United States of America v. Google LLC

UNITED STATES DISTRICT COURT

FOR THE DISTRICT OF COLUMBIA

_________________________________________

)

UNITED STATES OF AMERICA et al., )

)

Plaintiffs, )

)

v. ) Case No. 20-cv-3010 (APM)

)

GOOGLE LLC, )

)

Defendant. )

_________________________________________ )

_________________________________________

)

STATE OF COLORADO et al., )

)

Plaintiffs, )

)

v. ) Case No. 20-cv-3715 (APM)

)

GOOGLE LLC, )

)

Defendant. )

_________________________________________ )

MEMORANDUM OPINION

TABLE OF CONTENTS

INTRODUCTION......................................................................................................................... 1

PROCEDURAL HISTORY ......................................................................................................... 7

I. PRETRIAL LIABILITY PHASE ......................................................................................... 7

II. LIABILITY DETERMINATION ......................................................................................... 9

III. REMEDIAL PHASE ............................................................................................................ 11

A. Remedial-Phase Filings and Pretrial Proceedings ............................................................ 11

B. The Parties’ Remedies Proposals ...................................................................................... 13

FINDINGS OF FACT................................................................................................................. 16

I. GENERATIVE ARTIFICIAL INTELLIGENCE ............................................................. 17

A. Key Terms ......................................................................................................................... 17

B. AI and Search .................................................................................................................... 19

1. Integrating AI Features into Search .............................................................................. 19

2. AI Chatbots ................................................................................................................... 22

3. AI Assistants ................................................................................................................. 25

4. On-Device AI ................................................................................................................ 27

C. LLMs................................................................................................................................. 28

1. How LLMs Work (Greatly Simplified) ........................................................................ 28

2. The Limitations of LLMs.............................................................................................. 30

3. Grounding ..................................................................................................................... 32

D. The GenAI Market ............................................................................................................ 36

1. Participants .................................................................................................................... 36

2. Competition Among GenAI Companies ....................................................................... 40

3. GenAI’s Impact on GSE Usage .................................................................................... 43

II. SEARCH ACCESS POINTS ............................................................................................... 46

A. Access to Search Through GenAI Products...................................................................... 46

B. Circle to Search ................................................................................................................. 48

C. Google Lens ...................................................................................................................... 48

III. GOOGLE’S DISTRIBUTION CONTRACTS: AMENDMENTS & WAIVERS .......... 49

A. OEMs ................................................................................................................................ 49

i

1. Samsung ........................................................................................................................ 49

2. Motorola........................................................................................................................ 53

B. Carriers .............................................................................................................................. 55

C. Mozilla .............................................................................................................................. 57

D. Waiver Letters ................................................................................................................... 57

CONCLUSIONS OF LAW ........................................................................................................ 58

I. LEGAL FRAMEWORK ...................................................................................................... 58

A. General Principles ............................................................................................................. 58

B. Causation........................................................................................................................... 63

1. The Strength of the Causal Connection ........................................................................ 64

2. Causation Standard Applicable to Certain Remedies ................................................... 72

II. SUFFICIENCY OF THE LIABILITY-PHASE FINDINGS ............................................ 74

III. FRUITS OF GOOGLE’S UNLAWFUL CONDUCT ....................................................... 82

A. Freedom from Threats....................................................................................................... 82

B. Scale .................................................................................................................................. 89

C. Revenue............................................................................................................................. 95

D. Google’s Objection ........................................................................................................... 97

IV. THE INCLUSION OF GENERATIVE AI PRODUCTS ................................................. 99

REMEDY-SPECIFIC CONCLUSIONS OF LAW ............................................................... 104

I. ADEQUACY OF PROHIBITORY INJUNCTIVE RELIEF ONLY............................. 104

II. STRUCTURAL REMEDIES ............................................................................................. 111

A. Chrome Divestiture ......................................................................................................... 112

B. Contingent Android Divestiture ...................................................................................... 118

III. ADDITIONAL “CORE REMEDIES”.............................................................................. 119

A. Payment Ban ................................................................................................................... 119

B. Data-Sharing Remedies .................................................................................................. 128

1. Justification for Data Sharing ..................................................................................... 129

2. Search Index................................................................................................................ 136

3. Knowledge Graph ....................................................................................................... 148

4. User-Side Data Remedies ........................................................................................... 151

5. Ads Data...................................................................................................................... 164

ii

C. Syndication Remedies ..................................................................................................... 168

1. Search Syndication...................................................................................................... 168

2. Search Text Ads Syndication ...................................................................................... 180

3. Contingent Search Text Ads Syndication ................................................................... 186

D. Choice Screens ................................................................................................................ 187

IV. ADDITIONAL BEHAVIORAL REMEDIES .................................................................. 191

A. Ad Transparency & Advertiser Control Remedies ......................................................... 192

1. Search Query Reports ................................................................................................. 192

2. Keyword Matching ..................................................................................................... 196

3. Access to Data Reports ............................................................................................... 198

4. Search Text Ads Auction Changes ............................................................................. 199

B. Public Education Fund .................................................................................................... 201

C. Publisher-Related Remedies ........................................................................................... 203

V. ANTI-CIRCUMVENTION, ANTI-RETALIATION, AND ADMINISTRATIVE

REMEDIES ......................................................................................................................... 206

A. Anti-Circumvention and Anti-Retaliation Remedies...................................................... 206

B. Technical Committee ...................................................................................................... 210

C. Investment Notification Requirement ............................................................................. 213

D. Self-Preferencing Prohibitions ........................................................................................ 215

VI. EFFECTIVE DATE AND TERM OF FINAL JUDGMENT ......................................... 219

VII. REMAINING PROVISIONS ........................................................................................... 221

A. Definition of “Google” ................................................................................................... 221

B. Fees and Costs................................................................................................................. 222

CONCLUSION ......................................................................................................................... 222

APPENDIX ................................................................................................................................ 224

iii

INTRODUCTION

Last year, this court ruled that Defendant Google LLC had violated Section 2 of the

Sherman Act: “Google is a monopolist, and it has acted as one to maintain its monopoly.” The

court found that, for more than a decade, Google had entered into distribution agreements with

browser developers, original equipment manufacturers, and wireless carriers to be the out-of-the

box, default general search engine (“GSE”) at key search access points. These access points were

the most efficient channels for distributing a GSE, and Google paid billions to lock them up.

The agreements harmed competition. They prevented rivals from accumulating the queries and

associated data, or scale, to effectively compete and discouraged investment and entry into the

market. And they enabled Google to earn monopoly profits from its search text ads, to amass an

unparalleled volume of scale to improve its search product, and to remain the default GSE without

fear of being displaced. Taken together, these agreements effectively “froze” the search

ecosystem, resulting in markets in which Google has “no true competitor.”

Much has changed since the end of the liability trial, though some things have not. Google

is still the dominant firm in the relevant product markets. No existing rival has wrested market

share from Google. And no new competitor has entered the market. But artificial intelligence

technologies, particularly generative AI (“GenAI”), may yet prove to be game changers.

