavailable.

1564 Second, Epic introduced a new Fortnite season two weeks post-removal, which likely changed patterns of Fortnite demand and spend. The new season created an asymmetry in an important factor before and after the removal, making Professor Asker’s estimated impact of removal less reliable. For these reasons, Professor Goldfarb assessed diversion over longer time periods that would be less impacted by these two factors.

1565 Third, according to Professor Goldfarb, Professor Asker incorrectly included the week of removal in his regression. As removal occurred on a Friday, users spent as normal for the majority of the week. Including this week as a post-removal week mis-attributed pre-removal spending to the post-removal period.  So, it understated the loss in Play Store spending directly following removal and similarly renders Professor Asker’s results unreliable.

1566 In Professor Goldfarb’s opinion, Professor Asker’s reliance on a shorter 4-week analysis period post-removal and inclusion of the removal week and the two weeks following the removal week may have impacted the measured rate of diversion.

1567 Professor Goldfarb tested the impact of including these weeks on Professor Asker’s analysis. The overall diversion rate to non-Android channels remained largely consistent when Professor Goldfarb updated Professor Asker’s analysis. I will address his updated analysis in a moment.

1568 Professor Goldfarb also considered the analysis presented by Professor Asker to justify the 4-week post-removal analysis period. Professor Asker sought to justify the 4-week analysis period in part to avoid introducing survivorship bias into his analysis.

1569 Professor Goldfarb explained that survivorship bias, also known as attrition bias, occurs when a study using data that tracks individuals over time only focuses on individuals who passed some selection process and are observed in the data through the end of the sample. When the probability of attrition is correlated with the exogenous variable or other unobserved variables that affect the outcome, this would lead to biased results.

1570 But according to Professor Goldfarb, survivorship bias is not a risk owing to the design of Professor Asker’s analysis and using survivorship bias as a justification for a 4-week analysis period is illogical. Professor Asker selected users that pay prior to removal and followed these users over time. Users who do not have another observation remained in the dataset and were included in the analysis. This removed the possibility of biasing the sample to consider only those users that had an observation in the post-removal period. In Professor Goldfarb’s analysis, he also avoided survivorship bias by following all users in his selected sample post-removal in the same manner as Professor Asker.

1571 Professor Asker offered a second justification for using a 4-week window. He presented an additional analysis where he claimed to show that 70% of Android and Play Store users who spend on Fortnite spend again within 4-weeks. Professor Goldfarb examined this and I will return to this topic in a moment.

1572 The upshot of this according to Professor Goldfarb was that the use of a 4-week window provides less accurate results than those provided by a longer window post-removal. Professor Goldfarb considered a time window of assessment that is at least 13-weeks long to be more appropriate.

1573 Professor Goldfarb found that extending Professor Asker’s window of analysis to 13-weeks pre-removal and 13-weeks, 26-weeks or 52-weeks post-removal, there were consistently lower overall diversion rates to non-Android spending channels than those reported by Professor Asker.

1574 Further, Professor Goldfarb said that extending the time window out past 4 weeks also required the relaxation of an overly restrictive assumption that users must spend within the analysis window and spend or play before the analysis window. When using a pre-removal analysis window of 13-weeks, this entails that users most likely must have had played or spent prior to Fortnite becoming available on the Play Store.  This leads to a selection bias towards multi-homing users.

1575 I will set out some of Professor Goldfarb’s tabulated results in a moment.

1576 Professor Goldfarb concluded that, similar to usage, users do not treat the ability to spend through native apps on Play Store as substitutable with the ability to spend on other apps on desktop computers, laptop computers, or gaming consoles, except to a limited extent.

1577 Let me descend into more detail concerning Professor Goldfarb’s analysis and some aspects that I have just touched on above.

1578 First, let me explain why Professor Goldfarb said that Professor Asker’s primary diversion methodology reveals illogical diversion of Fortnite spending from Play Store to other channels.

1579 In the current context, a diversion estimate is the ratio of the additional usage or spending on a given channel, for example, PlayStation, following removal, relative to the loss in usage or spending on Play Store resulting from removal.

1580 To calculate the denominator, it is necessary to estimate the usage or spending but for removal. Professor Asker did not explicitly calculate this counterfactual spending. But one can observe his implicit counterfactual spending using his regression coefficient estimates of the removal effect.

1581 The average counterfactual spending levels indicated by Professor Asker’s analysis were negative for the PlayStation and Switch channels. These negative estimated spending amounts mean that Epic would pay users to transact for such channels but for the removal. This is an unrealistic result.

1582 According to Professor Goldfarb, to understand why Professor Asker’s primary diversion methodology is illogical, one can break down the two conceptual components of a diversion ratio in more detail.

1583 The first measure, the denominator, in a diversion ratio is an estimate of the usage or spending that was lost overall from Fortnite on the Play Store because of the removal.

1584 The second measure, the numerator, in a diversion ratio is the portion of the usage or spending following removal that was gained or lost because of the removal for each Play Store alternative channel, for example, PlayStation, PC, Xbox or Switch.

1585 Both measures are based on the estimation of a counterfactual usage or spending amount for each channel. This counterfactual is the amount of usage or spending that would have occurred in the absence of the removal. It is the usage or spending that would have occurred in the counterfactual world.

1586 In Professor Goldfarb’s own diversion analysis, he explicitly calculated this counterfactual amount while Professor Asker did so implicitly with his selected methods.

1587 Second, Professor Goldfarb said that Professor Asker estimated a negative counterfactual spending amount for some channels.

1588 Professor Asker and Professor Goldfarb differed in their approach to calculating this counterfactual for non-Play Store channels. Professor Asker’s primary methodology used a levels difference-in-differences approach to estimate counterfactual channel spending for non-Play Store channels. He assumed that had the hotfix not occurred, the change in spending relative to the pre-period for the treatment group would have been by the same amount as the change observed for the control group.

1589 In contrast, Professor Goldfarb estimated counterfactual treatment group spending based on the percentage change in spending from pre-removal to post-removal realised in the control group. His analysis was more appropriate given the difference in spend levels between the treatment and control groups evidenced in his table 5.

1590 Professor Goldfarb replicated and added to Professor Asker’s primary regression results in table 5 to demonstrate the main issue with his underlying counterfactual spending estimates. He presented average spending rates for the treatment and control group pre- and post-removal, and the difference between pre- and post-removal for each group in columns (A) to (F). Pre-removal, the overall average spending for the control group at $8.89 (column A) is lower than that of the treatment group at $13.30 (column D) on average per user per week. He also showed that the treatment group played more hours per week on average. The control group had noticeably higher spending on PlayStation, Xbox, and Switch pre-removal, while the treatment group had noticeably higher spending on Android channels. These differences limit the applicability of using a levels-change difference-in-differences approach to understand the impact of the removal.

1591 In column (G), Professor Goldfarb replicated Professor Asker’s reported regression estimates from exhibit 61. This coefficient is the absolute increase in spending due to removal for the treatment group, estimated through his regression. In column (I), he calculated the counterfactual average spending amount implied by Professor Asker’s levels difference-in-differences regression approach for each channel. This is estimated as the observed post-removal average spending amount for each channel (column E) less the estimated increase due to removal (level-change regression coefficient from column G). Column (I) revealed that Professor Asker’s underlying counterfactual spending estimate on non-Play Store channels yields illogical, negative spending for Switch and PlayStation. Clearly Epic would not pay users to transact for such channels.

1592 According to Professor Goldfarb, the error lies in Professor Asker’s reliance on a levels-change (absolute value) difference-in-differences applied to highly divergent spending amounts for treatment and control groups for certain channels.

1593 A difference-in-difference estimation takes the difference in average spending for the treatment group before and after removal less the difference in spending for the control group before and after removal.

1594 For PlayStation, the control group average pre-removal spending was $4.53. This is more than three times greater than the pre-removal treatment group spending average, at $1.37. Average spending on PlayStation dropped an average of $1.51 for the control group post-removal, to $3.02. The same drop for the treatment group leads to a negative spend amount, as $1.51 is more than the average pre-removal spending amount of $1.37.

1595 Professor Asker’s approach implied negative spending for Switch for the same reasons. Professor Asker’s regression model controls for additional individual and time specific effects and estimates that removal caused a $1.75 increase in spending for the treatment group on average per user for a PlayStation user.  Column E shows that actual spending in the treatment group on PlayStation was $1.49 on average per user. This implies that spending on PlayStation for the treatment group would have been approximately -0.25 on average per user in the absence of removal.

1596 In Professor Goldfarb’s view, the difference in pre-removal spending for Professor Asker’s treatment and control groups meant that a proportional change approach was more appropriate.

1597 A proportional change approach may take the form of a regression specification using the log of spending as an outcome variable. This is a common technique used in economics. A proportional approach may also take the form of estimating a counterfactual based on the percentage change in the control group based on raw numbers. Professor Asker performed this percent difference-in-differences analysis in exhibit 158 as a robustness test against his primary results and found noticeably lower diversion rates. To show why this approach renders more logical results, in column (H) Professor Goldfarb replicated the same methods that he used to calculate a counterfactual average spending amount. He based this on Professor Asker’s main regression modelling assumptions. When using a proportional change approach, as in column (H), Professor Goldfarb found counterfactual estimates that are always above zero.

1598 Third, Professor Goldfarb said that Professor Asker’s regression methodology biased his diversion estimates towards channels with high spending pre-removal in the control group.

1599 Professor Asker used the estimated regression coefficients from his levels difference-in-differences regression approach, with underlying illogical counterfactual spending estimates, to calculate diversion. This led to biased results for several reasons. One such reason is the bias in his numerator described above.  Another reason is that the denominator and numerator in his diversion calculation are internally inconsistent. The numerator Professor Asker used was a level change in spending that overstated gains in spending on Play Store alternatives. His denominator was the impact on Play Store spending scaled by the percentage change observed for the control group. These two numbers were estimated through different assumptions and have minimal relation to one another.

1600 Professor Asker’s methodology overstated the overall estimated rate of diversion.

1601 Professor Asker’s use of a levels difference-in-differences regression specification, which causes negative counterfactual estimates for some channels, will overstate the increase in spending for channels with much higher pre-removal spending in the control group versus the treatment group. Larger differences in pre-removal spending averages for the treatment and control groups will generate more disproportionate estimates of the change in spending in the treatment group attributable to removal. This effect in particular skewed Professor Asker’s diversion estimates towards PlayStation, Switch, and Xbox, since in these channels the control group average spending is more than double that of the treatment group.

1602 To demonstrate the impact of using a percentage change approach, Professor Goldfarb left Professor Asker’s modelling assumptions unchanged, except he applied a percentage change approach to estimate diversion in table 6.  Using this approach, the loss of Play Store spending to non-Android channels was 26.4% (or 29.7%) using the assumption that 100% (or 50%) of EDP spending is through direct download.

1603 The results in Professor Goldfarb’s table 6, together with Professor Asker’s exhibit 158, indicate large differences in estimated diversion using a levels difference-in-differences compared to a percent difference-in-differences approach. But Professor Asker surmised that the results in exhibit 158 reveal that his methodology and Professor Goldfarb’s methodology generate very similar estimates for diversion to non-Play Android stores and PC stores.  Professor Goldfarb did not agree with this statement.

1604 The estimated diversion to non-Android alternatives using the appropriate percentage change diversion methodology is half that of Professor Asker’s diversion calculation based on illogical counterfactual post-removal spending estimates. The largest differences are for those based on negative estimates of spending in Professor Asker’s counterfactual, such as PlayStation and Switch. Using the percentage change diversion methodology, non-Android alternatives represent a significantly lower proportion of diverted spending.

1605 Fourth, Professor Goldfarb said that Professor Asker’s use of a short period to assess consumer switching behaviour biased his results.

1606 In addition to Professor Asker’s incorrect diversion estimates, Professor Goldfarb considered the period adopted by Professor Asker to study diversion of spending to be unreliably short and impacted by several sources of bias.

1607 In his analysis, Professor Asker obtained results using a 4-week window before removal and a 5-week period after removal, which included the week of the removal and the 4-weeks after that week. The post-removal window was influenced by several factors that rendered these results unreliable.

1608 As I have already said, in Professor Goldfarb’s own removal analysis studying usage, he presented results assessing diversion from 13-weeks prior to removal to 13-weeks, 26-weeks, and 52-weeks post-removal, representing short, medium, and long-term windows of analysis.

1609 Professor Asker stated that longer time intervals, for example, greater than 4 weeks, are inappropriately long as they may be impacted by factors other than the removal.

1610 The influence of other factors is important and something that Professor Goldfarb considered when performing his removal analysis. He found that a longer time window corrects for factors that render a four-week window uninformative and does not introduce additional significant bias.

1611 Fifth, Professor Goldfarb said that Professor Asker’s short analysis period led to biased channel diversion ratios.

1612 Professor Asker estimated diversion of spend, following the removal event, from the Play Store to other channels. He employed a 5-week post-removal window in this analysis, being the week of removal and an additional four weeks after.  There are several factors that cause results based on a 5-week post-removal window to be unreliable. When Professor Goldfarb adjusted Professor Asker’s analysis to correct for these factors, he found noticeable differences in individual channel diversion rates.

1613 The first factor causing Professor Asker’s results to be less reliable is that users could still use the Play Store Fortnite app with full functionality until 27 August 2020, two weeks following the removal event. Any spending from those users during that time period would be included in Professor Asker’s estimate of the treatment group’s post-removal spending through the Play Store. It is likely that consumer use and spending behaviour may have been impacted directly following the removal. But Professor Asker’s estimate of the loss in spending owing to removal included a period where users could continue to use the Play Store Fortnite app with full functionality, leading him to underestimate the loss in spending once users could no longer spend via the Play Store.

1614 Further, Professor Asker’s 5-week post-removal window included the release of a new Fortnite season on 27 August 2020. This meant that Professor Asker’s analysis compared spending during the release of a new season in the five weeks post-removal to spending during the tail end of a season in the four weeks pre-removal.

1615 The direction of bias of this event on the results is not clear. If users belonging to the treatment group engaged more in the first weeks of a season compared to users belonging to the control group, this would bias Professor Asker’s diversion ratio estimate upwards and overstate diversion from Play Store to other channels. If users belonging to the treatment group instead engaged less in the first weeks of a new season compared to users belonging to the control group, then the estimated diversion rate would be understated.

1616 Finally, Professor Asker included the week of removal in his regression. To estimate the effect of removal on spending diversion using a difference-in-differences approach, Professor Asker must compare the difference in pre-removal average spending and post-removal average spending between the treatment group and control group.

1617 The removal occurred on a Friday, meaning that during the week of the removal, users spent as normal from Sunday to Thursday but were affected by the removal from end of day Thursday to Saturday. Professor Asker included the removal week in his post-removal period.

1618 Including the Play Store spending prior to removal on 13 August 2020 in the post-removal period adds noise to the estimation. This artificially inflated the average post-removal spending on Play Store amount for the treatment group.

1619 In table 7, Professor Goldfarb demonstrated the effect of correcting for these factors that affect Professor Asker’s analyses. In table 7, he presented percentage difference-in-differences diversion calculations, plus two iterations in which he made updates to his calculations. Columns A and B of table 7 replicate Professor Asker’s exhibit 158. These are based on his comparison of the four weeks pre-removal to the removal week and the following four weeks.

1620 Iteration 1, presented in columns C and D, provides results following Professor Asker’s sampling methodology but excluding the removal week from the calculation. In other words, iteration 1 compares the four-weeks pre-removal to the four-weeks post-removal, excluding the removal week.

1621 Iteration 2, presented in columns E and F, provides the results following Professor Asker’s sampling methodology but excluding the two weeks following the removal week as well as the removal week. Iteration 2 compares four-weeks pre-removal with the four-week period beginning 31 August 2020.

1622 Table 7 shows that the diversion to non-Android channels remains largely consistent across these iterations using a 4-week window of analysis.

1623 The diversion rates to the Galaxy Store and the diversion rates to direct download vary between exhibit 158, iteration 1, and iteration 2. One source of variation is the difference in share of EDP through direct download. Professor Goldfarb first considered this source of variation and then discussed the remaining variation arising from adjusting the time windows for Professor Asker’s analysis.

1624 Each iteration in table 7 provides two results, being one for 100% direct download and one for 50% direct download. This is to account for insufficient visibility into the relevant channel for EDP data subsequent to the removal. The results for each spending channel within each iteration in table 7 are largely consistent regardless of whether 100% or 50% direct download share is used, apart from direct downloads itself.

1625 To assess whether the 100% direct download or the 50% direct download result provides the better estimate in iteration 1 and in iteration 2, Professor Goldfarb considered the availability of EDP in the Play Store and direct download channels.

1626 Professor Asker provided an estimate that found that between 10 August and 13 September 2020, 43.9% of EDP transactions occurred via the Play Store app. This indicates that for exhibit 158, the estimate using the 50% direct download share is likely more accurate than the estimate using the 100% direct download share. For iteration 1, the estimate using the 50% direct download share, or an estimate between the 50% and 100% direct download share estimates, is likely more accurate than the estimate based on 100% direct download share.

1627 The Fortnite app on the Play Store lost significant functionality for users following the season release on 27 August 2020.  Following this date, it is likely that most transactions through EDP were related to spending through the direct download version of Fortnite.

1628 The share of EDP through direct download also likely increases over time. This means for iteration 2, where the post-removal period begins on 31 August 2020, the estimated diversion ratio using a 100% share of direct download is likely more accurate than the estimate using a 50% direct download share. It would also be more accurate to assume a 100% share for longer periods of analysis, including for an analysis that extends to 13-weeks post-removal or longer.

1629 Professor Goldfarb highlighted the more accurate results for exhibit 158, iteration 1 and iteration 2 in his table 7.  I will not set out the highlighted table, but would note that what he highlighted were all of columns B, D and E in the table set out above.

1630 The direct download diversion ratio varies more than the diversion ratio for other spending channels in part because changing the share of direct download payments affects both the numerator and denominator of the direct download diversion ratio. For other channels, the share of direct download payments affects only the denominator. Accounting for the more accurate rate of direct download reduces the variability in estimates for diversion to direct download.

1631 Table 7 shows that the diversion rates to the Galaxy Store changes between exhibit 158 and iterations 1 and 2, from 10.6% to 11.1% and 18.6%. This reflects the effect of adjusting the window for analysis to remove the removal week, being iteration 1, and the removal week plus the following two weeks, being iteration 2, and accordingly comparing the pre-removal spend to a more appropriate post-removal time period when considering a 4-week window.

1632 Fortnite users who downloaded via the Play Store on or prior to 13 August 2020 could still play with full functionality in Fortnite between the removal and 27 August 2020. Once the Fortnite app via the Play Store functionality diminished on 27 August 2020, users may have been more likely to switch Fortnite usage and spending to another channel with full functionality. Users who switched to another channel selecting the Galaxy Store rather than non-Android channels is consistent with users being motivated by the “device first” and “clicks are costly” concepts.  Indeed, the Galaxy Store has the highest diversion ratio in iteration 2 amongst the six spending channels displayed in table 7.

1633 Sixth, Professor Goldfarb said that Professor Asker’s reasons for choosing a four-week time window did not support his conclusion and in fact indicated that a longer time period was more appropriate.

1634 Professor Asker asserted that a time frame of four weeks was appropriate to assess diversion due to the potential for survivorship bias. He supported this assertion with two analyses of Fortnite churn.

1635 In one analysis, Professor Asker looked at the number of weeks in which Fortnite Android or Play Store users who engaged in paid transactions did so during the 16-week period in which Fortnite on the Play Store was available. He found that 36% of spending users spent in only one week out of the 16-weeks. Professor Asker said that this means that using a shorter time period for analysis corrects for the possibility of error due to survivorship bias.

1636 As I have already touched on, survivorship bias occurs when a study using data that tracks individuals over time only focuses on individuals who passed some selection process and are observed in the data through the end of the sample. Professor Asker asserted that short time frames are appropriate for studying substitution as longer time frames would narrow the analysis to a selected set of Fortnite users and be impacted by survivorship bias.

1637 But this concern is misplaced given how both Professor Asker and Professor Goldfarb carried out the removal analysis in practice. The analysis that both Professor Asker and Professor Goldfarb performed did not narrow focus on a selected set of users that continued to play or spend on Fortnite. They both instead required that a user play or spend prior to removal and track all users in the post-removal period. If a user does not spend in a week, they are recorded as having zero spend in that week, not dropped from the sample. If a user does not spend following the removal, the remainder of their recorded weeks are recorded with zero dollars of spending. They are not dropped from the sample.

1638 Given that the analysis design does not introduce survivorship bias, Professor Asker’s finding that one week is the most common number of weeks in which to have a paid transaction does not affect the choice of appropriate analysis window.

1639 In the second analysis, Professor Asker looked at the distribution of time elapsed between paid transactions, which he referred to as spending events, on Fortnite, for Australian Android and Play Store users who made more than one purchase across the full data period.

1640 Professor Asker found that the median time between spending events for spenders with two or more transactions is only two weeks for both Australian Android and Play Store users and that 70% of Android and Play Store users who spend on Fortnite would choose where to transact again within four weeks of a triggering event.

1641 Professor Goldfarb reviewed Professor Asker’s calculation methodology and found a more accurate way to calculate the statistic provided by Professor Asker.

1642 Professor Asker purported to show the amount of time that elapses between purchases for Australian Android and Play Store users. This is relevant given that both were interested in observing, for a user who purchased in the Play Store before the removal, where that user purchased following the removal. The typical interval between purchases helps one know how long after removal to look to see the user’s post-removal purchase, if any.

1643 But upon examining Professor Asker’s methodology to calculate the typical interval per user, it was apparent to Professor Goldfarb that Professor Asker had calculated that 70% of consecutive purchases are less than four weeks apart, not that 70% of users make transactions less than 4 weeks apart. Professor Asker’s calculation looked at the distribution of time between two consecutive purchases across users, not at the user-level. Professor Asker then calculated the average and other statistics of these transaction-distance observations. This introduced error in assessing the typical frequency of user purchases because users that purchase more frequently occur more often in the transaction-distance observations. If there is a group of Fortnite users who make frequent and numerous purchases during the data period, these players will skew the distribution towards the shorter intervals.

1644 In other words, Professor Asker calculated these statistics across transactions, which weights frequently transacting users more heavily. The more informative calculation is to take an average per user and then the average across users. This calculation indicated the typical gap between purchases for a user and consequently the time period should be examined post-removal.

1645 Professor Goldfarb replicated Professor Asker’s exhibit 53 but considered instead a distribution based on user-level average number of weeks between spending events. Professor Goldfarb’s table 8, which I will not reproduce, shows that the average user who spends on Android has 10.6 weeks between spending events. The average user that spends on Play Store has 10.8 weeks between spending events. Even in the 30th percentile of Android and Play Store users, the number of weeks between spending events is 4.3. His table 8 demonstrated that a window of analysis of only 4-weeks is too restrictive because more than 70% of users average more than 4 weeks between spending events.

1646 Based on these two results, Professor Asker argued that user attrition will bias estimates relying on a longer analysis window and that a shorter analysis window is reasonable given his calculation of purchase frequency.

1647 For the reasons stated above, Professor Goldfarb did not agree with Professor Asker’s conclusion. As most users who make paid transactions do so across more than one week and the time between transactions is long, a longer time period of analysis should be used. Professor Goldfarb assessed the effect of switching to a longer time period in Professor Asker’s analysis.

1648 Professor Goldfarb considered it appropriate to update Professor Asker’s analysis to consider diversion over a longer time period than four weeks, in order to obtain more accurate estimates of diversion rates that are not as prone to the biases detailed above. To do so, he had to first update his sample selection criteria. He then updated Professor Asker’s analysis to reflect a longer window.

1649 Seventh, Professor Goldfarb said that updating Professor Asker’s sample selection criteria was appropriate when extending the time window for analysis.

1650 Professor Asker imposed two conditions on his sample selection.

1651 The first condition required that a user must spend in Fortnite on any channel within the pre-removal period in order to be included in his sample. In Professor Asker’s analysis, this required that a user spends in the four weeks prior to removal, between 16 July 2020 and 13 August 2020.

1652 The second condition required that a user either plays or makes a purchase in Fortnite on any channel or spending channel prior to the pre-removal period, that is, between the start of the dataset on 26 February 2018 and 16 July 2020, being four weeks before removal, in Professor Asker’s analysis. Professor Asker imposed the second condition to include in his sample only users who are familiar with Fortnite.

1653 Professor Goldfarb considered that Professor Asker’s analysis would produce a more accurate result if the pre-period was extended. For that reason, he extended his pre-removal period from four weeks to thirteen weeks. This extension required the exclusion of the second condition imposed by Professor Asker for the following reasons.

1654 Fortnite was released on Play Store on 20 April 2020. Increasing the pre-period to thirteen weeks, covering a time period from May 2020 to August 2020, reduced the amount of time between the availability of Fortnite on Play Store and the beginning of the pre-period of the analysis to just three weeks, which was relevant to the second condition.  This provided therefore only a short time period to observe users who did not play or purchase in Fortnite through other channels but were attracted to adopt Fortnite on Play Store following its release. Given that such users did not already play or purchase in Fortnite through other channels, Professor Goldfarb’s expectation was that most users would single-home on Fortnite through the Play Store.  So, extending the time period, which he considered to be more reliable, required removing the second condition.  Otherwise, the inclusion of that condition would likely artificially bias the treatment group, being users who had ever paid in Fortnite on the Play Store, to Fortnite users who multi-home across Play Store and other channels.

1655 Professor Goldfarb assessed the possibility that imposing the second condition with a 13-week pre-period biases the treatment group towards multi-homing users. In table 9, he displayed summary statistics for two samples. The first sample was selected according to Professor Asker’s two conditions and a 13-weeks pre-removal period, whilst the selection of the second sample removed Professor Asker’s second condition, requiring that users have paid or played on Fortnite before the pre-period.

1656 Table 9 shows that imposing the second condition increases the share of multi-homing users in the treatment group and excludes many single homing Play Store users from the sample.  The share of treatment group users multi-homing in the 13 weeks before removal drops from 42% to 30% when removing the second condition.  The change is even larger when considering multi-homing over the full time period between 2018 and removal. This is because users in the treatment group necessarily have purchased in Fortnite on the Play Store and the second condition allows a much longer window to include users who have also paid or played on Fortnite in other channels (July 2018 to May 2020) compared to users for whom Fortnite on the Play Store is their sole channel between 20 April 2020 and May 2020.

1657 Professor Goldfarb assessed the effect that imposing the second condition and increasing the share of multi-homing users in the treatment group had on the analysis results. In table 10, he provided diversion estimates for both samples considered above, applying a 13-week pre- and post-removal analysis. In this analysis, he also applied the proportion change diversion estimate method, along with an exclusion of the removal week.

1658 Comparing the results with and without the second condition, he found lower diversion to each channel excluding the Galaxy Store. This lower rate of diversion for channels other than the Galaxy Store is likely because the sample without the second condition aptly included more single-homing users who were more likely to stop spending altogether after the removal.

1659 The Galaxy Store sees an increase likely because it is the most straightforward alternative for the single-homing Fortnite Play Store players. When applying a longer time window for analysis, Professor Goldfarb considered the analysis without the second condition to be more reflective of what occurred for consumers post-removal and less prone to a sample selection bias.

1660 Eighth, Professor Goldfarb said that extending Professor Asker’s analysis out to medium and long-term windows showed generally consistent rates of diversion to non-Android channels.

1661 Professor Goldfarb extended Professor Asker’s analysis by considering three longer windows of analysis, being 13 weeks before and 13 weeks after removal, 13 weeks before and 26 weeks after removal, and 13 weeks before and 52 weeks after removal. In his extensions he also employed a proportional diversion estimate and removed sampling the second condition.  He further presented results excluding the week of removal, being iteration 1, and removing both the week of removal and the following two weeks, being iteration 2, given that Fortnite was still available with full functionality up to 27 August 2020.  He considered excluding the removal week to be the correct approach. Understanding the impact of the additional two weeks following removal on diversion results is also important, though in practice the effect of these weeks is likely to be attenuated when using a longer analysis window.

1662 In table 11.A, he found generally consistent results across the analyses when using longer time periods and excluding the removal week. He found a total diversion to non-Android channels of 28.1% to 29.7% when measuring diversion in the 13 weeks post-removal. He found a total diversion to non-Android channels of 27.8% to 28.7% when measuring diversion in the 26 weeks post-removal. When measuring diversion in the 52-weeks post-removal, the total diversion to non-Android channels was slightly higher at 33.3% to 34.3%.

1663 So, consumers do not treat usage or spending on Android smart mobile devices as substitutable with laptop computers, and desktop computers, or gaming consoles, except to a limited extent.

1664 Since users were able to play Fortnite with full functionality on their Play Store Fortnite app until the new season was released on 27 August 2020, the full extent of switching to other channels may not have occurred immediately after the removal.

1665 Professor Goldfarb conducted a robustness analysis where he removed not only the week of removal, but also the following two weeks from the post-removal period. So, the post-removal period started at the third week after the removal, that is, from 31 August 2020, and is for 13-, 26- and 52-weeks thereafter. The pre-removal period started 11 May 2020 and ended the week before removal.

1666 Table 11.B presents results for this robustness extension with time windows of 13-weeks, 26-weeks, and 52-weeks post-removal. The results for the 13-week, 26-week, and 52-week periods are generally consistent with table 11.A.

Analysis

1667 In my view Professor Asker’s difference in differences analysis is unreliable and should not be used. Moreover, Professor Asker’s analysis inappropriately includes the weeks of 10, 17 and 24 August 2020 which leads to unreliably low estimates of diversion to the Galaxy Store.  I prefer Professor Goldfarb’s approach.

1668 A reasonable range of diversion estimates is provided by the highlighted estimates for 13-, 26- and 52-weeks in Professor Goldfarb’s table 11.A and the highlighted estimates in table 11.B, with a range of between 10.1 and 18.7% for diversion to the Galaxy Store.

1669 The best single estimate of diversion to the Galaxy Store is the highlighted estimate for 13 weeks in table 11.B, being 15.8%.

1670 On Professor Goldfarb’s figures, approximately half of all spending was lost as a result of the removal.

1671 Let me now address some specific topics.

What is preferable? A levels analysis or a percent approach?

1672 Now as I have said, Professor Asker used a levels difference in differences analysis to analyse the hotfix, which assumed that the expenditure of the treatment group would have changed by the same amount as the expenditure of the control group. Contrastingly, Professor Goldfarb used a percent approach, which assumed that the treatment group’s expenditure would have changed by the same proportion as that of the control group.

1673 Google says that Professor Asker’s use of the well-accepted and utilised levels difference in differences approach is orthodox. Let me set out some of its arguments.

1674 First, Google says that there are trade-offs associated with either approach. A trade-off of using a levels approach is that it can result in a negative counterfactual for some sources of diversion. But this is not a problem that ultimately matters for the following reasons.

1675 It does not impact estimates of diversion to the Galaxy Store, as the relevant counterfactual estimates are positive.

1676 One can bound the counterfactual to never be less than zero as Professor Asker did, and it does not change the economic substance of the results. Professor Asker explained that negative counterfactuals are a consequence of lower average spending by the treatment group than the control group and spending on Fortnite generally declining over time.

1677 Further, Google says that there are trade-offs in using the percent approach and it makes the following points.

1678 It says that when the treatment group has much lower spending than the control group, while a levels difference in differences approach could overstate the relevant change and so produce a negative counterfactual, a percent difference in differences approach could understate the relevant change.

1679 Further, it says that as a result of the proliferation of zero spending events in the relevant dataset, the percent approach cannot be implemented in a manner that allows for individual controls or any analysis where the individual is the unit of analysis.

1680 It says that Professor Asker explained how the inability to control for differences in spending patterns between recent adopters and long-time users renders the regression machinery inoperable in this environment under the percent approach.

1681 Google says that Professor Goldfarb appeared to accept that it was sensible if controlling for variables such as the duration of phone ownership. Professor Goldfarb did not control for those variables, because he did not consider them to make a meaningful difference to his conclusions about overall diversion.  So, he preferred a simpler analysis lacking appropriate controls.

1682 Google says that it is these trade-offs that explain why Professor Asker correctly viewed the two approaches as complementary for estimating the relevant parameters of interest.

1683 Second, Google says that despite not having conducted any diversion analyses of her own, Ms Edwards-Warren sought to proffer some views on this issue. She said that the parallel trend assumption, which is that the control group must be a reasonable approximation for the treatment group, does not hold for Professor Asker’s model, with the effect that his analysis cannot account for any confounding effect on users spend over time. But Google says that this contention is both wrong and irrelevant. It is said that her concerns about parallel trends did not extend to estimates of diversion to the Galaxy Store. Further, Google says that Professor Asker’s results were robust to using the levels and the percent approach, particularly with respect to the Galaxy Store, which indicated that his results were not being driven by issues with parallel trends. This is because the percent approach does not have a parallel trend assumption.

1684 Third, Google says that Epic misunderstands the parallel trend assumption. It says that there is no connection between Ms Edwards-Warren’s placebo analysis and the parallel trend assumption. Testing the parallel trend assumption requires analysing the treatment group’s and control group’s outcomes in the pre-period. Ms Edwards-Warren’s placebo test instead studied the outcomes of a different treatment group and control group in a different pre-periods.

1685 Fourth, Google says that the difference between the results of Professor Asker’s levels approach and Professor Goldfarb’s percent approach is primarily the degree of total diversion.

1686 But it says that total diversion was not used by Professor Asker in his critical loss analysis. Whilst for the purposes of that analysis Professor Asker was concerned with diversion to the Galaxy Store, which is quantitatively robust, his levels analysis highlighted the relevant importance of relative diversion to alternative distribution channels, for which there is a similar pattern under the percent approach in any event.

1687 Google says that having regard to these matters, Professor Asker’s levels approach is reliable, and indeed in some respects preferable to Professor Goldfarb’s percent approach.

1688 But in my view, Professor Asker’s difference in differences analysis of the diversion of Fortnite spending suffered from various flaws. And I agree with Epic that this caused him to overestimate the total amount of expenditure which was diverted following the Fortnite removal event and particularly diversion to consoles, and to underestimate the expenditure diverted to the Galaxy Store and Samsung.  I have accepted Epic’s position based upon the evidence of Professor Goldfarb and Ms Edwards-Warren, aspects of which I have set out.

1689 Professor Asker, in using a difference in differences analysis to estimate the rate at which Fortnite expenditure was diverted from the Play Store to other channels following the removal event, relied upon an assumption that, but for the removal, the expenditure of the treatment group, being Play Store users, would have changed by the same amount as the expenditure of the control group, being users who he considered were not affected by the removal.

1690 Contrastingly, Professor Goldfarb used a percent difference in differences approach, which assumed that the treatment group’s expenditure would have changed by the same proportion as that of the control group.

1691 Professor Asker’s approach was problematic because the average expenditure of Play Store users was substantially lower than the control group pre-removal. This meant that the decrease in control group spending from pre-removal to post-removal had an exaggerated effect on his estimate of the counterfactual expenditure of Play Store users.

1692 Professor Goldfarb illustrated this flaw by demonstrating that Professor Asker implicitly estimated that Play Store users would have had negative spending on certain channels if the removal had not occurred.

1693 Professor Asker accepted that these negative counterfactual estimates were a consequence of the treatment group’s lower average spend compared to the control group. He further acknowledged that the fact that his methodology led to negative counterfactual spending estimates was problematic, and conceded that it led to exaggerated estimates of the rate at which Play Store users’ Fortnite revenue was diverted to consoles, and to all other channels in aggregate.

1694 But he defended the use of a levels methodology to estimate diversion to the Galaxy Store on the basis that the Galaxy Store was not one of the channels for which his methodology estimated negative counterfactual spending, and that the differences between the Galaxy Store diversion estimates based on levels and percent were so small that it would give an undue sense of precision to treat them as being meaningfully different.

1695 But the fact that the Galaxy Store counterfactual spending happens to be non-negative does not mean that Professor Asker’s otherwise inappropriate methodology is reliable when applied to this channel. As Professor Goldfarb explained, the negative counterfactuals are a symptom of the problem rather than the problem itself.

1696 Further, Professor Asker’s dismissal of the significance of the differences in the Galaxy Store diversion produced by the two methodologies is unfounded, given that the results of his critical loss analysis are sensitive to small variations in input assumptions.

1697 Now Professor Asker said that a percent approach foreclosed the use of a linear regression, because this would involve dividing by zero, which would produce an undefined result. But Professor Goldfarb addressed this issue by implementing his difference in differences based on averages rather than a regression. He concluded that this difference in methodology did not impact his results.

1698 Further, in seeking to explain his reliance on the levels approach despite its calculation of negative counterfactual spending, Professor Asker argued that a percent approach could similarly produce strange results if applied in a situation where the expenditure of the treatment group was much larger than the control. Now depending upon the context, it is sometimes appropriate to measure the changes in levels and sometimes in percent. But in the present context, a percentage approach is more appropriate because expenditure by the control group is substantially greater than expenditure by the treatment group.

1699 A further matter supporting Professor Goldfarb’s percent approach relates to the parallel trend assumption, which for difference in differences analysis to be meaningful requires the control group to be a good approximation for the behaviour of the treatment group as I have previously touched on. A difference in differences analysis is used to compare outcomes for the treatment group to a control group to eliminate any potential confounding variables that affect the outcome of interest and occur at the same time.

1700 Further, Ms Edwards-Warren demonstrated that the parallel trend assumption does not hold for Professor Asker’s treatment and control groups, and as such, his analysis could not account for any confounding effect on user spend over time.

1701 Ms Edwards-Warren re-ran Professor Asker’s analysis using three randomly selected placebo dates on which there was no reason to expect Fortnite user diversion in spend. These placebo dates generated similar results to Professor Asker’s removal analysis, that is, they show high rates of diversion to consoles and lower diversion to the Galaxy Store.

1702 Professor Asker explained that the reason his regression method resulted in diversion to consoles on the placebo dates was due to underlying drift from mobile to console that underpins at least part of Epic’s business strategy. But his analysis should have controlled for this underlying drift, given that the aim of difference in differences analysis is to eliminate the effect of potentially confounding variables.

1703 But this defective choice of control group likely led to his finding of diversion from the Play Store to consoles prior to the new season release on 27 August 2020 which Professor Asker mistakenly relied upon to assert that these weeks contain information about diversion in response to the removal.

1704 Contrastingly, Ms Edwards-Warren gave evidence that the parallel trend assumption does hold for Professor Goldfarb’s analysis, on the basis that the percentage differences observed for the expenditure of the control and treatment groups track one another more closely.

1705 In my view, Professor Asker’s method is not appropriate in this case.  And I have given little weight to the estimates of diversion based on this methodology in favour of the percent difference in differences approach.

1706 Let me turn to another topic.

What is the relevant period for a proper analysis?

1707 Now Epic says that Professor Asker’s analysis is flawed because it includes the week of the hotfix, in which the hotfix occurred partway through the week, and the two weeks after the hotfix, which were before the new Fortnite season release on 27 August 2020.  And Epic says that users were unaffected by the hotfix until the new season was released two weeks later, and therefore any analysis should only begin after the new Fortnite season was released.

1708 But Google says that these criticisms are misconceived for the following reasons.

1709 First, it says that the data on user behaviour within the period immediately following the event being studied is valuable and informative as to diversion and ought not be ignored. So, it is appropriate and relevant to include the period immediately following the hotfix.