Today, tens of millions of people use GenAI chatbots, like ChatGPT, Perplexity, and Claude, to

gather information that they previously sought through internet search. These GenAI chatbots are

not yet close to replacing GSEs, but the industry expects that developers will continue to add

features to GenAI products to perform more like GSEs.

The emergence of GenAI changed the course of this case. No witness at the liability trial

testified that GenAI products posed a near-term threat to GSEs. The very first witness at the

1

remedies hearing, by contrast, placed GenAI front and center as a nascent competitive threat.

These remedies proceedings thus have been as much about promoting competition among GSEs

as ensuring that Google’s dominance in search does not carry over into the GenAI space. Many

of Plaintiffs’ proposed remedies are crafted with that latter objective in mind.

The question now is what to do about Google’s violations. Precedent requires fashioning

antitrust remedies that “effectively pry open to competition a market that has been closed” by a

monopolist’s “illegal restraints.” Int’l Salt Co. v. United States, 332 U.S. 392, 401 (1947).

Denying the fruits of the violation is a valid objective, and so, too, is ensuring that anticompetitive

behavior will not recur in the same or related ways. The court has broad discretion to impose

remedies to accomplish those aims.

Notwithstanding this power, courts must approach the task of crafting remedies with a

healthy dose of humility. This court has done so. It has no expertise in the business of GSEs, the

buying and selling of search text ads, or the engineering of GenAI technologies. And, unlike the

typical case where the court’s job is to resolve a dispute based on historic facts, here the court is

asked to gaze into a crystal ball and look to the future. Not exactly a judge’s forte.

Because of what is at stake, the remedies phase has been contentious. The parties have

seen eye-to-eye on very little. In Google’s view, there is little work to be done. The only

appropriate remedy is to enter an injunction that prohibits it from entering into exclusive

distribution agreements, which would include its GenAI products, the Gemini app and Google

Assistant. Google says it has already stripped from its existing contracts the provisions that the

court found had an exclusionary impact, so the market is now effectively unfettered and open to

competition. It also urges the court to do nothing more because the GenAI technology space is

2

highly competitive, and any further restrictions would unfairly hobble it in that fight. Google asks

that the final judgment be for a term of three years.

Plaintiffs believe that far more must be done. Simply enjoining Google’s bad acts, as

Google insists, would maintain the status quo. Plaintiffs therefore have put forward a

comprehensive slate of remedies, whose component parts, they claim, work together both to revive

competition in the general search market and to protect the GenAI market from suffering the same

fate as search. Plaintiffs’ proposals can be collected into three general categories: (1) structural

remedies; (2) behavioral remedies; and (3) administrative, anti-retaliation, and anti-circumvention

remedies. Plaintiffs ask that the judgment remain in place for up to 10 years.

Because of the number and complexity of the parties’ proposed remedies, the court does

not recite its conclusions and reasoning in detail in this introduction. But here are the top-line

determinations:

• Google will be barred from entering or maintaining any exclusive contract

relating to the distribution of Google Search, Chrome, Google Assistant,

and the Gemini app. Google shall not enter or maintain any agreement that

(1) conditions the licensing of the Play Store or any other Google

application on the distribution, preloading, or placement of Google Search,

Chrome, Google Assistant, or the Gemini app anywhere on a device;

(2) conditions the receipt of revenue share payments for the placement of

one Google application (e.g., Search, Chrome, Google Assistant, or the

Gemini app) on the placement of another such application; (3) conditions

the receipt of revenue share payments on maintaining Google Search,

Chrome, Google Assistant, or the Gemini app on any device, browser, or

3

search access point for more than one year; or (4) prohibits any partner from

simultaneously distributing any other GSE, browser, or GenAI product.

• Google will not be required to divest Chrome; nor will the court include a

contingent divestiture of the Android operating system in the final

judgment. Plaintiffs overreached in seeking forced divesture of these key

assets, which Google did not use to effect any illegal restraints.

• Google will not be barred from making payments or offering other

consideration to distribution partners for preloading or placement of Google

Search, Chrome, or its GenAI products. Cutting off payments from Google

almost certainly will impose substantial—in some cases, crippling—

downstream harms to distribution partners, related markets, and consumers,

which counsels against a broad payment ban.

• Google will have to make available to Qualified Competitors certain search

index and user-interaction data, though not ads data, as such sharing will

deny Google the fruits of its exclusionary acts and promote competition.

The court, however, has narrowed the datasets Google will be required to

share to tailor the remedy to its anticompetitive conduct.

• Google shall offer Qualified Competitors search and search text ads

syndication services to enable those firms to deliver high-quality search

results and ads to compete with Google while they develop their own search

technologies and capacity. Such syndication, however, shall occur largely

on ordinary commercial terms that are consistent with Google’s current

syndication services.

4

• Google will not have to present users with choice screens on its products or

encourage its Android distribution partners to do the same. Precedent

requires courts to avoid remedies that compel product design requirements,

and in any event, choice screens have not been shown to enhance

competition among GSEs.

• Google will not be required to share granular, query-level data with

advertisers or provide them with more access to such data. Nor will it have

to restore an “exact match” keyword bidding option. Plaintiffs did not

establish that these remedies would promote competition in the search text

ads market.

• Google will be compelled to publicly disclose material changes it makes to

its ad auctions to promote greater transparency in search text ads pricing

and to prevent Google from increasing prices by secretly fine-tuning its

ad auctions.

• Google will not have to underwrite a nationwide public education

campaign. That remedy does not fit Google’s violations and its terms are

too indefinite.

• Google will not have to modify its policies to offer website publishers more

choice in how Google uses their content. This remedy bears no relationship

to Google’s unlawful acts and is an improper demand to implement overly

regulatory requirements.

• Google will not be subject to an investment reporting requirement. It, too,

bears no relationship to Google’s anticompetitive conduct.

5

• Google will not be subject to anti-retaliation, anti-circumvention, or self-

preferencing provisions. The first two restrictions are too vague and do not comport

with the requirements of Federal Rule of Civil Procedure 65(d). There is no legal

or factual basis for the last.

The court will establish a Technical Committee to assist Plaintiffs and the court in implementing

and enforcing the final judgment. The term of that judgment will be six years, and it will become

effective 60 days after entry, except for those provisions relating to the Technical Committee,

which will go into effect immediately.

This opinion is organized as follows: The court begins by recounting the liability- and

remedial-phase histories of this case. In that same section the court sets forth the parties’

competing remedies proposals. The court then makes its Findings of Fact. Those factual findings

were drafted with the primary objective of updating the reader on market developments since the

close of the liability phase and therefore are less detailed than those in the liability opinion. The

court then turns to explaining its Conclusions of Law. That section sets forth the general antitrust

principles that will guide the court’s evaluation of remedies and addresses various remedies-related

disputes of law. Next are Remedy-Specific Conclusions of Law. That section will detail the

court’s rationale for adopting, rejecting, or modifying the parties’ proposed remedies. The court

makes extensive findings of fact in that section as well. Finally, the court concludes with a

directive to the parties to meet and confer and present by September 10, 2025, a joint revised final

judgment that is consistent with this Memorandum Opinion. As with the liability opinion, the

court includes as an Appendix a list of the names and titles of all witnesses whose remedies-phase

trial or deposition testimony is cited in this decision.

6

The court is grateful to all counsel for their professionalism and zealous advocacy during

these remedies-phase proceedings.

PROCEDURAL HISTORY

I. PRETRIAL LIABILITY PHASE

This case began on October 20, 2020, when the U.S. Department of Justice and 11 states

(“U.S. Plaintiffs”) filed an antitrust complaint against Google pursuant to authority conferred by

Section 4 of the Sherman Act, 15 U.S.C. § 4. See Compl., ECF No. 1; see also Am. Compl.,

ECF No. 94 [hereinafter Am. Compl.] (including 14 states). U.S. Plaintiffs asserted three

violations of Section 2 of that law, each corresponding to an alleged product market in the

United States. Am. Compl. ¶¶ 173–193. Those markets were general search services, search

advertising, and general search text advertising. Id. ¶¶ 88–107. The crux of U.S. Plaintiffs’

complaint was that Google illegally maintained a monopoly in these markets by entering into

exclusive agreements to secure default distribution for its GSE on nearly all desktop and mobile

devices in the United States. See id. ¶¶ 111–165. U.S. Plaintiffs sought a finding of liability, an

injunction against the challenged conduct, and structural relief necessary to cure any resulting

anticompetitive effects. Id. ¶ 194.

Roughly two months later, 38 additional states and territories (“Plaintiff States”), filed

State of Colorado v. Google, 20-cv-3715 (APM) [hereinafter Colorado v. Google Docket], under

Section 16 of the Clayton Act, 15 U.S.C. § 26. Compl., Colorado v. Google Docket, ECF No. 3

[hereinafter Colorado Compl.], ¶¶ 21–23. Their complaint largely mirrored U.S. Plaintiffs’ but

supplemented it in three primary ways. Plaintiff States: (1) alleged an additional advertiser-side

market for general search advertising but did not adopt U.S. Plaintiffs’ broader market for search

advertising, id. ¶¶ 56 n.3, 82–89; (2) asserted exclusionary conduct by Google that targeted

7

specialized vertical providers, which are internet companies that provide search services focused

on niche markets such as travel (e.g., Expedia or TripAdvisor), restaurant reservations

(e.g., OpenTable), or shopping (e.g., Amazon or eBay), id. ¶¶ 168–199; and (3) claimed that

Google had engaged in additional exclusionary conduct by using its proprietary advertising tool,

SA360, to harm competition in the relevant markets, id. ¶¶ 144–167. Plaintiff States likewise

sought declaratory and injunctive relief. Id. ¶ 233.

The court consolidated the two cases for both pretrial and trial proceedings. Order,

Colorado v. Google Docket, ECF No. 67; Status Conf. Tr., ECF No. 609, at 10–14. At the parties’

request, the court bifurcated the proceedings into liability and remedies phases. See Order,

ECF No. 264.

Following a lengthy period of merits discovery, the court granted in part and denied in part

Google’s motions for summary judgment in both cases. See generally United States v. Google

LLC, 687 F. Supp. 3d 48 (D.D.C. 2023). In United States v. Google, the court entered judgment

in Google’s favor with respect to those portions of U.S. Plaintiffs’ claims that related to Android

Compatibility Commitments and Anti-Fragmentation Agreements, Google’s voice-activated

assistant and other “Internet-of-Things” devices, and the Android Open-Source Project. Id. at 85–

87. In Colorado v. Google, the court found in Google’s favor with respect to Plaintiff States’

theory that Google’s exclusionary acts directed at specialized vertical providers—namely, placing

restrictions on the visibility of their content on Google’s search engine results page (“SERP”) and

coercing them to share data—had not caused anticompetitive effects in the proposed markets. Id.

at 78–83. The remaining claims proceeded to trial.

8

II. LIABILITY DETERMINATION

The liability trial commenced on September 12, 2023, and concluded on November 16,

2023. After receiving testimony from dozens of witnesses; reviewing thousands of pages of

exhibits and deposition designations; considering the parties’ post-trial briefs, proposed findings

of fact, and proposed conclusions of law; and hearing from the parties at closing arguments over

two days, the court issued its liability findings on August 5, 2024. The court held that Google had

violated Section 2 of the Sherman Act by maintaining a monopoly in certain product markets

through exclusionary conduct. See generally United States v. Google LLC, 747 F. Supp. 3d 1

(D.D.C. 2024) [hereinafter Google].

The court identified three relevant product markets: general search services, search

advertising, and general search text advertising.1 Id. at 109–16, 125–33, 136–38. The court did

not find a separate general search advertising market, as alleged by Plaintiff States. Id. at 139–42.

The court determined that Google had monopoly power in the general search services and general

search text advertising markets but not the search advertising market. Id. at 117–24, 133–36, 138–

39.

Having found that Google had monopoly power in two relevant product markets, the

court’s Section 2 inquiry proceeded to analyze whether Google engaged in exclusionary conduct

in those markets. See id. at 142–52. Plaintiffs challenged certain contracts between Google and

other companies as unlawful exclusive agreements. The court held that (1) agreements between

Google and browser developers, such as Apple and Mozilla, were exclusive insofar as they

established Google as the out-of-the-box default search engine; (2) mobile application distribution

agreements (“MADAs”) between Google and Android original equipment manufacturers

1

The parties agreed that the United States is the relevant geographic market. See Google, 747 F. Supp. 3d at 107.

9

(“OEMs”) were exclusive in practice; and (3) revenue share agreements (“RSAs”) between Google

and Android device distributors—both OEMs and wireless carriers—formalized the practical

exclusivity of the MADAs. Id. at 146–52.

Because merely characterizing Google’s challenged contracts as “exclusive” did not

answer whether those agreements violated Section 2, the court next evaluated whether they caused

anticompetitive effects in the relevant markets—they did. See id. at 152–81. First, the court held

that the agreements caused significant anticompetitive effects in the general search services

market. Id. at 152–71. The court found that Google’s distribution agreements (1) foreclosed a

substantial portion of the general search services market and impaired rivals’ opportunities to

compete, id. at 153–59; (2) denied rivals access to the volume of user queries and the resulting

data, or “scale,” needed to compete effectively, id. at 159–64; and (3) reduced other companies’

incentive to invest and innovate in search, id. at 165–71.

The court rejected Google’s procompetitive justifications for the distribution agreements.

It concluded that the contracts did not “(1) enhance the user experience, quality, and output in the

market for general [search] services, (2) incentivize competition in related markets that redounds

to the benefit of the search market, [or] (3) produce consumer benefits within the related markets.”

Id. at 171; see also id. at 171–77. Accordingly, the court held that Google violated Section 2 of

the Sherman Act by maintaining its monopoly in the general search services market through

exclusive distribution agreements. Id. at 177.

As for the general search text advertising market, the court reached similar conclusions.