1710 Second, Google says that Epic’s claim that users were unaffected by the hotfix is incorrect. It says that when Fortnite was removed from the Play Store, the Play Store users immediately lost the ability to download or update Fortnite via the Play Store and lost the ability to transact with Epic via Google Play Billing. This meant that they lost access to a transaction venue because the Play Store no longer facilitated sales of Fortnite in-app content, that is, it did not guarantee delivery of content, push updates, and facilitate discovery. The only way in which they could make purchases on Fortnite in the version of the app they had downloaded from the Play Store was via EDP. Professor Asker characterised this initial change due to the hotfix as a degradation in quality.

1711 Third, Google says that the raw data shows that the hotfix had an immediate effect on the Play Store user spending.

1712 Professor Asker analysed the spending of treatment group users, that is, those who spent via the Play Store version of Fortnite, before and after the hotfix. He found that spending via Google Play Billing went immediately to zero after the hotfix, which was consistent with these users losing access to Google Play Billing on their app, from an average of $9 per week in the week prior to the hotfix, but that there was not a commensurate increase in spending via EDP after the hotfix.

1713 Google says that what this means is that users experienced an immediate quality degradation in Fortnite that was borne out in their spending.  So, the appropriate start date to analyse diversion from the hotfix is the date of the hotfix.

1714 Google says that this reasoning is no different from evaluating the effects of a medical trial starting on the date that users receive medication. If users had adverse effects for the first two weeks after the start of the trial, and then positive effects thereafter, no scientifically valid study design would remove the first two weeks of trial as that would involve cherry-picking the data.

1715 Google says that Epic’s assertion that the two weeks after the hotfix should be removed from the analysis is similarly unsupportable.

1716 Fourth, Google says that as to Epic’s claim that Professor Asker’s inclusion of the week of the hotfix creates results that are biased in favour of his ultimate conclusions, Epic misrepresents the direction of this alleged bias.

1717 Google says that Professor Goldfarb’s analysis shows why excluding the week of the hotfix increases the diversion for some venues and decreases the diversion for other venues. But, overall, it has a minimal impact on the estimates, as Professor Asker stated.

1718 Similarly, Google says that Professor Asker described how diversion to the Galaxy Store increased after the new season was released, but that this effect was short-lived and dissipated over time.

1719 Consistently with this, Google says that Professor Goldfarb’s results revealed similar diversion to the Galaxy Store when using a 4-week period, including the two weeks after the hotfix, and when using a 26- or 52-week period, regardless of whether the two weeks after the hotfix were included.

1720 This means that, according to Google, Professor Asker’s results did not represent diversion during some outlier period, but rather captured typical diversion patterns for Fortnite users.

1721 Instead, Professor Goldfarb’s choice to eliminate the first two weeks after the hotfix, and focus on a time window when diversion to the Galaxy Store was artificially inflated, meant that his estimates captured atypical diversion patterns.

1722 Further and as I have said above, when analysing the hotfix, Professor Asker used a shorter period, being 4 weeks before the hotfix and 5 weeks after, while Professor Goldfarb analysed it using a longer period, being 13 weeks before the hotfix, and 13, 26, or 52 weeks after.

1723 Google says that Professor Asker’s approach was informed by his analysis of spending patterns of Fortnite, which indicated that a shorter period was more appropriate given the high churn in the Fortnite user base.

1724 Further, Google says that Professor Asker’s shorter period should be preferred for the following reasons.

1725 First, it says that Professor Goldfarb ultimately accepted in cross-examination that his preferred specifications were the longer ones, being 26- and 52-weeks. So it says that the 13-week specification should not be preferred over the other specifications.

1726 Second, it says that Professor Goldfarb sought to support his asserted preference for a post-hotfix period of at least 13 weeks by saying that it was the longest analysis period that had a symmetric pre- and post-period length.

1727 But it says that that symmetry alone does not justify a longer time horizon relative to a shorter one. Moreover, there is no econometric reason to prefer a specification with symmetric pre- and post-periods, consistent with Professor Goldfarb’s statements in the joint expert report and during his evidence in the Apple proceeding that he prefers a 26- and 52-week post-period.

1728 Third, Google says that there are challenges in drawing a causal inference from a single event, being the hotfix, over longer time periods because of confounding factors. There can be other differential shocks to the treatment group and control group in the post-period.

1729 Google says that this means that the Play Store users’ substitution many weeks after the hotfix may reflect a variety of other factors, whereas the Play Store users’ substitution immediately after the hotfix can likely be attributable to the effect of the hotfix itself, after controlling for pre-existing trends, as the difference in differences method does.

1730 Additionally, Google says that whilst as a general matter a shorter time period such as 4 weeks is more appropriate for analysing the hotfix, Professor Asker also explained that he was less bothered by the longer time periods. Rather, the more significant area of difference between him and Professor Goldfarb pertained to omitting the weeks immediately after the hotfix.

1731 But again, I have accepted the position of Epic and its experts.

1732 In my view, Professor Asker’s estimates of diversion are also based on an inappropriate time period which caused him to underestimate the rate of diversion to the Galaxy Store.

1733 Professor Asker primarily relied upon the diversion of spending by Fortnite users following the removal to draw conclusions about the extent to which users would divert their revenue to the Galaxy Store in response to a SSNIP on the prices of apps and in-app content on the Play Store. With this purpose in mind, the five-week post-period Professor Asker relied upon for his analysis is inappropriate for two reasons.

1734 First, Professor Asker’s post-period started on 10 August 2020, three days prior to the Fortnite hotfix.

1735 Second, Professor Asker’s post-period included two weeks between the Fortnite hotfix on 13 August and the Fortnite new season launch on 27 August 2020.

1736 All of Professor Asker’s difference in differences estimates of diversion included the week of 10 August 2020 in the post-period. Professor Asker’s inclusion of the week of 10 August 2020 in his post-period was inappropriate because it included several days which were unambiguously prior to the removal.

1737 Professor Asker indicated that he did not have a strong view as to whether this week should be removed from his analysis but claimed that its inclusion would have resulted in the calculation of lower diversion ratios and so was conservative for his purposes.

1738 But as Professor Goldfarb pointed out, Professor Asker primarily relied on his analysis to show that diversion to the Galaxy Store was below his critical loss threshold. As such, any dampening of his diversion estimate would have been biased in favour of broadening the market.

1739 Additionally, Professor Goldfarb found that including the week of 10 August 2020 from the post-period produced higher estimates of total diversion and lower estimates of diversion to the Galaxy Store.

1740 So, Professor Asker’s inclusion of this week produced results that were biased in favour of his ultimate conclusions. There was no sound basis for including the week of 10 August 2020.

1741 Further, given that Professor Asker primarily relied upon this analysis for estimating diversion to the Galaxy Store, the inclusion of the following two weeks of 17 and 24 August 2020 prior to the season release was also inappropriate.

1742 Following removal, users of Fortnite on the Play Store experienced the following changes to their experience.

1743 From 13 August 2020, they were given the choice between buying V-bucks through Google Play Billing for the usual price or using EDP for a 20% discount. This discount was also available to users who bought V-bucks in Fortnite on other platforms.

1744 By 18 August 2020 at the latest, the option to pay with Google Play Billing was removed.  Users could continue to spend through Play Store Fortnite using EDP.

1745 On 27 August 2020, Fortnite was updated to a new season. This update could not be downloaded on the version of Fortnite downloaded from the Play Store, meaning that users could not access the new season’s content, or play with users on other platforms.

1746 So, it is apparent that an assessment of the diversion of revenue from the Play Store to the Galaxy Store should focus on the change in expenditure following the season release, as until this point, the experience of playing and spending in Fortnite downloaded from the Play Store was largely unchanged. To the extent that there were changes to the Play Store version of Fortnite on 13 August 2020, Professor Asker acknowledged that these changes were smaller than those caused by the season release.

1747 In particular, prior to the season release on 27 August 2020, Play Store users would have had no incentive to divert any spending to the Galaxy Store. Professor Asker’s analysis shows that prior to the hotfix, only in the order of 1% of Play Store Fortnite users’ expenditure was through the Galaxy Store.  After the hotfix on 13 August 2020 but before the season release on 27 August 2020, Play Store users could still play their existing Play Store version of Fortnite on their Android device and could make purchases through EDP through their existing Play Store version of the app.

1748 The idea that Play Store users would have diverted spending to the Galaxy Store prior to 27 August 2020 is premised on the notion that between 13 August and 27 August 2020, rather than using EDP to make purchases within their existing Play Store version of Fortnite, the Play Store users would have chosen instead to undertake the additional steps of downloading the same version of the Fortnite app via the Galaxy Store onto their Android device, setting up Samsung Pay and making a Fortnite purchase through Samsung Pay.

1749 But the frictions involved in re-downloading the same app on the same device and the lack of any corresponding benefit in doing so point against any appreciable diversion to the Galaxy Store prior to 27 August 2020. This logic is borne out by two findings.

1750 First, Professor Asker’s finding that Fortnite user expenditure on the Galaxy Store in Australia was largely consistent for three weeks before increasing significantly on the day of the season release.

1751 Second, Professor Goldfarb’s finding that the rate of diversion to the Galaxy Store increases when the two weeks following removal are excluded from the analysis.

1752 Professor Asker’s argument that there was some observed diversion in the period after Google Play Billing ceased and 27 August 2020 misses the point. What is observed in that period is diversion to channels other than the Galaxy Store.

1753 Relevantly for Professor Asker’s analysis, during the period after Google Play Billing ceased and prior to the season release, users may have had an incentive to divert their spending to channels they already multi-homed on, because of the reduced frictions in using their existing payment channel on another device rather than setting up EDP. But as the data bears out, there was limited multi-homing prior to the hotfix between the Play Store and the Galaxy Store.

1754 The consequence is that in any analysis focused on diversion to the Galaxy Store, the inclusion of the weeks of 17 and 24 August 2020 is inappropriate because it will result in unreliably low estimates of diversion.

1755 It is particularly unreliable if one uses, as Professor Asker did, a 5-week post-treatment period in which three of the five weeks are purporting to measure an effect which has not yet occurred.

1756 Professor Goldfarb explained that removing these two weeks from the post-period is critical when relying on a 4-week analysis window. For 13, 26 and 52-week analysis windows, the effect of the inclusion of these two weeks is attenuated.

1757 Now Professor Asker criticised Professor Goldfarb’s preference for longer periods of analysis on the basis that Fortnite users generally exhibit a high level of churn, meaning that assessing substitution over a longer period is inappropriate, and an analysis based on longer periods is likely to be affected by factors other than the removal.

1758 Professor Asker’s preference for a 4-week period was based on a churn analysis by him suggesting that 70% of consecutive purchases on Fortnite on the Play Store and Android were less than 4 weeks apart.

1759 On the other hand, as I have already said, Professor Goldfarb assessed diversion by reference to a 4-week post-period compared to a 4-week pre-period, as well as 13, 26, and 52-week post-periods compared to a 13-week pre-period. He indicated that extending the analysis to longer, that is, 13, 26 or 52-week time periods was preferable because, for analyses which rely upon data before the season release, that is, the weeks of 17 and 24 August 2020, using a longer period of time attenuates the effect of the ambiguity associated with this period. He also explained that considering a longer window allows for more balance across Fortnite seasons. As to this point, because Fortnite was only made available in April 2020, 13 weeks was the longest period assessed which allowed for balanced pre- and post- periods.

1760 Consistently with these considerations, Professor Goldfarb considered his estimate of diversion based on a 13-week pre- and post-period excluding the weeks prior to the season release to be the most reliable.

1761 Further support for the 13-week period is provided by Professor Goldfarb’s reasoning and his correction to Professor Asker’s churn analysis.

1762 Professor Goldfarb accepted that the typical interval between purchases helps identify an appropriate window of analysis. But Professor Goldfarb identified that Professor Asker’s churn analysis was not considering the average time between purchases on a user level, and that Professor Asker’s result that 70% of consecutive purchases were less than 4 weeks apart could be biased by a small number of users making a large number of repeated purchases.

1763 Professor Goldfarb identified that when purchases are considered on a user level basis, in fact 70% of users had an average period of around 10 weeks between consecutive purchases, rather than 4 weeks.

1764 Accordingly, applying Professor Asker’s criterion would lead to the conclusion that a time window in the order of 10 weeks is an appropriate window of analysis to consider diversion. That gives further support to the symmetric 13-week periods selected by Professor Goldfarb.

1765 Further, Google did not attempt to engage with why, when measuring diversion to the Galaxy Store and Samsung, the weeks of 17 and 24 August 2020 should not be included.

1766 Further, there is nothing inconsistent with the proposition that nothing very much happened in the two weeks following the hotfix, but that effective removal after 27 August 2020 was an extreme event.

1767 Let me turn to another topic.

Professor Goldfarb’s diversion estimates

1768 Now Google says that Professor Goldfarb said in his opening statement that he preferred the yellow shaded columns of his table 11.A, save for the 4 week specification, and all specifications in table 11.B.

1769 And according to Google, in cross-examination, Professor Goldfarb’s position became further refined.

1770 It is said that he conceded that he was not putting forward any 4 week specification as preferred, even in table 11.B.  Further, it is said that he accepted that the 13 week specification was not preferred to the 26 or 52 week specification if the first two weeks after the hotfix are included, that is, the 13 week specification in table 11.A was not preferred.

1771 As to the 13 week specification in table 11.B, which excludes the first two weeks after the hotfix, Google say that Professor Goldfarb said he was more comfortable with this when those two weeks are removed.  But Google says that Professor Goldfarb also accepted that he viewed 13 weeks as a short period, had said in the joint report that he preferred longer periods than 13 weeks, and had described 52 weeks as the preferred time period in his Apple reply report.

1772 So, in context, Google says that when Professor Goldfarb said he was comfortable with the 13 week specification in table 11.B, he was not saying this was preferred.

1773 Google says that the effect of Professor Goldfarb’s evidence was that his preferred specifications were the 26 and 52 week specifications in tables 11.A and 11.B.  And so Google says that the result is that Professor Asker and Professor Goldfarb were in agreement that I should at least rely on the 26 and 52 week specifications in tables 11.A and 11.B.

1774 Google says that this has implications for the market definition analysis.

1775 It shows that the diversion ratio in spending to the Galaxy Store ranged between 10.1% and 12.1%, much lower than the 27.8% to 34.9% diversion ratio to non-Android sources, and lower than the diversion to Xbox or PlayStation on the 52 week specification.

1776 In other words, the regression analysis indicates that non-Android platforms, that is, app stores on gaming consoles and PCs, are closer substitutes than other Android app stores.

1777 But I do not accept Google’s position on this aspect. Again, I have accepted the position of Epic and Professor Goldfarb.

1778 Initially, Professor Goldfarb said that in the context of estimating diversion to the Galaxy Store and Samsung, he was comfortable with all of the estimates in table 11.B at 4, 13, 26 and 52 weeks, giving a range of approximately 10 to 18% and that his most preferred estimate was the 13-week estimate of 15.8%.  He explained that the 13-week estimate was best because it had an equal number of weeks on either side with one new Fortnite season on either side.

1779 In cross-examination, Professor Goldfarb also said that each of the 4-, 13-, 26- and 52-week estimates in table 11.B provided a reasonable range of estimates. Professor Goldfarb again said that the 13-week estimate was the best estimate.

1780 Now it was put to Professor Goldfarb that he had stated a preference in the joint expert report for 26- or 52-week estimates. But as  Epic says, Professor Goldfarb indicated that that was in a particular context, being in circumstances where the two weeks commencing 17 and 24 August 2020 were included, and in the context where Professor Goldfarb was focusing his attention on overall diversion rates.

1781 As Epic says, his evidence was that when you did not have those “ambiguous” two weeks, he was more comfortable with a 13-week analysis. This is consistent with his earlier evidence that the 13-week specification was the most preferred.

1782 I agree with Epic that in circumstances where Professor Goldfarb’s evidence was never challenged directly on that aspect, Google’s assertion that Professor Goldfarb’s preferred specification was for 26 or 52 weeks is not open.

1783 In any event, a symmetric period is preferable given that it avoids confounding events by having the release of one new Fortnite season on either side of the event.  And as Professor Goldfarb explained, in the Apple proceeding Professor Goldfarb had additional data so that symmetric 26 and 52-week pre and post periods could be constructed.

1784 Further, a 13-week estimate is an appropriate time period having regard to gross churn rates.

1785 Finally, it is not correct that Epic has sought to just fasten on the highest of Professor Goldfarb’s Galaxy Store diversion estimates. The highest estimate in table 11.B was 18.7%.  The 13-week estimate was 15.8%.

Conclusions

1786 Professor Asker said that he considered the best estimates to be those recorded in exhibit 158 of his first report, table 11.A in Professor Goldfarb’s reply report, and the 26- and 52-week specifications in table 11.B in Professor Goldfarb’s reply report.

1787 Professor Goldfarb said that he considered that the best estimates were the 13-, 26-, and 52-week specifications in table 11.A and all specifications in table 11.B of his reply report.

1788 Now Google made the following points concerning its preferred diversion estimates based on different regression specifications.

1789 First, it says that diversion to the Galaxy Store ranges from 9.4% to 13.2% across all specifications, which means that Epic’s candidate market fails the HMT on the 14.5% critical loss threshold calculated by Professor Asker.  It says that this conclusion remains true when much lower pass-through rates are used.

1790 Second, it says that diversion to a console transaction venue, particularly PlayStation and Xbox, is equal to or larger than diversion to the Galaxy Store in 5 out of the 8 specifications.

1791 Third, it says that diversion to a console transaction venue, particularly PlayStation and Xbox, is larger than diversion to sideloading in every specification.

1792 Fourth, it says that diversion to non-Android venues is larger than diversion to Android venues in every specification.

1793 But in my view and having regard to the evidence, the estimates of diversion which are based on a percent difference in differences are more reliable than those using levels.  Further, I am satisfied that the inclusion of the three weeks of 10, 17 and 24 August 2020 is inappropriate for an analysis which attempts to calculate diversion to the Galaxy Store, particularly when relying on a more time limited window.

1794 Finally, I have given the most weight to the Professor Goldfarb’s estimate of diversion based on an analysis window of 13 weeks, as this window provides for the most balance across Fortnite seasons and seasonal effects.

1795 So, I accept Professor Goldfarb’s evidence of the following matters.

1796 First, the most reliable estimates of diversion are the highlighted estimates for 13 weeks in table 11.B of Professor Goldfarb’s reply report, which calculate that 15.8% of revenue was diverted to the Galaxy Store.  I have already set out tables 11.A and 11.B.

1797 Second, beyond this, a range of reliable diversion estimates is provided by the highlighted estimates for 13, 26 and 52 weeks in table 11.A, and all the highlighted estimates in table 11.B, indicating that between 10.1 to 18.7% of revenue was diverted to the Galaxy Store.

1798 The results in table 11.B also show that approximately 50% of all revenue was lost following removal, that is, they range from 45.8 to 54.7%.

1799 Further, they show that there was substantially less diversion to consoles than is indicated by Professor Asker’s estimates.

1800 In summary then, Professor Asker’s conclusion that consoles are closer competitors to the Play Store than other Android app stores is not supported by the data.

Other matters

1801 Professor Asker said that the key insight from his Fortnite analysis was not just that consoles were a constraint, but that developer substitution possibilities to non-Android venues were meaningful and to some extent constrained the Play Store.

1802 But as Epic points out, there is an inherent asymmetry in the extent to which quantitative analysis of Fortnite and the Fortnite removal event can inform market definition. That asymmetry has two aspects as Epic correctly described.

1803 As Epic explained, the first aspect of the asymmetry relates to relative rates of substitution between Android apps and apps on non-mobile devices. Because Fortnite is an app that has an unusually high degree of cross-platform functionality, a conclusion that Fortnite users might substitute from Fortnite on Android to Fortnite on other devices says very little about substitution in relation to other apps where cross-platform functionality is limited. Contrastingly, if the analysis shows that even users of Fortnite are unlikely to substitute to other devices, then a fortiori the same result will hold for apps with more limited cross-platform functionality.

1804 In relation to this first aspect, Fortnite’s cross-platform functionality is unusually broad, even among games apps. In order for users to be able to substitute across platforms, in particular from Android to games consoles, games must be available on both platforms.

1805 Professor Asker’s exhibit 55 shows the top 25 games apps by 2021 Australian Play Store consumer spending. The exhibit focused on cross-wallet and cross-progression functionality across platforms. Having cross-wallet and cross-progression functionality across platforms permits spending and play across different platforms, but only across those platforms on which the game is available.

1806 As Epic pointed out, figure 13 of Ms Edwards-Warren’s reply report compared the availability of Fortnite with the 25 apps from Professor Asker exhibit 55. It showed that at the time of the removal, being 13 August 2020, Fortnite was available on Android, iOS, PlayStation, Xbox, Switch and PC. As of 29 February 2024, Fortnite was no longer available on iOS but was available on streaming services in addition to Android, PlayStation, Xbox and Switch.

1807 Of the top 25 gaming apps on the Play Store, only two out of 25 are available on console: PlayStation and/or Xbox. None are available on Switch.

1808 Five of the top 25 gaming apps on the Play Store are available on mobile platforms only: Android and iOS. Their users have no substitution options to console, and no substitution options to PC. These five games account for 20% of the top 25 games apps and [REDACTED] of Australian consumer spend on in-app purchases of digital content across the 25 game apps in 2021.

1809 The remaining 18 out of the top 25 are available on at least one PC platform as well as mobile platforms, being Android and iOS.

1810 So, as Epic said, Fortnite results with respect to substitution from mobile to console generalises to potentially only two out of the 25 top games apps on the Play Store.  I agree with Epic that with such limited cross-platform functionality for the highest revenue games, conclusions with respect to diversion to consoles cannot be extended to apply to games apps more broadly or to non-games apps, except insofar as the Fortnite results show limited diversion.

1811 Further, as Epic pointed out, the second aspect of the asymmetry relates to absolute levels of substitution in light of the Fortnite removal event.  Given that the Fortnite removal event is not a SSNIP, and is analogous to an infinite price increase, this entails that absolute diversion levels can only be informative in one direction.

1812 Low levels of substitution in response to complete removal logically suggests even lower levels of substitution in the case of a SSNIP. Contrastingly, even if there were high levels of substitution in response to complete removal, this would say nothing about the absolute level of substitution in the case of a SSNIP.

1813 In other words, Fortnite data and the Fortnite removal event can really only provide evidence in one direction.  For Epic, quantitative analysis of Fortnite can provide useful positive evidence in support of Epic’s proposed markets. But for Google the evidence is neutral.

1814 Now as I have already indicated, Professor Asker relied upon his calculation that following the removal event Australian Play Store users divert approximately 80% of their spending to other platforms, to conclude that other transaction venues are substitutes for the Play Store.

1815 But as I have already discussed, this was based on Professor Asker’s inappropriate levels difference in differences methodology, which produced much higher levels of diversion to consoles and other platforms overall.  The correct figure on Professor Goldfarb’s analysis, which I have accepted, was that approximately half of spending (45.8 to 54.7%) was lost following removal.

1816 But I agree with Epic that even if Professor Asker’s figure was right, it could not support any conclusion that non-mobile platforms were substitutes for Android apps in response to a SSNIP.

Logit demand model, transition probability analysis and counting of diversion estimates

1817 Now the purpose of the logit demand model is to generate diversion ratios, and the purpose of diversion ratios is to inform substitution, which is the basis for defining relevant markets.

1818 As Google explained, the logit demand model translates observed consumer choices among a viable choice set into estimates of diversion and substitution. Indeed Google went so far as to say that it is an economically rigorous way to perform that translation where the assumptions needed to draw inferences are made explicit.

1819 Now the architecture of the logit demand model means that diversion is proportional to spending shares. If consumers choose one transaction venue more than another out of the available choice set, the logit model says that there will be more diversion to that transaction venue.

1820 And as Mr Moore SC explained, the logit demand model formalises an underlying economic intuition, being that the realised choices consumers make among a set of possible alternatives is informative as to properties of consumer demand like substitution.  Indeed, usually such revealed preference analysis is preferable to an analysis of hypothetical choices such as may be the subject of a survey.

1821 Now Professor Asker used the logit demand model in the present context to derive a measurement of diversion to the iOS App Store which was not possible from available natural experiments like the hotfix.  Diversion estimates from Professor Asker’s logit model exercise were said by Google to corroborate diversion estimates from the analysis of the hotfix.

1822 In Professor Asker’s view diversion to non-Android transaction venues was greater than diversion to other Android transaction venues. In his view this also indicated that diversion to the iOS App Store was quantitatively meaningful.

1823 Professor Asker examined how Fortnite users that downloaded Fortnite from the Play Store spend on Fortnite. These are users for whom the Play Store is a demonstrably viable transaction venue. As is obvious, if they choose not to spend via the Play Store that means that they may instead be allocating their spending via a different transaction venue, such that the Play Store is losing out on potential business to that alternate transaction venue.

1824 Now Professor Asker presented the data on how Fortnite Play Store users spend on Fortnite. In his view, what the data reveals is that Fortnite Play Store users overwhelmingly spend outside the Play Store, and indeed primarily spend via non-Android venues like console stores.

1825 But, in summary, I agree with Epic that the logit model put forward by Professor Asker has some problematic features.

1826 First, it does not provide new evidence additional to spending data.

1827 Second, it is relied upon by Professor Asker to support his conclusions about market definition but it is itself driven by assumptions regarding the boundaries of the market.

1828 Third, the independence of irrelevant alternative (IIA) assumption, an essential technical condition for relying upon a logit model, does not hold.

1829 Now the logit model allocates diversion shares proportionately to market share.  Moreover, it was not in doubt that most user spending on Fortnite occurs on consoles and PCs.

1830 Professor Asker showed that the majority of the spending of Fortnite users who downloaded Fortnite on Android smart mobile devices occurs via consoles, with consoles accounting for 75.6% of the user spending of all Fortnite users on Android smart mobile devices in Australia between 20 April 2020 to 9 August 2020.  The equivalent figure for Fortnite users downloading Fortnite via the Play Store is 71.5%.

1831 But given that the logit model represents spending shares, as Epic rightly said it is unsurprising that when consoles are assumed to be in the market, one finds diversion to consoles.

1832 Accordingly, in my view the logit model as applied by Professor Asker is not informative about market definition in this case.  It does not independently provide information about market definition.  As Epic pithily put it, it simply represents market shares in a claimed market, the boundaries of which have been assumed. By assumption, it generates more diversion to devices with higher shares of spending such as PlayStation.

Professor Asker’s evidence

1833 Let me begin by setting out some evidence of Professor Asker.

1834 The logistic model, typically referred to as the logit model, is a popular model in modern industrial organisation economics. It belongs to a group of economic models called discrete choice models which economists use to explain how agents choose one option from a set of alternatives, often referred to as the choice set. So, an economist can use the logit model to analyse purchase data from the cereal aisle of a grocery store, including the prices, number of boxes bought, and characteristics about each cereal available, to estimate how price sensitive consumers are and the substitution patterns between cereals.

1835 Choice sets must exhibit several properties for a discrete choice model to be appropriate for a given setting.  First, the alternatives must be mutually exclusive, choosing one option means that none of the other alternatives are being chosen. Second, the choice set must include all relevant alternatives to the decision maker. Last, there must be a finite number of alternatives.  In practice, economists include an alternative referred to as the “outside option” in the choice set. The outside option typically represents the option not to purchase any available alternative.  So, it can represent a consumer choosing to not to spend their money or purchase another option not in the choice set.

1836 Now the logit model assumes that the utility provided to consumer m, from buying product i, takes the form:

1837 The index i =1..., J indexes items in the choice set. J is the total number of products in the choice set. x_j typically represents qualities of the product that provide utility to the consumers. p_j represents the price the consumer pays for product j.ε_i^m represents any taste consumer m has for good i that is not explained by good ’s characteristics and price.

1838 The coefficients α and β represent the taste for each characteristic and price sensitivity, respectively. It is implicitly assumed, through this utility function, that consumers only differ from each other through the ε_i^m term.  The logit model explicitly assumes that all ε_i^m terms are independent of each other. In other words, any additional taste beyond what is explained by xi and pi that consumer m has about good i is unrelated to their additional taste for good j and consumer m’s additional taste for any good is unrelated to all other consumers additional taste for that good.

1839 Because economists do not typically have information about ε_i^m, they must make assumptions about its distribution amongst individuals in the market. This distributional assumption has allowed economists to derive straightforward mathematical equations from the model, which is a large reason for the logit model’s popularity.

1840 Now the formulae for diversion ratios and price elasticity of demand, which are commonly used concepts in a quantitative application of the HMT, are as follows.

1841 The diversion ratio from product i to product j under the logit model is:

1842 The cross-price elasticity from product i to product j under the logit model is:

where sj represents the choice share of good j and α is the same parameter from the utility function defined above.

1843 In the above formulations, the share of good j (s_j) represents how often good  j is chosen, which can be measured using metrics such as the share of consumers that choose good j, when estimating the model on the aggregate level, or the probability that a given consumer  chooses good  j, when estimating the model on an individual level.

1844 Now the logit model can be written down in terms of market shares, which of course reflect the proportion of individuals that choose a particular product.  This can be aligned with the likelihood that an individual chooses a particular product.  Such a likelihood multiplied by the total number of consumers can provide a measure of the demand.  So, demand and market shares are related.

1845 Now another implication of the logit assumptions is the IIA property.

1846 The IIA assumption means that if a new alternative is introduced or removed from a choice set, the relative probability of choosing good j compared to good k, that is, the ratio of those probabilities, does not change. In other words, under the IIA assumption in a logit model, if consumers lose or gain access to one alternative, then the ratio of shares for all other products remains the same.

1847 Although this property implies substitution patterns that may not be realistic in some scenarios, it also gives economists a way to test if the logit model is appropriate in any given setting. In the economics literature, the substitution pattern that the IIA assumption leads to is known as the red bus – blue bus problem.  I will return to this later.

1848 If the data shows that the ratios of choice shares of the existing or remaining alternatives after entry or exit of an alternative are relatively similar, then the logit model is consistent with observed patterns of substitution. Professor Asker tested the IIA assumption in the context of Fortnite spending.

1849 Now the logit model relates diversion ratios to choice shares, but calculating a diversion ratio requires knowing the choice share on the outside option.

1850 But when applying the logit model to the Fortnite data, Professor Asker could not observe spending by Fortnite users on any outside option. He only observed their spending on Fortnite. So, he did not observe the full choice shares of Fortnite users because in the logit model any choice set includes the outside option. Instead, he could only observe the choice shares within Fortnite spending.

1851 Now Professor Asker estimated diversion ratios based on his difference-in-differences regression, which were based on the change in spending amounts and could therefore be estimated as standard diversion ratios.  Put differently, he only observed a normalised choice share that is the relative share excluding the outside option:

1852 In the above formula, s_i refers to the share of product i, s₀ refers to the actual share for the outside option, and ŝ_i refers the normalised share of product i.

1853 Professor Asker purported to show that under a logit model, the ratios of normalised shares that he did observe corresponded to a normalisation of diversion ratios.  He put forward the following analysis.

1854 The diversion ratio from product i to j implied by the logit model is:

1855 For convenience he denoted:

1856 Because this formula depends on the share spent on the outside option, he could not estimate it from the data he analysed. Instead, one can estimate a normalisation of diversion ratios, which eliminates the denominator.

1857 So, he let J represent the set of transaction venues, that is, all options excluding the outside option.  And he let J ⁄ i represent the set of transaction venues excluding i. Total diversion from i to all the inside options is then:

1858 With this, one can estimate, , the normalised diversion ratio from i to j as:

1859 He said that the normalised diversion ratios from product i to another product j are like the standard logit model diversion formula but using their normalised shares, which he observed in the data.

1860 He said that the interpretation of the normalised diversion ratio addresses the following question: if a firm increases the price of i, of the customers that leave but do not choose the outside option, what percentage will go to product j?  He said that this was informative about diversion within the inside options but could not be used to infer the extent of diversion to the outside option.

1861 So Professor Asker’s position was that if he examined spending by Fortnite Play Store users across transaction channels, he could infer normalised diversion ratios by using the distribution of their spending across transaction channels.

Ms Edwards-Warren’s evidence

1862 Let me at this point set out some evidence given by Ms Edwards-Warren.

1863 Ms Edwards-Warren considered that the logit diversion ratios calculated by Professor Asker were not informative about how Fortnite users of the Play Store substitute to other platforms, and hence about which platforms are in the same anti-trust market as the Play Store.

1864 She explained the following matters.

1865 First, she explained how the normalised diversion ratios implied by the logit demand model presented by Professor Asker have been calculated using only data on the expenditure shares of all platforms.

1866 Second, she explained that the logit diversion ratios calculated using these shares of expenditure can only be interpreted as actual diversion ratios, that is, diversion ratios that reflect the actual substitute patterns of Fortnite users, due to Professor Asker assuming that all platforms are substitutes and that Fortnite users substitute to all of these platforms in the restrictive manner assumed by the logit demand model.

1867 Third, she explained why the test of the IIA assumption of Professor Asker is not a reliable test of the IIA assumption made by the logit demand model, and why the IIA assumption does not hold.

1868 Further, her position was that although Professor Asker calculated the normalised diversion ratios implied by the logit demand model, these logit diversion ratios merely reflect the restrictive assumptions about Fortnite users’ behaviour made by the logit demand model used by Professor Asker, and so as a consequence, these normalised logit diversion ratios provide no evidence as to which platforms Fortnite users consider as substitutes for the Play Store as a platform, and so no evidence on which of these other platforms are in the same market as the Play Store.

1869 Let me say something more about Ms Edwards-Warren’s evidence as to how the logit diversion ratios have been calculated.

1870 In her view, Professor Asker’s logit diversion ratios are simply the result of assuming that Fortnite users substitute to platforms that Professor Asker assumed are substitutes for the Play Store, in the way assumed by the logit model.

1871 The logit demand modelling is essentially a transformation of the Fortnite spend data discussed by Professor Asker. The logit model of Professor Asker estimated diversion ratios from Fortnite spend on the Play Store to other transaction venues by assuming that users switch to other platforms in proportion to each platform’s market share of the assumed market.

1872 Now she articulated the steps in this analysis to illustrate what was occurring.

1873 The logit analysis of Professor Asker started with the assumption that the Play Store competes in a market that includes console stores, the Apple App Store, the Samsung Galaxy Store, and PC stores.  And it assumed a high proportion of Fortnite spend occurred on consoles (71.5%) during the period of analysis.  The analysis found that most switching from the Play Store goes to console stores.  This is because the analysis assumed that switching to alternatives occurs in proportion to market shares.

1874 Further, it assumed that a higher proportion of Fortnite spending occurred on the Apple App Store than on the Samsung Galaxy Store during the period of analysis. And on this basis, the analysis found that users are more likely to switch from the Play Store to the Apple App Store than to the Samsung Galaxy Store.

1875 Further, it assumed that a higher proportion of Fortnite spending occurred on PC stores than on the Samsung Galaxy Store during the period of analysis. On this basis, the analysis found that users are more likely to switch from the Play Store to PC stores than to the Samsung Galaxy Store.

1876 But according to Ms Edwards-Warren, this assumption driven analysis is circular.  The analysis conveys no information about market definition that has not been pre-assumed.  I have no doubt that she would agree that it is about as useful for real world analysis as an analytic statement is in philosophy where the predicate is necessarily and wholly included within the definition of the subject.

1877 As a consequence, she said that these logit diversion ratios provide no evidence as to which platforms Fortnite users who use the Play Store actually consider to be the largest and closest substitutes for the Play Store as a platform.

1878 She said that they can therefore provide no independent evidence as to which platforms are in the same relevant market as the Play Store because the answers are driven by the starting assumptions.

1879 Let me say something further about her views on the normalised diversion ratios of Professor Asker.

1880 As I have said above, Professor Asker presented normalised logit diversion ratios, which rely on the assumption that the logit demand model accurately captures how Fortnite users would substitute to other platforms when purchasing in-app content, particularly those Fortnite users who used the Play Store to purchase in-app content.

1881 As Ms Edwards-Warren said, the logit diversion ratios of Professor Asker were normalised diversion ratios because they did not take account of the possibility that some Fortnite users might substitute to an outside product in response to an increase in the costs of purchasing in-app content via a platform.

1882 Professor Asker explained that due to the restrictive assumptions about consumers’ preferences for substitute products made by the logit demand model, it is possible to calculate diversion ratios using only data on market shares. Professor Asker used each platform’s share of total expenditure of Fortnite users’ expenditure on in-app content to calculate the diversion ratios from the Play Store to each of the other platforms.

1883 Now the implied logit diversion ratios were calculated using the expenditure shares as follows:

where (𝐸𝑆_PS) is the expenditure share of the Play Store (𝑃𝑆) and (𝐸𝑆i) is the expenditure share of platform 𝑖.

1884 But as she said, Professor Asker did not estimate these normalised logit diversion ratios econometrically, and did not econometrically test the restrictive assumptions made by the logit demand model.

1885 Now using data for the period during which Fortnite was available on the Play Store, Mr Edwards-Warren’s table 30 presented the expenditure shares for each of the platforms and the implied logit diversion ratios calculated using them, being the implied logit diversion ratios shown in Professor Asker’s exhibit 64.

1886 Ms Edwards-Warren’s table 31 presented equivalent figures but used the expenditure shares for the four-week period before the Fortnite removal event. These are shown with the logit diversion ratios shown in Professor Asker’s exhibit 159.

1887 As the logit diversion ratios of Professor Asker are calculated using each platform’s share of Fortnite users’ total expenditure on in-app content, these diversion ratios merely reflect the relative importance of each platform in terms of this expenditure.

1888 Given this, Ms Edwards-Warren said that it is not surprising that PlayStation and Xbox have the largest diversion ratios since these platforms have the largest shares of total expenditure.

1889 Let me set out some further points made by Ms Edwards-Warren.

1890 Professor Asker interpreted these normalised logit diversion ratios as actual diversion ratios, because he assumed that Fortnite users who purchase in-app content, particularly those Fortnite users who used the Play Store to purchase in-app content, divert to assumed substitution platforms in the restrictive manner assumed by the logit demand model.

1891 Now Ms Edwards-Warren said that the logit demand model makes two key assumptions about consumers’ preferences in order for diversion ratios to be estimated using only data on market shares.

1892 First, it assumed that consumers consider all the products used to calculate the market shares as substitutes for each other.

1893 Second, it assumed that consumers’ preferences for these substitute products are such that consumers substitute to these products in proportion to these products’ shares of expenditure. This is the IIA assumption.

1894 Ms Edwards-Warren said that in the context of app distribution this means that not only did Professor Asker assume that the response of Fortnite users to an increase in the cost of purchasing in-app content via the Play Store is representative of the response of games app users in general to an increase in the costs of in-app purchases made via the Play Store, but in addition, he assumed the following.

1895 First, he assumed that as well as the Samsung Galaxy Store and the directly downloaded app, PCs, PlayStation, Xbox, Nintendo Switch and iOS are all considered by Fortnite users to be substitute platforms for the Play Store.

1896 Second, he assumed that in response to an increase in the costs of purchasing in-app content via the Play Store, marginal Fortnite users who use the Play Store would substitute to all of these assumed substitute platforms in proportion to these platforms’ relative shares of expenditure.

1897 Ms Edwards-Warren said that if Fortnite users do not behave in accordance with either of these two assumptions, then the diversion ratios implied by the logit demand model are not accurate or reliable estimates of the actual diversion ratios of Fortnite users.