It found that the exclusive agreements (1) foreclosed a substantial share of the market; (2) allowed

Google to increase text ads prices without any meaningful competitive constraint; (3) permitted

Google to degrade the quality of its text ads offerings; and (4) capped rivals’ advertising revenue,

10

thereby decreasing their ability to reinvest in quality improvements and attract more users and ad

dollars. Id. at 177–82. Google did not offer any procompetitive justifications beyond those the

court had already rejected in the general search services market. Id. at 181. The court therefore

held that Google’s exclusive agreements also violated Section 2 of the Sherman Act in the general

search text ads market. Id.

That left Plaintiff States’ additional claim concerning Google’s use of its search engine

management tool, SA360. SA360 allows advertisers to run online marketing campaigns across

multiple platforms all in one place. Id. at 181. According to Plaintiff States, Google promised

that it would not self-preference its own proprietary ad platform, Google Ads, on SA360, but it did

just that when it implemented certain features for Google Ads but not for Microsoft’s ad platform,

Microsoft Ads. Id. This lack of feature parity, Plaintiff States asserted, placed Microsoft at a

competitive disadvantage in the relevant markets. See id.

The court found in favor of Google. It concluded that, under Supreme Court precedent,

Google had no duty to deal with its rival Microsoft, and in any event, Plaintiff States had failed to

demonstrate the SA360 conduct resulted in any anticompetitive harm in the relevant markets. Id.

at 181–85.

Lastly, the court declined to make a finding of anticompetitive intent, as it is not an element

of a Section 2 violation, and it did not sanction Google for its failure to preserve certain evidence.

Id. at 186–88.

III. REMEDIAL PHASE

A. Remedial-Phase Filings and Pretrial Proceedings

The remedial phase began September 18, 2024, with entry of a scheduling order that set

aside approximately six months for discovery to be followed by a three-week evidentiary hearing.

11

See Order, ECF No. 1043. At the outset, the court ordered Plaintiffs to submit a high-level

framework of potential remedies, see id., which they did on October 8, 2024, see Notice of Pls.’

Proposed Remedy Framework, ECF No. 1052. Thereafter, Plaintiffs filed their initial proposed

final judgment on November 20, 2024. See Notice of Pls.’ Proposed Final J., ECF No. 1062.

Google submitted its initial proposed final judgment on December 20, 2024. See Notice of Def.’s

Proposed Final J., ECF No. 1108. Both parties then submitted revised proposed final judgments

in anticipation of the remedies trial. See Pls.’ Revised Proposed Final J., ECF No. 1184-1

[hereinafter Pls.’ RPFJ]; Def. Google LLC’s Proposed Final J., ECF No. 1185 [hereinafter

Google’s RPFJ]. Discovery concluded on April 9, 2025. See Order, ECF No. 1043.

The evidentiary hearing ran from April 22, 2025, to May 9, 2025. Each side presented

hundreds of exhibits, see, e.g., ECF Nos. 1220, 1228, 1233, 1234, 1238, 1268, 1283, 1339, 1340,

and the court heard from nearly 50 witnesses via live testimony and deposition

designations, see List of Anticipated Live Witnesses, ECF No. 1210-1; List of Anticipated

Witnesses by Dep. Designation, ECF No. 1210-2.

Following the evidentiary presentation, the parties made extensive post-hearing

submissions, including proposed findings of fact and conclusions of law and replies in support.

See Def.’s Proposed Findings of Fact, ECF No. 1346 [hereinafter Google’s PFOF];

Def.’s Proposed Conclusions of Law, ECF No. 1347 [hereinafter Google’s Br.]; Pls.’ Remedies

Post-Trial Br., ECF No. 1348 [hereinafter Pls.’ Br.]; Pls.’ Remedies Proposed Findings of Fact,

ECF No. 1349 [hereinafter Pls.’ PFOF]; Pls.’ Remedies Responsive Proposed Findings of Fact,

ECF No. 1364 [hereinafter Pls.’ RPFOF]; Pls.’ Remedies Responsive Post-Trial Br.,

ECF No. 1365 [hereinafter Pls.’ Reply]; Def.’s Responsive Proposed Findings of Fact,

12

ECF No. 1366 [hereinafter Google’s RPFOF]; Def.’s Responsive Proposed Conclusions of Law,

ECF No. 1367 [hereinafter Google’s Reply].

The court heard closing arguments on May 30, 2025. See Order, ECF No. 1362.

B. The Parties’ Remedies Proposals

Not surprisingly, the parties’ proposed remedies are dramatically different. The wide gulf

separating them reflects a fundamental disagreement over the scope of appropriate antitrust

remedies and the court’s authority to impose them.

Plaintiffs have offered what they say is a “comprehensive and unitary framework” of

remedies that will “reinforce each other to restore competition.” Remedies Hr’g Tr. at 13:20-24

(Opening Arg.) [hereinafter Rem. Tr.]. Their proposed final judgment covers the full spectrum of

equitable remedies available under the case law.

Most significantly, Plaintiffs demand structural relief. They seek an immediate forced

divesture of Google’s web browser Chrome. Pls.’ RPFJ § V.A. They also ask the court to include

a contingent divestiture of Google’s operating system, Android, in the event the initial remedies

fail or Google circumvents them. Id. § V.C.

Plaintiffs’ next class of proposed remedies, broadly speaking, are behavioral remedies.

These are remedies that would compel Google to undertake affirmative acts or bar it from certain

conduct beyond restrictions on contracting. The proposed remedies in this category are many.

They include (in order of their appearance in Plaintiffs’ RPFJ):

1. prohibiting Google from making any search-related payments to

distribution partners, with one limited exception, id. § IV.A–G;

2. requiring Google to provide advance notice of certain acquisitions and

investments, id. § IV.H–I;

13

3. forbidding Google from using proprietary assets like YouTube, Android,

and Gemini to self-preference its own search or advertising products, id.

§ V.B;

4. compelling Google to share search index data and certain user and

advertising data with “Qualified Competitors,” id. § VI.A, C–F;

5. ordering Google to allow website publishers and content creators to opt out

of Google crawling their web pages and domains for inclusion in Google’s

search index and for training its GenAI models and products, id. § VI.B;

6. requiring Google to grant licenses to “Qualified Competitors” to syndicate

search results and general search text ads from Google, id. §§ VII.A–G,

VIII.E;

7. compelling Google to provide advertisers enhanced performance data and a

bidding-control feature known as “exact match,” as well as disclose

material changes to its ad auctions, id. § VIII.A–D;

8. mandating choice screens on Google products that would allow users to

select the GSE they wish to use, and permitting Google to pay device

distributors (other than Apple) for offering choice screens on existing

devices, id. § IX.A–D; and

9. launching a nationwide public education campaign, underwritten by

Google, that would inform consumers about different GSE options and how

to switch GSEs. (This last remedy is sought solely by Plaintiff States, id.

§ IX.E.)

14

Plaintiffs also ask the court to establish a Technical Committee that would assist them in

enforcing the judgment and to require Google to appoint an internal compliance officer. Id. § X.A–

D. They additionally seek a general prohibition against retaliatory acts by Google, id. § X.E, and

a prohibition on conduct designed to circumvent the terms of the judgment, id. § X.F. Plaintiffs

propose that the judgment last for up to 10 years and that the court retain jurisdiction to enforce it.

Id. §§ XI–XII.

Google’s proposed remedies are far narrower. Google says the court can do little more

than enter a prohibitory injunction that bars it from entering or maintaining any unlawful exclusive

distribution agreement for the term of the judgment. See Google’s Br. at 3–4. Such an injunction

“ensures that partners and consumers get the benefit of competition, while prohibiting the forms

of contractual provisions the Court deemed exclusive.” Id. at 3. Google would have the court

enter an order that:

1. prohibits Google from bundling Google Search, Chrome, Google Assistant,

or the Gemini app with Google Play or other Google apps in their OEM

agreements, Google’s RPFJ § III.A–D;

2. bars Google from entering any contract with an OEM or wireless carrier

that prevents the partner from preloading or otherwise promoting alternative

GSEs, browsers, or AI assistants, id. § III.E–G;

3. prevents Google from conditioning payments to OEMs and wireless carriers

on the required installation of Google Search, Chrome, Google Assistant,

or Gemini on more than one search access point or device, id. § III.H–J; and

4. allows browser developers like Apple and Mozilla to set different GSEs as

defaults across various modes and platforms and to promote other search

15

services, and grants them the option to change default search settings

annually, id. § III.K–L.

Notably, Google also includes provisions to prevent exclusive distribution of two GenAI

products, Google Assistant and the Gemini app, on Android devices. See id. § III.C–D, G–J.

Google also proposes a compliance regime with annual reporting and an internal compliance

officer. Id. § IV. It would have the judgment expire in three years. Id. § V.

Plaintiffs agree that Google’s prohibitory equitable remedies are appropriate and generally

propose the same, see Pls.’ RPFJ § IV, but contend that such minimal relief will merely “maintain

the status quo” and do nothing to restore competition, Pls.’ Br. at 3.

FINDINGS OF FACT

The court’s primary purpose in this Findings of Fact section is to update the reader on

developments that have affected the relevant product markets since the liability trial concluded.

These findings are therefore far less extensive and detailed than those contained in the court’s

liability opinion. The court’s main factual findings are woven into the Remedy-Specific

Conclusions of Law section, as that is the more natural place to evaluate and weigh the evidence.

In this section, the court focuses on three main topics: (1) GenAI technology and products;

(2) new search access points; and (3) changes to Google’s search distribution agreements. As to

the first topic, the court covers the basics of GenAI technology, GenAI products that perform

functions akin to GSEs, and the main players in the GenAI space and the competition among them.

The court then discusses Google’s Gemini app as a search access point, as well as two new search

access points, Circle to Search and Google Lens. Last, the court updates the record as to Google’s

distribution contracts.

16

I. GENERATIVE ARTIFICIAL INTELLIGENCE

A. Key Terms

1. “Artificial intelligence is the science and engineering of getting machines, typically

computer programs, to exhibit intelligent behavior.” Google, 747 F. Supp. 3d at 52 ¶ 107 (citation

omitted).

2. Generative artificial intelligence (“GenAI”) is a type of artificial intelligence that

uses machine-learning techniques to generate new data, including text, images, sound, code, and

other media. See Rem. Tr. at 149:8-12 (Durrett) (GenAI is “a sort of sub-field of artificial

intelligence that uses machine-learning techniques to generate structured outputs, which might be,

for example, long chunks of text or images.”); id. at 3315:23–3316:2 (Collins) (“Generative AI is

about producing information, so producing content.”); id. at 4057:16-19 (Hitt) (stating that people

use GenAI to “generate some kind of new content or give [them] a new idea”); PXR0102 at -700

(“Generative AI goes a step further” than traditional AI “by creating new data (e.g., text, image,

sound) similar to its training data—pattern creation.”).2 Machine learning blends computer science

with statistics to learn how to solve problems based on exposure to data. See Rem. Tr. at 148:25–

149:7 (Durrett) (discussing PXRD003 at 6).

3. Large language models (“LLMs”) are a type of GenAI model that takes text or other

types of data as inputs and then generates text or other outputs based on predictions. See id. at

149:13-16 (Durrett). Language modeling is “the task of predicting the most likely next token in a

sequence given a prior sequence of tokens,” where one can think of a “token” as a short word or

2

This opinion uses the last three digits of Bates numbers on an exhibit to cite the specific pages that support a finding

of fact. The parties’ demonstratives are cited in the manner that they appear in the parties’ respective proposed findings

of fact. See Pls.’ PFOF; Google’s PFOF.

17

small unit of language. Id. at 152:16–153:8 (Durrett). The ability to predict the next token relies

on both the quality of the model and the amount of input data. See id. at 154:20–155:3 (Durrett).

4. The relationship among the above-discussed types of AI can be visualized as

follows:

PXRD003 at 6; see also Rem. Tr. at 148:16–149:16 (Durrett) (discussing this slide).

5. Most LLMs are “transformer” models. Rem. Tr. at 155:4-9 (Durrett) (“[T]he most

typical method of implementing large language models these days is a model called the

transformer.”). Transformers are a neural model—a computational model that attempts to mimic

the way the human brain works—that uses billions of parameters to predict the probability of the

next token. Id. at 155:9-22 (Durrett). Google released a paper in 2017 that ushered in the use of

transformers, and Google’s transformer architecture is now the backbone of modern LLMs. See id.

at 3318:17–3319:7 (Collins) (“Transformers was a model, an AI model architecture that Google

invented in 2017 that all modern AI models are based on.”); id. at 2447:25–2448:13 (Pichai)

18

(“[A]bout a decade ago, Google invented transformers and . . . they are now the backbone of what

are called large language models.”).

B. AI and Search

1. Integrating AI Features into Search

6. GenAI technologies have increasingly become incorporated into search products.

See Google, 747 F. Supp. 3d at 53–54 ¶¶ 109–113 (describing how AI technology has influenced

and invigorated search processes); Collins Rem. Dep. at 24:11–25:9 (discussing how Gemini

GenAI models are incorporated into Search-related features and Search-related enhancements);

id. at 28:20-21 (“Google Search will deploy Gemini models as part of building their product

. . . .”). Google and other companies incorporate GenAI technologies into their search products

today. See Parakh Rem. Dep. at 25:13–26:1 (describing Google’s effort to incorporate “a large

number of GenAI ideas” into Search through Project Magi, which eventually became AI

Overviews); Rem. Tr. at 3601:4-9 (Reid) (agreeing that Google is incorporating and has

incorporated AI technology and LLM technology into Google Search); id. at 2455:16-18 (Pichai)

(“[W]e have deeply used AI technologies for well over a decade across our most important

products. Obviously Search.”); id. at 1020:23–1021:5, 1024:16–1025:22 (Schechter) (describing

how Microsoft has incorporated GenAI technologies into Bing); see also id. at 2458:12–2459:5

(Pichai) (predicting that “AI technology is going to deeply transform Google Search”).