1898 She said that if Fortnite users do not consider PCs, PlayStation, Xbox, Nintendo Switch and iOS to be substitute platforms for the Play Store, for example, because Fortnite users consider these to be complementary platforms, then the logit diversion ratios of Professor Asker will underestimate the actual diversion ratios to the Samsung Galaxy Store and direct download, because by design the logit demand model assumes some diversion to PCs, PlayStation, Xbox, Nintendo Switch and iOS, even though Fortnite users do not consider these platforms to be substitutes for the Play Store.

1899 She said that if Fortnite users consider only the Samsung Galaxy Store and direct download to be substitutes for the Play Store, and substitute between these platforms in proportion to their relative expenditure shares, then the diversion ratios implied by this logit demand model would be 78% and 21% respectively.

1900 Similarly, she said that if Fortnite users consider some of these platforms to be closer substitutes for the Play Store than other platforms, that is, the IIA assumption of the logit demand model does not hold, then the logit diversion ratios of Professor Asker will underestimate the actual diversion ratios to the platforms that are closer substitutes and overestimate the diversion ratios to the platforms that are more distant substitutes.

1901 Ms Edwards-Warren considered that the analysis of critical loss and the Fortnite hotfix event of Professor Asker are flawed and cannot be relied upon for purposes of market definition.

1902 She said that Professor Asker provided no robust evidence to justify the assumption made in the logit demand modelling that Fortnite users who purchase in-app content, particularly those Fortnite users who used the Play Store to purchase in-app content, consider PCs, PlayStation, Xbox, Nintendo Switch and iOS to be substitute platforms for the Play Store.

1903 Further, she said that Professor Asker’s test for the IIA assumption is not a reliable test of the IIA assumption made by the logit demand model.

1904 Now Professor Asker claimed to test the IIA assumption made by the logit demand model, and concluded that the results of this test are consistent with the IIA assumption holding for Fortnite users who purchase in-app content.

1905 If Fortnite users consider all of the platforms to be substitutes, then the IIA assumption means that in response to a change in prices Fortnite users will substitute to rival platforms in proportion to the expenditure shares of each of these substitute platforms.

1906 Now to test whether the IIA assumption holds, Professor Asker compared the ratio of each platform’s share of expenditure to the expenditure share of PCs, before and after the launch of Fortnite on the Play Store in April 2020. If these before and after ratios are similar for each platform, then Professor Asker interpreted this as evidence that the IIA assumption holds for Fortnite users.

1907 But Ms Edwards-Warren said that there were several problems with this test such that it was not reliable.  As a consequence Ms Edwards-Warren did not consider that the test of the IIA assumption of Professor Asker provided reliable evidence that Fortnite users behave in a manner consistent with the IIA assumption.

1908 She said that the test is unable to distinguish between all the platforms being substitutes for the Play Store and the IIA assumption holding, and none of the platforms being substitutes for the Play Store. In both cases the launch of the Play Store will result in a pro-rata reduction of the existing platforms’ expenditure shares and hence leave the ratios of these shares to one of them unchanged.

1909 She said that more importantly, the results of the test are sensitive to which platform it uses as the base to calculate the ratios of the expenditure shares before and after the launch of Fortnite on the Play Store.

1910 Now Professor Asker used the expenditure share of PCs to calculate the ratios. But she said that using the expenditure share of iOS results in there being much larger differences between the ratios of expenditure for both PlayStation and Xbox before and after the launch of Fortnite on the Play Store.

1911 Ms Edwards-Warren’s table 32 and table 33 below contain the ratios of the expenditure shares of Australian Fortnite Play Store users, in the top half of the table, and Australian Fortnite Android users, in the bottom half of the table, for the non-Play Store platforms before and after launch of Fortnite on the Play Store using the expenditure share of PC to calculate the ratio, and the expenditure share of iOS to calculate the ratio for each group of users. Ms Edwards-Warren’s table 34 and table 35 contain the equivalent figures for global (excluding China) Play Store users and global (excluding China) Android users, respectively.

1912 Now using the expenditure share of PC to calculate the ratios, Professor Asker said that the differences in the ratios before and after the launch of Fortnite on the Play Store are minor for both groups of Australian Fortnite users, and interpreted this as evidence that the IIA assumption holds for Fortnite users. Professor Asker also found that the differences are minor for global (excluding China) Play Store and Android users.

1913 But using the expenditure share of iOS to calculate the ratios, Ms Edwards-Warren found that there are much larger differences between the ratios for both the PlayStation and the Xbox before and after the launch of Fortnite on the Play Store for both groups of Australian Fortnite users. For both the PlayStation and the Xbox the differences between the before and after ratios are at least four times larger using the iOS expenditure share to calculate the ratios compared with using the PC expenditure share. She found similar large differences for global (excluding China) Play Store and Android users.

1914 Ms Edwards-Warren said that as using a different expenditure share to calculate the ratios has a material effect on the difference between the before and after ratios, and Professor Asker used the size of this difference to test whether the IIA assumption holds, she did not consider the results of Professor Asker’s exhibit 160 to be informative about whether the IIA assumption holds for Fortnite users; I should note that Professor Asker’s exhibit 160 is presented a little later in one of Professor Goldfarb’s tables that I have reproduced.

1915 So, in summary she said that the logit diversion ratios reported by Professor Asker merely reflect the unsubstantiated assumption made by him that Fortnite users who purchase in-app content, particularly those Fortnite users who used the Play Store to purchase in-app content, behave in the manner described by the logit demand model, that is, Fortnite users consider all of the platforms to be substitutes for each other and would substitute to them in proportion to each platform’s expenditure share.

1916 Now I should at this point say something about Professor Goldfarb’s evidence, although I have covered some aspects in my discussion of Professor Asker’s and Ms Edwards-Warren’s views.

Professor Goldfarb’s evidence

1917 Professor Goldfarb explained that in a logit demand model, the estimated rate of substitution to each alternative is simply proportionate to the share of each alternative, so that the ratio of shares remains constant across alternatives.

1918 As he postulated, consider a set of four alternatives, each with a 25% share of current user spending. Under a logit model, if one of these alternatives is removed, users of the removed alternative will substitute to the other three remaining alternatives at equal rates. This is because the remaining three alternatives had equal shares prior to the exit of the other alternative. Prior to the loss of the fourth alternative, the share of each of the remaining three alternatives was 25%, and following the loss of the fourth alternative, the share of each of the remaining three alternatives is one-third. The relative share of the three options does not change. The shares occur in a 1-1-1 ratio before and after.

1919 Professor Goldfarb said that because of the way that the logit demand models consumer choice, it is important that proportionate substitution to alternatives holds in the particular context in which the model is applied. As I have already said, this is the IIA assumption.  Under the IIA assumption in a logit model, if consumers lose or gain access to one alternative, then the ratio of shares for all other products remains the same.

1920 Now Professor Goldfarb said that the red bus - blue bus problem is instructive in understanding why IIA is a strong assumption that does not always hold.  I have mentioned this problem above, but let me explain it here.

1921 Suppose a commuter has a choice between commuting via a car or via a red bus. The IIA assumption implies that the introduction of a blue bus, being a blue bus that is the same as the red bus apart from colour, would reduce the usage of the car and the red bus in proportion to their respective shares, with the result that their relative share remains unchanged.  But in reality, the introduction of a blue bus would reduce the usage of the red bus much more than usage of the car, as the blue and red buses are much closer substitutes than the blue bus and the car.

1922 In A Cameron and P Trivedi, Microeconometrics: Methods and Applications (Cambridge University Press, 2005) it is said (at p 503):

As an extreme example, the conditional probability of commute by car given commute by car or red bus is assumed [under IIA] to be independent of whether commuting by blue bus is an option. However, in practice we would expect introduction of a blue bus, which is the same as a red bus in every aspect except color, to have little impact on car use and to halve use of the red bus, leading to an increase in the conditional probability of car use given commute by car or red bus.

This weakness … is known in the literature as the red bus-blue bus problem, or more formally as the assumption of independence of irrelevant alternatives. It can be tested by a Hausman test (see Hausman and McFadden, 1984). …

1923 Now the IIA assumption may not be realistic in some scenarios.  So Professor Asker aimed to test whether the IIA assumption held in relation to Fortnite user behaviour by comparing the relative shares of the channels for Fortnite before it was available on Play Store and while Fortnite was available on the Play Store. He provided his results in exhibit 160 in his report, which are replicated in Professor Goldfarb’s table 12.A.

1924 Upon reviewing Professor Asker’s results, Professor Goldfarb noted that there were large decreases to the relative shares of mobile channels, being a decrease in normalised share of 50% or more for the Samsung Galaxy Store.

1925 Professor Goldfarb noted that there were much smaller decreases to the relative share of consoles. The normalised share of PlayStation decreased by less than 5% for Australian Play Store users and increased by about 5% for Australian Android smart mobile device users.

1926 On the basis of these results, Professor Asker concluded that there were generally similar ratios of normalised shares before Fortnite launched on the Play Store and while Fortnite was available on the Play Store.

1927 But Professor Goldfarb said that on the contrary, he observed disproportionate changes in normalised shares following the launch of Fortnite on the Play Store, indicating a disproportionate pattern of substitution and indicating that the IIA assumption does not hold.

1928 To further test the changes in relative shares, Professor Goldfarb provided in his table 12.B below a reconstruction of exhibit 160 of Professor Asker’s report. In this reconstruction he provided shares normalised relative to the Samsung Galaxy Store, rather than PC, to highlight the decrease in the relative share of Samsung Galaxy Store as compared to PlayStation, Xbox, PC, and Switch. The Samsung Galaxy Store has a lower overall share than PC and using a channel with a lower share can make any relative share changes easier to observe. Professor Asker said that the IIA assumption is valid if the normalised are generally similar before the launch of Fortnite on Play Store and while Fortnite was available on Play Store.

1929 Table 12.B illustrates that the shares of PlayStation and Xbox, relative to other channels, considerably increased following the launch of Fortnite on the Play Store. This holds for both Australian Play Store users and Australian Android users. The normalised shares of PlayStation among Australian Play Store users more than doubled, increasing from 19.8 to 54.1, whilst the PlayStation shares among Australian Android smart mobile device users changed by over 30%, increasing from 19.3 to 26.4.

1930 According to Professor Goldfarb, these results further confirm that the IIA assumption does not hold.  So Professor Asker’s application of the logit demand model is inappropriate.

1931 Finally, Professor Goldfarb noted that although there are known extensions of the logit model which allow it to be used in situations where the IIA assumption does not hold, Professor Asker indicated that he did not have sufficiently detailed data to apply these extensions of the logit demand model.

1932 In summary, Professor Goldfarb found that Professor Asker’s logit analysis was flawed and concluded that his associated results were not informative.

Google’s arguments

1933 It is convenient at this point to address Google's arguments.

1934 Now Epic says that the logit model analysis should not be attributed any weight, because it is driven by assumptions regarding the boundaries of the relevant market, but Google says that this suggestion is wrong for the following reasons.

1935 First, Google says that contrary to Ms Edwards-Warren’s claim that Professor Asker assumed the set of products over which to estimate the diversion ratios, Professor Asker’s logit model analysis was informed by the empirical evidence.

1936 Professor Asker’s analysis focused on Play Store Fortnite users and Android Fortnite users and evaluated what transaction venues those users used when spending on Fortnite.

1937 So, his approach was to focus on Fortnite users who could spend via the Play Store and evaluate where they choose to spend. The set of transaction venues included in the analysis are  those where users of interest who could spend via the Play Store chose to spend instead.  So these are venues to which the Play Store lost potential revenue, and so represent the issue considered in the market definition exercise.

1938 Second, Google says that were Fortnite Play Store users to have only spent via Android app stores or the Play Store, then those would have been the only transaction venues included in the logit analysis.

1939 According to Google, Ms Edwards-Warren’s critique amounts to dismissing the fact that Fortnite users who could have spent via the Play Store chose not to.

1940 Third, Google says that Play Store Fortnite users also spend on Fortnite via non-Android transaction venues. There is no evidence that these realised choices by consumers who could otherwise spend on the Play Store are uninformative of their choice preferences.

1941 Further, Google says that Professor Asker’s logit model satisfies the IIA assumption.

1942 Now as Google pointed out, the assumption needed for the logit model to translate spending patterns into diversion estimates is the IIA assumption. As I have already said, essentially, the assumption states that the likelihood of choosing between options A and B is unaffected by the ability to choose option C.

1943 The IIA assumption can be tested by asking whether observed spending patterns across options in the choice set change when a new option is added.

1944 Now Epic says that the IIA assumption does not hold, and that as a consequence, no weight should be attributed to the logit modelling. But Google says that Epic is wrong for the following reasons.

1945 First, Google says that Professor Asker tested whether the IIA assumption holds by asking whether relative spending shares across existing transaction venues meaningfully changed when Fortnite was first published to the Play Store. They did not.

1946 Professor Goldfarb said that the relative spending shares did change if one was to calculate them relative to the Samsung Galaxy Store. But as Professor Asker explained, that approach simply captures noise due to dividing by small numbers. The issue stems from few people wishing to use the Samsung Galaxy Store to spend on Fortnite even prior to the launch of Fortnite on the Play Store.  Professor Goldfarb’s criticism in this respect is misplaced.

1947 Second, Google says that Professor Goldfarb suggested that the removal of Fortnite from the Play Store due to the hotfix can also be used to test IIA, and that what that removal shows is that there is a large change in the share of spending on the Samsung Galaxy Store relative to pre-removal shares.  Google says that this point should be rejected, because it is not relevant to IIA.

1948 It says that by removing the Play Store as an option, the share of spending on almost every venue increased. The standard test for IIA accounts for this by comparing the ratio of shares. Professor Goldfarb did not examine the ratios of shares. In contrast, Professor Asker found that these ratios held approximately constant before and after the launch of Fortnite on the Play Store. Professor Asker’s assessment focused on the launch of Fortnite on the Play Store rather than on the hotfix.

1949 Contrastingly, because the hotfix affected both the Play Store and the iOS App Store, it cannot be used to test the IIA assumption without further assumptions being made as to how to treat the removal of two, instead of just one, potential choices. Professor Goldfarb conceded this by observing that “[t]he problem with doing that is that we’re not going to be able to deal with Apple”. But he then provided no explanation of how he would account for this problem. So, Epic’s position in this regard is not maintainable.

1950 Google says that as Professor Asker explained, the basic pattern of how Fortnite Play Store users spend is so stark that even extensions of the logit model that relax the IIA assumption in specific ways are unlikely to yield qualitatively different conclusions as to the importance of different transaction venues as competitive constraints on the Play Store.

Analysis

1951 Let me deal with some central questions.

1952 As I have said earlier, in order for the logit model to hold, the IIA assumption also needs to hold.

1953 As Professor Asker stated, if the IIA assumption is not satisfied, the diversion ratios coming out of the standard logit model are going to be incorrect. Whilst there are extensions to the standard logit model which can be implemented, Professor Asker indicated these would have been impossible to implement because doing so requires a richness of data and variation which he did not have.

1954 In this case, the IIA assumption does not hold. This is reflected in the fact as I have said above that even on Professor Asker’s estimates, diversion from the Play Store to the Galaxy Store following the removal far exceeded what was predicted by Professor Asker’s model.

1955 As Professor Goldfarb explained, it is also evident from a comparison between the graphs in Professor Asker’s exhibits showing Fortnite platform shares.  Prior to the Fortnite removal event the share of Play Store user spending for “other Android” is in the order of 1% and is the smallest share. Following the removal the “other Android” share is 31% and is the largest share.

1956 That is a violation of the IIA assumption. The failure of this assumption means that Professor Asker’s logit modelling will not generate correct results and is uninformative about substitution or removal.

1957 Despite these limitations, Professor Asker said that he would not expect his conclusions to change were he to have access to sufficient data to instead use those richer versions of the logit framework. Professor Asker’s assertion is just that.

1958 I reject this claim, particularly considering Professor Goldfarb’s evidence that based on the data he had, he had no information about how those alternatives would play out.

1959 For these reasons, I do not attribute Professor Asker’s logit modelling analysis any weight with respect to substitution or market definition.

1960 Now Google says that Professor Goldfarb’s results disproving the IIA assumption simply result from noise due to dividing by small numbers. But Professor Asker never established that it was due to noise, rather than a disproportionate diversion in violation of the IIA assumption.

1961 Further, if the IIA assumption were true, then upon removal of Play Store, diversion from the Play Store should occur in proportion to the market shares of the other channels assumed to be in the market.  But Professor Asker’s relevant exhibit shows that not to be the case.

1962 As I have already said, prior to removal, “other Android” had 0.8% of total “market” share, and was the smallest of all channels. If, following removal of the Play Store, diversion to “other Android” was in proportion to market share, then it should have remained the smallest of all channels.  In fact, after removal of Fortnite from the Play Store, “other Android” became the largest channel.

1963 In my view the logit diversion ratios presented by Professor Asker add no new evidence as to which platforms Fortnite users who use the Play Store actually consider as the largest and closest substitutes for the Play Store as a platform.

1964 So, these logit diversion ratios are not informative about market definition in this case. This is because the results of the logit model are driven by the assumptions made on the set of products over which to estimate the diversion ratios.

1965 So, the high diversion to consoles is solely driven by the fact that the majority of user spend on Fortnite is on consoles.

1966 And the assumption that the market includes consoles is based on the results of the hotfix analysis which was not properly conceived.

Transition probability analysis

1967 Let me turn to another topic concerning estimates of diversion using transition probabilities.

1968 Now Professor Asker supplemented his diversion estimates from the hotfix and logit analyses with diversion estimates based on transition probabilities.

1969 According to Google, diversion estimates based on transition probabilities are another way of interpreting baseline switching rates in the raw data.  It is said that the value of this exercise is to complement existing diversion estimates by interpreting them through a fresh lens, especially given that the one relevant natural experiment available does not allow for measuring diversion to the iOS App Store.

1970 Google says that the transition probabilities analysis confirms Professor Asker’s findings from other analyses that console transaction venues are the closest competitors to the Play Store for the facilitation of gaming content. Professor Asker considered that the diversion estimates based on transition probabilities are informative when it comes to evaluating the totality of the evidence.

1971 Professor Asker calculated the diversion ratios based on transition probabilities using the observed spending patterns of Fortnite users.

1972 Professor Asker analysed pairs of successive spending events on Fortnite to calculate transition probabilities.  According to his evidence, these probabilities answered the question: after a user spends on a transaction venue A, what share of their next spending occurs via transaction venue B?

1973 Professor Asker used these transition probabilities to calculate the implied normalised diversion ratios. These were calculated by assuming that Play Store Fortnite users can no longer spend next via the Play Store, and then using the magnitudes of their transition probabilities to estimate where their spending would most likely divert.

1974 According to his evidence, the identifying assumption behind this analysis is that users divert spending from the Play Store proportionally to where they transition spending in the observed data, which means that users switch in response to a triggering event proportionally to where they were already switching.

1975 Now Professor Asker purported to calculate diversion ratios based on the proportion of transactions which were performed on each non-Play Store channel by users whose previous transaction was on the Play Store.

1976 But I agree with Epic that this analysis does little more than provide an indirect measure of multi-homing amongst Fortnite users prior to the removal. It does not provide any insight into the likely diversion rates which would occur in response to a price or quality change on the Play Store.

1977 In particular, it significantly underestimated the likely diversion to the Galaxy Store in response to such a change, given that one would not expect users who had downloaded the game from the Play Store to also have access to the Galaxy Store version, consistent with Professor Asker’s finding that Play Store users spent the least through “other Android channels”.

1978 Further, transition probabilities do not measure any observed behaviour in response to a price or quality change.

1979 Further, the probabilities calculated in Professor Asker’s exhibit 146 would not be expected to be observed in response to such a change in this case.

1980 As Epic says, the transition probabilities in question are based on an average of immediately consecutive transactions over a period of time. They do not show how spending changes over time.

1981 Further, the fact that the transition probabilities are asymmetric merely shows that console users are more likely to make their next purchase on a console than mobile users.

Counting of diversion estimates

1982 Finally, let me turn to the topic of Professor Asker’s counting of diversion estimates.

1983 Now Professor Asker also provided a meta-analysis in the joint expert report, which summarised the diversion estimates advanced by him and Professor Goldfarb.

1984 Whilst Professor Asker acknowledged in cross-examination that his “80 out of 86” and “70 out of 75” summaries would give equal weight to every estimate and would be problematic if interpreted in the manner put forth by Epic, Google says that properly understood, his main compilation of diversion estimates in the joint expert report, rather than these introductory statistics, is informative.

1985 First, Google says that Professor Asker explained in the concurrent evidence session that the purpose of compiling diversion estimates in the table in the joint expert report was to understand where the experts agree and differ, or where the body of the estimates lie, with respect to the key qualitative takeaways that matter for market definition.

1986 Google says that because there were more than 80 different diversion estimates advanced in Professor Asker’s first report and in Professor Goldfarb’s second report, such a summary, being the figure at page 87 of the joint expert report, helped to clarify the key points of disagreement between them and helped to interrogate the borderline results to understand what drives them.

1987 Google says that what the summary reveals is that the key point of disagreement that matters for market definition pertains to whether to include the two weeks after the hotfix in the post-period or not. This is because almost all other specifications yield similar qualitative results.

1988 Second, Google says that compiling areas of agreement and difference across a range of econometric specifications is important because of the econometric concept of robustness. It says that conclusions that are robust to multiple, correct specifications that vary samples, time periods, and controls are more likely to yield correct inferences than those that are sensitive to these dimensions.

1989 Google says that what Professor Asker’s compilation of diversion estimates revealed is that his key findings are robust.  So, it is said that they support his position that diversion to at least one console store is typically as large if not larger than diversion to the Samsung Galaxy Store, with consoles taken together accounting for much larger diversion. Further, it is said that they support his position that diversion to non-Android transaction venues is usually as large if not larger than diversion to Android transaction venues.

1990 Third, Google says that although the compilation included duplicate and intermediate results, as well as including specifications that Professor Goldfarb did not analyse, this is inconsequential for the following reasons.

1991 It is said that Professor Asker’s diversion estimates based on the logit analysis and the transition probability analysis are informative about substitution, and should be considered as part of the totality of evidence.  Further, it is said that Professor Asker’s analysis using the levels method are valid estimates of diversion, and preferable to the percent method.  Further, it is said that Professor Asker’s analysis, which relied on an assumption of 0% of spending via EDP on Android occurring through sideloading, is valid, especially in the context of the 5-week post-period that Professor Asker analysed.

1992 Google also says that Epic incorrectly said that Professor Asker regarded these estimates as inaccurate.  Professor Asker stated that he considers the 0% assumption informative, but that the 50% assumption is likely more accurate than the estimates assuming 0% or 100%.  This is different from framing it as inaccurate.

1993 Google says that Epic overlooks that the 0% assumption is Professor Asker’s conservative specification since assuming lower spending via sideloading means that the diversion to the Samsung Galaxy Store is higher. That is evidenced in the results.

1994 Further, it is said that Professor Asker did not include any duplicate or intermediate results in his presentation of diversion estimates in the said figure of the joint expert report, only in the preceding summary of the total number of estimates to which he had regard. The duplicate and intermediate results arose from Professor Goldfarb’s method of presenting the same analyses in multiple tables, or presenting analyses that he viewed as uninformative as to the correct diversion estimate.  So it is said that this means that the 46 estimates that Professor Asker provided from his first report are not affected by this criticism and the fact of them being recorded in a spreadsheet is irrelevant.

1995 Further, Google says that removing duplicate and intermediate entries has no effect on the qualitative takeaways of the exercise. Removing the relevant entries does not change the proposition that under most specifications presented, diversion to the Samsung Galaxy Store is too low for Ms Edwards-Warren’s market to be well-defined, diversion to at least one console store is higher than diversion to other Android transaction venues, and diversion to non-Android transaction venues is higher than diversion to Android transaction venues.

1996 Further, Google says that whether or not Professor Goldfarb analysed a particular specification is irrelevant.  Professor Asker analysed diversion in spending in his first report without reference to any prior estimates of diversion in spending put forward by Epic’s economic experts, of which there were none.

1997 Further, Google says that Professor Asker’s synthesis of the totality of the evidence through the relevant figure in the joint expert report reflects an effort to assist me by distilling the areas of agreement and disagreement between the experts.

1998 But clearly he had to acknowledge that the vast majority of the summary figures were not robust and could be disregarded.

1999 I agree with Epic, largely for the reasons it advanced, that Professor Asker’s counting exercise was unreliable and did not provide any useful information.

2000 First, this analysis combined and afforded equal weight to estimates of diversion based on methodologies of differing reliability and relevance to the case at hand.

2001 So, Professor Asker included 10 estimates of diversion which are based on the logit and transition probability analysis, which he acknowledged are less useful for measuring diversion than directly analysing user behaviour following the removal.

2002 Further, he included 30 estimates of diversion based on his levels difference in difference methodology which produces nonsensical estimates of negative counterfactual expenditure, 10 of which assumed that 0% of spending through EDP after the removal went through the sideloaded app, something he regards as inaccurate.

2003 Further, he included six duplicate estimates which were individually counted as separate estimates of diversion.

2004 Additionally, as well as being unrealistic or irrelevant, many of the specifications underlying the 46 diversion rates Professor Asker relied upon from his own report were not considered at all by Professor Goldfarb.

2005 So, out of the first 30 estimates Professor Asker included from his own report, 20 related to parameters or assumptions which Professor Goldfarb did not consider, including those directed to diversion globally (excluding China) or based on an assumption of 0% EDP spend through the sideloaded app.

2006 The inclusion of these estimates biased the outcome of Professor Asker’s counting exercise towards alignment with his own results.

2007 The consequence of what I have said is that Professor Asker’s characterisation of the weight of the available evidence in the joint expert report was influenced by the inclusion of diversion estimates which are irrelevant, duplicative, and unreasonable.

2008 Generally, the more appropriate approach was to consider the estimates of diversion which were calculated based on the most appropriate methodologies. Professor Asker’s exercise of taking every diversion estimate and including it provided little assistance.

2009 Further, in the joint expert report Professor Asker purported to use his counting exercise to form direct quantitative conclusions concerning market definition based on his critical loss analysis, rather than attempting to show areas of agreement and difference and what the quality estimates were.

2010 He wrote:

Hence, at this conservative set of inputs, 70% of available estimates indicates that Ms Edwards-Warren’s Android mobile app distributions market fails the HMT.

What this tells me is regardless of the inputs chosen … the weight of evidence on diversion from the Play Store indicates that the Android mobile app distribution market is too narrow. …

Hence, the weight of evidence in this case points to the Android mobile app distribution market failing the HMT.

2011 Professor Asker’s reference to the “weight of evidence” was to the 70% of his useless compilation of estimates, including duplicates and estimates using unreliable methodologies.

2012 In cross-examination, Professor Asker conceded that the counting exercise was not useful in terms of the precise numbers. Once that is accepted, the counting exercise ceases to be relevant. An imprecise counting exercise is worthless.

2013 Further, in relation to the 0% sideloading assumption, Professor Asker was counting analyses which he did not think were accurate and which Professor Goldfarb did not perform. There was nothing conservative about including them, because ultimately the 0% estimates did not lead to a higher number of estimates exceeding Professor Asker’s relevant threshold.

2014 On the other hand, if Professor Goldfarb had calculated estimates based on a 0% sideloading assumption, a greater number of Professor Goldfarb’s estimates would have exceeded Professor Asker’s threshold.

2015 So, the inclusion of the 0% estimates merely biased the counting exercise in Professor Asker’s favour. This highlights the inherent bias in Professor Asker’s method by including estimates in his counting exercise that had nothing comparable in Professor Goldfarb’s analyses.

2016 Further, the results in figure 1 of the joint economic report  are based on Professor Asker’s list of estimates which included duplicate and intermediate results.

2017 Further, once one only includes estimates based on the most reliable methodology for estimating diversion to Samsung Galaxy, one is left with Professor Goldfarb’s estimates in table 11.B, which I have reproduced in the previous section of my reasons dealing with the Fortnite removal event and diversion ratios.

Critical loss analysis including the Lerner index

2018 As Professor Asker explained, in the absence of complete data on cross-price elasticities and margins required for a complete quantitative HMT, an empirical tool that economists use to determine whether a candidate market is a well-defined market, but underpinned by the principles of HMT, is a critical loss analysis.

2019 A critical loss analysis first establishes a critical loss threshold that would make a price increase by the hypothetical monopolist unprofitable, and then determines whether actual losses from that price increase would be higher or lower than the critical loss threshold.

2020 In essence, critical loss answers the question: if the hypothetical monopolist raises the price of one product in the candidate market by X% , what percentage of the sales of that product would need to be lost, that is, not recaptured by the hypothetical monopolist’s other products, for the price increase to be unprofitable?

2021 In its standard formulation, critical loss is expressed as a threshold that is a function of the percentage increase in price, X, and the firm’s margins, M.  The standard formulation of the critical loss threshold is expressed as a fraction where the numerator is X and the denominator is X plus M.  This means that the critical loss threshold increases with a higher price increase and decreases with a higher margin.  Actual loss answers the question: if the hypothetical monopolist raises the price of one product by X%, what percentage of sales of that product would be lost due to the price increase?

2022 Let me also couch this concept in profit maximising terms as Professor Asker has done.

2023 Critical loss analysis starts by asking the following question: would a hypothetical monopolist of the candidate market find it profit-maximising to increase the price of one of its products, say Product A, by at least X %? The profit-maximising price increase is at least X % if an increase of twice the size (2X %) is profitable. This occurs when the actual loss in sales, for a 2X % price increase, is less than what is called the critical loss threshold.

2024 To determine the actual loss in sales, one must balance two competing effects. On the one hand, by increasing the price, the firm will lose customers from Product A. On the other hand, the firm, as the hypothetical monopolist, can recapture some of those lost consumers through the other products it owns, that is, everything else in the candidate market except Product A. The amount of lost Product A sales can be calculated using consumer elasticity, where the elasticity can be imputed using the Lerner index. The amount of recapture can be calculated using an aggregate diversion ratio, that is, how much is diverted to all other products in the candidate market.

2025 Critical loss analysis is typically evaluated by estimating an aggregate diversion ratio.

2026 An aggregate diversion ratio answers the question: if the hypothetical monopolist raises the price of one product, what fraction of lost sales of that product does the hypothetical monopolist recapture by other products it owns?

2027 In the present context then, calculating an aggregate diversion ratio for a hypothetical monopolist of Android app stores would answer the question: what fraction of sales that divert from the Play Store in response to an increase in the Play Store’s prices are recaptured by other Android app stores?

2028 Now Professor Asker said that the advantage of a critical loss analysis is that it requires a parsimonious set of parameters to understand when a candidate market is a well-defined market being the aggregate diversion ratio from the product of interest to other products in the candidate market, the margin for the product of interest, and the hypothetical percent increase in the price of the product of interest.

2029 According to Professor Asker, a critical loss analysis usually evaluates a 10% price increase by the hypothetical monopolist, as that corresponds to the hypothetical monopolist finding it profit-maximising to increase prices by at least 5%.  I will put to one side for the moment the debate as to the difference between the profitable step approach and the profit maximising step approach when one is dealing with the application of the HMT.

2030 Now a standard critical loss analysis that compares aggregate diversion ratios to the critical loss threshold is predicated on the Lerner index. The Lerner index relates a firm’s margins to the elasticity of customer demand it faces.  Obviously the more price-sensitive consumers are, the smaller margins a firm can sustain.

2031 Now the standard Lerner index usually assumes a single set of customers.  But this does not apply to the Play Store as the Play Store connects two sets of customers, being developers and consumers, with developers setting the price consumers pay and the Play Store monetising by a fraction of the transaction value.

2032 So, to conduct a critical loss analysis for the Play Store, Professor Asker needed to first derive the relevant Lerner index for the Play Store, and then derive the relevant critical loss formulae.

2033 Now whilst both consumers and developers might substitute away from the Play Store in the event of a price increase, Professor Asker’s resulting platform critical loss analysis focused on substitution by consumers.

2034 The use of critical loss analysis to evaluate markets in which the Play Store competes only allowed for consumer substitution, not developer substitution.  So, this means that such analysis is inherently conservative because it does not account for how developers can respond to worsening terms from the Play Store. It assumes that the only way developers can respond is to adjust the price of the content they set for consumers on the Play Store whilst continuing to use the Play Store at the same level of intensity.  Now because of this reason, when Professor Asker deployed critical loss analysis, he supplemented it with a discussion of developer substitution possibilities that could further discipline any hypothetical monopolist.

2035 Now compared to a standard critical loss analysis, Professor Asker’s platform critical loss included an additional parameter reflecting the different way that the Play Store monetises. That parameter is the elasticity of developer price with respect to the platform’s service fee, which is also known as the pass-through elasticity.

2036 Now developer pass-through depends on the price elasticity of consumers. The more price elastic consumers are, the more likely developers are to absorb service fee increases instead of passing them through.

2037 Now without directly estimating developer pass-through, an alternative according to Professor Asker is to perform the platform critical loss analysis under different assumptions of developer pass-through. The most conservative approach for purposes of market definition analysis is to consider the value of developer pass-through that corresponds to the lowest consumer price elasticity being full developer pass-through. The platform critical loss analysis therefore seeks to understand whether consumer substitution alone is sufficient to discipline a hypothetical monopolist from raising prices.

2038 Now to perform a critical loss analysis Professor Asker used the diversion ratios calculated based on the hotfix regression specifications, which is a method of conducting the HMT quantitatively.  But it only considers the hypothetical monopolist raising prices on one product, rather than all products in the market, as would be required under the HMT.

2039 Both Ms Edwards-Warren and Professor Asker agreed that one way of conducting a critical loss analysis is to follow the approach described in M Katz and C Shapiro, “Critical Loss: Let’s Tell the Whole Story” (2003) 17(2) Antitrust 49.

2040 This approach compares the aggregate diversion ratio, which is the share of sales lost by Google Play that would be recaptured by other products in the candidate market in response to a price increase, to a critical loss threshold. When the diversion ratio is below the threshold, the candidate market fails the HMT.

2041 It was also agreed that in a platform context, the critical loss analysis requires four key inputs being, first, Google Play’s economic margin, second, the hypothetical price increase, third, the developer’s pass-through elasticity and, fourth, the aggregate diversion ratio.

2042 For the first input, Professor Asker used Professor Davila’s estimates of Google Play’s gross margin, which ranged between 74.9% and 80.7% during the relevant period, as the profit input. This usage did not appear to be contentious between the economic experts as best as I could tell in this context.

2043 As to the fourth input, the Samsung Galaxy Store diversion ratio from the hotfix regression analysis was used as the aggregate diversion ratio, because Samsung Galaxy Store was the only Android app store on which Fortnite was available after the hotfix.  This did not seem contentious in this context, although it was said by Google that this made the analysis “conservative” so far as it was concerned; this resonated with the line that it continued to push that the Galaxy Store was a weaker substitute than other modes that it said were stronger substitutes.

2044 Now as to the second and third inputs, there was disagreement between Professor Asker and Ms Edwards-Warren.

2045 As to the second input being the hypothetical price increase, Professor Asker’s view was that the relevant price increase to consider for the purposes of conducting a quantitative HMT is 10%, rather than 5% or something between.

2046 He said that this was because a 10% price increase being profitable means that a hypothetical monopolist would find it economically rational (profit-maximising) to increase prices by at least 5%.  Professor Asker also used a developer pass-through assumption of 100%.

2047 If those assumptions are made, according to Professor Asker the relevant platform critical loss threshold ranges between 14.5% and 15.4% for the relevant period.

2048 Further, according to Professor Asker and on these assumptions, diversion ratios to the Samsung Galaxy Store based on the 26 week and 52 week specifications in Professor Goldfarb’s table 11.A and table 11.B were well below the critical loss threshold. This meant according to Professor Asker that a 10% price increase by a hypothetical Android app store monopolist would not be profitable.

2049 In other words, on Professor Asker’s analysis Epic’s pleaded Australian Android mobile app distribution market failed the HMT on a 10% price increase.  And that being so, the next closest substitutes needed to be included in the market, which according to Professor Asker were gaming consoles or non-Android platforms based on the diversion ratios calculated from the hotfix regression analysis.

2050 Further, according to Professor Asker, he disagreed with Ms Edwards-Warren that the next closest substitute was direct downloads.  According to Professor Asker this was because the diversion ratio to direct downloading was far smaller than Xbox and PlayStation and the total for non-Android platforms.

2051 Let me at this point say something about Ms Edwards-Warren’s position.

2052 Ms Edwards-Warren’s main point concerning the second input was that the critical loss analysis should not be done only with a 10% price increase, and other price increases between 5% and 10% should be considered.

2053 Now it was common ground that if a 5% price increase were used, the critical loss threshold would be exceeded, and the candidate market would pass the HMT.  Ms Edwards-Warren said that this result, that is, passing at 5% but failing at 10%, indicates that the candidate market definition is “borderline”.

2054 I should note that the basis for the disagreement over which notional price increase should be used principally arose from a difference of opinion concerning the proper application of the HMT when conducting it quantitatively.

2055 As to the third input, as I say, Professor Asker used a developer pass-through assumption of 100%.

2056 Contrastingly, Ms Edwards-Warren’s view was that a 100% developer pass-through is unrealistically high, but she offered no estimate of her own.  Instead, she re-ran the critical loss analysis using pass-through rates ranging between 5% and 100%, and Google Play gross margins ranging between 70% and 95%, which were beyond Professor Davila’s estimates.

2057 Ms Edwards-Warren listed the critical loss thresholds for a 10% price increase at every pass-through rate in 5% increments for gross margins between 70% and 95%.

2058 But according to Google, even if the lowest pass-through rate of 5% is assumed, the diversion ratios to the Samsung Galaxy Store based on the 26 week and 52 week specifications in Professor Goldfarb’s table 11.A and table 11.B all remain below the threshold when using a realistic gross margin of 80%.

2059 Further, if one were to use the lowest pass-through rate that Professor Asker considered might be within the reasonable range, that is, 20%, the Samsung Galaxy Store diversion ratios based on the 26 week and 52 week specifications in Professor Goldfarb’s table 11.A and table 11.B are below the threshold on whichever gross profit margin is assumed. In other words, so Google said, Ms Edwards-Warren’s criticism based on pass-through does not materially change the result.

2060 Let me turn to some literature which sets out some useful themes.

Critical loss theory — some literature

2061 It is convenient at this point to set out some extracts from M Katz and C Shapiro (2003) although I have touched on some of the themes above; there has been some paraphrasing by me.

2062 Let me begin with the authors discussion of the SSNIP test and the concept of a critical loss calculation.

2063 Consider a proposed merger between two companies manufacturing prescription sleeping pills. If a single firm controlling all brands of prescription sleeping pills would find it profitable to impose a SSNIP for at least one of the brands sold by the merging parties, then prescription sleeping pills constitute a relevant product market. If not, the next-best substitute, for example, non-prescription sleeping pills, is added to the candidate relevant market and the test is repeated.

2064 As a matter of arithmetic, the effect of a SSNIP on the hypothetical monopolist’s profits depends upon the prevailing profit margin earned on each unit sold and on the percentage of unit sales that would be lost as a result of the price increase.  This can be called the actual loss.  The maximal percentage of unit sales that can be lost for the price increase to be profitable is known as the critical loss. If the actual loss from a price increase would be greater than the critical loss, the price increase would be unprofitable.