7. One way Google incorporates GenAI technologies in Search is through a feature

known as “AI Overviews.” Rem. Tr. at 2456:25–2457:4 (Pichai) (“And recently we have launched

something called AI [O]verviews. It uses a custom Gemini-based model to give users for any

query overall context, a summary, and then helps them explore sources on the web, and it’s been

one of the most popular features we’ve launched in Google Search.”); id. at 3609:2-6 (Reid)

19

(agreeing that AI Overview incorporates GenAI into Search). It was introduced in 2024.

See PXR0037 at -229.

8. When a user enters a query in Google Search, AI Overviews “will take the search

results that come back and use a[n] [LLM] to produce a summary of those results” that can be seen

at the top of the SERP “as a kind of brief natural language description of what the search engine

found.” Rem. Tr. at 149:21–150:4 (Durrett); see also id. at 4539:11-14 (Jerath) (describing AI

Overviews as “a new kind of section that Google has included in its search engine results page,

and it shows basically an AI summary response to a query”); id. at 3550:2-8 (Reid) (discussing

how AI Overviews is part of the main search results page and includes “an overview in response

to the user’s question and pulls together relevant webpages about it”).

9. A depiction of AI Overviews on a mobile Google SERP is illustrated below:

PXRD005 at 10; Rem. Tr. at 646:14–647:19 (Hsiao) (discussing this slide).

20

10. AI Overviews are not triggered in response to every search query. See Rem. Tr. at

3550:5-15 (Reid) (stating that an AI Overview appears “whenever we think that it is both high

quality information and a net add to the page”); Parakh Rem. Dep. at 81:1–83:8, 83:13–84:7

(whether AI Overviews appears on a page depends on relevance signals and other signals generated

from Search results); see also N. Fox Rem. Dep. at 188:17–192:22 (discussing AI Overview–

eligible queries); accord Rem. Tr. at 3617:22–3618:16 (Reid) (agreeing that she believes the

percentage of queries triggering an AI Overview response “will continue to increase over time”

and that she reported as much to Google’s Board of Directors). Their introduction has had a

generally positive effect on Search—Google has seen an increase in both consumer satisfaction

and volume of queries. Rem. Tr. at 3615:12–3616:2 (Reid) (agreeing that people who use AI

Overviews use Search more and are more satisfied with their results); id. at 3616:6–3617:4 (Reid)

(stating that Google Search queries in the United States have increased 1.5% to 2% since the

introduction of AI Overviews); PXR0038 at -303 (“[P]eople who use AI Overviews actually use

Search more and are more satisfied with their results.”); see also Rem. Tr. at 2459:6-13 (Pichai)

(observing that users are now asking “longer” and “more complex” questions due to Google’s

incorporation of GenAI search features). Some evidence suggests that placement of features like

AI Overviews on the SERP has reduced user interactions with organic web results (i.e., the

traditional “10 blue links”). See Parakh Rem. Dep. 193:19–194:15, 194:18-25, 195:16-20

(agreeing that there is a sentiment within Google that Google’s first-party search features reduce

users’ interactions with third-party results); PXR0158 at -910 (“[W]hen AI Overviews appear,

pages that appear as a corroborating link [in AI Overviews] get more clicks than if the page had

appeared as a traditional ‘blue link’ listing for that query.”); see also PXR0001 at -612 to -613

21

(discussing a drop in interactions with organic web results where Google’s WebAnswers appears

on the SERP); Parakh Rem. Dep. 21:12-14 (describing WebAnswers as another GenAI product).

11. Google recently integrated a new GenAI feature in Search called “AI Mode.”

Rem. Tr. at 2490:14–2491:21 (Pichai); see also id. at 3352:4–3355:12 (Reid) (describing AI Mode

as one of the “modes on Search that allow people to sort of dig deeper on some aspect of Search”).

Google’s early experience with AI Mode shows that consumers are asking longer questions than

in traditional Search. Id. at 2490:25–2491:5 (Pichai). Google “expect[s] it to . . . become a deeper

part of the Search experience.” Id. at 2491:9-21 (Pichai).

2. AI Chatbots

12. Another type of GenAI consumer-facing product is a “chatbot,” which is available

on both desktop and mobile devices. See id. at 2461:21-23 (Pichai) (“There are many companies,

both big and small, which are both building models, as well as using models to build chatbots or

Generative AI applications.”); id. at 1025:23–1026:5 (Schechter) (discussing how Microsoft’s

Copilot is accessible both on mobile and desktop); id. at 4062:25–4064:13 (Hitt) (“[A]t least for

the data we’ve seen from OpenAI, desktop is—I think at least the majority of the queries come

through desktop.”). Some examples include OpenAI’s ChatGPT, Anthropic’s Claude, xAI’s Grok,

Microsoft’s Copilot, and Google’s Gemini. Id. at 4044:18–4045:19 (Hitt) (discussing RDXD-

32.020).

13. Chatbots are based on the LLMs described above. They serve different purposes

than GSEs albeit with some overlap. Id. at 2457:5–2458:9 (Pichai) (describing the “areas of

overlap” between AI chatbots and Google Search but noting that there are “entirely different use

cases as well”); id. at 385:18–388:8 (Turley) (describing the relationship between Google Search

and ChatGPT as a Venn diagram (discussing PXR0176 at -126)); id. at 3541:18–3543:15 (Reid)

22

(recognizing that the use cases for GSEs and GenAI chatbots “are not identical but they do overlap

in a number of places” like “a Venn diagram”); N. Fox Rem. Dep. at 74:20–76:8, 76:20-21, 76:23–

77:11 (describing the relationship between GSEs and GenAI chatbots as “a Venn diagram” that is

“quite overlapping” but not “fully overlapping”); see also Rem. Tr. at 3373:16–3376:2 (Collins);

id. at 205:10–206:15 (Durrett). When a user submits a query to a chatbot, the underlying GenAI

model makes a prediction about the answer, drawing upon the data used to train the model.

Compare Rem. Tr. at 155:4-22 (Durrett) (“[The transformer] takes as input this sequence of tokens

and outputs this probability distribution over all of these words in the vocabulary.”), with Google,

747 F. Supp. 3d at 38–39 ¶¶ 27–32 (describing how a GSE works by responding to queries via

retrieving and ranking websites responsive to the query from an information index).

14. Like a GSE, consumers can interact with AI chatbots by entering information-

seeking queries. See Google, 747 F. Supp. 3d at 39–41 ¶¶ 33–39 (discussing types of queries

presented to GSEs); Rem. Tr. at 382:5-17 (Turley) (stating that “asking questions” is “one of the

core-use cases that people came to ChatGPT for”); id. at 2457:12-14 (Pichai) (“[F]or certain types

of queries, you can look—you can either ask a search engine or you can go to a chatbot and ask a

similar question.”); id. at 656:8-13 (Hsiao) (“There is definitely truth to the fact that people use

[GenAI] products to do what we would call informational tasks.”). Thus, chatbots perform an

information-retrieval function like that performed by GSEs. Compare Google, 747 F. Supp. 3d at

38–39, 41 ¶¶ 27–32, 41–42 (explaining how a GSE works and depicting a sample SERP), with

Rem. Tr. at 636:8–647:23 (Hsiao) (walking through how a sample query and response would look

in the Gemini app (discussing PXRD005)); id. at 695:16–700:5, 701:7–704:3, 706:23–707:22

(Shevelenko) (walking through a sample query and response on Perplexity’s website (discussing

23

PXRD006)); see also infra Findings of Fact [hereinafter FOF] ¶¶ 36–46 (discussing “grounding”

an LLM in search data).