2065 A critical loss calculation can usefully frame the empirical estimation of demand responsiveness for the purpose of delineating relevant product markets. Critical loss analysis is commonly used, both by economists for private parties and by economists in the US Department of Justice and Federal Trade Commission. But critical loss analysis can also be misused.

2066 A common but incomplete and potentially misleading argument based on critical loss analysis has been used to justify broad relevant markets, and it runs as follows. Because the suppliers’ profit margins are high, any lost sales have a large adverse impact on profits, and so even a hypothetical monopolist controlling a group of products could not profitably raise price.

2067 This story is incomplete because high margins also tend to imply that the actual loss is small, and so a price increase might be profitable even when the critical loss is small.

2068 The authors explain a simple approach that uses the aggregate diversion ratio, being the percentage of the total sales lost by a product when its price rises that are captured by all of the other products in the candidate market, to make greater use of the available market evidence.

2069 Their central result is that an aggregate diversion ratio greater than the critical loss creates a presumption that the candidate product market is in fact a relevant antitrust market. This implies that, all other things being equal, higher pre-merger margins, which lead to a low critical loss, tend to support a finding of narrower markets.

2070 Now the HMT examines the profitability of a SSNIP for at least one product, including at least one product of one of the merging firms. One begins by asking the traditional question posed by critical loss analysis: would a given SSNIP yield higher profits for the hypothetical monopolist than do the pre-merger prices?

2071 Consider the effects of raising the price of just one product, Product Z, which is produced by one of the merging parties. A hypothetical monopolist would consider the effect of that price increase on the profits earned on all of the products in the candidate market, not just Product Z. When the price of Product Z rises, some of the sales lost by Product Z will shift to other products in the candidate relevant market. These sales are not lost to the hypothetical monopolist. Indeed, if the difference between price and incremental cost is larger for these products than for Product Z, such diversion of sales actually boosts the monopolist’s profits, thereby making the price increase more attractive on a relative basis. To keep the analysis as simple as possible, one can focus on the case in which all of the products in the candidate market have the same difference between price and incremental cost, P – C.

2072 With this simplification, one can focus attention on the sales actually lost by the hypothetical monopolist when raising the price of Product Z, namely the sales that are lost by Product Z and not gained by any of the other products in the candidate relevant market. For this purpose, it is useful to define the aggregate diversion ratio for a given Product Z. For a given price increase for Product Z, the aggregate diversion ratio is the fraction of overall sales lost by Product Z that are captured by, or diverted to, any of the other products in the candidate relevant market. Suppose, for example, that a 5% increase in the price of Product Z causes the sales of Product Z to fall by 200 units, but the sales of other products in the candidate market to rise by 90 units.  The aggregate diversion ratio is then 45%, being 90 divided by 200 and then multiplied by 100.

2073 The aggregate diversion ratio must lie between zero and 100%.

2074 With this definition, the hypothetical monopolist considering whether to raise the price of Product Z will capture a fraction D of any lost sales through increased demand for its other products.   So, the hypothetical monopolist effectively loses only a fraction (1 – D) of the sales that are lost by Product Z. The greater the aggregate diversion ratio, the fewer the sales lost by the monopolist, and the more profitable is a price increase. Making use of some algebra, this fact leads to a powerful and surprisingly simple finding when gross margins are large. If and only if the aggregate diversion ratio is larger than the critical loss, then the actual loss is less than the critical loss and so a hypothetical monopolist would find a SSNIP profitable.

2075 Although the comparison of the actual loss with the critical loss tells one whether a given price increase would be profitable, the HMT asks whether the hypothetical monopolist’s profit-maximising [or profitable] SSNIP would be at least as large as the chosen threshold, say 5%.  As a rule, increasing the price of Product Z above its pre-merger level will cause the profits of the hypothetical monopolist to rise, reach a maximal level, and then decline. The price that maximises profits is not as large as the price at which profits fall all the way back down to their pre-merger level. As a good working approximation, the profit-maximising price increase is half as large as the maximal price increase that yields profits above their pre-merger level.  So, if a 10% price increase would cause the hypothetical monopolist’s profits to be higher than their pre-merger level, then the profit maximising price increase is at least 5%.

2076 This approximation supports the following presumption when gross margins are large:

Given the pre-merger gross margin M, calculate the critical loss associated with a 10 percent price increase.  If and only if the aggregate diversion ratio associated with a group of products is at least as large as the critical loss, then this group of products forms a relevant market using a 5 percent price increase threshold.

2077 This is a cautionary result.

2078 For example, with the 90% gross margin used in that SunGard case, the critical loss associated with a 10% price increase would be 10%.  Based on this analysis, and assuming the firms were not engaged in coordinated pricing prior to the merger, there would have been a presumption that a hypothetical monopolist controlling all products in the relevant market alleged by the government would find it profitable to impose a 5% price increase for at least one product if and only if the aggregate diversion ratio were at least 10%.

2079 In other words, there would have been a presumption that the candidate market was in fact a relevant market if a total of at least 10% of any sales lost by one product whose price was increased would have been captured collectively by the other products in the candidate market.

2080 But the presumption can be rebutted. For instance, the sample calculations are only an approximation based on the firm’s perceived elasticity of demand at the pre-merger price, in conjunction with the standing assumption that the firms are pricing independently.

2081 Further, all of this analysis concerns a horizontal merger. I should also note that I have discussed the debate concerning profit maximising as compared with profitable elsewhere.

2082 Let me also set out an extract from P Davis and E Garcés, Quantitative Techniques for Competition and Antitrust Analysis (2009, Princeton University Press) at pp 210 to 212, with a little paraphrasing by me.

2083 Critical loss analysis is conceptually closely related to the HMT. It also uses information about demand and in particular the own-price elasticity of demand to make inferences about the price constraint exerted by substitute products.

2084 As I have indicated earlier, the question asked in critical loss analysis is the following: how much do sales need to drop in order to render an X% price increase unprofitable?

2085 In the context of a benchmark homogeneous product model, this question is answered by the following formula:

2086 If the quantity demanded falls by more than 7.7% following the 5% price increase, the price increase is not profitable and the candidate market must be expanded.

2087 At least three issues commonly emerge in applying a critical loss test.

2088 First, the fact that a 5% price increase is not profitable does not mean that a 50% price increase is not profitable. Yet one is interested in market power and would clearly wish to draw narrow market boundaries if one found that a hypothetical monopolist could raise prices by 50%.

2089 Second, it may be argued that the critical loss is likely to be far smaller than the drop in sales that would be experienced by a 5% price increase and therefore a 5% price increase would be unprofitable. When accepting evidence of actual sales declines following price increases, one needs to be careful about the potential endogeneity of price and sales changes.

2090 Third, when considering critical loss calculations, it is very important to bear in mind that if pre-merger margins are high, each unit less of sales is associated with a large fall in profits and so will produce a critical loss in sales that is small. To illustrate, in the case of a 5% price increase the authors obtained the values for the critical loss shown in their table 4.8.

2091 This issue is related to the cellophane fallacy because if the margin is high, it means market power is probably already being exercised and so one must be careful to rely on the effect of price changes on an already supra-competitive price level when drawing conclusions about substitutability and market definition. If the firm has market power, it will increase price up to the point where margins are high and therefore the critical loss appears small.

2092 The fallacy in this analysis is to treat the elasticity and the margin as if they were independent from each other. In fact, according to the benchmark model, margins tells one about the own-price elasticity before the price increase. If margins are high, it implies a low price elasticity and that in turn suggests, perhaps even strongly, there will be low actual losses due to a price increase.

2093 Firms sometimes argue that because their critical loss is small, their actual loss is probably larger and the market should be large. Such arguments should not be accepted uncritically, but rather the firm must explain why they would have a low elasticity of demand evidenced by the high margins and relatively large actual losses of demand following a price increase.

2094 More generally, this is one example of a tension between the pieces of data, being the margin and the likely actual loss resulting from a price rise, and the model which states that the Lerner index is inversely related to the own-price elasticity of demand.

2095 Whenever a model and pieces of data are difficult to reconcile, one must question each. The apparent tensions may be reconciled by the finding that one or more pieces of data are “wrong”,  or alternatively that the data is right but the benchmark model is not the correct one for the relevant industry.

2096 It is important to note that the exact form of the critical loss formula depends explicitly on the monopoly model being used to characterise the industry. So, table 4.8 captures the results only of one particular type of critical loss exercise.

2097 Let me now delve into more detail concerning Professor Asker’s approach.

Critical loss analysis for a platform – Professor Asker’s approach

2098 Now Professor Asker first presented the derivations necessary for a standard critical loss analysis as derived by M Katz and C Shapiro (2003) as I have referred to above.  He then proceeded to re-derive them to aid with intuition when deriving the modified equations needed for a platform like the Play Store.  He then presented the derivations for the critical loss formulae for a platform like the Play Store.

2099 Let me say something about these derivations, re-derivations and the standard critical loss threshold by reference to Professor Asker’s evidence.

2100 As I have indicated earlier, the critical loss in sales is the amount of sales that would have to be lost to make a price increase by the hypothetical monopolist of a candidate market unprofitable. Suppose there is a hypothetical price increase from p₀ to p₁, where the quantity sold also changes from q₀ to q₁. Further suppose that the firm’s marginal cost stays fixed at c. Under this setup, the price increase is profitable if:

2101 Professor Asker defined the following terms:

(a) the increase in price (in percent terms) is defined as:

(a) the decrease in quantity (in percent terms and in absolute value) is defined as:

(b) the initial gross margin for the product is:

2102 The above inequality can be rearranged to express the critical loss threshold, which is the maximum possible value of L for the price increase to be profitable:

2103 Suppose the hypothetical monopolist of products in a candidate market increases the price of one product i. This results in a loss of sales from i, but part of this is recaptured by the hypothetical monopolist with other products in the candidate market. One can denote the share of recaptured sales as D∈[0,1], which is the aggregate diversion ratio. The higher the aggregate diversion ratio, the fewer sales lost by the hypothetical monopolist, which makes the price increase more profitable.

2104 Using this setup, one can then express the hypothetical monopolist’s actual loss in terms of the other parameters.  Professor Asker derived the expression for the actual loss as follows.

2105 Under the standard Lerner index, the elasticity of product i (ϵ) and the margin for product i (M) have the following relationship:

2106 For an X percent increase in price, the (absolute) percent change in quantity sold is shown by the following formula, but note that the percentage change in quantity sold is approximated by Xϵ for small values of X. This is because  represents the percentage change in quantity in response to a 1 % change in price, while  represents the percentage change in price, so their product is the percentage change in quantity in response to a X percent change in price:

2107 Because of the recapture, the actual loss for the hypothetical monopolist is:

2108 Using the critical loss formula, this means that the price increase is profitable if A, that is, the actual loss is less than the critical loss, which happens if and only if D>L, that is, the aggregate diversion ratio is greater than the critical loss.  This can be seen as follows:

2109 Let me turn to the platform critical loss threshold.

2110 As Professor Asked explained, the difference between the standard critical loss analysis and that for a platform amounts to correctly accounting for developer pass-through, but the overall intuition remains the same.

2111 A platform charges a service fee τ onto the price charged by developers p. For simplicity, assume there is one good on the platform with a representative developer that charges a price of p. These assumptions are reasonable as accounting for developer heterogeneity would not change Professor Asker’s conclusions.

2112 The platform raises its service fee from τ₀ to τ₁. This results in a change in the price set by developers from p₀ to p₁ and the quantity sold changes from q₀ to q₁. In a market that passes the HMT, the service fee increase would be profitable:

2113 Professor Asker defined the following terms:

(d) the increase in the service fee in percent terms is defined as:

(e) the increase in the platform’s marginal revenue (in percent terms) is defined as:

(f) the decrease in quantity, in percent terms and in absolute value, is defined as:

(g) the initial gross margin for the product is defined as:

2114 First, he derived the platform critical loss threshold.

2115 To start, he subtracted τ₀p₀q₁ from both sides of the inequality above and rearrange terms.

2116 He divided each side by τ₀p₀ and substituted the terms above:

2117 Then he added and subtracted X_pq₀ to the right-hand side and re-arranged further:

2118 Professor Asker noted that the net price increase can be expressed in terms of elasticities:

2119 Next, Professor Asker derived the actual loss formula for the platform.

2120 Under the platform Lerner index, the consumer price elasticity (ϵ_pq), the developer pass-through elasticity (ϵ_τp), and the margin for product i (M) have the following relationship:

2121 For an X percent increase in the service fee, the (absolute) percent change in quantity sold is:

2122 Because of the recapture, the actual loss for the hypothetical monopolist is:

2123 Using the critical loss formula, this means that the price increase is profitable if A, which occurs when:

2124 In the formula, μ is an adjustment term that is small, positive, and a function of the developer pass-through elasticity.

2125 This means that actual loss is less than the critical loss if the aggregate diversion ratio is greater than a term that is slightly larger than the critical loss threshold.

2126 In M Katz and C Shapiro (2003), the authors recommend using a 10% increase for the critical loss analysis.  The logic for this thought experiment is as follows.

2127 The HMT asks if the profit-maximising price increase is at least above the usual threshold of 5%.  But a given price increase can be profit-increasing without being profit-maximising.

2128 Under some assumptions, for example, linear demand or close approximations, the profit-maximising price increase is half as large as a price increase that would leave the hypothetical monopolist’s profits unchanged.  This is because beyond the profit-maximising price increase, the hypothetical monopolist’s profits will decline up until a point that it reaches the initial profits.

2129 This means that if a 10% price increase is profit-increasing, then the profit-maximising price is at least 5%.

2130 This can be illustrated through an example.

2131 Consider a hypothetical monopolist of two products, which has the following profit function:

2132 Suppose one had a linear demand of the form:

2133 Call the current prices observed as p₁* and p₂*. Next, consider that the hypothetical monopolist of both products wishes to raise the price of p₁ only. The firm would choose the profit-maximising value of p₁, while holding  constant at p₂*, by maximising the following:

where the coefficients A,B,C are a combination of the other constant terms.

2134 This formulation makes it clear that the profit function to be maximised is quadratic in p1. Call the solution to this p₁**.

2135 Given that one knows the hypothetical monopolist’s profit-maximising price is larger than the current prices, that is, p₁**>p₁*, one knows that the firm’s profit function looks as shown in Professor Asker’s exhibit 112 below:

2136 This is because a firm that owns the two rival products would have an incentive to set prices higher than a firm that only owns one of the products.  In particular, due to the concavity of the profit function, one knows that there is a price p₁M>p₁**>p₁* such that:

2137 Moreover, due to the symmetry of the parabola, one also knows that p₁** is halfway between p₁* and p₁M.

2138 Now, consider a 10% price increase. If this price increase is profitable, this means that the new price must be between p₁^* and p_M*. This means that there are two cases, both of which imply that the profit-maximising price is greater than 5%.

2139 If the 10% price increase falls between p1*and p1**, we know that p1** is greater than a 5% increase.

2140 If the 10% price increase falls between p₁** and p₁M, we know that a 5% increase falls between p₁* and p₁**. This means that p₁** is greater than a 5% increase.

2141 A similar logic also applies to the platform case, according to Professor Asker. This means that when analysing the platform critical loss analysis, it is also appropriate to use a 10% price increase.

2142 Consider a hypothetical monopolist of two platforms, which has the following profit function:

2143 Suppose one had a linear demand of the form:

2144 In addition, suppose that the developer price function is linear:  pi= fi+gi τi. The service fee τi only enters the consumer demand function through . Call the current service fees observed as τ₁* and τ₂*.

2145 Next, consider a hypothetical monopolist of both products who wishes to raise the price of τ₁ only. The firm would choose the profit-maximising value of τ₁, while holding τ₂ constant at τ₂*, by maximising the following:

2146 In the formula, the coefficients A,B,C,D are a combination of the other constant terms.

2147 This formulation makes it clear that the profit function to be maximised is cubic in τ₁. Call this solution τ₁**.  Due to diminishing returns, economists typically assume that profit functions are concave, which guarantees a solution to the firm’s profit-maximisation problem, that is, firms do not earn infinitely higher profits by setting higher prices. This property ensures that the cubic function set out has a local maximum.

2148 Whilst a cubic function is not symmetric around its local maximum as a quadratic function is, as M Katz and C Shapiro (2003) recommend, one can use a quadratic approximation of the function around τ₁**. A common approach for approximating any function is called a Taylor series, which more accurately approximates a function the higher the order of the Taylor series. This means that a quadratic approximation, that is, a second-order Taylor series approximation, will closely approximate a cubic function, that is, a third-order polynomial. This allows one to use a similar logic and apply the 10% threshold as the basis for the critical loss test.

2149 Now putting to one side questions of inputs, the underlying assumptions and other issues between the parties such as the distinction between a profitable step and a profit maximising step, there did not seem to be an issue between the parties concerning Professor Asker’s derivations including his algebra.  On a lighter note I should add that although Professor Asker professed to be an Aristotelian in his methods, I suspect that he is a closet Platonist given his predilection for such mathematization and his attraction to pure forms of economic theory such as prioritising the significance of profit maximising steps over profitable steps.  Let me turn to the Lerner index.

The Lerner index — general

2150 Professor Asker gave the following evidence concerning the Lerner index.

2151 Determining what constitutes a “high” margin requires an understanding of the institutional features of the market. And what is high in one context may be unremarkable in another.

2152 Now economists may study whether a margin is “high” by examining what the margin implies about customer elasticities using a relationship called the Lerner index.

2153 The Lerner equation states that if a product is priced to maximise the profits from that product, then the proportional gross margin, m, will be the inverse of the elasticity of demand facing the product ε: that is, ε = 1/m.

2154 As I have indicated earlier, the Lerner index reflects the relationship between a firm’s margins and the elasticity of customer demand the firm faces. So, when consumers are less price-sensitive, firms can charge a higher price and enjoy higher margins. What this means is that when a firm’s customers have many alternatives that they can easily substitute to, that is, when customers are very price sensitive so the firm faces very elastic customer demand, then the firm cannot earn high margins.

2155 When economists view high margins as indicative of market power, it is because these high margins are understood to imply low customer price elasticity, indicating that the firm has pricing power.

2156 Professor Asker said that the traditional Lerner index does not apply in this setting because Google in respect of the Play Store does not set prices for customers that choose where to allocate the quantity of transactions from which Google through the Play Store monetises. Instead, the relevant profits depend upon two prices, being the service fee Google sets for the Play Store and the price developers set in response to that service fee. This means that Google sets its prices by considering both how developers respond to its service fee changes as well as the knock-on effect of how consumers then respond to the developer’s price change.

2157 Now according to Professor Asker, the Lerner index traditionally, and the version that he re-derived, cannot account for all possible substitution possibilities by consumers and developers. Instead it accounts for all substitution possibilities by consumers, who choose where to spend their money, but restricts developer responses to adjusting their price for Play Store content.

2158 In this way, even the re-derived Lerner index understates the competitive constraints facing the Play Store.

2159 To take into account these features of the Play Store, Professor Asker re-derived the Lerner index for a platform. The key property of this platform Lerner index is that it relates the Play Store’s margins to two elasticities being the developer pass-through elasticity and the elasticity of consumer demand with respect to prices.

2160 So, Professor Asker needed information on developer pass-through elasticities to draw inferences about consumer elasticities from the Play Store’s margins. Absent this information, he made a conservative assumption of a high developer pass-through elasticity. This is because for a given value of the Play Store’s profit margins, he derived the lowest estimate of consumer elasticity consistent with that margin and therefore, the strongest indicator that the Play Store has market power, were he to assume the highest value of developer pass-through elasticity.

2161 Professor Asker gave the following derivation of the Lerner index for a platform.

Derivation of the Lerner index for a platform

2162 The Lerner index embodies or reflects the relationship between a firm’s margins and the elasticity of customer demand (ϵ_pq) it faces.

2163 The standard Lerner index for a firm setting a price p for a good it produces with marginal cost c is:

2164 is the total revenue earned by the firm. This is equal to pQ, where  is the per-unit price of the good and Q is the quantity sold. TVC is the total variable cost incurred by the firm. This is equal to cQ, where  is the per-unit cost of the good and Q is the quantity sold. ϵ_pq is the consumer price elasticity for the good.

2165 The equation relates the per-unit margin the firm charges to the inverse of the price elasticity of customer demand facing the individual firm. The less price elastic a firm’s consumers, that is, the larger ϵ_pq is, the higher the per-unit margin the firm can charge.  This means that when consumers are less price sensitive, firms can charge a higher price and earn higher margins. So, if the firm operated in a perfectly competitive environment, it would face a perfectly elastic demand curve, that is, consumer price elasticity would be infinite, and so would set a markup of zero.

2166 Now as I have already indicated, Professor Asker explained that the traditional Lerner index does not apply in this setting because the Play Store does not set prices for consumers. Instead, the Play Store sets the service fee faced by developers (τ); the developers then set the price faced by consumers (p).

2167 So, Professor Asker re-derived the appropriate Lerner index for this setting so that the Play Store’s margins relate to two elasticities being the elasticity of consumer demand in response to developer pricing (ϵ_pq), and the elasticity of developer pricing in response to the service fee (ϵ_τp). The latter is also known as the elasticity of pass-through which poses the question: by what percent does developer pricing change in response to a given percent change in service fees?  He called his formula the platform Lerner index being the following:

2168 Now this formula is more complex than the traditional Lerner index, and no longer relates high margins to necessarily low elasticities of consumer demand. As the platform Lerner index shows, high Play Store margins are consistent with high elasticities of consumer demand whenever pass-through elasticities are low.  And for a fixed level of platform margins, the lower the value of the pass-through elasticity, the higher the value of the implied consumer elasticity. This is a direct implication of the platform Lerner index that Professor Asker derived.  It is appropriate to set out his derivation, which if I might say so was an impressive piece of work in theory although it had its limitations in the present context which I will discuss later.

2169 Consider the profit function (Π) of a platform such as the Play Store:

2170 is the set of digital content that is distributed via the platform, which are indexed by k. τ is the platform’s service fee.  is the platform’s marginal cost for distributing digital content. pk is the price set by the developer for content k. q_k is the quantity sold of content k.

2171 Under this profit function, the platform chooses the service fee (τ) to maximise its profits according to the following first-order condition:

2172 Unlike the standard Lerner index, this formulation takes into account that the price set by developers (p_k) and the quantity of digital content sold (q_k) can be affected by the platform’s service fee.

2173 Now for ease of notation, Professor Asker assumed that there is one good and suppressed the k subscript. This is consistent with assuming that all digital content is distributed by a representative developer, which implies that the developer’s pricing power for their content with respect to consumers is uniform across all content on the platform. He largely made this assumption for mathematical clarity. Allowing for heterogeneity in developer pricing power would make it necessary to keep track of more elasticities, but the intuition of the end result would remain the same. Further in this context, when only a small number of developers make up the majority of the Play Store’s spending, this would be equivalent to focusing only on the content distributed by these competitively meaningful developers.  This simplifies the first-order condition to:

2174 Next, he rearranged terms to express this equation as follows:

2175 To simplify this further, he assumed that the quantity of content sold q is only affected by the service fee  through the price of the content p. In other words, this assumption means that consumers are only affected by the service fee via the price they pay for the digital content. This assumes, for example, that developers do not change the quality of their product or remove their product from the platform in response to a change in the service fee.  Professor Asker said that this assumption was conservative for his purposes. It allowed him to provide intuition that high platform margins are consistent with elastic consumer demand but does not capture the full range of substitution possibilities available to developers. This allowed him to express the above equation as follows:

2176 To express this in similar terms to the standard Lerner index, he noted the following definitions:

(c) TR = τpq is the total revenue of the platform.

(f) TVC = cq is the total variable cost of the platform.

2177 The consumer price elasticity of demand with respect to the price of the content is:

2178 The developer’s price elasticity with respect to the platform’s service fee, that is, the pass-through elasticity is:

2179 Using these definitions, Professor Asker expressed the above equation as the platform Lerner index:

2180 Now according to Professor Asker, this platform Lerner index reveals a key difference from the standard Lerner index. The per-unit mark-up charged by a platform depends on both the consumer price elasticity of demand and the developer’s pass-through elasticity. This means that a platform having high margins can be consistent with consumers having high price elasticities and developers having sufficiently low pass-through elasticities.

2181 And according to Professor Asker, in the case of the Play Store, high consumer price elasticities are consistent with the various estimates of the Play Store’s margins even under an assumption of 100% developer pass-through.

2182 Let me now set out some evidence of Ms Edwards-Warren.

Ms Edwards-Warren’s evidence

2183 Now as should be apparent from what I have set out earlier, Professor Asker used two critical loss equations, being the standard critical loss equation and a platform critical loss equation.  The latter accounts for the fact that the Play Store charges a service fee, being a percentage commission, rather than a unit price.

2184 But according to Ms Edwards-Warren, neither is well suited for a market like app distribution which is characterised by high fixed costs, close to zero marginal costs and large economies of scale.

2185 Further, for such markets, Ms Edwards-Warren said that the Lerner index on which the critical loss is based is an inadequate tool for approximating price elasticity.

2186 Further, Professor Asker assumed a 100% pass-through rate from developers to users which was said to be conservative, but according to Ms Edwards-Warren this is incorrect and the test is more likely to be passed with a lower pass-through rate.  Ms Edwards-Warren considered it unlikely that the pass-through rate is very high or close to 100%.

2187 Further, she said that Professor Asker used a diversion ratio from the Play Store to the Samsung Galaxy Store calculated from the Fortnite hotfix analysis, which is flawed.

2188 Further, she said that Professor Asker did not consider whether diversion ratios from the Play Store to other Android app stores are suppressed because of the relevant Google conduct.

2189 Let me deal with another point.

2190 The results of the analysis of Professor Asker show that the aggregate diversion ratio is higher than the critical loss threshold when applying a 5% SSNIP, which suggests that the market for Android app stores is well defined. The results show that the aggregate diversion ratio is lower than the critical loss threshold when applying a 10% SSNIP, which suggests that the market should be widened.  And Professor Asker argued that these results show that the Android mobile app distribution market fails the HMT and one should rely on the 10% SSNIP threshold.

2191 But Ms Edwards-Warren said that even if the diversion ratios used by Professor Asker were well estimated, which she did not believe they were, she considered that these results were at best borderline, and did not point strongly either to the market being too narrow or too wide. Further, she said that the failure of the test at the 10% level could be an indication of the cellophane fallacy.

2192 In her view, because of these flaws in the design, implementation and interpretation of the test, it is not appropriate to place any weight on the results of the critical loss analysis of Professor Asker.

2193 But even if one accepted the methodology and inputs, she said that the results do not universally suggest that the market should be wider than the Android mobile app distribution market.

2194 Furthermore, the implication of the analysis is that, if one were to accept that the market should be widened, such that the test is re-run to include direct downloading onto Android devices in the market, the test passes under all scenarios.

2195 So, Ms Edwards-Warren said that at most the analysis suggests that the market should include direct downloads in the market for Android mobile app distribution, but not that it should be widened further.

2196 And according to Ms Edwards-Warren, this slightly wider market definition would have a limited impact on her analysis, for example, by slightly changing her market share calculations, and would not have affected her overall conclusions on the effect or likely effect of the relevant Google conduct.

2197 In summary, according to Ms Edwards-Warren, to the extent that I place weight on Professor Asker’s results, the correct interpretation is that the market for Android mobile app distribution might include direct downloads, and that it should not be widened any further. Ms Edwards-Warren explained in detail the reasons for her views and it is appropriate to set out some aspects.

Critical loss equations – standard and platform

2198 Now as she said, a standard critical loss analysis is sometimes used as part of the market definition exercise in competition inquiries. It seeks to determine whether actual losses from a price increase would be higher or lower than the critical loss threshold. If the test finds that the actual losses are higher, that means that the price increase is unprofitable and so the market needs to be widened.

2199 The standard formula to calculate the critical loss threshold is 𝐶𝐿=𝑥/(𝑚+𝑥), where 𝑚 is the percentage margin and 𝑥 is the percentage price increase.

2200 Ms Edwards-Warren’s table 11 below sets out illustrative critical loss thresholds for different combinations of percentage margins for both 5% and 10% price increases. It illustrates how the critical loss threshold decreases as the margin increases.

2201 If a firm earns a margin of 50%, a 5% price increase results in a 9% critical loss threshold. That is, if the actual lost revenue from a price increase over a set of products is larger than 9%, the price increase is unprofitable, which implies that the relevant market is wider than this set of products. If the margin is instead 75%, the actual loss only needs to be larger than 6% to render a 5% price increase unprofitable.

2202 Now as I have explained, Professor Asker extended the standard critical loss formula to account for the sensitivity of developer pricing in response to an increase in the service fee. Professor Asker derived a platform Lerner index to calculate the critical loss threshold.  To do so, and as I have explained, he adjusted the standard Lerner index formula to additionally account for the elasticity of developer pricing in response to an increase in service fee. He argued that this adjustment was needed as the Play Store sets service fees for developers, who in turn set prices for consumers, hence a measure of pass-through should be included in the platform Lerner index.

2203 The platform Lerner index formula derived by Professor Asker therefore reflects the relationship between a platform’s margins and the consumer price elasticity of demand and developer pass-through elasticity. It accounts for the fact that the service fee is not imposed directly on consumers.

2204 Using this index, Professor Asker demonstrated that a high margin can be associated with high elasticity of smart mobile device user demand for digital content. Professor Asker used this finding to support the argument that the high margin reported by Professor Davila cannot be used as evidence of Google’s market power.

2205 But according to Ms Edwards-Warren, neither the use of the standard Lerner index nor the specific modification that Professor Asker derived is well suited for use in the present case.

2206 Ms Edwards-Warren said that the Lerner index has significant limitations in the digital world where there are high economies of scale/scope and close to zero marginal costs.

2207 When estimating economic margins and therefore the Lerner index, the only relevant costs are variable costs.  Ms Edwards-Warren said that this is a well-known flaw with the Lerner index. A high Lerner index may not reflect market power, but rather a firm with high fixed costs.

2208 Where firms have high fixed costs, they may price substantially above their marginal costs to cover their fixed costs. These substantial fixed costs will lead to a high Lerner-index, whether the market is highly competitive or monopolistic.

2209 The markets relevant to this case are characterised by these features, being high fixed costs, close to zero marginal costs, and large economies of scale.

2210 In E Lindenberg and S Ross, “Tobin’s q Ratio and Industrial Organization” (1981) 54(1) Journal of Business 1, the authors examine the adequacy of using the Lerner index in such markets with high fixed costs and large economies of scale and conclude that (at 28):

… In these cases, the Lerner index is inadequate in explaining market power because it does not recognize that some of the deviation of [price] from [marginal cost] comes from either efficient use of scale or the need to cover fixed costs and does not contribute to market value in excess of replacement cost. …

2211 App distribution is characterised by high fixed costs, close to zero marginal costs and large economies of scale.  For such markets, the Lerner index is an inadequate tool for approximating price elasticity.

2212 Further, Ms Edwards-Warren said that the platform Lerner index derived by Professor Asker leads to inconceivably high elasticities under various assumptions.

2213 In particular, exhibit 85 of Professor Asker’s report shows that where the assumed margin is lower, say 34.7% in 2013, the implied price elasticity of demand for the Play Store is 12.5, significantly higher than any of the examples produced by Professor Asker.  This high elasticity estimate would imply that demand for the Play Store is more elastic than that of Kellogg’s corn flakes, Mountain Dew, automobile models and Milwaukee’s Best beer, which are all the examples used by Professor Asker.

2214 Given the evidence of the Play Store's very high share of app downloads presented in her first report, Ms Edwards-Warren considered the elasticities inferred from the platform Lerner index to be inconceivably high, reflecting difficulties with relying upon the platform Lerner index.

The inputs into the critical loss equation

2215 Now as Ms Edwards-Warren discussed, the standard critical loss analysis of Professor Asker requires an estimate of Play Store margins.

2216 Professor Asker considered that robust economic margin estimates are not available. He did not have the data that would enable him to accurately estimate them. Nonetheless, to conduct his analysis he proceeded using the accounting margins estimated by Professor Davila and Mr Holt. This was reasonable given the lack of other estimates on file, and notwithstanding that accounting margins and economic margins may vary, with the latter taking into account opportunity cost.

2217 The platform critical loss threshold of Professor Asker required an estimate of Play Store margins, and an estimate of the developer pass-through elasticity which measures the extent to which developers pass on a change in the service fee to users.

2218 To calculate the developer pass-through elasticity, Professor Asker divided the product of an assumed pass-through rate, the percentage service fee and the percentage increase in the service fee, by the percentage increase in service fee.

2219 This calculation reflects the percent by which developer pricing changes in response to a given percent change in service fees.

2220 Professor Asker assumed that the pass-through rate is 100%, that is, that developers pass-through service fee increases in full to users.

2221 But Ms Edwards-Warren did not consider this to be supported by the evidence. Further, Professor Asker said that this assumption was conservative, but Ms Edwards-Warren was of the view that that was incorrect.

2222 She said that the critical loss threshold is lower when a lower pass-through rate is assumed. Given the way that the test was applied by Professor Asker, being an assessment of whether aggregate diversion to Samsung is higher than the critical loss threshold, a lower threshold makes it more likely that the test will pass and the Android mobile app distribution market will be well defined.

2223 Conversely, assuming 100% pass through, as Professor Asker does, makes it more likely that the test will fail and the Android mobile app distribution market will not be well defined.

The actual losses as compared with the critical loss threshold

2224 Now as Ms Edwards-Warren explained, the usual application of a critical loss analysis is to compare the critical loss threshold, which is the percentage lost revenue which would make a price increase unprofitable, and actual losses, to assess whether they are greater or less than the critical loss.

2225 If actual losses exceed the critical loss, then a price increase is not profitable. The HMT fails and the market should be widened.

2226 Professor Asker did not use estimates of actual losses to compare to the critical loss threshold. Instead, he compared the critical loss to estimates of the aggregate diversion ratio from the Play Store to other Android app stores.

2227 Professor Asker relied on a finding in the economic literature that the actual loss is less than the critical loss, and the market is correctly defined, if the aggregate diversion ratio to other products in the candidate market is more than the critical loss. Professor Asker used aggregate diversion ratios instead of actual loss estimates.

2228 Professor Asker therefore compared the critical loss threshold, which is the percentage lost revenue which would make a price increase unprofitable and the proportion of switching from the Play Store to other products in the candidate market being the Samsung Galaxy Store to assess whether losses are greater or less than the critical loss.

2229 If the aggregate diversion ratio is higher than the critical loss, then a price increase is profitable. The HMT passes and the market is well defined. So, if the diversion ratio from the Play Store to the Samsung Galaxy Store exceeds the critical loss threshold, Professor Asker would consider the market for Android mobile app distribution to be well defined.

2230 Ms Edwards-Warren said that she had not seen the test applied this way in practice, but she understood that it is a tool that has been used in merger assessments in the US.  She queried Professor Asker’s application of the aggregate diversion ratios approach to a conduct case such as that before me.  I will return to this topic later.

2231 Professor Asker also did not consider whether diversion ratios from the Play Store to other Android app stores were suppressed because of the relevant Google conduct.  Moreover, the diversion ratio estimate from the Play Store to the Samsung Galaxy Store that Professor Asker relied upon is the one estimated from the Fortnite hotfix analysis.

2232 Ms Edwards-Warren considered these diversion ratios to be incorrectly estimated.

The results of Professor Asker’s analysis

2233 Now Professor Asker found that the standard critical loss threshold for a 10% price increase during the period 2017 to 2022 is between 11.0% and 11.8%.  Further, he found that the platform critical loss threshold for a 10% price increase during the period 2017 to 2022 is between 14.5% and 15.4%.

2234 Moreover, as these thresholds are higher than the estimated aggregate diversion ratio to Samsung (9.4%), Professor Asker found that the test fails, and that the market should be wider than the Android mobile app distribution market.

2235 Now as Ms Edwards-Warren noted, Professor Asker did not present results for a 5% price increase. He explained in a footnote that the test passes if conducted using a 5% price increase:

… the critical loss analysis that I conduct indicates it is not profitable for a hypothetical monopolist of Android app stores to raise prices by 10 percent, but it is profitable for them to do so by 5 percent under the conservative assumptions embodied in the critical loss test I evaluate… …because my specific application of critical loss analysis is conservative… …I do not view a critical loss analysis with a 5 percent threshold as dispositive for market definition.

2236 Now as to the interpretation of these results for the definition of the market, Ms Edwards-Warren said that putting to one side the errors in Professor Asker’s analysis and taking the results presented by Professor Asker as given, the results from the critical loss analysis would suggest that the relevant market should be wider than the Android mobile app distribution market.

2237 But the next appropriate step is to include the next closest substitute in the market and re-run the analysis. In her view, direct downloads were the only other option in the data/analysis available on the same device for Fortnite users on Android. Other options involved switching devices.

2238 Ms Edwards-Warren said that conducting the same critical loss analysis on an Android mobile app distribution market and including direct downloads results in a limited expansion of the Android mobile app distribution market which is sufficient to define a relevant market.  And in her view there was no further need to expand the market to include any other alternative platform.

2239 Specifically, this limited expansion of the Android mobile app distribution market results in an aggregated diversion ratio of 21.2%. Further, this aggregated diversion ratio is higher than any standard or platform critical loss threshold estimated by Professor Asker between 2017 and 2022.

2240 So, Ms Edwards-Warren said that a market combining both the Android mobile app distribution market and direct downloads would be well defined.

2241 Let me turn to another topic concerning the interpretation of the difference in results under 5% and 10% price increases.

2242 Professor Asker asked whether a hypothetical monopolist of Android mobile app distribution would find it profit-maximising to increase the Play Store service fee by at least 5%.  I have discussed elsewhere that one should consider whether a hypothetical monopolist would find it profit-increasing to raise prices.

2243 Now Professor Asker found that it is not profitable for a hypothetical monopolist of Android app stores to raise prices by 10%, but is profitable for them to do so by 5%.  That is, on this version of the test, while the Android mobile app distribution market fails Professor Asker’s critical loss test using a 10% price increase, it passes using a 5% price increase.

2244 Ms Edwards-Warren said that such a result could be consistent with the Android mobile app distribution market being the relevant market and the failure of the test at a 10% price increase being the result of the cellophane fallacy.

2245 Professor Asker did not consider such interpretations of the result that the Android mobile app distribution market passes the test using a 5% increase but fails using a 10% increase.

2246 Instead, he dismissed the results using a 5% price increase.  First, he said that the result that a 10% price increase is not profit-increasing means that it is not profit-maximising to raise prices by 5%.  Second, he said that given the conservative assumptions applied, a 5% price increase being profit-increasing is not dispositive for market definition.

2247 Now in Ms Edwards-Warren’s view, a more appropriate interpretation is that these results are borderline at best, even if one accepted that Professor Asker’s inputs into the calculation were accurate. That is because the test passes under a 5% price increase and fails under a 10% price increase.

2248 Furthermore, Ms Edwards-Warren said that the results are sensitive to the inputs relied upon.  Let me elaborate on her evidence in that respect.

2249 Now to further test the sensitivity of the critical loss analysis to the value of the inputs assumed by Professor Asker, Ms Edwards-Warren computed critical loss thresholds based on a range of values for each input.

2250 Now the standard critical loss threshold requires an estimate of the Play Store margin.  The platform critical loss threshold requires an estimate of the Play Store margin and an assumed pass-through rate of the increase in service fee by developers to consumers.  Both thresholds were to be compared to an estimate of the aggregate diversion ratio to Samsung. The test was carried out for both a 5% and 10% price increase.

2251 Ms Edwards-Warren examined the range of standard and platform critical loss thresholds that arise from different combinations of these inputs and her table 12 below summarised the range of values which she selected for each input to the critical loss threshold calculations and explained why those ranges were selected.

2252 Ms Edwards-Warren then compared the standard and platform critical loss thresholds estimated from the range of inputs above to estimates of the aggregate diversion ratios for the Samsung Galaxy Store using diversion ratios estimated by Professor Asker as well as those estimated by Professor Goldfarb based on a similar four-week time period for comparability.

2253 Ms Edwards-Warren’s figure 8 below presents the results. It compares the standard critical loss thresholds estimated based on a range of inputs to diversion ratios for the Samsung Galaxy Store from Professor Asker and from Professor Goldfarb.

2254 The horizontal axis shows the increase in the service fee, that is, the SSNIP being approximated. It ranges from 5% to 10%.

2255 Further, the vertical axis shows the critical loss threshold that results depending on the different combinations of inputs. The grey region on the graph shows all possible outcomes for the critical loss threshold depending on the assumptions with respect to margin and pass-through rate.

2256 Further, the red region identifies the range where diversion to the Samsung Galaxy Store would be insufficient for the Android mobile app distribution market to pass the HMT, and therefore the Android mobile app distribution market is not well defined, under any standard critical loss threshold.

2257 Further, the green region identifies the range where diversion to the Samsung Galaxy Store would be sufficient for the Android mobile app distribution market to pass the HMT, and therefore the Android mobile app distribution market is well defined, under any standard critical loss threshold.

2258 Further, the red dashed lines reflect diversion ratios estimated by Professor Asker.

2259 Further, the blue dashed lines reflect diversion ratios estimated by Professor Goldfarb based on a similar four-week time period as used by Professor Asker.

2260 Ms Edwards-Warren said, and I accept, that the figure shows that, depending on the assumptions made about pass-through and margins, there is a range of critical loss thresholds.

2261 If one posits a 5% price increase, the test passes under all possible combinations of critical loss threshold and diversion ratio.

2262 It is only when one posits a price increase of more than 8% that the test passes under some of Professor Asker’s diversion ratio estimates under some scenarios reflecting certain assumptions regarding margins and pass-through.

2263 The results using the diversion ratios of Professor Asker do not universally point to the Android mobile app distribution market being poorly defined.

2264 The results using the diversion ratios of Professor Goldfarb universally point to the Android mobile app distribution market being well defined.

2265 Ms Edwards-Warren’s figure 9 below presents a similar comparison to the platform critical loss thresholds estimated based on a range of inputs outlined in table 12 above.

2266 Similarly, the figure above shows that diversion ratios estimated by Professor Asker are sometimes above and sometimes below the range of critical loss thresholds, and they do not universally point to the Android mobile app distribution market being well defined or not.

2267 The diversion ratios to the Samsung Galaxy Store estimated by Professor Goldfarb, based on a similar four-week time period used by Professor Asker, are higher than all platform critical loss thresholds calculated, thus universally pointing to the Android mobile app distribution market being well defined.

2268 Let me turn now to Google’s principal arguments based upon the evidence that I have just summarised given by Professor Asker and some criticisms of the evidence of Ms Edwards-Warren.  Some of what follows will concern matters I have already touched upon in my discussion of Professor Asker’s and Ms Edwards-Warren’s views.

Google’s arguments

2269 Google says that Professor Asker demonstrated using a critical loss analysis that the proposed Android mobile app distribution market fails the HMT with a 10% price increase based on the substitution patterns he observed in the data.

2270 Couching the matter in profit-maximising terms, Google says that critical loss analysis starts by asking the following question: would a hypothetical monopolist of the candidate market find it profit-maximising to increase the price of one of its products, say Product A, by at least X %? The profit-maximising price increase is at least X % if an increase of twice the size (2X %) is profitable. This occurs when the actual loss in sales, for a 2X % price increase, is less than what is called the critical loss threshold.

2271 To determine the actual loss in sales, one must balance two competing effects. On the one hand, by increasing the price, the firm will lose customers from Product A. On the other hand, the firm, as the hypothetical monopolist, can recapture some of those lost consumers through the other products it owns, that is, everything else in the candidate market except Product A. The amount of lost Product A sales can be calculated using consumer elasticity, where the elasticity can be imputed using the Lerner index. The amount of recapture can be calculated using an aggregate diversion ratio, that is, how much is diverted to all other products in the candidate market.

2272 Google says that the benefit of the critical loss analysis is that it provides a parsimonious way of implementing a quantitative HMT, which in this case only requires four necessary inputs being margins, pass-through, diversion estimates and a hypothetical price increase. Let me elaborate a little on what I have already said.

2273 First, as to margin, this input is the margin for the product of interest, which is part of the Lerner index formula. For this, Professor Asker relied on gross profit margins as estimated by Professor Davila. Between 2017 and 2022, Professor Davila’s estimates for margins were approximately 75% to 81%. Epic relies on a similar range.

2274 Second, as to developer pass-through rate, this input measures how much developers pass on an increase in the service fee to consumers. This is relevant for the platform Lerner index that Professor Asker derives, which adapts the standard Lerner index to account for the fact that the Play Store does not set consumer prices directly.

2275 Google says that rather than estimating the pass-through rate, Professor Asker conservatively assumed full 100% pass-through as that implies the lowest possible consumer price elasticity. In other words, he assessed whether consumer substitution alone can discipline the hypothetical monopolist.

2276 Google says that because the structure of the critical loss test constrains developers to only substitute by passing through service fees, when assuming lower levels of pass-through it would be necessary to adapt the critical loss framework to allow for other ways developers can substitute.

2277 Now as I have already indicated, Ms Edwards-Warren disagreed with the assumption of full pass-through, but Google says that she provided no greater specificity than saying it is highly unlikely that there is full or even significant pass-through. Moreover, Google says that there is no empirical evidence to support Epic’s claim that pass-through is likely to be below 20%.

2278 Google says that in the absence of any evidence, all possible pass-through values between 0% to 100% should be considered to test the robustness of Epic’s proposed market, as Professor Asker did by evaluating the critical loss test using the standard formulae, which are equivalent to 0% pass-through.

2279 Third, as to diversion, this input is the aggregate diversion ratio, which in this case means the diversion from the Play Store to all other app stores in the proposed market. For this input, Professor Asker used his estimates of diversion to the Samsung Galaxy Store based on the hotfix analysis given that the Galaxy Store is the largest Android app store in Australia other than the Play Store and the only one that had Fortnite. As to the relevant diversion ratio estimates, his range was 10.1% to 13.2%.

2280 Fourth, as to price increase, this input is the hypothetical price increase in the service fee that is being tested to see if it is profitable.

2281 As I have already indicated, Professor Asker assessed whether a 10% price increase is profitable to test if the profit-maximising price is above 5%. I have said elsewhere that the relevant test should not focus on the profit-maximising price. Ms Edwards-Warren only required that the hypothetical monopolist find it profitable to increase its price by 5%.

2282 Now Google says that the Lerner index is appropriate and reliable in the circumstances of this case.

2283 As I have already said, underpinning the critical loss analysis is the Lerner index. The Lerner index is a standard formulation in economics that captures the relationship between margins and elasticities.

2284 Google says that the traditional Lerner index does not apply in this setting because the Play Store does not set prices for consumers. Consumer prices are set by developers. Professor Asker re-derived the Lerner index for this setting, which captures the fact that Play Store margins are related to both consumer elasticity and developer pass-through.

2285 Epic says that the Lerner index is not appropriate in this setting and that Professor Asker’s platform Lerner index is inappropriate. But Google says that Epic is wrong for the following reasons.

2286 First, Epic says that the Lerner index does not apply to Google because it does not act in a profit-maximising manner and therefore bias would be introduced. This claim is based on the suggestion that because of regulatory and political scrutiny, Google does not act in a profit-maximising manner.

2287 Google says that Epic points to no evidence to support its claim that Google is not profit-maximising, and nor did it put this proposition to any Google lay witness. Further, there is no evidence or analysis provided by Epic as to what direction the bias would go, if any, or to explain how or why regulatory or political pressure would somehow mean that Google is not acting rationally and in a profit-maximising manner.

2288 Second, Epic claims that the Lerner index is inappropriate in markets with high fixed costs, close to zero marginal costs and large economies of scale.

2289 Now Google challenges the high fixed costs characterisation of the Play Store’s position. But even if Google’s fixed costs are high, Google says that Epic and Ms Edwards-Warren have misinterpreted the academic literature. As I have already indicated, Professor Asker said that a well-known flaw of the Lerner index is that a high Lerner index may not reflect market power, but rather a firm with high fixed costs. This means that the Lerner index is conservative for estimating price elasticity and market power when analysing an industry with high fixed costs and close to zero marginal costs.

2290 Third, Google says that Epic mischaracterised the consumer elasticities implied by the platform Lerner index. Epic says that Professor Asker’s platform Lerner index is unreliable because it leads to inconceivably high elasticities that would imply price elasticities for Google Play Store that are greater than everyday consumables, such as soft drinks or breakfast cereals.

2291 It is said that Epic provides no explanation or reasoning for why demand for the Play Store should be more inelastic than demand for a box of cornflakes or a bottle of a carbonated beverage.

2292 In fact, the Play Store having more elastic demand is consistent with the numerous ways that consumers can transact for digital content outside of Google Play.

2293 Moreover, Epic also says that developer pass-through is low. But Google says that low pass-through is consistent with high consumer price elasticities, that is, developers would not pass on a price increase because consumers are highly responsive to price changes.

2294 Let me turn to another matter.

2295 As I have said, the critical loss analysis requires as an input the aggregate diversion in response to the hypothetical price increase. Professor Asker used the best estimate for this, being the diversion from the Play Store to the Samsung Galaxy Store after the hotfix.  Google says that this is a reasonable estimate for the aggregate diversion ratio for the following reasons.

2296 First, the hotfix is a natural experiment that can be used to estimate substitution in response to a SSNIP.  Google says that this is not an extreme event, but an informative one.  Both Professor Asker and Professor Goldfarb considered the hotfix to be a natural experiment that is useful for measuring substitution.

2297 It is standard practice for market definition by economists and competition authorities to rely on the exogenous entry or exit of a firm’s offering in a candidate product market to analyse diversion.

2298 For example, in P Davis and E Garcés (2009) the authors use the example of a plant closure to illustrate how to use a natural experiment to identify diversion ratios. Other natural experiments that have been studied include unexpected plant closures due to extreme weather conditions, the entry or exit of a single grocery store, and the entry or exit of a single petrol station.

2299 The diversion ratios estimated from these analyses are regarded as reasonable approximations to those estimated from a SSNIP.

2300 Second, despite Epic challenging Professor Asker’s analysis in cross-examination, Google says that Epic has not identified any evidence that the diversion ratios estimated from the hotfix would differ from those estimated from a smaller change in price or quality.

2301 Ms Edwards-Warren says that the diversion ratios could differ to the extent that the Play Store spender who diverts spending in response to a SSNIP (the marginal consumer) is different from the users who spent via the Play Store prior to the hotfix (the average consumer).

2302 But Google says that Ms Edwards-Warren did not evaluate that possibility despite the availability of the necessary data. Ms Edwards-Warren did not provide any evidence of the claimed difference, how large that purported difference is, and in what ways the estimates would be biased as a result of that purported difference.

2303 Third, Google says that Ms Edwards-Warren’s assertion is inconsistent with the economics of how marginal and average consumers would substitute in response to a SSNIP and the data available in this case. Epic simultaneously claims that the hotfix was extreme while also claiming that users were largely unaffected until the new Fortnite season was released two weeks after the hotfix.

2304 That would mean that the period immediately after the hotfix is more informative as to how marginal consumers behave rather than average consumers, consistent with Professor Asker’s view that the initial hotfix period is closer to a SSNIP than the new season release because it was a smaller quality degradation.

2305 During that period, Professor Asker illustrated that diversion was primarily to console stores and diversion to the Samsung Galaxy Store was limited. It was only after the new season release that diversion to the Samsung Galaxy Store increased, though this was short-lived.

2306 This indicates that, if anything, the marginal consumer is more likely to divert to console stores than the average consumer, being the opposite outcome to that asserted by Ms Edwards-Warren.  Let me move to another topic.

2307 Google says that the critical loss analysis is informative in conduct cases as well as in merger cases and has made the following points.

2308 The critical loss analysis has long been a widely used tool in competition analysis and critical loss analysis can be informative about market definition. Epic criticised Professor Asker’s use of the aggregate diversion ratios formulation of a critical loss test pioneered by Professor Katz and Professor Shapiro, and incorrectly said that the aggregate diversion ratios formulation is only appropriate in merger analysis and not in conduct cases. But Google says that Epic is wrong for the following reasons.

2309 First, the use of the aggregate diversion ratios formulation of critical loss analysis is to implement a quantitative HMT in a parsimonious way.  By Epic’s logic, so Google says, the aggregate diversion ratios formulation of the critical loss analysis would be appropriate to analyse a hypothetical merger between all Android app stores, but it would be inappropriate to analyse the conduct of a hypothetical monopolist of all Android app stores.

2310 Under Epic’s view, so Google says, the only valid critical loss analysis in a conduct case requires measuring actual losses rather than estimating those losses via the aggregate diversion ratios formulation. This is economically nonsensical. The standard aggregate diversion ratios formulation of critical loss is used because mathematically there is a one-to-one correspondence between the aggregate diversion ratios and actual losses. Actual losses exceed critical losses if and only if the aggregate diversion ratio is below the critical loss threshold. There is no suggestion that the underlying mathematical foundation presented in M Katz and C Shapiro (2003) is in dispute.

2311 Second, Epic relied on Ms Edwards-Warren’s assertion that the critical loss analysis has been criticised and that it is not a universally accepted tool. But Google says that Ms Edwards-Warren did not provide detail as to what these criticisms were and whether they entail that the critical loss analysis cannot be applied to a conduct case.

2312 Third, Epic said that it is inappropriate to use the aggregate diversion ratios formulation of critical loss analysis in a conduct case because the alleged conduct may have an impact on competition and also impact the level of observed diversion to products in the candidate market. To correct for this, Epic claims that a buffer should be built into the analysis.  But Google says that this contention is misconceived for several reasons.

2313 Google says that Epic neither explains how large the purported buffer should be nor provides any evidence that absent the conduct diversion to the Samsung Galaxy Store would be higher or by how much.

2314 It also says that Epic does not grapple with the widespread availability of the Samsung Galaxy Store.  Epic is merely speculating that including such a buffer would mean that its proposed market passes the HMT.

2315 Further, in any event, Google says that there are many ways in which Professor Asker’s critical loss analysis is already conservative and so builds in a buffer.

2316 For example, the critical loss analysis cannot measure diversion to the iOS App Store or to the web given that the hotfix also affected iOS, and given that Fortnite did not distribute content via the web at the time of the hotfix.

2317 Google says that at some level, Epic’s concern appears to be a variation on the theme of the cellophane fallacy discussed by Ms Edwards-Warren, which Professor Asker had already accounted for in his analysis. He did so by conducting the critical loss analysis using margins estimated by Mr Holt under what Mr Holt claims is a competitive service fee of 15%.

2318 Google says that Epic’s proposed market fails even under these margins, which indicates that the cellophane fallacy is not a concern.  Let me move to another matter.

2319 Google says that Epic’s criticisms regarding developer pass-through rates are unfounded.

2320 Epic criticised Professor Asker’s initial assumption of 100% developer pass-through and said that it is likely as low as 20%.  But Google says that I should reject this criticism.

2321 Google says that Epic has presented no evidence on plausible pass-through rates. Ms Edwards-Warren did not comment on current or reasonable values of pass-through and made it clear that she did not have an answer on the precise figures of pass-through.

2322 Whilst Ms Edwards-Warren said that it is unlikely for there to be significant pass-through, she provided no explanation of what “significant” means.

2323 Professor Asker explained that he considered low pass-through to be unlikely and chose a somewhat arbitrary cut of 20% to explain his reasoning. So, the assertion of developer pass-through being as low as 20% is simply without basis.

2324 In any event, Google says that Professor Asker considered a range of pass-through assumptions.

2325 Generally, Google says that Professor Asker’s critical loss analysis empirically demonstrates that the posited Android mobile app distribution market fails the HMT.

2326 The Android mobile app distribution market fails the HMT if the diversion ratio to the Samsung Galaxy Store is below the critical loss threshold. The values estimated for these two statistics are as follows.

2327 In respect of the diversion ratio to the Samsung Galaxy Store, the relevant diversion ratio estimates have been discussed by me earlier, and as to those on which Professor Asker relied, the range is 10.1% to 13.2%.

2328 Google says that in respect of the critical loss threshold, the plausible critical loss thresholds for a 10% price increase range from 11.3% to 15.4% when examining developer pass-through rates from 5% to 100% and gross margins from 75% to 80%.  And it says that on Epic’s claim that pass-through could be as low as 20%, the thresholds range from 11.3% to 12.5%.

2329 So, Google says that using a 100% developer pass-through rate and 80% margin, the critical loss threshold is 14.5%.

2330 With this threshold, Google says that Epic’s proposed market fails the HMT under every estimate that Professor Asker prefers.

2331 Further, Google says that using even a 5% developer pass-through rate and 80% margin, the critical loss threshold is 11.3%.

2332 So, Google says that Epic’s proposed market fails the HMT under 6 out of 8 estimates that Professor Asker recommended adopting, so it cannot be relied upon to conclude that the proposed market would pass the HMT. This is so particularly given the several conservative features of Professor Asker’s critical loss analysis, including that diversion to iOS is not taken into account by the hotfix analysis, that the analysis does not consider closer substitutes before weaker ones, and that the hypothesised price increase applies only to the Play Store rather than to all products in the candidate market.

2333 As a result, Google says that Professor Asker’s critical loss analysis demonstrates that, even when adopting the inputs advanced by Epic’s own experts, Epic’s Android mobile app distribution market does not have robust empirical support for a 10% SSNIP.

2334 Google says that it is not sufficient to find one specification where the market passes. Epic is required to show that its proposed market is well-defined, which would necessitate it passing critical loss under a range of reasonable assumptions and specifications.

2335 Google says that Epic has not discharged that onus. It says that it is incorrect to characterise these results as borderline, as Ms Edwards-Warren does, including for the following reasons.

2336 First, even if that characterisation were true, what is borderline are not the results, but the candidate market. The exercise is asking whether the market has robust support. It does not.

2337 Second, the characterisation of the results as borderline relies on using the 5% price increase as an informative input value. But it is not an informative value input, because such a formulation is inconsistent with the standard HMT.

2338 Third, the results indicate that the proposed market fails the HMT under a range of reasonable estimates. These results suggest that the market is too narrowly defined.

2339 Fourth, even if Professor Goldfarb’s preferred range of 10.1 to 18.7% diversion to the Samsung Galaxy Store were accepted, Google says that the plausible critical loss thresholds range from 11.3% to 15.4%. As such, Google says that it cannot be concluded from Professor Goldfarb’s diversion estimates that Epic’s market is well-defined. To do so would require putting greater precision on Professor Goldfarb’s estimates than he was willing to do.

2340 In summary, Google says that given that the critical loss analysis shows that the market fails the HMT, the candidate market must be expanded.

Analysis

2341 Now in my view Professor Asker’s critical loss analysis has several methodological difficulties as Ms Edwards-Warren pointed out and I have accepted.

2342 But even if those difficulties are put to one side, adopting reasonable estimates of relevant parameters leads to the conclusion that there is a well-defined Android mobile app distribution market. Let me descend into the detail, but noting at the outset that some of what follows concerns concepts or criticisms that I have already touched on.

2343 Now I accept that a critical loss analysis is a method of implementing a quantitative SSNIP test with limited inputs. It involves comparing a critical loss threshold with an estimate of actual loss. And the critical loss threshold depends on the SSNIP percentage being examined, such as 5 or 10%, and a percentage gross margin.

2344 If actual loss exceeds the threshold, the candidate market fails the HMT. This means that the market is not well defined and should be widened.

2345 As I have already indicated, estimates of actual loss following a SSNIP are not available in this case, and so in place of estimates of actual loss, Professor Asker relied upon an alternative methodology proposed by economists M Katz and C Shapiro (2003) in the horizontal merger context, which explains how aggregate diversion ratios can be used in place of estimates of actual loss to implement a critical loss analysis.

2346 In the aggregate diversion ratios approach, if aggregate diversion is lower than the critical loss threshold, the candidate market fails the HMT and the market is not well defined and should be widened. But if aggregate diversion is higher than the critical loss threshold, the market is well defined.

2347 Now as Ms Edwards-Warren explained, the essential concept in the merger context was to consider the hypothetical monopolist of products sold by two merging firms and to ask if prices were increased on one of the products sold by the hypothetical monopolist (Product A), whether the aggregate diversion in sales from Product A to the hypothetical monopolist’s other products would be sufficient to make up for the losses on Product A.  The aggregate diversion ratio is the share of diversion captured by other products within the candidate market.

2348 But in the present context, Professor Asker adapted the analysis by considering a hypothetical monopolist of all Android app stores and asking whether if the hypothetical monopolist increased its prices on the Play Store there be sufficient diversion to other Android app stores, specifically the Samsung Galaxy Store, to make up for the losses in sales to the Play Store. Professor Asker estimated the aggregate diversion using diversion ratios calculated from the Fortnite analysis.

2349 And using the aggregate diversion ratios approach, Professor Asker applied the price increase to just one product in the relevant candidate market, being the commission rate of the Play Store, rather than all products within the candidate market, being the commission rate of all Android app stores. Professor Asker considered this to be conservative.

2350 But as Ms Edwards-Warren explained, whilst that is true in theory, Professor Asker’s critical loss implementation is very close to asking about the hypothetical monopolist in the candidate market increasing its prices over all of its products, due to the prominence of the Play Store. Professor Asker accepted that due to the prominence of the Play Store, the distinction between applying the price increase to one versus all products in the candidate market is potentially small.

2351 Now as I have indicated, Professor Asker developed a platform critical loss threshold which assumed that developers pass through a percentage of a service fee change to users. At 0% pass through, the platform critical loss threshold is equal to the standard critical loss threshold. It sets the lower bound to the platform critical loss threshold.  And at pass-through rates above 0%, the platform critical loss threshold is always higher than the standard critical loss threshold.

2352 But Professor Asker’s adoption of the aggregate diversion ratios approach to a conduct case such as the present is questionable for the reasons that Ms Edwards-Warren explained.

2353 First, Professor Asker’s application of the aggregate diversion ratios approach to a conduct case is not supported by the academic literature that he relied upon. Professor Asker accepted that M Katz and C Shapiro (2003) concerned market definition in the context of a horizontal merger. Professor Asker also accepted that the authors did not suggest that it should be adopted in a case where there was alleged to be pre-existing anti-competitive conduct.

2354 As I have indicated, Ms Edwards-Warren noted that whilst she has not seen the aggregate diversion ratios approach applied in her experience, it is a tool that has been used in merger assessments in the US.  But I accept what she says that the aggregate diversion ratios approach has been criticised and it is not a universally accepted tool.

2355 Second, Professor Asker’s application of the aggregate diversion ratios approach to a conduct case has not been accepted by a US court in a non-merger case.

2356 Third, in a case where there is alleged anti-competitive conduct, the aggregate diversion ratios approach requires building into the analysis a buffer when using the diversion ratios for the firm affected by the alleged conduct. In such a case, the diversion to the second firm would be limited if the impugned conduct is found to be made out.

2357 Now Epic’s case is that Google has engaged in conduct which has had the effect of impairing the competitiveness of other Android stores, but Professor Asker did not build in any such buffer.

2358 Now Professor Asker said that no such buffer was needed, as his analysis omitted websites and iOS, which are factors taking away options that may otherwise attract people away from the Samsung Galaxy Store to other options. But I agree with Epic that Professor Asker has provided no evidence to suggest that these omissions would have a greater impact than the effect of the relevant Google conduct on the Galaxy Store.

2359 Let me turn to another question as to whether the diversion following the removal of Fortnite necessarily reflects what would occur in response to a small change in price or quality.

2360 Ms Edwards-Warren said that a methodological issue with the critical loss analysis is that, in relying upon the patterns of diversion following the removal to draw conclusions about how users would substitute in response to a small change in price or quality, Professor Asker assumed that the marginal customer behaves in the same way as the remainder of the consumer base.

2361 But as to the appropriateness of this assumption, Ms Edwards-Warren said that it depends on whether the marginal consumers who would respond to a price increase would respond in exactly the same way that all consumers responded to the Fortnite removal event. This is because when one is examining an extreme event, like the removal event, it is possible to conceptualise the event as an infinite price increase whereby no user can afford to purchase the product any longer, and so every user has to switch and make a switching decision.

2362 Now rather than presenting any analysis to support the reasonableness of this assumption, Professor Asker defended his position on the basis that he had seen no evidence to indicate that diversion ratios estimated from the removal of Fortnite would differ from those estimated from a smaller change in price or quality.  But I agree with Ms Edwards-Warren and Epic that this response is problematic given that his analysis of diversion in response to the removal event underpins his conclusions regarding market definition.

2363 Let me now turn to the Lerner index. As Epic points out and as Ms Edwards-Warren discussed in her evidence which I have already referred to, a difficulty with the critical loss analysis relates to the Lerner index.  Professor Asker’s Lerner indices were inputs into his critical loss analysis. Professor Asker relied upon the standard Lerner index and his modified platform version to calculate the standard and platform critical loss thresholds relevant to the Play Store.

2364 Professor Asker’s platform Lerner index leads to unrealistically high elasticities under various assumptions. So, for example, demand for the Play Store would be more responsive to price changes than demand for everyday consumables, such as soft drinks or breakfast cereals.

2365 Now exhibit 85 of Professor Asker’s report, which used Professor Davila’s margin estimates, shows that where the assumed margin is lower, say 34.7% in 2013, the implied price elasticity of demand for smart mobile device users is 12.5. This high elasticity estimate would imply that demand for the Play Store is more elastic than that of Professor Asker’s examples being Kellogg’s corn flakes and Milwaukee’s Best beer of all things.

2366 But given the Play Store's very high share of app downloads, Ms Edwards-Warren considered the elasticities inferred from the platform Lerner index to be inconceivably high, reflecting problems with relying on that index.

2367 Further, Ms Edwards-Warren analysed the price elasticity estimated from the Lerner index as a function of the service fee. Her figure 32 below plots the price elasticity of smart mobile device user demand as a function of the service fee, assuming a 77% profit margin, which is the 2022 margin reported in Professor Davila’s evidence.

2368 If one assumes that the service fee is lower than 30%, say at 15% being a level applied by Google in some contexts and cases, figure 32 shows that at a 15% service fee the implied elasticity of smartphone demand is around 10, being significantly higher than say Kellogg’s corn flakes.

2369 Further, the relevant literature establishes that the Lerner index is an inadequate tool for approximating price elasticity in relation to markets with high fixed costs, close to zero marginal costs and large economies of scale, therefore making the Lerner index not appropriate for this case.

2370 Now I have discussed the Lerner index, but let me address some other inputs into the critical loss analysis for a given price increase being the gross margin percentage and the developer pass through rate.

2371 Now as to the gross margin percentage, Google’s own records show a gross margin percentage for the Play Store between 2017 and 2022 of between [REDACTED].  This is consistent with Professor Davila’s calculations of the Play Store’s gross margins. Professor Easton did not challenge Professor Davila’s estimates of the Play Store’s gross margin in his report.

2372 Now as to pass-through and as I have already indicated, Professor Asker’s critical loss analysis assumed that developers pass-through 100% of the hypothetical price increase on the Play Store service fee to users.

2373 Professor Asker said that this was a conservative assumption, but this was not correct for the reasons given by Epic.  The higher the pass-through rate, the higher the platform critical loss threshold. The higher the pass-through rate, the less likely Epic’s proposed market would pass the HMT. The evidence supports a conclusion that a pass-through rate as low as 20% is a likely outcome.

2374 Now Professor Asker said that he applied a pass-through rate of 100% as it acted as an upper bound, while his standard critical loss threshold, which is the equivalent of the platform critical loss threshold with a pass-through of 0%, acted as a lower bound.  But there is nothing to suggest that an extreme pass-through rate of 100% is appropriate in this case.

2375 Professor Asker considered that economic theory indicates that pass-through of less than around 20% was unlikely, but explained that he did not have a basis to determine what pass-through might be above this somewhat arbitrary threshold.

2376 Ms Edwards-Warren considered that the rate is likely to be less than complete in industries with ad valorem pricing and low marginal costs. Ms Edwards-Warren considered it is highly unlikely that there is full or even significant pass-through.

2377 On the basis of her evidence, which I accept, I can reasonably regard a developer pass-through rate as low as 20% as a likely outcome.

2378 Further, a problem with Google’s position is that it is a consequence of its profit-maximising approach that the hypothetical monopolist must be able to profitably implement a 20% increase in prices.  But the SSNIP is usually referred to as a price increase of 5 to 10%.

2379 Further, as to the results of Professor Asker’s critical loss analysis, once it is accepted that Professor Goldfarb’s 10.1 to 18.7% range of Samsung diversion estimates are reasonable, and the 15.8% estimate is the most accurate, Google’s position becomes problematic.  As I have said earlier, I have accepted Professor Goldfarb’s estimates.

2380 In summary, I agree with Epic that Professor Asker’s methodological errors and assumptions bias his results towards finding a broader market. Professor Asker’s calculation of diversion to the Samsung Galaxy Store, being the aggregate diversion ratio, is flawed and unreliably low.

2381 Further, I agree with Epic that Professor Asker erred in assuming that 100% developer pass-through is conservative.

2382 Further, I agree with Epic that the aggregate diversion ratios approach is not well suited for non-merger cases where the aggregate diversion ratio may be suppressed by anti-competitive conduct.

2383 Now as Epic points out, one can correct for a number of these errors or model different parameters.  Professor Goldfarb corrected for Professor Asker’s errors in calculating the Samsung diversion ratio and found a diversion ratio of between 10.1 to 18.7%.  His preferred diversion ratio was 15.8%.  Ms Edwards-Warren calculated the platform critical loss thresholds at various combinations of gross margin percentage and developer pass-through rate, and a 5 to 10% price increase range.

2384 So, in relation to a 5% price increase, even on Professor Asker’s preferred diversion figures or on Professor Goldfarb’s figures, if the gross margin is between 70 and 80% and even on the assumption of 100% pass-through, a 5% increase would be profitable for the hypothetical monopolist under a critical loss analysis.

2385 Further, in relation to a 9% price increase, and assuming Professor Goldfarb’s preferred diversion ratio of 15.8% was correct, even on the assumption of 100% pass-through and between 70 and 80% gross margin, a 9% platform increase would be profitable for the hypothetical monopolist under a critical loss analysis.

2386 Further, in relation to a 10% price increase, and assuming the top of the diversion range that Professor Goldfarb suggested of 18.7%, the price increase would always be profitable, even on the assumption of 100% pass-through and a gross margin percentage of 70%, under a critical loss analysis.

2387 Further, in relation to a 10% price increase, but using Professor Goldfarb’s preferred 15.8% diversion figure, it would only be unprofitable if the gross margin was 70% and the developer pass-through rate exceeded 85%.

2388 What this all shows, as Ms Edwards-Warren correctly demonstrated, is that given Google’s gross margins, the likely developer pass-through rates, Professor Goldfarb’s estimates of diversion and the real possibility of anti-competitive conduct reducing the observed diversion rates, the critical loss analysis supports the conclusion that the Android mobile app distribution market is well-defined, even on a 10% SSNIP.

2389 Drawing the threads together, let me summarize where I have got to based upon Ms Edwards-Warren’s evidence and Epic’s position, which I have accepted.

2390 To recap, the results of the analysis by Professor Asker show that the aggregate diversion ratio is higher than the critical loss threshold when applying a 5% SSNIP, which suggests that the Android mobile app distribution market is well defined. But the results show that the aggregate diversion ratio is lower than the critical loss threshold when applying a 10% SSNIP, which suggests that the market should be widened.

2391 Now Professor Asker argued that these results show that the Android mobile app distribution market failed the HMT and one should rely on the 10% SSNIP threshold.

2392 But even if the diversion ratios used by Professor Asker were well estimated, these results are at best borderline, and do not point strongly either to the market being too narrow or too wide. The failure of the test at the 10% level could be an indication of the cellophane fallacy, as Ms Edwards-Warren pointed out.

2393 But because of the flaws in the design, implementation and interpretation of the test, it is not appropriate to place significant weight on the results of his critical loss analysis in any event for the following reasons in summary.

2394 First, Professor Asker used two critical loss equations, but neither equation is well suited for a market like app distribution which is characterised by high fixed costs, close to zero marginal costs and large economies of scale.  And for such markets, the Lerner index, on which the critical loss is based, is an inadequate tool for approximating price elasticity.

2395 Second, Professor Asker assumed a 100% pass-through rate from developers to users which he states is conservative.  But the test is more likely to be passed with a lower pass-through rate. And it is unlikely that the pass-through rate is very high or close to 100%.

2396 Third, Professor Asker did not use estimates of actual losses to compare to the critical loss. He compared the critical loss to estimates of the aggregate diversion ratio from the Play Store to other Android app stores.

2397 Fourth, Professor Asker used a diversion ratio from the Play Store to Samsung Galaxy Store calculated from the Fortnite hotfix analysis, which is flawed.  And he did not consider whether diversion ratios from the Play Store to other Android app stores were suppressed because of the relevant Google conduct.

2398 But even if one accepted the methodology and inputs, as I have said, the results do not universally suggest that the market should be wider than the Android mobile app distribution market.

2399 Further, Professor Asker erroneously considered whether a SSNIP is profit-maximising rather than profit-increasing, which is the equivalent to doubling the magnitude of the price increase and only considering a 10% price increase.

2400 Finally, whatever the various difficulties with Professor Asker’s critical loss analysis, his analysis shows that even on his lower estimates of diversion to the Samsung Galaxy Store, a hypothetical monopolist of the Android mobile app distribution market could profitably implement a 5% price increase.

2401 But as I have said, if one adopts Professor Goldfarb’s estimates of Samsung diversion and reasonable assumptions of developer pass-through, Professor Asker’s critical loss analysis shows that a hypothetical monopolist of the Android mobile app distribution market could profitably implement a 10% price increase.

2402 So, and as Epic rightly said, irrespective of the debate about whether the HMT involves a profit maximising or profit increasing criterion, Professor Asker’s critical loss analysis suggests that the Android mobile app distribution market is well-defined.  The analysis indicates that a SSNIP of 10% would be profitable to the hypothetical monopolist of the Android mobile app distribution market, and accordingly a SSNIP of 5% would be profit-maximising.

What are the close substitutes, if any?

2403 Given that I have identified the relevant focal product, what, if any, close competitive constraints exist on the supply of that product?  That exercise is undertaken by reference to the likelihood of substitution to other products.

2404 To be relevant to the analysis, substitution between services involves a shift by developers and/or users, or in some cases suppliers, from one service to another in response to a small but significant change in relative prices or quality. To be included in the same market, two services must be close substitutes in this sense.

2405 In considering the close substitutes for Android app distribution services, the task is to assess what, if any, alternatives users and developers would switch to in response to a small but significant change in the relative price or quality of those services.

2406 Google says that the Play Store is exposed to competition within the Android ecosystem. This includes competition from other Android app stores, pre-installed apps on a mobile device, and sideloading of apps by an end user. It is also said that the Play Store faces competition from non-mobile platforms and the use of gaming consoles and PCs.

2407 Let me summarise the evidence as to whether there are any such close substitutes.

2408 I should say upfront that I agree with Epic that other Android app stores are the only close substitutes for the services supplied by the Play Store.

2409 Further, Professor Asker centred much of his evidence of substitution around whether products could be substitutes in theory.  According to Professor Asker, web apps and native apps could be substitutes if there exists apps with relatively higher use of web apps or browsers over native apps.  Web apps and native apps could also be substitutes if developers ever create both a web app or native app, or only a web app.  And he presented some evidence on consumers switching between Android smart mobile devices and iOS devices as evidence of substitution between apps on these devices.

2410 Professor Asker’s evidence of theoretical substitution did not show whether consumers or developers consider the use of one product to be reasonably interchangeable with another.  Moreover, evidence of some users ever using more than one app channel or developers having apps on more than one channel is consistent with these app channels being independent products.  The same point holds true for payment channels.

2411 But I should say that in summary I am in substantial agreement with Google concerning the available choices save that they do not rise to the level of being close substitutes for market definition purposes, save for other Android app stores.  Moreover, the existence of such technical or non-close substitutes does not deny or negate Google’s substantial market power as I have discussed elsewhere.

2412 Let me now go through some of the possibilities that have been discussed in the evidence.

Android app stores other than the Play Store

2413 Clearly, the distribution services supplied by other Android app stores are substitutable for the services supplied by Google through the Play Store.

2414 This includes app stores on Android devices other than the Play Store, including OEM app stores such as the Samsung Galaxy Store, Huawei App Gallery, Xiaomi Market, Vivo App Store, and Oppo App Market, and third-party app stores such as Uptodown, the Amazon App Store, Aptoide, F-Droid and One Store.

2415 This extends to app stores however they come to appear on an Android device, including whether they are pre-installed or directly downloaded.

2416 Other Android app stores and the Play Store are substitutes. The services supplied by other Android app stores on mobile devices are substitutable for the services supplied by Google through the Play Store, and they are constraints.

2417 First, other Android app stores are functionally comparable to the Play Store in their distribution of developer apps and, like the Play Store, can be used to distribute a large number of apps to Android users.  Further, the Samsung Galaxy Store is pre-installed on two-thirds of all Android devices in Australia.

2418 Second, competitive pressures from other Android app stores not only exist, but are increasing.  Google’s internal documents make many references to the competitive threat posed by other Android app stores.

2419 They record that there is increased competition arising from Samsung promoting its own products, launching a rewards program on the Galaxy Store and aggressively driving exclusive games, such that it is a key competitor that has taken aggressive actions to grow market share. These actions include pursuing gaming on its store and growing its services business, differentiating hardware, and appealing to younger users.

2420 Further, they record that the Amazon Appstore was emerging as a challenge to the Play Store in gaming globally, including by acquiring top game titles and engaging high value users to grow its market share. It has also aggressively pursued gaming on its store by funding in-app purchase discounts of up to 20% to attract users, growing its services business, differentiating hardware, and appealing to younger users.

2421 Further, alternative Android app stores can each pursue an area of focus that is unique and thereby put pressure on the Play Store to follow.

2422 Further, key competitors have taken aggressive action to grow their user bases.  OneStore has lowered revenue share to court developers, and ecosystem partners such as Samsung, OneStore, and the Amazon Appstore have invested in game distribution and exclusive content.  Further, Amazon in Japan is an illustration of the severity of the threats that Google faces.  This app store has used heavy discounts to attract the Play Store’s high value users. Google faces the risk of becoming a “showroom” for Amazon or other app stores.

2423 Further, there is a risk of major developers de-prioritising the Play Store or distributing less on the Play Store, in the place of pursuing gaming opportunities with competitors like Samsung and the OneStore.

2424 Third, Google has responded to the competitive threats from other Android app stores with competitive programs, such as Google Play Points, which is a loyalty program aimed at high value users.