15. Chatbots often include citations and links to websites when responding to

information-seeking queries. Rem. Tr. at 405:8–406:18 (Turley) (“We do that by allowing users

to see high-quality links inside ChatGPT for areas that they may want to read more about.”); id. at

696:13-25, 698:15–699:7, 701:7–702:11 (Shevelenko) (Perplexity’s answer engine responses

present users with LLM-generated responses, and within those responses, Perplexity provides

weblink citations and a list of web sources).

16. A sample chatbot response to a query is illustrated below.

PXRD005 at 3; Rem. Tr. at 637:23–638:5 (Hsiao) (discussing the Gemini results page displayed

on this slide).

24

17. But chatbots have many use cases that traditional GSEs do not, including

composing text, generating code, and creating novel images and video. See Rem. Tr. at 4058:1-4

(Hitt) (“There’s a lot of applications that GenAI can do that are not suitable for Search, and a lot

of things Search does that may not be well suited for GenAI, at least the way [that] consumers are

using it now.”); id. at 2457:5–2458:9 (Pichai) (observing that “there are areas of overlap” between

Google Search and GenAI chatbots but that “there are completely . . . different use cases as well,”

and identifying coding, image generation, and video generation as “typical use cases which Search

hasn’t done in the past”); id. at 3541:18–3543:15, 3629:19-24 (Reid) (explaining that “what is the

right tool for a question evolves over time,” and listing homework help, companionship, and

brainstorming as use cases for which people typically choose chatbots over search). A person may

enter a short prompt and receive various forms of media rather than the traditional “10 blue links”

returned by a GSE. Compare Google, 747 F. Supp. 3d at 41–42 ¶¶ 41–46 (explaining what appears

on a SERP and depicting a sample SERP), with Rem. Tr. at 2457:17–2458:2 (Pichai) (“If you’re a

developer trying to build an application, you can go on and have these chatbots write you large

sections of code . . . . In the Gemini app, you can go type in a short prompt and it will create a

video for you, it will generate a whole video for you, and those are novel use cases. You can go to

Chat GPT and generate images.”); PXR0176 at -125 (“ChatGPT already provides answers without

10 blue links, in a simple experience, with natural contextual conversational follows ups.”).

3. AI Assistants

18. In an earlier phase of this case, the court discussed Google’s voice assistant product.

See United States v. Google LLC, 687 F. Supp. 3d at 86 (explaining at summary judgment that

Google Assistant “is a virtual assistant that can respond to voice commands to perform various

tasks” (internal quotation marks and citation omitted)). Google has been upgrading that product

25

to the Gemini Assistant, which incorporates LLM technology and GenAI functionality. See Rem.

Tr. at 668:15–670:13 (Hsiao) (discussing the upgrade of Google Assistant to Gemini Assistant).

The Gemini Assistant comes preloaded on certain Android devices. Id. at 3900:14-20 (Samat).

19. Other companies have voice-assistant capabilities built into their chatbots.

See, e.g., id. at 744:3-25 (Shevelenko) (discussing voice-assistant capabilities built into

Perplexity’s app).

20. Google’s GenAI Assistant can be accessed on a smartphone through both a voice

command (a “wake word” or “hot word” like “hey Google”) or a “long press” of the power button

for several seconds. See id. at 635:6-20 (Hsiao) (explaining what a hot word is and that it can

activate the Gemini app on Android devices); PXR0571 at -389 (Gemini–Samsung Commercial

Agreement calling out “Hey Google” and “Hey Gemini” as hot word invocations); Kim Rem. Dep.

at 173:14–174:2.

21. Today, new Android devices come with a default assistant, which gets triggered

when a user hits a hardware button known as the action button. Rem. Tr. at 710:17-19

(Shevelenko).

22. Over the longer term, GenAI companies are striving to transform chatbots into a

kind of “[s]uper [a]ssistant.” Id. at 375:2-8, 375:22-24 (Turley). A super assistant would be able

to help perform “any task” requested by the user. Id. at 375:2-8 (Turley); id. at 387:12–388:8

(Turley) (providing example of a request to buy shoes and the super assistant identifying options

and completing the transaction); id. at 666:22–667:12 (Hsiao) (Google developing the Gemini app

as an AI assistant to “help people solve their everyday problems”). Search is a necessary

component of this product vision. Id. at 388:9–389:6, 424:3-8 (Turley).

26

4. On-Device AI

23. AI functionality on a smartphone can be delivered through a cloud-based AI model

or an on-device AI model (or both). Id. at 2518:5–2519:24 (Nieh). An on-device AI model is

stored entirely on a device and data is not transmitted to the cloud, meaning that an application

running an on-device model can provide faster responses and enhanced privacy. See id. Google’s

on-device LLM is called Gemini Nano. Id. at 3956:10-15 (Samat). Google’s Gemini app runs on

cloud-based models, not Gemini Nano. Id. at 3956:16-19 (Samat).

24. AICore is an Android software module developed by Google that supports the use

of on-device LLMs such as Gemini Nano. Id. at 3955:25–3956:4 (Samat); Collins Rem. Dep. at

55:1–56:12. AICore currently only supports Google’s on-device Gemini Nano model. Rem. Tr.

at 1557:6–1558:1 (Mickens) (opining that models running inside AICore “have to be blessed by

Google”); Collins Rem. Dep. at 56:13-22; see also Rem. Tr. at 3966:19-23 (Samat) (describing

AICore as “kind of a container around Google’s Gemini Nano model”). To run Gemini Nano on-

device in an accelerated way, AICore takes advantage of specialized smartphone hardware

accelerators known as tensor processing units (“TPUs”) and neural processing units (“NPUs”),

which can execute the necessary mathematical computations with efficiency and speed. Rem. Tr.

at 3961:13–3964:2 (Samat); see also id. at 1553:5–1554:19 (Mickens) (explaining that

NPUs/TPUs result in “an increase in model execution efficiency”).

25. AICore does not prevent other models from accessing a device’s NPU or TPU. Id.

at 3964:19–3965:19 (Samat). Other on-device model providers could work with OEMs and chip

manufacturers to access a device’s accelerators. Id. at 3965:23–3966:11 (Samat). But an OEM

may not want to add a second system service alongside AICore and may prefer “to simply ship

with a system service that allowed for plug-and-play use of models.” Id. at 1558:2-9 (Mickens).