2425 The Play Store has a very high share of app downloads to Android devices, being roughly 97.9% in Australia and 91.4% globally, and pre-installations on Android devices.

2426 But whilst Google’s high share of Android app downloads shows that the Play Store is a highly popular app store within the Android ecosystem, it is by no means evident that Android device users lack genuine alternatives to the Play Store for obtaining apps and, importantly, for purchasing app-based digital content. There is no sense in which Android device users are locked in to using the Play Store.  This is for several reasons.

2427 First, even on their Android device, users have a range of readily available and convenient alternative options for obtaining apps and making purchases of app-based digital content. In fact, over 84% of the top 50 apps by spend on the Play Store in Australia are available on Android devices by a means other than the Play Store. The options available to, and utilised by, users are as follows.

2428 In respect of users using another Android app store, there are a range of Android app stores from which users can obtain apps on their device. These include OEM stores such as the Samsung Galaxy Store or the Vivo app store, the ONE Store developed by Korean carriers and third-party app stores such as the Amazon Appstore.

2429 In Australia, roughly 72% of Android devices come with an alternative app store pre-loaded including 67% of devices having the Samsung Galaxy Store pre-installed. And insofar as an app store is not pre-installed, it can be sideloaded. At least 5% of Android devices have at least one sideloaded app store. The pre-installed app stores are easy to access, typically being placed on the device home screen and not being subject to any unknown source warnings.

2430 Further, the Android app stores to which users have access typically carry many of the apps published on the Play Store.  For example, the Amazon Appstore carries 72% and the Samsung Galaxy Store carries 34% of the top 50 apps by spend on the Play Store.

2431 In summary, other Android app stores provide a competitive constraint on the Play Store to such an extent that they are substitutes.

Direct downloading/sideloading

2432 Another possible substitute relevant to Android devices is the possibility of users directly downloading individual apps to their devices without the use of an app store.

2433 Ms Edwards-Warren and Professor Asker agreed that there is a measure of substitutability between direct downloading of individual apps and the service supplied by an Android app store, but they disagreed as to the extent.

2434 Google says that sideloading is a close substitute for an Android app store.  Now I disagree but let me identify some of its arguments.

2435 Now although Epic says that there are functional limitations with sideloading, including the fact that sideloaded apps cannot be automatically updated, Google’s witness Mr Cunningham said that sideloading is a straightforward process and in his experience takes merely an estimated 15 to 30 seconds for a user to follow installation flows and successfully sideload an app. Epic itself has described the process as being simple and easy to navigate.

2436 Further, in October 2021, Android 12 introduced the ability for non-preloaded sources of apps to install app updates without requiring user involvement.  So sideloaded apps can be updated automatically.

2437 Further, although Epic says that users and developers do not consider sideloading to be a close substitute for Android app stores, Google says that there is evidence that users and developers in practice can and do sideload.

2438 Further, Google says that Professor Asker’s analysis of the Fortnite spending data before and after the hotfix is consistent with users viewing sideloading as a close substitute.  Now I accept that there is evidence that for at least some users, sideloading is a viable substitute.

2439 Further, although Epic says that there are significant switching costs associated with sideloading for users and developers, Google says that Epic has provided no quantification of what those costs may be. Further, in any event, from the developer’s perspective, sideloading would be a cheaper option to deploy because it would avoid paying app store commissions.

2440 Further, the suggestion that there are prohibitive switching costs that prevent sideloading is inconsistent with the evidence of Android Fortnite user spending patterns after the hotfix.

2441 Further, although Epic says that sideloading is a form of self-supply of distribution services, and therefore not a close substitute, Google says that if a developer is readily able to bypass the Play Store by self-supplying distribution with sideloading, it will have significant countervailing power. If so, this constitutes a constraint on Google.

2442 So, Google says that sideloading is a close constraint on the Play Store.  But I disagree and on this aspect largely agree with Epic’s position.

2443 Now Ms Edwards-Warren considered that there is only limited substitutability, and it is unlikely to be an option for developers who rely on the Play Store or another Android app store for discoverability.

2444 Direct downloading of individual apps relies upon the user being aware of the specific developer’s app and the ability to download that app from the specific developer’s website. For that reason, direct downloading is not a sufficient constraint on distribution via Android app stores to be in the same market as Android app stores.

2445 Contrastingly, Professor Asker considered direct downloading of individual apps as sufficiently substitutable to impose equivalent or greater constraints on the Play Store as services supplied by third-party Android app stores.

2446 I prefer Ms Edwards-Warren’s view. The direct downloading of individual Android apps from a website on an Android device is not a close substitute for distribution through an Android app store.

2447 First, there are inherent functional limitations with direct downloading for both developers and users, which prevent a meaningful amount of switching from occurring.

2448 Second, users and developers do not consider direct downloading to be a close substitute to downloading from an Android app store.

2449 Third, there are significant switching costs associated with directly downloading apps for users and developers.

2450 Fourth, for direct downloading, developers distribute their own app directly to users without an intermediary. This is a form of self-supply of distribution services.

2451 Now direct downloading has functional limitations which prevent a meaningful amount of switching.

2452 There are significant limitations on users acquiring apps by direct download as compared to an Android app store. This in turn reduces the desirability of distributing apps via direct download from the perspective of developers. Given network effects, if app developers anticipate that only a minority of users will sideload apps, it will make them less likely to make direct downloads available, which in turn makes sideloading a less commonly sought alternative by users.

2453 First, there are significant discoverability challenges associated with distributing apps via direct download. This method of distribution requires users to navigate through their web browser to a developer’s website for each app they wish to download. Users are only likely to discover apps available via direct download if they are specifically looking for the app. App stores, by comparison, contain a library of categorised and searchable apps from multiple developers.

2454 Now direct downloading might be an option for popular apps that have high awareness among Android smart mobile device users outside of the Play Store, but it is unlikely to be an option for apps that rely on the Play Store or alternative app stores for discoverability.

2455 Mr Sweeney’s evidence was that it would be more difficult for Epic to attract users to the Epic  Games Store on Android if Epic only distributed the Epic Games Store on Android devices via direct download.

2456 This is because the process of direct downloading an app from a website involves more time and effort than downloading it from an app store. It requires the user to navigate to the developer’s website using their mobile web browser app. In contrast, storefronts such as the Play Store and the App Store have search functions, provide users with suggested and recommended apps and group apps by category to enable apps to be more easily discovered if the user were not actively looking for the specific app.  Further, many Android device users have become accustomed to downloading apps only from app stores, and not directly from developers’ websites.

2457 Further, Mr Weissinger explained that Android users would be unlikely to discover an app that was made available outside the Play Store through direct download or via the Galaxy Store, unless they were specifically looking for that app and were aware that there were alternative options to the Play Store which they could use to download the app.

2458 Second, users must endure the install flow that is imposed by the technical restrictions, which materially degrades the direct download experience as compared to that of downloading and installing an app through an app store with privileged install permissions.

2459 The preponderance of the evidence is that additional steps and frictions in the download and install process act as a disincentive to Android device users from directly downloading apps, as I have discussed elsewhere.

2460 Further, the frictions have a compounding effect, because users who have experienced the frictions associated with direct downloading once will then anticipate the frictions on subsequent occasions, reducing the share of users opting for direct downloads from a website in the future.

2461 Further, as well as encountering frictions during the downloading process, users who directly download apps encounter additional frictions when updating those apps. Those frictions are generally not present when users download apps from Android app stores.  Many app stores as an additional service automatically update apps that users have installed from their stores. But Google does not make its automatic update functionality available to directly downloaded apps. So, if users only access an app by sideloading, rather than obtaining it through the Play Store, the process of updating that app is more difficult as the user will not receive updates through the Play Store.

2462 Further, users and developers do not consider direct downloading to be a close substitute to Android app stores.

2463 Let me say something about the developers.

2464 Now only six of the top 50 apps (or 12%) by consumer spending on the Play Store in Australia in 2021 are available for direct download. This small percentage of developers making their apps available for direct download indicates that developers do not consider distributing their apps via direct download to be a viable alternative to app store distribution.

2465 Mr Grant’s understanding was that the vast majority of developers do not make their mobile apps available for direct download.  And even where users can navigate to a developer’s website and attempt to install an app, the result is often that the website link will redirect users to an app store.

2466 Further, the fact that some developers offer their apps for direct download in addition to distributing their apps through an app store is not evidence that, in response to a SSNIP on the commission payable to the app store, developers would stop distributing via Android app stores and exclusively offer their app for direct download. It is also not evidence that users would switch to directly downloading their apps.

2467 Epic’s unsuccessful attempt to distribute Fortnite on Android outside of the Play Store demonstrates the lack of substitutability for developers as between Android app store distribution and direct downloading.

2468 As Mr Weissinger explained, the Fortnite usage statistics on Android from the launch of Fortnite in around August 2018 to March 2020 demonstrate that Epic was unable to acquire a significant user base by making the game available outside of the Play Store via direct download from Epic’s website.

2469 This was so despite Epic’s significant endeavours to promote the launch of Fortnite via direct download and the Galaxy Store when it launched in August 2018, including by, in the months following the launch, allocating over USD 7 million a month to advertisements aimed at encouraging users to directly download Fortnite.

2470 This demonstrates that even when developers attempt to remedy the discoverability and friction issues associated with direct downloads by deploying large and sophisticated marketing campaigns, distribution via direct download is not a viable alternative to app store distribution.

2471 Let me say something about the users.

2472 The low take-up of direct downloading reveals that users do not consider direct downloading of individual apps to be a substitute for downloading apps via Android app stores.

2473 In 2020, no more than 3% of apps in Australia, and 12% globally, were distributed to Android devices via direct download. In contrast, 75% of apps in Australia, and 80% globally, were distributed through an app store.

2474 The data Ms Edwards-Warren analysed recorded both downloads made either directly from an app developer’s website or via directly downloaded third-party app stores, described as an “unknown source” download in the dataset.

2475 So, the 3% figure calculated for direct downloads is the upper bound, and the true proportion of apps distributed via direct download is likely to be lower.

2476 Google’s own documents confirm that direct downloading of individual apps by users is limited.

2477 One undated internal presentation titled “Special Topic: Off-Play Installs (a.k.a. Sideloading)” showed that off-Play Store installs accounted for 12% of all installs globally and came from three main sources, the largest being the “Package Installer” source, which includes direct downloading and apps downloaded from third-party app stores. “Package Installer” accounted for 79% of all off-Play Store installs.

2478 Ms Edwards-Warren considered this document and concluded that “Package Installer” accounts for approximately 9.5% of all installs globally. The document also identified India, Indonesia, Iran and Iraq as countries with over 15% of installs occurring via off-Play Store installs, whereas off-Play Store installs accounted for 3.6% of downloads in Japan and 4.4% in the United States. Data specific to Australia was not included.

2479 These statistics support the conclusion that Android device users have likely grown accustomed to searching for and downloading apps through the Play Store, this being a further barrier to substitution.

2480 Further, there are significant switching costs associated with directly downloading apps for users and developers.

2481 For Android device users, the cost of directly downloading an app onto an Android device in place of downloading via an Android app store is closely linked with the functional differences, in particular the additional steps required and frictions introduced by the technical restrictions. This is supported by the data on the limited prevalence of direct downloading.

2482 Further, a developer wishing to distribute apps via direct downloading would need a website and a mechanism for promoting and marketing its apps to Android users.

2483 It follows that if a developer does not already have an appropriate website, then the cost of switching away from an Android app store to distribution via direct downloading is likely to be significant.

2484 Developers also incur additional costs in reaching users when switching to distribution via direct download.

2485 Epic had to spend millions of dollars on marketing and advertising targeted at acquiring users via a direct download distribution channel, which it was ultimately unable to recoup. Many developers are not able to make such a financial outlay.

2486 It is too great a barrier for small developers and a sizeable disincentive for large developers to switch from Android app store distribution to direct downloading. To go from the use of an app store for distribution to direct downloading essentially involves the developer having to self-supply its own distribution service, which is likely to carry with it its own inherent costs.

2487 All of these matters further support the conclusion that direct downloading of individual apps is not a close substitute for distribution via Android app stores for either users or developers.  Moreover, such a possibility does not deny or negate Google’s substantial market power.

2488 Let me turn to another topic.

The possibility of pre-installation

2489 Android apps can be installed onto OEM devices before those devices are sold to users, such that the app is available to the user "out of the box". This method of app distribution is referred to as pre-installation, and an app that is installed in this way is known as a pre-installed app.

2490 Pre-installation of individual apps is not a substitute for distribution via Android app stores for a large number of developers.

2491 Now the only area of dispute between the economists is whether pre-installation of individual apps is a substitute for distribution via Android app stores for a small number of large developers.

2492 Ms Edwards-Warren said that pre-installation of individual apps does not constrain Google, and so ought not be included in her Android mobile app distribution market.

2493 Contrastingly, Professor Asker was of the view that pre-installation of individual apps is a substitute for distribution via Android app stores for large developers. But he said that there was a lack of data necessary to quantitatively evaluate whether substitution by large developers to pre-installation, in conjunction with other substitutes, would on its own be sufficient to discipline a hypothetical monopolist of Android app distribution.

2494 Accordingly, the narrow area of disagreement between the economists has no impact on the question of market definition.

2495 Now whilst Professor Asker said that only a few large developers would need to substitute to distributing their apps via pre-installation in response to a price increase by the hypothetical monopolist to discipline that hypothetical monopolist, he did not ultimately identify sufficient evidence to lead him to conclude that pre-installation of individual apps ought be included in the Android mobile app distribution market.

2496 Let me turn to Google’s arguments.

2497 Google says that the issue concerning pre-installed apps is whether they are a substitute for distribution via Android app stores for enough large developers to constrain the Play Store.

2498 Now Epic says that they are not, because it is costly to negotiate pre-installation agreements with developers, OEMs are reluctant to pre-install apps, and pre-installed apps encounter discoverability issues. But Google makes a number of points.

2499 First, Google says that large developers have successfully utilised pre-installation, indicating that consumers are willing to follow them to the pre-installed app.  Out of over 5 million new Android smartphones active in Australia in 2021, the number of new pre-installations was 3.7 million for Microsoft OneDrive, 3.3 million for Facebook and 2.7 million for Netflix. This does not reflect reluctance on the part of OEMs or developers with respect to pre-installation.

2500 Second, Google says that Epic has provided no evidence to support its proposition that pre-installed apps face discoverability issues, particularly in the case of well-known apps.

2501 Third, Google says that Epic itself reached pre-installation agreements for the Fortnite installer or the Epic Games App stubs with Samsung, Huawei, Sony and LG. And this was at a time when Fortnite accounted for less than 0.1% of the Play Store in-app purchases revenue in Australia when it was on the Play Store.

2502 So, there is little basis for thinking that large developers would have difficulty negotiating pre-installation agreements with OEMs, particularly when apps developed by major developers like Microsoft (Microsoft OneDrive, Microsoft Outlook), Meta (Facebook), Netflix (Netflix) and Amazon (Amazon Kindle) are distributed on Android devices in Australia in substantial numbers via pre-installation.

2503 Fourth, Google says that it would only take a few of Google’s top 25 developers, who account for nearly half of the Play Store’s service fee revenues, to substitute to pre-installation to meaningfully constrain the Play Store.

2504 But on this aspect of the case, in my view the evidence establishes that pre-installation of individual apps is not a close substitute for distribution via Android app stores, even for large developers.

2505 I accept Ms Edwards-Warren’s conclusion about that for the reasons she gave.  It is inefficient and costly for developers to negotiate pre-installation agreements for individual apps.  Further, OEMs are reluctant to pre-install third-party apps.  And further, pre-installed apps encounter discoverability issues and account for a small fraction of total Android app installs.

2506 In any event, Professor Asker conceded that he had insufficient information to conclude that pre-installation would discipline a hypothetical monopolist of Android apps distribution. Accordingly, the narrow area of disagreement between the economists has no impact on the question of market definition.

2507 But even if I preferred Professor Asker’s view on the issue of substitution for a small number of large developers, I would not conclude on the available evidence that pre-installation is a sufficiently strong competitive constraint on distribution via Android app stores to bring pre-installation within the Android mobile app distribution market.

2508 Further, it is inefficient and costly for developers, including large developers, to reach Android users at scale by pre-installing individual apps. That is because doing so requires developers to reach pre-installation agreements with numerous OEMs and carriers, typically including individual agreements with local sales teams for each OEM in each local market.

2509 Clearly, a large majority of developers that distribute through the Play Store lack the resources to successfully negotiate pre-installation deals with OEMs.

2510 But even if an agreement is negotiated, it is unlikely to apply to all geographic markets.  Further, it may only apply to a limited set of devices or models including, typically, only new devices. Further, it could be time limited.

2511 Further, any agreement with an OEM as distinct from a carrier will typically only apply to unlocked devices, as opposed to locked devices, which are sold by carriers as part of retail mobile contracts and plans. A separate pre-installation agreement would likely need to be negotiated with the carrier in respect of locked devices, which typically differ from country to country.

2512 By comparison, distributing an app on an Android app store is much easier for developers because it provides a single and accessible mechanism by which developers everywhere can distribute their apps to a global audience.

2513 The technical infrastructure to efficiently distribute apps and app updates on billions of devices in around 190 countries around the world on an ongoing and timely basis is complex, and is one of the services that the Play Store provides to developers.

2514 App developers would therefore incur significant switching costs in switching from distribution via an Android app store to distributing via pre-installation.

2515 Further, as I have said above OEMs are reluctant to pre-install third-party apps.  Moreover, the challenges encountered by developers when negotiating pre-installation agreements with OEMs are compounded by the fact that the number of spaces available for pre-installed apps on Android devices is limited. Google, the OEM and the carrier typically pre-install a large number of their own apps on Android devices.

2516 There is also evidence of user resistance to pre-installed apps. There is a perception among device users that pre-installed apps are bloatware, meaning that they are not useful and unnecessarily take up storage and screen space on the device.  Having a number of pre-installed apps serving similar purposes creates a bad user experience, and tends to lead to devices being filled with bloatware that users do not want.

2517 A 2015 internal Google presentation titled “Application discussion” recorded that 66% of Samsung users were dissatisfied or very dissatisfied with having multiple pre-installed apps on their devices which serve similar purposes.

2518 That document also recorded Samsung’s “[p]reload app reduction target” and sets out the number of pre-installed apps across various devices. In each case, the device is dominated by pre-installed apps owned by Samsung, Google and the carrier, and the reduction target contemplates reducing the number of third-party pre-installed apps from between seven and 11 down to between two and four.

2519 The conclusion recorded is that “[r]educing lots of preloaded apps by focusing on high-value/usage can help Samsung provide a better, uncluttered Android experience”.

2520 In contrast, an Android app store can distribute notionally an unlimited number of apps. The number of apps available on Android app stores is orders of magnitude higher than the number of apps typically pre-installed.  Moreover, pre-installation is not an option for many app developers given that it would be impossible to pre-install the millions of apps available via Android app stores.  There are over two million apps on the Play Store.

2521 The inherent limitations mean that pre-installation of individual apps is an option for a limited number of apps only.

2522 Further, given that OEMs are driven by user demands, user resistance to pre-installed apps constitutes another barrier to app distribution via pre-installation.

2523 Further, pre-installed apps suffer from issues with discoverability unless they are placed prominently on a user’s device, for instance, on the default home screen (DHS).  DHS placement maximises the discoverability of the app as users can more readily see and access the app, and are therefore more likely to open and use it. But discoverability is limited if an app is on the “plus one” screen, meaning it is less likely that users will notice and open the app.  But it is challenging for developers to secure DHS placement.

2524 In the case of Epic's negotiations with LG and Sony in relation to the pre-installation of the Fortnite installer and the Epic Games App, Epic was not able to place the Fortnite installer app stub on the DHS of those OEMs’ devices. Rather, it was and would have been placed on the plus one screen of the relevant devices.

2525 On Samsung devices, the Fortnite installer stub was located within Samsung’s Game Launcher, which was located in the “app tray”, requiring users to swipe on their device to locate the app, rather than the DHS of the relevant Samsung devices.

2526 So, pre-installation of individual apps is not a comparable distribution method from a device user's perspective given the passive nature of app acquisition and lower app discoverability.  Further, pre-installed apps are a small fraction of total app installs.

2527 Moreover, pre-installation is not an option for many developers.

2528 Now Professor Asker maintained that a number of large developers with substantial consumer spending could viably use pre-installation to distribute content. Professor Asker identified 10 examples of such large developers in support of that proposition.  But even if Professor Asker is correct in suggesting that his 10 identified large developers could feasibly substitute to pre-installation, substitution by those developers would not discipline the Play Store.

2529 Of the 10 developers identified by Professor Asker, only five of those apps contribute to Play Store revenue, as five are consumption-only apps.  As Ms Edwards-Warren explained, the five apps listed as having high spending via the Play Store being Microsoft OneDrive, Facebook, Dropbox, Linkedin and Evernote each accounted for a maximum of 0.31%, and 0.81% in aggregate, of Google’s revenues from the Play Store in Australia in 2021.  So, even if these apps could reasonably substitute all consumer spending towards pre-installed versions of their app, it would unlikely be sufficient to discipline a hypothetical monopolist.

2530 In summary, the possibility of pre-installation goes nowhere in terms of market definition.  Moreover, to state the obvious, that possibility does not deny or negate Google’s substantial market power.  Let me now deal with some other topics.

App stores on gaming consoles and PCs

2531 Google says that the fact that users and developers can and do choose between Android and non-mobile platforms respectively for purchasing app-based digital content and developing apps means that the Play Store faces a significant threat to its revenue from other platforms. And Google says that given that its service fee revenue is highly concentrated amongst users and developers, this revenue risk from other platforms represents a competitive constraint.

2532 It is said that there are four key sources of evidence which demonstrate this substitution, and therefore a competitive constraint, between the Play Store and app stores on gaming consoles and PCs.

2533 First, there are internal Google documents recording Google’s perception of competitive constraint from gaming consoles and PCs.

2534 Second, there are communications and agreements between Epic and gaming console operators.

2535 Third, there is the Fortnite spend data and the spending patterns that it shows, which I have discussed elsewhere.

2536 Fourth, there is the quantitative analysis by Professor Asker and Professor Goldfarb which I have separately discussed in detail.

2537 Let me deal with the first and second aspects in this part of my reasons.

2538 Now Google says that there is evidence of Google’s perception of competition from consoles and PC.

2539 Google’s internal documents record perceptions of competition with, and competitive threat from, PC and gaming console platforms. For example, the following may be noted.

2540 A February 2018 presentation titled Project Hug: Risk & Leakage Model, in the section headed “Market dynamics”, lists “PC/Console” operators alongside mobile app stores and describes their competitive activities. Another slide states “[d]evelopers have invested in PC/Console and don't value Play's position; need to give partners more back because they’ve invested”.

2541 A further February 2018 presentation titled Project Banyan, Hug and RSA / Play Kicker: Risk & Leakage Models records the same perception as the Project Hug: Risk & Leakage Model presentation. Further, on a slide titled “Mitigation effectiveness”, which refers to Project Hug and the “Samsung Deal” (Banyan), it rates mitigation effectiveness of Hug with respect to the Epic Store, Samsung Store, Amazon Store and PC-Dev Store. The reference to “Epic Store” and “PC-Dev Store” are references to PC stores.

2542 Google says that this all reflects a perception on its part that there is some competition with PC and console operators as well as mobile app stores.  Moreover, there is little doubt that Google’s perception is that lines between platforms are becoming blurred such that there is competition across them.

2543 Further, Google itself has been interested in how Android and the Play Store compare with other platforms including gaming consoles, which is consistent with perceiving gaming consoles to be competitors to some extent

2544 Further, Google says that there is evidence of console operators’ perception of competition from other platforms. Major gaming console operators perceived a real and substantial competitive threat to their in-app transaction commission revenue from other platforms including Android.

2545 Google says that these fears were expressed and acted upon when Epic sought to include cross-progression and cross-wallet functionality in the versions of Fortnite on the Sony PlayStation, Microsoft Xbox and Nintendo Switch consoles.  Google says that such fears reflect the essence of competitive constraint between platforms for revenue-generating transactions.

2546 First, Google says that in the case of the Sony PlayStation, there is a “most favoured nations” clause in the cross-platform policy schedule to Sony’s PlayStation Global Developer & Publisher Agreement with Epic concerning Fortnite, both in the 2019 policy schedule and the current 2022 policy schedule.

2547 The policy schedule essentially permits Fortnite users to engage in cross-platform gameplay, being online play with users on other permitted platforms, cross-platform progression, being the transfer of progress in the game across permitted platforms, and cross-platform commerce, being the ability to use digital content and virtual currency purchased on PlayStation on other permitted platforms and vice versa.  The permitted platforms are PC (Windows PC, Mac OS or Linux OS), mobile (iOS or Android), console (Xbox and Nintendo Switch) and streaming.

2548 And the relevant effect of the “most favoured nations” provisions of the schedule is that Epic must offer Sony the opportunity to sell on the PlayStation equivalent digital content and virtual currency made available on other platforms, but limited to non-exclusive content in the 2022 version.

2549 Further, Epic must offer Sony a wholesale price no higher than the wholesale price for equivalent digital content and virtual currency on other platforms, or where Epic controls the retail price, a wholesale price that is no higher [REDACTED] of the retail price of equivalent digital content and virtual currency on other platforms.

2550 Further, if the ratio of PlayStation’s cross-platform revenue share to its gameplay share is below [REDACTED] Epic must pay Sony [REDACTED] of the difference between [REDACTED] [REDACTED] [REDACTED] [REDACTED] [REDACTED]

2551 Now it would seem that the commercial object of such provisions is to ensure that PlayStation is not competitively disadvantaged in price or quality terms on Fortnite in-app purchases relative to other platforms, and compensate Sony if users switch their spending (but not use) to other platforms.

2552 Google says that the inference that arises from the inclusion of these restrictions is that Sony was concerned that cross-wallet and cross-progression functionality in Fortnite would allow users to switch their spending to other platforms and that this represented a strong competitive threat to Sony’s PlayStation commission revenue.

2553 Second, Google says that in the case of Microsoft’s Xbox gaming console, emails between Epic and Microsoft executives record similar competitive concerns held by Microsoft with respect to cross-progression and cross-wallet in Fortnite. In particular, the following may be noted.

2554 When Mr Sweeney in January 2018 raised the possibility of Xbox participating in Fortnite cross-progression and cross-wallet across platforms including iOS and Android, Mr Spencer of Microsoft expressed concern about digital items on Xbox being uncompetitively priced relative to other platforms, and the revenue loss to Microsoft that this would cause.

2555 In March 2018, Microsoft’s proposal for Fortnite cross-progression between Xbox and other platforms, but not cross-wallet, was that Epic must offer content parity across platforms on digital items, and must not intentionally drive purchases of V-Bucks and digital content to one platform versus another.

2556 More recently in June 2020, when Mr Sweeney made a proposal to Mr Spencer to use XCloud for its games in exchange opening up mobile platforms and connecting app stores, Mr Spencer expressed the view in an email to Mr Sweeney on 10 June 2020 that allowing third-party app stores to sell their games on Xbox “breaks the model”, as there would be no revenue back to Microsoft.

2557 These communications reflect a perception on Microsoft’s part that when cross-progression and/or cross-wallet functionality across platforms is enabled within a game, users will treat different participating platforms including mobile as substitutable transaction venues.

2558 The fact that Microsoft was not prepared to allow cross-progression or cross-wallet without mitigating the threat to its revenue gives rise to an inference that Microsoft perceived this form of substitution to be a competitive constraint.

2559 Third, Google says that in the case of the Nintendo Switch, in July 2018 Epic and Nintendo discussed the inclusion of cross-wallet functionality in the Nintendo version of Fortnite. In that context, a Nintendo executive expressed concern that allowing users to buy V-Bucks on other platforms and use them on the Nintendo Switch could cannibalise Nintendo’s commission revenue. This represented a perception on Nintendo’s part that cross-wallet functionality allows users to substitute between platforms for making in-app purchases.

2560 Now I agree with Epic that developers and users would not switch away from using Android app distribution services to obtaining app distribution services on PCs or consoles as a result of small but significant changes in the relative price or quality of Android app distribution services.

2561 Pointing out the obvious, Epic says that mobile devices and non-mobile devices serve quite distinct purposes.  They have distinct user bases and, insofar as those user bases overlap, the use to which each device type is put is complementary or supplementary.

2562 Moreover, developers who wish to have their apps accessed and used in the contexts in which Android devices are used will need to develop their app for Android, and have it distributed using Android app distribution services.

2563 And clearly, the distinct roles of mobile and non-mobile devices are reflected in several observable characteristics about the industry. Many mobile apps are not available on PCs or consoles, and vice versa. And a material proportion of Android device users do not have a PC and/or console, and are unlikely to obtain one in response to an increase in the price of Android app distribution services.

2564 As Epic says, from a user’s perspective, if the app they wish to use is not available on non-mobile platforms, they will not switch to using those alternative platforms.  And even where an app is available on both mobile and non-mobile platforms, users are unlikely to consider these as close substitutes.

2565 Further, there are important technical differences between mobile and non-mobile devices that mean that they lend themselves to use in different contexts or for different purposes.

2566 The considerations discussed by Professor Goldfarb under the term “device first” explain this phenomenon. As he explained, in the case of mobile devices, usage is often a function of situational convenience. So, should the user’s desired app be unavailable on their mobile device, a user is more likely to switch to using a different app on that device, or engage in a different activity entirely, rather than switch to a non-mobile device to access the app.

2567 Because users do not substitute PC and console devices for Android devices, developers must develop Android apps to access Android users. Consequently, developers are unlikely to consider distribution of their app on non-mobile devices as a substitute for distribution on Android. Further, many apps developed for mobile are not developed for PCs or consoles, limiting the ability or willingness of users to switch between the two.

2568 Now central to Professor Asker’s view on this topic was his econometric analysis of switching by Fortnite users from the Play Store to PCs and consoles.  But as I have discussed elsewhere, Professor Asker’s analysis had various problematic aspects.  But in any event, most other apps cannot be used across platforms in the same way as Fortnite. So the conclusions which Professor Asker draws from his analysis of Fortnite do not have general application.

2569 Now I accept that Google has genuinely perceived from time to time competitive constraint from gaming consoles and PCs. But none of this amounts to close substitution or would deny or negate Google’s market power as I have discussed elsewhere.

2570 Further, I also accept that communications and agreements between Epic and gaming console operators reflect a perception of competition from other platforms.  But again, none of this amounts to close competition or would deny what I have said elsewhere on market power.

2571 Let me now say something about out-of-app payments.

Out-of-app payments

2572 Now Professor Asker said that developers that pay the Play Store’s service fee can transact with consumers that download their app from the Play Store outside of the app for content that consumers then use within the app. As a result, he said that relevant substitution possibilities that should be considered include “rival transaction venues”.

2573 But this is all a variation on Professor Hitt’s and Mr Houston’s theory that I have discussed in my reasons in the Apple proceeding that there is a “transaction” market.  It has little substance.  And importantly, it has nothing to do with substitution for distribution services.

2574 Epic made the correct and obvious point that the ability of users to acquire content outside of an Android app, but use it within the Android app, is dependent on them having obtained the native app via a distribution method. That is ordinarily done through an Android app store, involving the provision by Google of its app distribution services. So, the postulated constraint involves no substitution away from those distribution services.

2575 But in any event, as Epic says, the possibility of users purchasing digital content outside of an Android app, then consuming it in an Android app, does not constrain the hypothetical monopolist in its provision of Android app distribution services.

2576 First, for those apps that do not offer digital content outside of the developer’s Android app, out-of-app purchases are not possible.

2577 Second, many developers do not set different prices on different channels.

2578 Third, even for apps where out-of-app purchases are possible, users may not be aware of that possibility.

2579 Fourth, Google’s anti-steering rule prohibits developers from even telling their users about an out-of-app solution via in-app promotions or user-interface flow.

2580 Fifth, even where users are aware of the possibility, users are unlikely to divert unless they are aware that alternative payment solutions are cheaper.  But it is often difficult for consumers to obtain comparative information about pricing of app content across different sources.

2581 But in any event, even if all those difficulties were overcome, users are unlikely to engage with the process of conducting out-of-app payments instead of in-app purchases on native apps.

2582 But even if the ability to purchase digital content outside of an Android app then consume it in the Android app were to impose some constraint upon the hypothetical monopolist’s app distribution services, that would not warrant treating these other “rival transaction venues” as a substitute for distribution services.

2583 As Epic says, the constraint would simply be a product of the hypothetical monopolist’s present monetisation strategy, to which it cannot be assumed to be wedded.

2584 The hypothetical monopolist’s app distribution services are not the subject of any constraint from out-of-app purchases, particularly in circumstances where the hypothetical monopolist could change how it monetises its services.

2585 Finally, if it is necessary to state the obvious, Apple’s App Store and the services it provides is not a close substitute.  I have dealt with the topic of competition with iOS and the App Store in a separate section of my reasons.

2586 Now let me approach questions of substitution and choice from another angle by focusing more specifically on, first, the perspective of developers and, second, the perspective of users. But contrary to what Google has put to me, these choices, save for other Android app stores, do not amount to substitutes for market definition purposes and nor do they negate or deny Google’s market power.

2587 Now I should say that I have addressed web apps and cloud streaming services in my reasons in the Apple proceeding and adopt what I have said there.

Other possible choices for users

2588 As Google has said, in respect of users using a web app, website or streaming app, another option for using apps and purchasing app-based digital content on Android devices is to use the device’s web browser.

2589 Further, there are many apps which offer out-of-app purchases on websites for which the Play Store receives no service fee revenue when users subscribe outside the app.

2590 The Play Store also earns no service fee revenue when users purchase in-app content via the developer’s website, which is common among some gaming apps. This too is an option for purchasing app-based digital content that users have available to them.

2591 Further, for Android users who own other kinds of devices, such as an iOS device, a PC or a games console, there are many apps that provide the option of using the app and/or purchasing app-based digital content on a non-Android device instead of an Android device, being of course multi-homing.

2592 Professor Asker’s analysis demonstrated that of the top 50 apps by spend on the Play Store in Australia, 92% offer an alternative non-mobile transaction venue. Professor Asker’s analysis demonstrated that of the top 25 game apps by spend on the Play Store, 22 can be played on non-mobile platforms, 19 have cross-progression functionality and 14 have cross-wallet functionality.

2593 Ms Edwards-Warren extended this analysis to show which other platforms could be used for these apps. Her analysis demonstrated that all 25 can be used on iOS, 17 can be used on PC in native form and 20 can be used on PC if browser-based formats are included. Further, 3 can be used on game consoles and 1 can be used on a cloud game streaming platform. Her analysis also demonstrated similar numbers for cross-wallet functionality, with digital content for 12 apps additionally available for purchase on a web shop.

2594 Professor Asker’s analysis also demonstrated that of the top 25 non-game apps by spend on the Play Store, 23 can be used on non-mobile platforms, 21 have cross-wallet functionality and 18 offer cross-platform subscription functionality.

2595 Now the ability for multi-homing users to switch between platforms is also not limited to switching completely to another platform for the purpose of using an app or purchasing app-based digital content.

2596 The cross-wallet functionality means that for many apps the user can de-couple their app use from their app spend. This allows partial switching between platforms, because a user can play on one platform, and spend on another.

2597 Google faces the risk of users purchasing virtual currency and digital content outside the Play Store and its service fee being bypassed.  App users can and might choose to purchase content on other platforms notwithstanding that they consume that content on mobile.

2598 But I do agree with Epic that the choices available to multi-homing users are limited. This is supported by the fact that the estimates of iOS and Android user multi-homing, and user multi-homing across mobile and non-mobile devices, are low.

2599 But I also agree that the analysis of the Fortnite spending data confirms that Android device users both have the choice and actually exercise the choice to purchase in-app digital content on alternative platforms notwithstanding that they use the Fortnite app downloaded from the Play Store.

2600 Epic’s own internal analysis shows that users who first started playing Fortnite on Android predominantly purchase Fortnite in-app digital content on other platforms. There is 21.6% of spend on Android versus 34.9% of spend on PlayStation 4, 19% of spend on Xbox, 4.8% of spend on Nintendo Switch, 14.6% of spend on PC and 5.2% of spend on iOS. This spending pattern only makes sense if there is significant multi-homing.

2601 Further, there is some evidence that some users might primarily play Fortnite on Android but deliberately choose to boot up their console to make in-app purchases.

2602 Now Epic said that the top games on the Play Store typically are not available on gaming consoles or PC app stores, and vice versa.  And it relied on a s 50 summary of screenshots of app store websites showing the extent to which top free, grossing and selling games on the Play Store are available on PlayStation, Xbox and the Microsoft PC store, and the extent to which top selling games on Steam, top games on PlayStation and most played games on Xbox are available on the Play Store.

2603 But Google pointed out that the majority of the top free, grossing and selling games on the Play Store are available on PC using an emulator of the native app or via a web browser. Likewise, about half of the top PlayStation and Xbox games are available on the Play Store as mobile versions, and more than half are available to play on Android devices via cloud streaming services. So, users have a choice of platforms with respect to top games on Android.

2604 Further, all Android device users have the option of switching to iOS when it comes time to replace their device, which typically has an expected life of 3.2 years.

2605 Based on survey data identified by both Ms Edwards-Warren and Professor Asker, a significant proportion of Android users in fact switch to iOS when they replace their device. Surveys identified by Ms Edwards-Warren show that on average 11% of Android smartphone users and 29% of Android tablet users switch to iOS. Surveys identified by Professor Asker show that between 8% and 25% of Android smartphone users switch to iOS. So, switching between device operating systems is an option for some users.

2606 Now Epic relies principally on the views of Ms Edwards-Warren and Professor Dhar concerning the significance of user switching from Android devices to iOS devices in response to a change in app price or quality, or in-app digital content price or quality.  I have discussed Professor Dhar’s evidence elsewhere.

2607 Ms Edwards-Warren expressed a view to the effect that a small increase in the service fee would not make much difference to Android device users having regard to the purchase price of a new iOS device.

2608 Ms Edwards-Warren’s proposition was that a 10% increase in the service fee, if fully passed on, would amount to a 3% increase for the user in app-based digital content spend, which is insufficient to justify purchasing a USD 900 iOS device based on the average user spending USD 235 per year on the Play Store.

2609 I should say here that I have accepted Ms Edwards-Warren’s position on this aspect.

The choices available to developers

2610 Let me say something about the choices available to developers which Google has identified and with which I agree, although I would say that although I have found that developers have some options available to them, these do not amount to a close competitive constraint upon Google for market definition purposes.

2611 Like Android device users, app developers have viable choices available to them for distributing and monetising their apps. These choices are also not necessarily all or nothing, because developers can pursue several different distribution and monetisation options simultaneously.  But there are limits to how many options a developer can pursue at once, and faced with finite resources, developers typically must at least prioritise some options over others, if not choose between them.

2612 Now the choices for developers start with options as to the platform for which they make or through which they distribute their apps. These options include mobile platforms, that is, Android and/or iOS, gaming console platforms, PC, web apps, websites and streaming services. Sometimes, the extent to which these options are attractive to a developer may depend on the nature of the app they intend to make.

2613 In the gaming app context, a high fidelity AAA game app may be more appropriate for a gaming console and PC than a mobile or web-based format, and developers may target the applicable platform accordingly. Likewise, whilst a casual game designed for a mobile environment may work well in a web-based format, it might not be appropriate for a gaming console or PC. Some developers therefore may choose to develop their game for a specific format.

2614 High fidelity role-playing games, shooter games and car racing games now can be played on mobile devices and not just gaming consoles and PC. This has been driven by improvements in technology, particularly for mobile, and users becoming more comfortable with using virtual thumb sticks on a glass screen.

2615 Further, whilst some casual games made for mobile are not made for gaming consoles, they can appear on PCs. Indeed, Epic’s s 50 summary regarding the platform availability of top games, and Google’s addendum to that s 50 summary, show that the top games on the Play Store typically are available on at least one other non-mobile platform.

2616 Likewise, most top games on gaming consoles or PC are available on Android via cloud streaming services, and half are available in a mobile version on Android.

2617 Further, in the non-gaming app context, although gaming consoles would not be a relevant platform choice, developers still have a choice as to other platforms for which to make their app such as mobile, and within that Android and/or iOS, PC, or web-based, whether as a web app or website.

2618 The availability of such choices is confirmed by the evidence which shows that developers often choose to make their non-game app for multiple platforms. For example, Professor Asker’s analysis found that all but two of the top 25 non-game apps on the Play Store are available on non-mobile platforms or via the web including webstores and have cross-wallet or cross-progression functionality.

2619 Generally, I accept that to some extent developers can in many cases choose between platforms when deciding whether to make an app for a platform, or the extent to which they will invest in a particular platform.

2620 Let me say something further about the distribution options within the Android ecosystem.

2621 If a developer chooses to develop an app for Android devices, the Play Store is not the only means by which it can make its app available to Android device users. The evidence shows that developers have several viable distribution options within the Android ecosystem other than the Play Store.

2622 First, there are several Android app stores through which developers can distribute their apps.  These app stores are a realistic option for developers in Australia given that roughly 72% of Android devices come with an alternative app store pre-loaded, including 67% of devices having the Samsung Galaxy Store pre-installed, and that at least 5% of Android devices have at least one sideloaded app store.

2623 Second, just as sideloading is an option for users, sideloading is an option for developers to distribute their apps.  But how viable this option is was a matter in contest between the parties.  I tend to agree with Google that this option is more viable than Epic would suggest.  Nevertheless, this possibility does not take Google far in terms of market definition.  It is still an inferior option.  Certainly it does not go close to suggesting close substitutability, although Ms Edwards-Warren was not as dogmatic as I am.

2624 Further, for some developers, particularly significant ones, pre-installation is an alternative for making an app available to users. But not all app developers would be able to do a deal with an OEM to pre-install their apps on Android devices.

2625 Further, developers have the option of making their app available as a web app or website, and many do so.  Many prominent developers offer their app both as a native app and a web app or website.  So, developers can and do choose between native apps and web based apps, and often use both.

2626 Further, game developers have the option of making their games available to Android device users via cloud game streaming. This is a viable option, even for high fidelity or high quality games.

2627 Let me now at this point say something about the available choices as to monetisation strategy. The options available to developers are not limited to the platforms on which they make their apps, or the distribution channels within those platforms. Developers also have various options with respect to how they monetise their apps and games, as Google points out and as I have also described elsewhere.

2628 First, there is the “free to play” or “freemium” model, whereby the app or game is free to download, but digital content within the app or game including perhaps extended game play requires an in-app purchase.

2629 Second, there is the “advertising” model, whereby the developer displays paid advertisements to users within the app or game and earns revenue on those advertisements from the advertiser.

2630 Third, there is the “premium” or “pay upfront” model, whereby the developer charges a price to download the app.

2631 Fourth, there is the “subscription” model, which involves offering access to digital content within the app if the user pays a regular subscription.

2632 Fifth, there is participation in Google’s Play Pass, which is a subscription service sold by Google which provides users with access to a bundle of more than 1,000 apps.

2633 Sixth, one can have a combination of any one or more of the above monetisation strategies.

2634 Now not all of these monetisation strategies result in revenue for the Play Store.  So, if a developer chooses the advertising model, the Play Store does not receive any service fees on the ads displayed within the app.  Further, under the subscription model, the developer has the option of making their app consumption only, which means that users can only access digital content within the app if they pay for digital content or a subscription outside the app.  So, the digital content or subscription is not offered as an in-app purchase.

2635 Developers have options as to the transaction venues that they use for monetising their apps.  The fact that developers can sell access to app-based digital content via a subscription or one-off purchase outside the app on a website means that developers can choose to distribute their app through the Play Store but avoid paying service fees by offering out of app purchases.

2636 Further, as Google says, the fact that app developers can offer cross-play, cross-progression and cross-wallet functionality in their apps means that developers can choose to distribute their app through the Play Store, but avoid paying service fees when users choose to purchase in-app content on a different platform, such as in a version of the app on a gaming console or PC.

The concentrated nature of the Play Store’s revenue

2637 Let me finally deal with one more topic raised by Google concerning what it says is the risk presented by the concentrated nature of the Play Store’s revenue.

2638 Now the evidence establishes that the Play Store’s apps and games revenue is heavily concentrated in numerous respects.

2639 So, a May 2019 Project Magical Bridge document records that apps and games revenue is highly concentrated in the following ways: by category, with 88% of revenue coming from games rather than non-game apps; by developer business model, with 94% from in-app purchases; by geography, with 29% from the United States, 22% from Japan, 13% from Korea; by developer, with less than 1% of developers accounting for 90% of spend; and by user, with 3% of users being high value users who account for 50% of revenue.

2640 So, the Play Store’s revenue is vulnerable to advances by other participants in the ecosystem, who may naturally seek to target the highest value developers on the Play Store, and to developer and user choices.  Perception of this risk is also reflected in Google’s internal documents.

2641 So, it may only take a small number of prominent developers to make choices that deprive the Play Store of revenue to have a significant impact on Google’s profitability.

2642 So, if a handful of prominent developers threaten to switch to alternative means of distribution, Google may have to respond by enhancing its value proposition for them.  Likewise, if high value users are switching to alternative app stores in order to obtain their favourite games, Google may have to respond by enhancing its value proposition for them. There are subtleties to the developer position which I will discuss later when I address Project Hug and “beacons of the ecosystem”.

2643 Undoubtedly this is relevant to any competitive constraint that Google may face, although it has little if anything to do with market definition as such given that on any view users and developers do not act as a close constraint.

Summary

2644 In terms of the question of substitution, my principal conclusions are the following.

2645 First, the only close substitute for Android app distribution services is other Android app stores.  Android app stores do provide a competitive constraint on the Play Store such as to be a close substitute.

2646 Second, there is some substitutability between sideloading of apps and Android app distribution services, but it is not a close constraint, and users and developers do not consider it to be a close substitute.

2647 Third, the pre-installation of apps is not a close substitute for Android app distribution services.

2648 Fourth, there is some competition with PC and console operators and mobile app stores, but there is not close substitution.

2649 Fifth, out of app payments would not constrain the hypothetical monopolist of Android app distribution services.

2650 Now I should note that the fact that I have found that various other products and services are not close constraints and therefore are not part of the relevant product dimension of market definition does not entail that the availability of such products and services do not provide some form of competitive constraint.  In my view they do provide a competitive constraint to some extent, although not such as to deny Google’s substantial market power in the relevant market.

Is there any competitive tension between Google and Apple?

2651 Now I accept that Google faces some competitive constraints from Apple and the App Store with respect to Android OS and the Play Store.

2652 First, Google perceives a threat of losing developers wholly or partially to iOS and has in some respects sought to mitigate this threat.

2653 Second, Google perceives a threat of losing users wholly or partially to iOS and has in some respects sought to mitigate this threat.

2654 Now in my view Epic understates the level of competition and Google overstates such a level.  But on any view, Apple does not provide any close competitive constraint.

2655 I have discussed this question to some extent in my reasons in Epic v Apple in considering the question from the perspective of Apple.  Let me now look at this from the perspective of Google and users or developers concerning Android apps.

2656 Now in my view services for the distribution of native apps on iOS are not a close substitute for Android app distribution services for users or developers, largely for the reasons that Epic has pointed out.

2657 First, native iOS apps will not function on Android devices, such that users would have to switch from an Android device to an iOS device before they could make use of services for the distribution of native apps on iOS.

2658 Second, developers who switch to services for the distribution of native apps on iOS would be unable to reach Android device users through those services.

2659 Third, it is uncommon for Android device users to multi-home and users face significant barriers to switching from Android to iOS devices.

2660 Fourth, in deciding what mobile device to purchase, users do not place any real weight on the price of apps and in-app purchases or the number and variety of apps. I note that that proposition is reasonably justified by Professor Dhar’s surveys.

2661 Fifth, Android device and iOS device users represent distinct user bases.  Developers do not distribute iOS apps instead of Android apps.

2662 Now I have discussed the detail of this elsewhere. Let me turn to some more specific aspects of Google’s position and the various points that it has made, many of which have some merit putting to one side for the moment that it overstates its position at times.

Google’s arguments

2663 Google says that the evidence demonstrates that there is actual and potential substitution of users and developers, including in terms of relative investment and release prioritisation particularly by developers of popular apps, between Android and iOS devices. And this is amplified by the presence of same-side and cross-side network effects between users and developers.

2664 Further, Google says that competition with iOS is a principal driver of Google’s strategy for Android and the Play Store, its constant investment in the development of innovative features for Android devices, and the pricing and quality of the services provided by the Play Store.

2665 Further, Google says that there are strong and enduring economic incentives for Google to partner with OEMs to ensure that Android devices are developed, promoted and remain attractive for users and developers. That is because Google can only monetise on Android when users actively engage with their devices.

2666 Further, Google says that unlike iOS, Android is a free and open-source operating system which is available for use and modification. This means that Android is susceptible to fragmentation. But fragmentation is not a problem for iOS devices because Apple does not face this externality by virtue of its vertically integrated model.

2667 Google says that whilst this means that Android’s openness and flexibility contrasts with Apple, it also means that the success of Android is dependent on the following matters.

2668 First, Google’s success is dependent on attracting OEMs to build Android-compatible devices, particularly high-end devices that are closely competitive with Apple’s devices.  Google says that its primary motivation in partnering with OEMs is to compete with Apple and to prevent users choosing iOS over Android.

2669 Second, Google’s success is dependent on securing developers to create apps for Android and incentivising them to invest in and prioritise their Android apps. This is one of the ways in which Google competes with Apple.

2670 Third, Google is ultimately dependent on users purchasing and continuing to purchase Android devices.  And the initiatives implemented by Google to attract iOS users and retain Android users is another way in which Google and Apple compete and is reflected in the rates of switching between these ecosystems.

2671 Google says that it competes with Apple to draw users and developers to its platform, including in respect of features of the operating system, security functions and ease of use. Mr Gennai gave evidence that competition with Apple for both users and developers is a “fundamental driver” for its investment in the Android ecosystem.

2672 Further, Google says that the empirical evidence suggests a high degree of switching between users of Android devices and iOS devices, with the Android-to-iOS switching rate having reached 25% in 2015.  Professor Asker also estimated that the net switching rate from 2019 to 2021 ranged between 5% to 13% in the direction of iOS.

2673 Further, Google says that Epic’s contention that the switching rates should be considered in the context of users who were already changing their smartphones oversimplifies the switching process.  Google says that the focus on the small window in which users are actively looking for a new device is misplaced, as a decision to switch in that period will be informed by the user’s experience of their existing device over its lifetime.

2674 Further, Google says that in recognising both the prospect and reality of a user switching between Android and iOS devices, Google has undertaken several initiatives designed to achieve the following objectives.  It made the following points.

2675 First, Google has sought to increase the number of users switching to Android devices from iOS devices.  So, in November 2020 Google entered into a “Go-To-Market” agreement with Samsung whereby it agreed to [REDACTED]  [REDACTED]  [REDACTED] [REDACTED] [REDACTED]   [REDACTED]   [REDACTED]   [REDACTED].  The agreement noted the desire of Samsung and Google to collaborate to promote the sales of Samsung’s Android phones and tablets. For a device to qualify for a payment under the agreement, it was required to have [REDACTED] [REDACTED] [REDACTED] [REDACTED] [REDACTED] [REDACTED]  [REDACTED] [REDACTED]  [REDACTED]  [REDACTED]  [REDACTED]  The agreement reflects two matters.  Google was incentivising Samsung to compete for iOS users, [REDACTED][REDACTED] reflecting the competition which Android devices face from iOS devices.  Further, it was in Samsung’s commercial interests to collaborate with Google in the promotion of Samsung devices, in competition with iOS devices.

2676 Second, Google has sought to invest in improving the quality and affordability of Android, together with enhancing the switching experience from iOS to Android, and highlighting new features, form factors and the perception of “newness” of Android devices. Apple also uses the marketing strategy of perceived “newness”.

2677 Third, Google has sought to retain Android users generally, and otherwise reduce the level of churn to iOS. Google has partnered with OEMs for the purpose of increasing the competitiveness of Android devices relative to iOS devices.  Google’s promotion of a consistent, premium and high-quality out-of-the-box and continuing experience for Android users is to ensure that Android devices remain competitive with, and to attempt to reduce switching to, iOS devices.

2678 Fourth, Google has sought to retain and win high value users (HVUs) from iOS, including by developing and introducing a loyalty reward program on the Play Store as a direct response to competition from other app stores in Japan for HVUs.

2679 Fifth, Google has sought to retain and win high value gamers, and in that respect has recognised that Android’s overall health could be improved by reducing iOS churn.

2680 Further, the purchase of an Android device by a user is necessarily a loss of sale opportunity for Apple’s iOS and vice versa.  Further, if OEMs lose Android device sales to iOS devices, Google will lose Android users and earn less from the Android OS.

2681 Further, Google says that it competes with Apple for developers.

2682 Google faces competition to distribute apps, including from iOS.  Often, developers do not develop apps for both Android and iOS by default. They often make a choice between the two and prioritise developing apps for certain platforms or ecosystems, with Epic and many others developers having prioritised launching on iOS rather than Android.

2683 Further, the Play Store’s strategy is based on the fact that user take-up of Android devices will be adversely affected by any significant difference between the quality and diversity of apps available on the Apple’s App Store.

2684 Further, Google says that it has recognised that the impact of popular apps launching exclusively on iOS, or on iOS before Android, is amplified by same-side network effects. The concern for sim-shipping partially stemmed from the technical debt that historically existed between Android and iOS, where developers, particularly gaming developers, faced technical issues in creating games for Android as opposed to iOS.

2685 Google says that this asymmetry between apps available on Android and iOS impacts Google’s ability to compete. As such, Google recognised that attracting a wide range of quality developers to Android would enable Google to close the gap between it and iOS at a faster rate.

2686 Let me turn to the question of fee changes.  Google says that it has revised its service fee in direct response to Apple’s reduction of subscription rates from 30% to 15% in 2016.

2687 Mr Feng’s evidence was that Google’s price reductions were founded on a concern that if Google did not make the same changes as Apple, it would make the Play Store less competitive and less attractive to developers for investing in Android apps distributed via the Play Store.

2688 Google has reduced its fees to 15% on the first USD 1 million in revenue for all developers in the United States, in direct response to Apple’s introduction of its small business program, which implemented reductions for only those developers who made less than USD 1 million per year.

2689 In this way, Google says that it was not simply mimicking Apple. It was actively seeking to undercut iOS. As Mr Gennai says, had Google failed to respond to Apple’s initiative, it would have jeopardised developers’ investment in the Play Store and, ultimately, user and OEM participation in the Android ecosystem. This is supported by the contemporaneous Google documents.

2690 For example, in an 18 November 2020 internal Google email Mr Samat said:

As you’ve seen today apple made their announcement about 15% for devs making $1m or less

…

Do we need to follow suit with a change and how quickly? My sense is we will have pressure from these same sized devs and thus need to do something. But we shouldn’t necessarily do the exact same thing as our challenges are different. For example, I could see us going to zero percent for devs less than $100k and then 15% up to $1m and then 30%.

…

2691 Google says that this communication, which was sent on the same day that Apple introduced its small business program, sits at odds with Epic’s suggestion that Google would have responded at a quicker pace if the App Store and the Play Store were truly substitutes.

2692 The evidence further demonstrates that the Play Store price changes were not immediate because Apple’s changes prompted a thorough and detailed pricing and business model review by Google. The reality was that the issue was complex and took some time to resolve.

2693 Further, Google says that it has also continually initiated, and responded to, pricing changes on the App Store. Exhibit 40 of Professor Asker’s report, which is worth setting out here, sets out a timeline showing both Apple’s and Google’s changes to service fees in the period between July 2008 and March 2022:

2694 As is apparent from exhibit 40, there were multiple price changes with each of Apple and Android as the first initiators.

2695 Further, Google says that the criticism that Apple did not respond to certain service fee changes made by Google goes nowhere. The important point to focus the competition inquiry on is why Google made that change, which was to better compete against Apple and iOS.

2696 Further, Google says that it competes with Apple on device quality and price.

2697 Further, Google says that it continually introduces new features to Android in response to developments on iOS and to front-run iOS, with the Play Store seeking to outperform the App Store with new innovations.

2698 Further, Google says that there is competitive rivalry between Google and Apple with respect to app store pricing.

2699 Now Epic says that Google does not compete with Apple because the only three service fee changes that Google identifies with respect to Apple were not driven by the need to compete with Apple, but were rather to address regulatory concerns.

2700 But Google says that the evidence shows that its perception of the need to compete with Apple was a significant motivating factor in each change to its service fee.

2701 The first change was a reduction in the service fee from 30% to 15% for subscriptions that had been active for more than one year, effective 1 January 2018.

2702 Google says that what prompted it to consider this change was that Apple had reduced its commission for subscriptions active for more than one year from 30% to 15% in June 2016, and the concern that if Google did not follow suit, the Play Store would be less competitive and less attractive to developers investing in the Play Store.

2703 In a 31 July 2017 internal presentation titled “Subscriptions V2 Follow-Up” that considered the potential change for subscriptions, Google referred to the app store commissions then charged by Apple and Amazon, and described potential competitive responses from both Apple and Amazon.

2704 Google says that when it came to make the service fee change, bringing the service fee into line with Apple was one of five reasons for doing so.

2705 In cross-examination, Mr Feng gave evidence that this was an instance of Google competing closely with iOS, and Google had received a substantial number of inquiries about whether and when it would at least match Apple.

2706 The time it took Google to make this change does not gainsay that it was in part a response to Apple. Mr Feng denied that Google waited, and said that “sometimes things take a long time at Google”.

2707 The second change was a reduction in the service fee to 15% for the first USD 1 million earned each year by all developers, effective 1 July 2021.

2708 This service fee change was the ultimate conclusion of Project Runway, which was a project led by Mr Feng that had the purpose of considering how Google could respond to Apple’s reduced 15% commission for small business developers announced on 18 November 2020.

2709 Google says that this evidence is corroborated by the fact that a November 2020 presentation titled “Project Runway”, just after Apple announced its 15% commission for small business developers, was focused on options to respond to Apple, and viewed “Do Nothing” as “Not recommended” and “Match Apple” as “Insufficient”.

2710 It is also supported by an 18 November 2020 internal email from Mr Samat, which was sent on the same day as Apple introduced its small business program, in which he said:

As you've seen today apple made their announcement about 15% for devs making $1m or less

…

Do we need to follow suit with a change and how quickly? My sense is we will have pressure from these same sized devs and thus need to do something. But we shouldn't necessarily do the exact same thing as our challenges are different. For example, I could see us going to zero percent for devs less than $100k and then 15% up to $1m and then 30%. …

…

2711 Google says that its service fee change went further than Apple’s because it applied to all developers, not just those earning less than USD 1 million, and Mr Feng regarded this as a “potentially subtle, but very important difference”. This is competition at work.

2712 Mr Feng also gave evidence that he considered this change to be critical for ensuring that the Play Store remained competitive with the App Store, and maintained this view in cross-examination.  And whilst Mr Feng acknowledged that one motivation for the change was responding to regulatory pressure, he disagreed that it was the primary reason or that the change otherwise would not have been made.

2713 The third change was a reduction in the service fee for subscriptions to 15% in the first year, effective 1 January 2022. Mr Gennai gave evidence that this change was made to ensure the Play Store would remain competitive with the App Store.  Google says that Epic has not pointed to any document which contradicts that rationale for this particular service fee change. With the change, the Play Store undercut the fee rate applied by the App Store which applies a 30% rate to subscriptions in the first year.

2714 Further, there are several app categories for which the Play Store charges are lower service fee than Apple, which is also consistent with the Play Store facing competitive constraint from the App Store.

2715 Let me turn to another topic.  Google says that it closely monitors the quality of user and developer experiences, and the competitiveness of Android as against iOS.   It has made the following points.

2716 First, Google says that it has monitored the gap between Android and iOS, particularly in respect of app quality, high value gaming and monetisation. Google has sought to compare performance on Android and iOS, and track results over time in respect of game quality metrics. It has also noted that top games do not perform as well on Android, and that the monetisation gap was caused in part by high value gamers favouring iOS devices over Android devices.

2717 Second, Google says that it has monitored the performance of the Play Store as against the App Store, including the range and flexibility of distribution of apps, number of installs, total consumer spend, top gaming releases and sim-ships on each platform.

2718 Third, Google says that it has monitored the revenue share on the Play Store as compared to the App Store and Apple’s revenue share for subscriptions.

2719 Fourth, Google says that it has monitored developments with Apple gaming initiatives such as Apple Arcade, and has perceived the need for Google to develop its cross-screen capabilities in light of the stronger alternatives.

2720 Fifth, Google says that it has monitored Apple’s initiatives with developers, including the reductions in service fees and the way it implemented the small business program. These were not only identified and reported internally within Google immediately, but steps were taken to respond swiftly.

2721 Further, Google says that it has frequently benchmarked itself against Apple and iOS and has made the following points.

2722 First, Google had regard to Apple’s responses arising from the setting of Google’s subscription fees in 2017. In particular, Google compared the revenue share models of Apple and Amazon.

2723 Second, the change in Apple’s service fees prompted Google to consider how Google could respond to Apple’s initiative. Contrary to Epic’s assertion that Google’s strategy vis-à-vis Apple was to engage in “mimicry”, as part of continually re-evaluating its business model, Google determined that such a strategy (“match Apple”) would be “insufficient”, instead opting to undercut Apple. Google also assessed Apple’s potential response if Google was to create its own independent business model for the Play Store.

2724 Third, Google has conducted surveys of users to ascertain perceptions of its brand, product quality and general attractiveness to users, as compared to iOS and how to better improve the Android experience to draw iOS customers.

2725 Fourth, Google has commissioned surveys of developers who distributed their app or game through the Play Store in respect of their perceptions of brand quality, safety, privacy, value and developer experiences generally, including metrics which compare their experiences against the experiences on the App Store.

2726 Further, Google pointed to the following witnesses’ evidence concerning Google’s perception of competition with iOS for developers.

2727 Mr Gennai gave evidence that Google perceives a risk that developers will not spend time creating apps for Android at all, or will make apps for iOS first, or will not invest in sufficient quality on Android apps, leaving Google at a competitive disadvantage to iOS. Developers often make a choice between Android and iOS and prioritise certain platforms or ecosystems.

2728 Mr Rawles gave evidence that Google is principally concerned with ensuring that game developers choose to develop for Android alongside iOS, that is, they do not choose to only develop for iOS or iOS first.

2729 Ms Kochikar gave evidence that attracting a wide range of quality developers to Android would enable Google to close the gap between it and iOS at a faster rate.

2730 Google says that this evidence was supported by its internal documents.  It referred to the following documentary material.

2731 A 9 April 2015 presentation titled “Reinventing Strategic Merchandising” identified as a high priority the need to encourage developers to invest in apps for Android, to sim-ship on Android and iOS and to invest in new features and updates on Android on par with iOS. It recorded that a significant proportion of apps launched on iOS were launched first on iOS or only on iOS. It also recorded what action Google considered it needed to take for apps that ship first or only on iOS, namely to ensure the same performance as iOS so that developers have incentives to develop for Android and do so at the same time as iOS.

2732 An April 2017 presentation titled “Android Games Strategy Review” recorded that Android users at that time had a “suboptimal gaming experience” versus iOS, and that there were “gaps in premium titles & indies” and “big simship misses”. In other words, developers were either not launching significant titles on Android, or releasing them after iOS. Developers preferred to launch on iOS due to better revenue growth and monetisation, and made inferior quality apps for Android OS due to performance issues.

2733 An internal email dated 15 May 2018 from Ms Kochikar recorded the concern that “Indies” were continuing to launch exclusively on iOS, reinforcing iOS as “the gaming platform”, and stated Google’s perception that it needs to “combat” Apple. It also recorded that high fidelity games were “gated by lots of tech issues across Android,” creating “the risk of an iOS only launch”. Ms Kochikar’s “nightmare scenario” was a big game from a large developer launching first on iOS.

2734 In an internal Google email on 8 February 2019, Mr Gennai said that “if an Android user has to think even twice about whether a game is available on their Android phone, we’ve already lost a battle to Apple”.

2735 A November 2020 presentation titled “App Quality Gap 1. Initial thoughts & learning plan” records Google’s analysis of whether an app quality gap to iOS exists. It shows not only a consumer perception of an app quality gap, but that a quality gap continues to exist as a matter of fact. It also records developer feedback that developers follow an iOS “first business strategy” and design first for iOS.

2736 Another November 2020 presentation titled “The Gen Z Effect: Tech, trends, and truths for a new(er) generation” records that the Play Store does not have some of the apps relative to the App Store which Generation Z users care about.

2737 Further, Google says that its concern about sim-shipping on Android OS reflects its perception that it is in a contest with Apple’s iOS.

2738 Further, Google says that there is a perception of competition between iOS and Android for users. It referred to the following documents.

2739 A 22 July 2016 presentation titled “Close-the-Gap” records Google’s concern about a widening app store revenue gap between the Play Store and the App Store. It observes that past the Play Store revenue growth versus iOS was due to improvements in the Play Store’s app catalogue, which reflects the need to incentivise developers to make apps for Android, and in the Play Store’s billing infrastructure. But at that time, there was a gap in app store revenue and in the apps available on Android versus iOS. Google identified strategies to mitigate this gap, including 19 action plans.

2740 An April 2017 presentation titled “Android Games Strategy Review” also recorded a monetisation gap versus iOS with respect to the Play Store, and that high value gamers favour iOS. The document drew a connection between high value gamer preference and the fact that Android lags iOS on tablet and has a frustrating user experience when games do not work on Android devices.

2741 In a 17 April 2018 email, Google’s director of Android Play Partnerships, Mr Richard Turner, described an “existential threat” from iOS, because the iPhone had won market share in premium devices and was threatening to win Android users of mid-range devices. The plea to the recipient of the email, Mr Rosenberg, was to increase focus and efforts on “defending Android share in the mid-range from iOS”.

2742 A 4 December 2019 presentation titled “JP Play OS Platform Gap” recorded that despite Android having a 50/50 OS share with iOS in Japan, consumer spend was heavily skewed to iOS and Android’s share had decreased from 46.9% to 38.8% over the past four years. The issue causing this was a significant gap in Japan between iOS and Android amongst HVUs and young gamers. The presentation also recorded a concern about the potential for a similar gap to develop in the United States and identified potential pilot programs to mitigate this loss of share.

2743 Further, Google pointed out that an important cohort of first-time smartphone purchasers are young users. Two-thirds of first-time smartphone purchasers in 2018 were so-called millennials. Further, Google was concerned about Android’s low popularity amongst Generation Z users.

2744 Further, it said that 4 to 6% of smartphone purchases in Australia in 2021 were from first-time purchasers. Accordingly, the possibility of first-time smartphone purchasers choosing iOS devices poses an additional constraint on Google.

2745 Further, Google made reference to the following other evidence.

2746 Mr Feng’s view was that the Play Store competes with the App Store, as well as with other distribution channels available on Android and other digital content distribution platforms such as gaming consoles.

2747 Further, Ms Kochikar said that Google’s objective in entering into a deal with Samsung in April 2019 for the provision of exclusive content to Samsung users via the Play Store was to allow Android devices to compete with iOS.

2748 Further, Mr Gennai’s evidence was that iOS is Android’s closest competitor in competing for developers, primarily because the large number of iOS users makes iOS a highly attractive platform for developers to develop apps for. Mr Gennai also gave evidence that there are some apps that exist on both Android and iOS, but with quality differences. He said that these differences were a regular focus of app reviews and some media commentary, often with a sentiment favouring iOS. Mr Gennai said that this has been a longstanding concern of the Play Store team and is a reason why Google works hard to assist developers and encourage them to invest in developing for Android. Further, Mr Gennai gave evidence concerning the reasons why developers may launch on iOS instead of Android.

2749 I should say that I have reflected on all such evidence although I do think that Ms Kochikar and Mr Gennai were prone to exaggeration as to the level of competition between iOS and Android.

Analysis

2750 Now as I have discussed in my reasons in the Epic v Apple proceeding, Apple says that Google and Apple compete for the business of developers, including the transactions that take place in respect of developers’ apps.  Further, I have discussed the parties’ various positions concerning competing app stores.

2751 I would incorporate by reference in these reasons what I said in my Apple reasons concerning the relevant evidence, arguments and findings.  Let me deal with some additional matters or other dimensions to matters that I have already discussed.

2752 Now as Epic points out, the long-standing commercial agreements in place between Google and Apple are the opposite of what one would expect to find if they were close competitors. In truth, Apple and Google describe themselves as partners, and recognise that their relationship is mutually beneficial.  That is not the language of rivalrous firms seeking to replace each other.  Let me elaborate here, although I have touched on these issues elsewhere.

2753 Although Apple is not an Android device OEM, Google has entered into agreements with Apple which are relevant to Apple and Google’s contentions that they compete with one another.

2754 On 20 December 2002, Google and Apple entered into a contract known as the information services agreement.

2755 The information services agreement contains terms under which Apple has agreed to make certain Google services, including Google Search, available on Apple smart mobile devices.  It has been amended at least nine times.

2756 Pursuant to amendment eight to the information services agreement, effective as of 30 September 2016, the following may be noted as Epic has pointed out.

2757 First, Apple agreed to make Google Search the default search engine in all geographies for use in its web browser software on iOS and other operating systems made generally available by Apple, or the Microsoft Windows operating system. “Default” here means that Google Search will automatically be used for responding to search queries initiated from Apple’s web browser software, unless the user selects a different search service.

2758 Second, Apple agreed to provide [REDACTED] [REDACTED] [REDACTED] [REDACTED] [REDACTED] [REDACTED] [REDACTED] [REDACTED] [REDACTED] [REDACTED] [REDACTED] [REDACTED] [REDACTED] [REDACTED] [REDACTED] [REDACTED] [REDACTED]

2759 Third, Google agreed to pay Apple a percentage of its net ad revenue, [REDACTED] [REDACTED] [REDACTED] [REDACTED] [REDACTED] [REDACTED] [REDACTED] [REDACTED] [REDACTED] [REDACTED] [REDACTED] [REDACTED] [REDACTED] [REDACTED] [REDACTED] [REDACTED] [REDACTED] [REDACTED]

2760 Further, Google agreed to pay Apple a percentage of its “Google app” search application/Chrome net ad revenue [REDACTED] [REDACTED] [REDACTED] [REDACTED] [REDACTED] [REDACTED] [REDACTED] [REDACTED] [REDACTED] [REDACTED] [REDACTED] [REDACTED] [REDACTED] [REDACTED] [REDACTED] [REDACTED] [REDACTED]

2761 So, Apple receives a significant sum from Google per year. A Google presentation titled “Apple Deal Staples” records that payments to Apple were expected to grow 41% in 2018, from USD 6.7 billion to USD 9.4 billion. Those payments are now well over USD 10 billion per annum.  And Google earns billions of dollars per year in advertising revenues from being the default search engine in Apple’s web browser.

2762 Further, on 15 May 2014, Google and Apple also entered into a joint cooperation agreement. Under the terms of this agreement, [REDACTED] [REDACTED] [REDACTED] [REDACTED] [REDACTED]  This is hardly the stuff of vigorous competition, with [REDACTED][REDACTED][REDACTED] being the usual weapon of choice, deploying hordes of feral [REDACTED]

2763 Let me move to another topic.  As Epic says, Google’s principal concern is to quash competition from other Android app stores.

2764 Now Google says that it nonetheless competes closely with Apple to attract a greater share of developers’ finite investment resources. Google says that it seeks to encourage game developers to sim-ship games on iOS and Android, as well as to build higher quality games for Android.

2765 But I agree with Epic that sim-shipping does no more than ensure that the Play Store receives releases at the same time as the App Store. Such measures would not bring about any switching from Apple to Google and do not show Google seeking to compete with Apple for developers or users.

2766 Further, Google says that its strategy for the Play Store is premised, in part, on the proposition that consumers’ take-up of Android devices will be adversely affected by any significant difference between the quality and diversity of apps available on the App Store relative to the quality and diversity of apps available on the Play Store. Yet, as Epic points out, it is curious that Google distributes a number of its proprietary apps through the App Store. Those apps include Google Maps, Google Chrome, Search, YouTube, Gmail, Google Drive, YouTube Music, and Google Photos.

2767 I tend to agree with Epic that Google’s conduct in making its highly popular apps available on the App Store is inconsistent with the premise described.

2768 Further, I am here not concerned with any direct or indirect competition for the supply of devices in any mobile devices market. I am here concerned with discussing the question of app distribution, although, undoubtedly, direct or indirect competition for the supply of mobile devices is a relevant contextual matter.

2769 Further, I agree with Epic that there has not been meaningful price competition between Google and Apple in respect of app distribution. The following matters are sourced to the evidence that I have discussed elsewhere as correctly summarised by Epic.

2770 Now with the exception of developers offering subscriptions, Google has charged developers a 30% commission on earnings over $1 million for approximately 13 years with no change. It has persisted with the 30% commission rate despite being aware of significant developer dissatisfaction towards it.

2771 Further, on the two occasions when Google has modified its commission following changes in Apple’s pricing, those modifications did no more than match, rather than undercut, Apple’s pricing, and would seem to have been motivated by regulatory and reputational aspects as opposed to competitive pressures.

2772 Further, as I have discussed elsewhere, the 30% commission is not a competitive price. It is not calculated by reference to Google’s costs of supply and it is not set at a level which would entice developers away from competing app stores.

2773 Moreover, Google’s internal documents observe that there is no real rationale for the quantum of the commission other than matching the fees charged to developers by Apple.

2774 Now Mr Feng sought to justify the 30% commission broadly on the grounds of the value that Google considers that developers derive from the Play Store. But I agree with Epic that that reflects a supplier setting a price unilaterally and then seeking to justify the price that it has already set.  Developers may derive value by reason of being able to distribute Android apps and earn 70% of the price which users pay. But that says nothing about whether that rate is competitive.

2775 Further, Google has not changed the price it charges developers on their earnings over $1 million in about 15 years, and those are the developers from which Google derives the most revenue. The equivalent position applies in respect of Apple.

2776 Further, I agree with Epic that Google and Apple do not exhibit the sense of urgency that typifies rivalrous behaviour.

2777 As Ms Edwards-Warren said, if consumers do in fact switch readily in response to changes in price or quality, one would not expect to see the types of lags in price responses that she observed.

2778 Further, as Epic has said, Apple’s and Google’s changes to their commissions in respect of small developers has primarily benefited the long tail of small app developers or, in the case of the Play Store, the first $1 million of an app developer’s revenues, and has not affected the bulk of the App Store and Play Store’s revenues.

2779 But as Professor Wright observed, competitive pressure, if it exists, generally incentivises firms to:

… offer discounts to their largest customers on the supplier side given these suppliers may be in a better position to switch their business to competitors if they do not get a discount.

2780 So the pattern observed as to Apple’s and Google’s charges and behaviour is not consistent with strong competitive rivalry.

2781 Indeed, as Epic says, Apple and Google’s approach stands in contrast to Steam’s response to price competition from the Epic Games Store which resulted in the largest developers benefitting from the largest commission reductions consistent with a rational response in the face of competition.

2782 Further, as Epic correctly points out, the evidence also shows that the Play Store and App Store do not compete on quality.

2783 Now Professor Hitt and Mr Houston listed developments which they said showed that Apple strives to provide a high quality of service. Similarly, Professor Asker identified what he called competitive responses based on Google’s introduction of similar innovations on Android. Again, parallel technical development does not amount to close competition.

2784 But as Epic pointed out, ten of the 15 innovations identified by Professor Asker have a lag time of more than one year.  It makes the obvious point that if Google and Apple were investing in the Play Store and the App Store in response to competitive pressure from each other, one would expect to see short lag times between Google and Apple’s respective innovations and responses to prevent user and developer switching.

2785 Moreover, as Professor Wright pointed out, the investments made by Google and Apple are equally consistent with investments made by a dominant firm in each case.  And as he said, it is well established that monopolists can have an incentive to invest. And it is problematic to in effect reason, say, that an entrenched monopolist like Google does not have market power in the market for online searches because they have implemented several innovations to their search engine.

2786 Further, Ms Edwards-Warren observed that a hypothetical monopolist would invest if the return on its investment meant that it was going to increase its profit. So, it does not follow that just because Google invests to attract developers and users to the platform, it must be facing competitive constraints. I tend to accept her view that Google would make those investments regardless of whether it was facing competitive constraints or not.

2787 Finally, I agree with Epic that most references to iOS within Google documents can best be understood as a form of benchmarking exercise.  But as Ms Edwards-Warren said:

… even if there is some competition between the iOS and Android ecosystems as a whole, and benchmarking of specific components within each ecosystem, this does not imply that Android Smart Mobile Device users would view the Apple App Store as a viable alternative to the Play Store (or other Android app stores) in response to a small increase in app distribution service fees.

2788 Generally, as I have said in my reasons in the Epic v Apple proceeding, there is no evidence that Apple and Google closely compete to distribute apps and parallelism is not competition.

2789 In summary, the evidence does not demonstrate that Google is significantly constrained by Apple within the  relevant market.

Market power – Android mobile app distribution

2790 Epic says that Google has and had during the relevant period a substantial degree of power in the Android mobile app distribution market, but Google does not accept this.

2791 Now the principles relating to the concept of substantial market power are largely uncontroversial. But Google emphasised three matters.

2792 First, substantial market power is the capacity of a firm to act persistently in a manner substantially unconstrained by the conduct of competitors.  So, if the firm has the capacity to raise prices above the supply cost without rivals taking away customers in due time, or to persistently exclude competition with restrictive conduct, this constitutes substantial market power.

2793 Google says that it did not have and does not have that capacity with respect to supplying Android OS licenses or with respect to the services which the Play Store supplies, even in Epic’s pleaded markets.

2794 Second, a necessary condition of substantial market power is the existence of barriers to entry. Absent barriers to entry, even a monopolist will be effectively constrained due to the standing credible threat of entry by rational firms if it were to raise prices and start making supra-competitive profits.

2795 Now Google says that this is important with respect to the Play Store.  It says that the evidence does not support a conclusion that there are high barriers to entry. It says that there are plenty of Android app stores and web stores. Epic itself is now launching its Epic Games Store on mobile, and so is Microsoft.  Google says that this is fatal to Epic’s claim that Google has substantial market power with respect to the Play Store.

2796 But in my view Google’s position as the first mover has fortified or entrenched its position to a significant extent.

2797 Third, matters such as high market shares and high profits are merely factors that may or may not be indicative of market power, but they are not determinative.  I should note that forensically speaking, these factors favour Epic’s position on market power.

2798 Now Professor Asker gave relevant evidence on this topic, most of which I accept at the level of generality with which he expressed himself.  Let me synthesise and set out some of his themes.

2799 The key economic concept behind market power is that a firm can elevate prices or worsen quality relative to competitive levels because its customers lack available substitutes.

2800 In that sense, the question of whether a firm has market power is closely related to defining relevant markets and the HMT.

2801 Typically, in a competition inquiry, economic analysis proceeds by first defining relevant markets and then assessing the extent of a firm’s market power within that market.

2802 Economists typically look at several indicia to assess the extent of a firm’s market power, including market shares, investment or innovation, ease of entry, and profit margins.  None of these indicia on their own are dispositive as to market power, and economists instead rely on the weight of evidence by examining multiple indicia.

2803 First, market shares are commonly used to infer market power. Inference is done either by examining shares directly, by computing various market concentration metrics, or both.  Market shares are generally measured as revenue shares or quantity shares.  The idea is that a firm with a high market share in a relevant antitrust market may face a limited number of strong competitors or, equivalently, compete against fewer compelling substitute products.

2804 There is no bright-line threshold for market shares that separates the absence of market power from its existence. But market shares below 30% are generally viewed as insufficient for conduct to give rise to any competition concern.

2805 High market shares can go together with a lack of market power. This can occur when a firm outcompetes existing rivals but faces potential entrants or fringe competitors that can readily enter or expand and compete successfully.

2806 When this occurs, even a firm with high market share can be constrained by potential entrants and potential expansionists to whom its customers could substitute in the event the firm raised price or worsened quality. Typically, for this to occur the market needs to exhibit low barriers to entry or expansion relative to the returns a firm can earn by entering or expanding were the incumbent to raise its prices.  A market with these characteristics is often referred to as contestable.

2807 Second, when a firm continues to make substantial investments in its product(s), economists recognise this as reflecting competitive rivalry.

2808 Investment in developing or providing better products reflects an effort to respond to the presence of competitively relevant substitutes.

2809 In certain contexts, it is important to distinguish investment designed to grow a market versus investment that is part of competitive rivalry. This is because firms with market power can continue to invest in a product to attract new customers that are not purchasing rival products.

2810 The way to distinguish these types of investments is to understand what types of customers the firm is trying to attract or retain. If the firm is trying to prevent a rival from poaching customers, or to attract a rival’s customers, the investment is consistent with competition rather than growing the market.

2811 Third, the competitiveness of a market is determined by both the rivalry of firms active in the market and by the prospect of entry by competitors should prices be too high, or product quality be too low.

2812 So, investigating the scope for the entry of new products, firms, or forms of product delivery, for example, can be relevant to the assessment of market power.

2813 Fourth, profit margins can be informative as to market power.  But inferring market power from profit margins is complicated by measurement issues, by determining the relevant competitive benchmark, and by understanding what might constitute a high profit margin.  I have discussed these matters and the relevant theory in detail in my reasons in Epic v Apple.

2814 Economic profits account for opportunity costs facing a firm, that is, costs a firm incurs by not having deployed its resources elsewhere and earned a return from doing so.  The cost of something is what you give up to get it. The opportunity cost of an item refers to all the things that must be forgone to acquire that item.  When economists speak of a firm’s cost of production, they include all the opportunity costs of making its output of goods and services as well as all explicit and implicit costs.

2815 An economist measures a firm’s economic profit as the firm’s total revenue minus all the explicit, implicit and opportunity costs of producing the goods and services sold. Contrastingly, an accountant measures the firm’s accounting profit as the firm’s total revenue minus only the firm’s explicit costs.

2816 These measures of accounting profits have value to firms and other relevant parties, but do not capture the relevant economic information of interest.  Accountants are often focused on the profitability of the business overall, often from a retrospective viewpoint and often shaped by compliance with government regulation.

2817 Now determining what constitutes a high margin requires an understanding of the institutional features of the market.  But another way of determining what constitutes a high margin is to compare a firm’s margins to a competitive benchmark.

2818 Economists can seek to find a competitive benchmark or benchmarks that may serve as proxies for what margins would be in the but for world.  The relevant competitive benchmark needs to reflect the dynamics of the industry that would prevail in the but for world, some of which would be the same as in the actual world.

2819 So, it should account for ex ante risk and the need to cover fixed costs of production and product development being overhead costs without which the product would not exist.  It should also reflect the competitive dynamics that would reasonably be expected to arise in the but for world. So, it is inappropriate to use a commodity market for a benchmark when the but for world is likely to include meaningfully differentiated products.

2820 Now a failure to use an appropriate competitive benchmark can incorrectly result in inferences of high margins that may, instead, reflect a necessary condition for competition to exist in the market in question, for example, risk-adjusted return on investments in fixed costs.

2821 Let me now turn to some specific topics.

Are there high barriers to entry?

2822 Google says that Epic needs to establish high barriers to entry for rival Android app stores before one can conclude that Google has substantial market power with respect to the Play Store.

2823 Now Epic says that high barriers exist because the OEM agreements give the Play Store prominent placement on Android devices and make it difficult for rival apps stores to obtain distribution via pre-installation, sideloading or the Play Store.  Further, Epic says that the Project Hug agreements prevented rival app stores obtaining exclusive content from prominent app developers.  But Google says that this is incorrect regarding the OEM agreements and Project Hug.

2824 Further, as to barriers to entry, Google says that Epic ignores the evidence of extensive past, present and future entry by numerous rival Android app stores, which demonstrates that barriers to entry are marginal to non-existent.  It points to the following matters.

2825 First, it says that there are many services that exist to facilitate app developers creating webstores as an alternative means of selling app-based digital content on Android, including Xsolla and ID3. This means that the cost of entry by developer web stores is low.

2826 Second, it says that the low barriers to entry for webstores are corroborated by the fact most of the top 10 apps within the top five 5 categories of apps use webstores.  So, of the top 10 apps in each category, 9 game apps have a webstore, 10 entertainment apps have a webstore, 6 social apps have a webstore, 10 health and fitness apps have a webstore and 7 lifestyle apps have a webstore.  And across all categories, 19 of the top 25 app developers in Australia on the Play Store use a webstore.

2827 Third, it says that several Android app stores have entered the marketplace since 2009, including Samsung Galaxy Store, Aptoide, Amazon Appstore and Huawei AppGallery, and many operate today.

2828 Fourth, Google says that Epic is launching the Epic Games Store on Android no matter what the conditions, and even if the technical restrictions and the prohibition on app stores being distributed via the Play Store remain in place. It says that such a proposal or intention is inconsistent with the existence of any barrier to entry or expansion for rival Android app stores.

2829 Fifth, Microsoft announced that it would be launching its own mobile app store in July 2024.

2830 Now I agree with Epic that there are high barriers to entry.

2831 There have been and there remain formidable barriers to entry and expansion in the Android mobile app distribution market by reason of the various contractual arrangements that Google insists upon.  It is not in doubt that the ability of a firm to prevent entry by potential competitors by reason of contractual arrangements is a matter to which I can and do have regard in considering the extent of Google’s market power (s 46(4)(b)), and irrespective of whether such contractual arrangements or their imposition separately give rise to competition contraventions.

2832 First, by reason of the MADA terms, the Play Store is pre-installed on all Android devices, and placed on the default home screen (DHS) or in the case of certain Samsung devices in the device hotseat. The placement requirements under the MADA assist in blocking the viability of rival Android app stores, such as Amazon’s app store, by impinging on their discoverability by users. Securing DHS placement exclusivity for the Play Store limits discoverability.

2833 The requirement for DHS placement for the Play Store has existed since 2014.

2834 The DHS placement drives users to select the Play Store over a rival app store, and ensures that the Play Store remains the obvious and simple choice for Android device users looking for apps, which is likely to continue driving widespread usage of the Play Store.

2835 Given the MADA requirements for both pre-installation and DHS placement on all Android devices, developers are and will remain highly unlikely to forgo distribution through the Play Store in favour of a rival Android app store.

2836 Second, although OEMs have the capacity to pre-install their own app stores and place them on the DHS, they can only do so on their own devices, whereas the Play Store is pre-installed on all Android devices.  And they cannot do so exclusively but only together with the Play Store, due to the MADA terms.

2837 Now Samsung users have, since about January 2019, had the Galaxy Store appearing on the DHS. But that has made little difference to the Play Store’s dominance. This is due to Google’s brand salience and perceived quality perceptions of the Play Store due to the fact that the Galaxy Store was not available on the DHS prior to 2019.

2838 Third, whilst it is possible for rival app stores to obtain distribution via pre-installation deals with OEMs, Google’s conduct hinders this method of distribution, and this method can be costly and time-consuming.

2839 Fourth, by reason of the RSA3s and MIAs which Google entered into during the relevant period, a number of OEMs were incentivised to refrain from pre-installing or promoting any alternative Android app store and to increase the Play Store’s revenues.

2840 Fifth, by reason of clause 4.5 of the DDA, as well as clause 4.1 of the device and network abuse policy, a rival app store cannot obtain distribution services from the Play Store. In other words, the most effective Android app distribution channel is closed to alternative app stores.

2841 Sixth, although a developer of a rival Android app store might distribute the app store from a website via sideloading, this method presents challenges for such an app store being discovered by users.

2842 Further, until October 2021, Google did not permit directly downloaded app stores to perform a critical function, which was to update apps automatically.

2843 Further, even if a user does manage to find a rival app store on the web and seeks to sideload it to their Android device, they will confront the technical restrictions.  This will likely cause a significant portion of users to abandon the installation process and the rival app store altogether.

2844 Seventh, during the relevant period, by reason of the GVP and AVP agreements, rival Android app stores were unable to offer exclusive or differentiated content, or even early access to content, from the major developers who were party to those agreements with Google.  This is a real impediment to the success of a rival Android app store, since, without access to exclusive or differentiated content from key app developers, it will be even more difficult to entice users and developers away from the Play Store.

2845 Eighth, the Play Store benefits from network effects. Its value and utility to both users and developers increases as more users download, and more developers distribute, through the store.  In contrast, rival Android app stores generally do not enjoy this benefit.

2846 As Google itself noted in a 4 September 2014 presentation titled Project Gabby:

3rd-party app store strategy was based on competing on completeness, but suffered from chicken-and-egg problem: Not enough eyeballs to get the catalog, didn’t have the catalog to get the eyeballs.

2847 Ninth, establishing and attempting to grow a rival Android app store capable of competing with and constraining the Play Store is likely to be expensive. A developer of a prospective alternative app store faces significant set up and scaling costs, including costs of servers and data centres, app curation, security, and customer acquisition including advertising. Most expensive of all will be the cost of persuading developers to distribute through a store which at least initially will have few users.

2848 Let me turn to another topic.

Does Google have a large market share and if so what is its relevance?

2849 Google says that the Play Store’s high market share with respect to Android app downloads and monthly active users is not an indication of market power in circumstances where there are low barriers to entry.  But as I have explained, I have rejected Google’s assertion that there are low barriers to entry.

2850 Further, Google says that the competitive constraint imposed on the Play Store by rival Android app stores is substantial.  Moreover, it says that Google not only perceives rival Android app stores to be a strong competitive threat, but regularly has responded with competitive conduct to meet this threat.

2851 So, Google Play Points was a response to the threat of high value users switching to the Samsung Galaxy Store and Amazon Appstore, as well as to Apple’s App Store.

2852 Further, as to Project Hug, Google says that its provision of substantial and costly incentives for large developers to remain on the Play Store would be irrational if it had substantial market power and was not exposed to competitive threats.

2853 But in my view Google has a very large market share.  Moreover, I agree with Epic that Google’s competitors are weak relative to Google’s position.

2854 Now Google says that unlike for iOS, app developers can choose to distribute their apps via other Android app stores. It says that there is strong competition provided by other Android app stores.  And it points to some statistics to the effect that a large number of apps are available across a range of Android app stores.

2855 But these statistics ignore the extent to which Android device users actually use the Play Store, both in terms of time spent and the proportion of app downloads, relative to the various other Android app stores.

2856 Clearly the Play Store has a dominant position in respect of the proportion of Android devices on which the Play Store is installed relative to competing Android app stores. Unless I state otherwise, I am using the concept of dominance here and in other parts of my reasons in an informal sense rather than addressing the economic concept of market dominance in a formal sense.

2857 This dominance can be seen from the following different metrics.

2858 First, as to pre-installations, from December 2017 to September 2022, the Play Store was pre-installed on between 97.7 and 99.6% of active Android devices in Australia, and between 97.2 and  99.2% of active Android devices worldwide excluding China.

2859 Its closest rival on this measure is the Galaxy Store, which was pre-installed on about 66% of Android devices shipped in Australia and about 40% globally excluding China.

2860 No other app store has been pre-installed on more than 5% of Android devices shipped in Australia, or more than about 13% globally excluding China, since 2016.

2861 Second, as to installations, app store visits, and app store time, an analysis prepared by Google in October 2019 revealed the following.

2862 As to installations, according to a 31 October 2019 presentation titled OEM App Store Share Analysis: Android Ecosystem Analytics, the Play Store was “installed on ~100 percent of 28DA devices in each region”, while no other OEM app store was installed on more than 50% of devices globally or in Southeast Asia.  Further, of the OEM app stores, the Galaxy Store was installed on 41% of devices globally, with the next highest install shares being Xiaomi Market at 7%, Oppo Apps at 4%, and Vivo App Store at 3%.

2863 As to app store visits, the same presentation noted that “[f]or each region, the majority (>90%) of all app store visits in a month are to the Play Store”. Globally, that figure was 94%.  Further, globally, only 2% of all app store visits in a month were to the Galaxy Store, while 1% were to Xiaomi Market, 2% were to Oppo Apps, and 1% were to Vivo App Store. Further, “[d]espite users with Chinese OEM app stores being MAU, the visits in a month make up only a tiny fraction of the region’s total app store visits (<3% of Play Store total visits)”.

2864 And as to app store time, the same presentation noted that “[f]or each region, the majority (>92%) of monthly app store time spent is in the Play Store”.  Globally, that figure was 96%.  Globally, only 2% of monthly app store time spent was in the Galaxy Store, while 1% was in Xiaomi Market, 1% was in Oppo Apps, and 1% was in Vivo App Store.

2865 Third, as to the volume of downloads, Ms Edwards-Warren estimated that in 2020, some 97.9% of new app downloads to Android devices in Australia that occurred through an app store were undertaken via the Play Store, while globally excluding China some 91.4% of new app downloads from an app store occurred via the Play Store.

2866 Its closest rival in Australia being the Galaxy Store had a share of only 1.5%.  Its closest rival globally, being the Oppo Store had a share of only 2.5% and the Galaxy Store’s global share was 1.5%.  The Play Store’s share of app downloads was very similar in 2019 and 2021.

2867 Fourth, as to the time spent by users and considering the position of the Play Store’s so-called closest competitor, the Galaxy Store and its use by Samsung users, being those one would expect to use the Galaxy Store the most, Google noted in a 19 March 2019 presentation titled “Samsung Galaxy Store Landscape” that, globally, the following position applied.

2868 About 83% of all Samsung monthly active users (MAUs) visit the Play Store for more than 10 seconds per month, while only about 19% visit the Galaxy Store for more than 10 seconds per month.

2869 Further, about 77% of all Samsung MAUs visit the Play Store for more than 10 seconds two or more times per month, while only about 12% visit the Galaxy Store for more than 10 seconds two or more times per month.

2870 Further, those MAUs who do visit the Galaxy Store for more than 10 seconds two or more times per month only stay within the Galaxy Store for about 2% of the time that they stay within the Play Store.

2871 Fifth, data compiled by Google in 2017 indicated the following.  In Japan, the Amazon App Store’s MAUs were 1.4% of the Play Store’s MAUs, while Amazon’s app installs were only 0.02% of the Play Store’s app installs.  In South Korea, OneStore’s MAUs were 20.8% of the Play Store’s MAU, while OneStore’s app installs were only 0.7% of the Play Store’s app installs.  In India, 9Apps’ MAUs were 16.4% of the Play Store’s MAUs, while 9Apps’ app installs were only 3.6% of the Play Store’s app installs.  Globally, Samsung’s MAUs were 13% of the Play Store’s MAUs, while Samsung’s app installs were only 0.2% of the Play Store’s app installs.  And in the US, UK and Germany, the Amazon App Store’s MAUs were 15.1% of the Play Store’s MAUs, while Amazon’s app installs were only 0.01% of the Play Store’s app installs.

2872 If other methods of Android app distribution besides app stores are included, the picture does not materially alter.

2873 In 2020, the Play Store accounted for 73% of new app downloads in both Australia and globally.  OEM and third-party app stores accounted for only 2% of new app downloads in Australia and 7% globally. Further, direct downloading, or downloads from directly downloaded app stores, accounted for 3% of new app downloads in Australia and 12% globally.

2874 A 2016 Google presentation titled Special Topic: Off-Play Installs recorded that installs from outside the Play Store represented 12% of all installs and derived from three main sources, being sideloading, pre-installed third-party app stores, and device hacks. For ten different countries, sideloading accounted for less than 12% of installs in six of them, less than 20% in the next three, and 69% in Iran; users in Iran cannot download paid apps or make in-app purchases via the Play Store. In other words, in all bar 4 countries for which a breakdown is given, sideloading accounted for less than 12% of off-Play Store installs in 2016.

2875 In a February 2018 presentation titled Project Hug: Risk & Leakage Model, Google identified Samsung as “the only OEM with sufficient share to plausibly build its own store in key Play markets”.

2876 Nonetheless, according to a February 2019 presentation titled Project Banyan Phase 1: Ecosystem Overview, Google found it “[c]hallenging to see Galaxy Store becoming attractive enough to developers to be a successful, sustainable stand-alone business” and considered that the Galaxy Store would need to acquire “exclusive rights” that would “build a user base” in order to succeed.

2877 The Galaxy Store is installed on more Android devices than any other rival Android app store. Yet Google concluded in a July 2019 presentation titled Samsung Update that “the cannibalization of Play Store revenue due to Galaxy Store is none to minimal”, that the Galaxy Store was likely “operating under net loss”; and “in absence of ‘exclusive hit-titles’ on Galaxy store, the projected trajectory is in red as well”.

2878 Further, even Samsung users who have the Galaxy Store pre-installed on their devices spend substantially more of their time within the Play Store than they do within the Galaxy Store.

2879 In my view the evidence demonstrates that the Play Store is not subject to strong competition from other Android app stores. The Play Store enjoys dominance in the Android mobile app distribution market and has significant market power.

2880 Let me turn to another topic.

Is there any meaningful or significant countervailing power of users and developers?

2881 Now ss 46(4) and (6) of the CCA provide:

(4)    In determining for the purposes of this section the degree of power that a body corporate or bodies corporate have in a market:

(a)    regard must be had to the extent to which the conduct of the body corporate or of any of those bodies corporate in that market is constrained by the conduct of:

(i)    competitors, or potential competitors, of the body corporate or of any of those bodies corporate in that market; or

(ii)    persons to whom or from whom the body corporate or any of those bodies corporate supplies or acquires goods or services in that market; and

(b)    regard may be had to the power the body corporate or bodies corporate have in that market that results from:

(i)    any contracts, arrangements or understandings that the body corporate or bodies corporate have with another party or other parties; or

(ii)    any proposed contracts, arrangements or understandings that the body corporate or bodies corporate may have with another party or other parties.

…

(6)    Subsections (4) and (5) do not limit the matters to which regard may be had in determining, for the purposes of this section, the degree of power that a body corporate or bodies corporate has or have in a market.

2882 Google says that app developers have significant countervailing power.

2883 It says that developers have several options available to them for wholly or partially bypassing the Play Store’s service fee, particularly with respect to paid in-app transactions for digital content.

2884 These options include offering a webstore for purchases of app based digital content, converting their app to a consumption only app, distributing their app via rival Android app stores or direct download or, for some app developers, pre-installation.

2885 But in my view, and as Epic correctly points out, users and developers lack significant countervailing power and, moreover, Google behaves as if it is unconstrained by competitors.

2886 There is an absence of countervailing power on the part of users and developers.

2887 Further, Google has been able to and has acted in a manner unconstrained by competitors, including by imposing tying arrangements, insisting on contracts of adhesion and  unilaterally altering contractual terms.

2888 Google is not constrained by countervailing power held by Android device users.

2889 Now one way in which users might exercise any countervailing power is by switching to alternative app stores or direct download. Such alternatives are available, but are generally not taken by users. Alternative app stores had a share below 2% in Australia and below 3% globally, and direct downloading only constitutes 3% of new app downloads in Australia.

2890 Further, Google is not significantly constrained by countervailing power held by developers.  Relatedly, Google behaves towards developers as if unconstrained by competitors. This is demonstrated by various matters.

2891 First, Google has for many years insisted that all developers must enter into a standard form DDA. Further, as a business matter, Google does not negotiate the DDA, even for large developers. All developers are on the same DDA with no exceptions. Moreover, the vast majority of developers perceive that they have no real choice but to accept Google’s terms.

2892 Second, by the terms of the DDA, Google has a number of powers which, in combination, place developers at its mercy.

2893 Google may alter the terms of the DDA unilaterally at any time with 30 days’ notice to the developer. If developers do not agree with Google’s changes to the DDA, their sole and exclusive remedy is to terminate their use of the Play Store.

2894 Google may also terminate its DDA with a developer for no reason at all on 30 days’ notice. Termination of the DDA will cause the developer to lose the ability to distribute or update their apps previously distributed through the Play Store. Moreover, current users of their apps will not be able to make any in-app purchases within those apps.

2895 These rights provide Google with very significant leverage over developers.

2896 Third, Google has unilaterally amended the DDA despite complaints and protests from developers.

2897 In particular, in January 2021 Google unilaterally amended its payments policy to remove the external consumption exemption, which required thousands of developers to use Google Play Billing and to pay a service fee on in-app purchases of digital content that, until then, had been exempt from these requirements. Google was able to do this despite the fact that many developers did not wish to use Google Play Billing, and had voiced significant complaints about Google Play Billing to Google.

2898 Further, in 2014 Google unilaterally amended the DDA to tighten clause 4.5 of the DDA to prohibit developers from using the Play Store to distribute “any product that facilitates the distribution of [Android] apps”.  Google imposed this change even though some developers, including large developers such as Amazon and Facebook, wanted to continue distributing Android apps through their Play Store apps.

2899 Fourth, there is no evidence of material switching by developers from the Play Store to rival Android app stores.

2900 Fifth, most developers could not make a credible threat to leave the Play Store given that doing so would involve forgoing access to the large and valuable Play Store app user base and would be likely to reduce their profits. And where a developer does pose a credible threat of leaving the Play Store, the evidence concerning Project Hug shows that Google can and will offer it a sweetheart deal.

2901 Clearly, these larger developers do not constrain Google and do not “protect” the smaller developers. Moreover, the larger developers are identifiable to Google and can be offered better terms without requiring Google to offer better terms to all developers. This is borne out by what has been described elsewhere in these reasons as Mr Gennai’s 2019 Play problem statement, which discloses that Google’s tactic with Project Hug was not to compete with new entrants on price across the board, but to offer special benefits to a small number of “strategic partners”, and so lowering the “effective rev-share” for those few whilst maintaining the prevailing rate for every other developer.

2902 In respect of the millions of developers across the world who are not “beacons of the ecosystem”, there is no evidence that Google feels constrained to compromise its demands under the DDA.

2903 In any case, Google’s conduct is inconsistent with it being constrained by any countervailing power of larger developers. If Google was so constrained, one would expect to see a pricing structure which contained discounts for large developers. There is only one example of this outside of Project Hug in evidence being Spotify.

2904 Spotify pays a 4% service fee when Google Play Billing is used and 0% when its own payment solution is used.

2905 A July 2020 Google presentation titled Spotify Economic Review described the rationale for this deal as being to address “Spotify’s compliance with Google Play payments policy”, “[r]egulatory risk mitigation”, “subscription revenue” and “public advocacy”.

2906 In fact, Google’s pricing structure is the opposite of what one would expect if it were constrained by the larger developers.  Clearly, Google in respect of its rates has not given larger customers larger discounts on the rates as compared with smaller customers, which is what one would normally expect if the larger customers had significant countervailing power.

2907 Sixth, Google price discriminates. It selectively determines which apps, and therefore which developers, have to pay the service fee. It exempts for example sellers of physical goods and services.  This is a further indication of Google’s freedom from competitive constraint.

2908 Seventh, Google’s ability to engage in the conduct concerning the parsimonious reductions to its commissions in the context of developer only billing and user choice billing and how the reduced commissions were set well demonstrates its substantial market power in both the Android app distribution market and the payment solutions market.  I have discussed the detail of this elsewhere.

2909 Finally, it is convenient to note here that other products and services do not provide significant competitive constraints and do not negate or diminish Google’s substantial market power to any significant extent.  I have discussed substitution questions elsewhere when addressing market definition.  There are some substitutes as I have previously discussed.  But these are not close substitutes and do not significantly constrain Google’s market power.

The service fee and its variation(s)

2910 Clause 3.4 of the DDA provides that a service fee, which may be revised by Google from time to time, will be calculated and charged on the sales price of the developer’s products.

2911 The service fee is only payable on apps and in-app products sold through the Play Store’s billing system or through an alternative billing system, the latter being a system permitted under Google’s user choice billing (UCB) or developer only billing (DOB) programs.

2912 The service fee payable by app developers under the DDA has varied over time to time, as summarised in the following table which I have already set out but it is convenient to again reproduce here:

2913 As the table shows, the present service fee varies depending on the developer’s revenue and the type of digital content offered for sale.

2914 Generally, developers now pay a service fee of 15% on the first USD 1 million of revenue earned by the developer each year; a service fee of 30% of revenue in excess of USD 1 million of revenue each year; and a service fee of 15% of revenue for automatically renewing subscriptions regardless of the revenue earned by the developer.

2915 Developers eligible for special programs pay a 15% or lower service fee. These programs generally target media companies, or developers offering video, audio or music subscriptions, or digital books.

2916 In addition, Google charges a varied service fee in circumstances where UCB or DOB apply.

2917 Outside of special programs for eligible developers, UCB, and DOB, Google has made only limited changes to the service fee that it charges for in-app purchases of digital content over time. When in-app purchases were first introduced to the Android Market in 2011, the service fee for such purchases was 30%.  Some 13 years later, the service fee is still 30%, save for the first USD 1 million in earnings each year and subscriptions, both of which are charged at 15%.

2918 Google’s service fees are similar to those generally charged by Apple, although there are two differences.  First, Apple charges a 15% fee to developers whose earnings were under USD 1 million in the prior year, as opposed to charging all developers 15% on their first USD 1 million.  Second, Apple’s fee for subscriptions is 30% in the first year and 15% thereafter.

2919 Now the first change to Google’s subscription pricing occurred with effect from 1 January 2018, when Google reduced its service fee for subscriptions that had been active for more than one year from 30 to 15%.

2920 Google witnesses said that this change was driven by a need to remain price competitive with Apple, which had reduced its fee on subscriptions active for more than one year from 30 to 15% some 18 months earlier, with effect from 13 June 2016.

2921 But the document in which this change was proposed, a document that Mr Feng co-presented, gave five reasons for the change, only one of which related to Apple.

2922 Mr Feng acknowledged that these were the five “main justifications for the change”.

2923 The other four reasons for the change listed in the document included encouraging subscription developers to adopt Google Play Billing and easing the path to a potential Play Store payment policy change, being the removal of the external consumption exemption. As to Apple, the document merely said, “Bring Play’s pricing for subscription SKUs in line with market – Apple offers 30% for 12 months, 15% thereafter”.

2924 Bringing one’s pricing into line with the market is not competing on price. Mr Feng agreed that there was no attempt to do any better than Apple at this time.

2925 Google’s behaviour did not amount to competing with Apple on price, or on any other metric. That is a fortiori when the price adjustment was made some 18 months after that of Apple. And it is especially true in this case, given the other reasons for the change.

2926 As far back as March 2017, Google had considered that a “Rev share change needs to be part of the mix to help land [the] policy change”, that is, the removal of the external consumption exemption. That consideration is expressly identified as a reason for this change in the October 2017 document which Mr Feng helped to prepare and present.

2927 Google felt that improving the price of its subscription offering would lay the groundwork for the policy change, in the face of the adverse developer and regulatory reactions which were anticipated. This was not an instance of price competition vis-à-vis Apple.

2928 Let me turn to some other changes.

2929 On 18 November 2020, Apple announced a small business program, whereby its service fee was reduced to 15% for those developers who earned less than USD 1 million in the prior year.

2930 With effect from 1 July 2021, Google reduced its service fee to 15% for the first USD 1 million earned each year by all developers. In October 2021, Google reduced the service fee for subscriptions to 15% in the first year.

2931 One of the motivations for making the first of those changes was the pressure Google was facing from regulators. But neither Mr Feng nor Mr Gennai accepted that these changes were made largely due to regulatory pressure and for PR reasons, and instead insisted that they were examples of competition with Apple.

2932 Their evidence is not supported by the contemporaneous documents.

2933 In an internal document dated 16 November 2020 titled Project Runway: Proposal for changes to Play business model, Google set out its hypotheses, including that “[w]e have an a la carte menu of additional levers to pull in a set of forecasted PR/regulatory scenarios for further discussion”, and “[w]e should proactively make moderate changes in 2021 to forestall more drastic required changes”.

2934 Under the heading “Scenario Planning”, the following scenarios were contemplated.

2935 The first stated:

* Should we be REQUIRED to adjust our service fee, we would need to re-imagine the store promotional strategy and GPB strategy.

* If government mandated requirement to make play billing optional, we respond with aggressive store changes and perhaps lower service fee …

2936 The second stated “[i]f government mandated price for play billing (say 20% or cost + 5%) (WIP, needs more thinking)…”.

2937 The third stated:

* We c/should VOLUNTARILY adjust our service fee now to improve optics, align value delivered with fees charged, and reduce agitation --

…

* We recommend the first magical bridge option (30% for the first year after user purchase, 15% thereafter) …

…

* Match Apple’s discount for small developers and perhaps be more aggressive by tiering it lower in emerging markets.

…

2938 The last scenario essentially describes the course that Google ultimately took.

2939 Mr Gennai tried to dismiss this document as a “group of five people” who were just speculating. Like much of his evidence, this was advocacy. The document does not contain speculation. It contains a record of essentially what Google did in the following year.

2940 And it is significant that the document makes no reference to a supposed competitive dynamic with Apple as somehow necessitating or justifying its proposals. Rather, the focus is on “forestall[ing] more drastic required changes” by way of regulation and “optics”.

2941 On 21 November 2020, the day of Apple’s small business announcement, Mr Samat prepared a document which commenced with a summary of Apple’s small business program announcement, including reactions from the press and developers. The document then provides an overview of the reaction from regulators, which includes that Apple’s announcement had been “informally described … as too little too late and will not help Apple’s position in the Digital Markets Act debate in Europe” and that Google was “expecting to receive” a request from South Korean legislators to “follow Apple’s lead”.

2942 The document then set out Google’s “take”, as follows:

Waiting for Apple to make the first move has been our plan. Our strategy has been to wait to make any public revenue share changes until Apple moved first. We did not want to move downward on rev share while Apple stayed at 30%, nor did we want to give them the opportunity to “one-up” any announcement we made and cause us to need to react again.

We have a very similar program scoped out…

We will take the next several weeks to gain more feedback on how Apple’s changes landed. The benefit of going second here is we have the opportunity to be more thoughtful and take into account the reaction. …

2943 So, Google wanted Apple to lower its prices and was not at all concerned that this would cause Google to lose customers.

2944 Mr Gennai denied the plain meaning of this document. It undermined his affidavit evidence and Google’s contention that it was engaging in price competition with Apple.

2945 Google’s motives are confirmed by a 2 December 2020 document titled Runway January 2021 Announcement Proposal and Status, which records that Google had “decided to announce in January 2021 an adjustment to our fee schedule, focused on providing relief to small developers and defusing some of the regulatory pressure experienced in several countries”.

2946 Defusing regulatory pressure was the real reason for this change, though Mr Gennai denied this.

2947 The following “goals” are then set out in the document:

Defuse immediate regulatory pressure

Be seen as helping people build their business…

Reduce the new perceived gap between Apple and Play’s billing policies

Continue to allow Apple to make first moves on service fee structure changes

2948 The penultimate “goal” was not about price competition but about the regulatory pressure that had arisen due to the “gap between Apple and Play’s billing policies”. That it was not about price competition is confirmed by the last goal.

2949 The document then outlines the “Desired PR and Regulatory Landings” and says “our message will be evaluated based on how many devs pay less. Press will line us up vs Apple”. Again, it was all about “PR and regulatory landings”, not winning customers from Apple.

2950 Tellingly, a drafting comment records:

… we need a headline PR/reg message – esp given that the message is the primary goal in this effort … Arguably the most believable, as well as most helpful reason is competitive pressure/response…Helpful to our case – if we’re trying to create the meme intense competitive pressure.

2951 In other words, competition was a “message” or “meme” Google was “trying to create”. If competition were the reality, this would not have been necessary.

2952 In cross-examination, Mr Feng sought to minimise this document as merely reflecting the views of its author, Mr Bankhead. I do not accept that evidence. It was self-serving and inconsistent with the extensive comments, reflecting real engagement among various contributors, including one as senior as Mr Samat. Further, the import of this document is similar to the earlier two documents.

2953 Let me say something about the possible changes to the business model and removing the payments tie.

2954 Now some developers have complained that they could not pay the 30% commission because of the impact on their ability to do business. Additionally, Google has been concerned about regulatory pressure over the level of the service fee.

2955 The service fee has been repeatedly described within Google’s own documents as “arbitrary”. That is an apt characterisation, since the 30% fee bears no discernible relationship to Google’s costs.

2956 In an August 2018 memorandum, Ms Kochikar’s subordinate described Google’s then service fee of 30% for all as “Status quo. It’s what Apple does” and said it had “No rationale. Unfeasible for many business models”.

2957 Similarly, in its March 2019 presentation titled Exploring New Business Models, Google observed that the 30% rate has “No rationale, other than copying Apple” and is “untenable for many apps/games developers”.

2958 Indeed, Google was aware that that 30% commission rate was out of proportion to the value certain developers were receiving through Google Play Billing.

2959 Now Project Magical Bridge considered alternative business models that included charging developers separately for different services, including for payment solutions.

2960 Google engaged in a lengthy examination of billing optionality, whereby developers would have the opportunity to acquire their own payment solution.

2961 There was never any dispute about whether it was technically possible to separate the billing system from other aspects of Google Play Billing. From the launch of the Play Store until 29 March 2011, Google did not offer a payment solution for in-app purchases and developers adopted their own payment solutions during that period. And in the 15 or so countries where Google Play Billing is not offered, developers incorporate alternative payment solutions into their apps.

2962 Further, Google has separated the billing system in South Korea. In jurisdictions where UCB is available, the alternative billing system is a separate and distinct service.

2963 I have discussed the detail of all of this elsewhere.

2964 Generally, outside of special programs for eligible developers, being UCB and DOB, Google has made only limited changes to the service fee that it charges for in-app purchases of digital content over time.

2965 When in-app purchases were first introduced to Android Market in 2011, the service fee for such purchases was 30%.  Some 13 years later, the service fee is still 30%, save for the first USD 1 million in earnings each year and subscriptions, both of which are charged at 15%.

2966 Moreover, where Google has made changes to its commission rates, those changes were not principally driven by price competition or non-price competition, but by regulatory pressure or other factors.

2967 Google has continued to charge the 30% commission rate despite significant developer dissatisfaction, with some developers describing the fee as a “tax” and an “exorbitant tax”. Moreover, some developers have complained that they could not pay 30% because it was impacting the viability of their businesses.

2968 I agree with Epic that Google’s ability to largely maintain prices that are deeply unpopular with developers all over the world, for some 13 years, and not have to adjust them even once due to the forces of competition, is evidence of a very substantial degree of market power.

Service fees – their comparability with “competitors”

2969 Now Google says that the service fees that the Play Store charges are not significantly above competitive benchmarks.

2970 It says that as demonstrated by Professor Asker’s analysis in exhibit 106 of his first report, the Play Store’s service fees are the same or lower than its competitors across several important indicia.  Exhibit 106 is as follows:

2971 First, compared with mobile app stores, the Play Store’s headline rate of 30% is the same as the headline rates charged by Apple’s App Store, Samsung Galaxy Store, [REDACTED] and Huawei AppGallery.

2972 Further, the Play Store’s service fee on subscriptions of 15% is lower than the fees for subscriptions on Apple’s App Store, Samsung Galaxy Store, [REDACTED] and Huawei AppGallery.

2973 Second, compared with PC app stores, the Play Store’s service fee is the same or lower than the headline rates charged by Mac App Store, [REDACTED] and GOG.

2974 The Play Store also has a lower service fee than Steam on subscriptions and also the first USD 1 million of the developers’ revenue, and the same service fee on transactions where the developer’s revenue is USD 1 million to USD 10 million.

2975 Third, compared with game console app stores, the Play Store’s service fee is the same or lower except for developers earning more than USD 10 million on Steam or for game app subscriptions on the Microsoft store on Xbox.

2976 Now Google says that there are three main exceptions to its service fee being comparable or lower than its rivals, being the Epic Games Store on PC, the Microsoft Store on PC, and itch.io on PC. But for the reasons identified by Professor Asker in his first report, these app stores are likely of much lower quality than the Play Store, given their much smaller reach, amongst other factors.

2977 So, Google says that the suggestion that its service fee is not comparable to competitive benchmarks should be rejected.

2978 Further, Google says that the Play Store’s service fees also have not increased since they were first introduced at a time when the Play Store manifestly had no market power. And instead, there have been several rate reductions.

2979 But in my view Google’s market power is further demonstrated by the fact that its service fee bears no relationship to supply costs and has not once been varied due to competitive pressure.

2980 And Google has had an ability to charge prices well above the supply cost without rivals taking away customers.

2981 The Play Store’s service fees are higher than those of other mobile app stores or PC app stores besides Apple, which is a monopolist.  Developers pay 15% for earnings on subscriptions and on the first USD 1 million in revenue per annum. Thereafter, developers pay a 30% service fee on non-subscription revenue.

2982 There are several PC app stores that charge less to developers.

2983 The Epic Games Store charges a commission of 12% on paid apps and in-app purchases of digital content. And in response to the Epic Games Store’ entry, in December 2018, Steam reduced its commissions such that it charges a 30% commission on the first USD 10 million in app developer revenues, 25% on app developer revenue between USD 10 million and USD 50 million and 20% commission on app developer revenue over USD 50 million.

2984 Now although the Steam percentages appear similar to the service fees imposed by Google, the difference lies in the reversal of the scale for the commission structure as against the Play Store’s service fees. For Steam, the percentage payable decreases as developer revenue increases, meaning that large developers are charged materially less than they are charged by the Play Store, with a gap of up to 10%.

2985 Further, in March 2019, about four months after the Epic Games Store launch, Microsoft lowered its commission rate for all non-gaming apps distributed on PC through the Microsoft Store from 30 to 15%.  In April 2021, Microsoft announced and then subsequently lowered its commission rate for PC gaming apps to 12%, matching the Epic Games Store’s commission rate.

2986 Since 1 July 2018, the ONE Store has charged developers a 20% commission rate for paid apps and in-app purchases, while each of  [REDACTED] and Samsung Galaxy Store charge lower commission rates upon negotiation.