27

26. The capacity to run multiple on-device AI models can be limited by the device’s

storage and random-access memory (“RAM”). See id. at 3966:12–3968:6 (Samat) (stating that

“how many different models can exist [on a single device] kind of depends on how big the models

are” but that “right now there would be no problem on these premium devices having multiple

models on storage” and that “practically speaking, . . . when you’re using one, you have it in RAM,

and then when you’re not using it, you swap it out and you swap the other one in”). A single device

may still have multiple on-device models—Samsung devices, for example, have both Gemini

Nano and Samsung’s own model, Gauss, on the device. Kim. Rem. Dep. at 69:24–70:13; Rem.

Tr. at 3959:19–3960:3, 3968:7-15 (Samat).

C. LLMs

1. How LLMs Work (Greatly Simplified)

27. As discussed, language modeling is “the task of predicting the most likely next

token in a sequence given a prior sequence of tokens,” where one can think of a “token” as a short

word or small unit of language. Rem. Tr. at 152:16–153:8 (Durrett). As an example, a user may

enter a query into an LLM that asks it to complete a sentence, such as “Once upon a time, there

was a ____,” or to answer a specific question, such as “When was Abraham Lincoln born?” Id.

at 154:3-19 (Durrett) (discussing PXRD003 at 10). Without human supervision, the model can

predict the next token—for example, that “Once upon a time, there was a war between two

kingdoms” or that Abraham Lincoln was born in 1809. See id. (Durrett) (discussing PXRD003 at

10 (noting these answers)). The ability to predict the next token relies on both the quality of the

model and the amount and quality of the data inputs. See id. at 154:20–155:3 (Durrett); id. at

163:15-17 (Durrett) (“One very important process of the training process for base large language

models is to remove a lot of . . . data that is not helpful.”).

28

28. The development of a successful LLM begins with an untrained model. See id. at

156:1-6 (Durrett). Such a model is “pre-trained” by exposing it to large amounts of data. See id.

at 155:23–156:16 (Durrett); id. at 4024:24–4026:9 (Hitt). By exposing the model to so much data,

the parameters of the model can be updated to enable it to make predictions that reflect the

information in that data. See id. at 156:12-15 (Durrett). That pre-training creates a base (or

foundation) model. See, e.g., id. at 155:23–156:16 (Durrett) (discussing PXRD003 at 12–13); id.

at 3348:3-12 (Collins); id. at 4017:8–4018:25 (Hitt) (discussing RDXD-32.006)

29. Data for pre-training LLMs is gathered from many sources, primarily from public

web pages. See id. at 4025:6–4028:22 (Hitt) (discussing RDXD-32.009 (identifying “[p]ublicly

accessible data,” “[l]icensed data from third parties,” and “[h]uman evaluation data” as data used

by Google and its competitors to pre-train models)); id. at 156:1-6 (Durrett) (“So starting from

a[n] essentially untrained version of the transformer model, we can expose it to large amounts of

text which are typically gathered from the web.”). For example, Google uses its Google Common

Corpus (“GCC”) to pre-train its Gemini GenAI models. See id. at 184:10-15 (Durrett); id. at

3346:8-17 (Collins). GCC is a dataset that involves large amounts of information scraped from

the web and stored in a repository called Docjoins, which is “a data structure that Google uses to

store URLs.” See id. at 183:25–185:6 (Durrett) (discussing PXRD003 at 38); see also PXR0185

at -116 to -117 (“The main corpus of Docjoins is a large repository of the documents publicly

available on the web and visited at least once by Googlebot in the last few months. It currently

consists of over documents. . . . By comparison, the external Common Crawl corpus is much

smaller, with only a bit over 3 B in the latest release.”); Rem. Tr. at 224:18–227:13 (Durrett)

(discussing PXR0185).

29

30. After a model is pre-trained, “[t]here is a second stage called post-training where

models are exposed to data that imparts the different capabilities that [the creators] want them to

have.” Rem. Tr. at 159:21-23 (Durrett); id. at 3341:12-15 (Collins) (“Pretraining is the first process

in model training. And so—it occurs once. And the subsequent training process is called post

training which is modifying, further modifying that pretrained model.”). For instance, if the

model’s creators want the model to be good at writing code or answering questions, then they can

post-train the model on data particular to these areas. Id. at 159:23–160:15 (Durrett) (discussing

PXRD003 at 15); id. at 3350:24–3351:2 (Collins) (agreeing that “an example” of post-training

“would be fine-tuning a foundation model to be able to answer Q and A”). All-purpose models

can be post-trained to accomplish more than one task through exposure to large numbers of

datasets encompassing broad collections of information. See id. at 160:18–161:1 (Durrett)

(discussing PXRD003 at 16).

31. One way to ensure high-quality training data is to filter it or remove unnecessary or

unhelpful data. Id. at 4028:13-22 (Hitt) (“[N]ot all data that’s out there winds up being useful for

training. It is a common business practice to filter your models in various ways, remove duplicates,

remove spam, garbage, inappropriate content. There’s a lot of stuff out there that you don’t want

to put into your models . . . .”); id. at 163:15–165:15 (Durrett) (describing how one data corpus

was filtered down to only .14% of open-source data); id. at 165:25–166:25 (Durrett) (describing

how filtering improves model performance).

2. The Limitations of LLMs

32. Even with high-quality filtered data, there are still limitations as to how accurately

an LLM can respond to prompts. Id. at 167:6–168:11, 172:23–173:21 (Durrett) (discussing

limitations of LLMs); id. at 382:18–383:14 (Turley) (discussing limitations with ChatGPT’s

30

functionality). The capabilities of an LLM are limited by the content of its training data and when

that training occurs, and their accuracy is similarly circumscribed by content and timing.

Importantly, (1) if information is not in an LLM’s training data, it will not be able to produce a

factual response to a query that seeks such information; (2) LLMs can provide reasonably accurate

responses for frequently seen data but struggle to do so for lesser-known information; and

(3) LLMs can forget the content of data they have previously been exposed to and thus cannot

store information in a “lossless way.” See id. at 158:24–159:19, 161:5–162:6, 171:13–172:21

(Durrett) (discussing these issues).

33. As noted, one key limitation is that LLMs have a knowledge cutoff. Id. at 167:9–

168:7 (Durrett) (discussing PXRD003 at 21). An LLM trained on data last updated in October

2024, for example, could not answer queries asking about Taylor Swift’s 2025 engagement to

Travis Kelce. See, e.g., id.; see also id. at 384:4-13 (Turley) (discussing how LLMs “do not contain

current events” as “[t]hey’re only ever limited to things that were . . . present in the training data

at the time of its training, and they don’t update live”). Retraining an LLM on an updated dataset

“takes weeks or months” and thus the LLM would be unable “to answer a user’s query about stuff

that had happened the same day using this kind of mechanism.” Id. at 167:24–168:7 (Durrett).

Moreover, training is costly. See id. at 167:9–168:7 (Durrett) (discussing PXRD003 at 21 (noting

that adding more recent documents to pre-training data is expensive)).

34. Another limitation is the problem of so-called “hallucinations,” facts made up by

GenAI products “that are maybe probable to be true but not actually true.” Id. at 383:2-9 (Turley);

id. at 1035:13-15 (Schechter) (“A hallucination is essentially a statement that might sound factual

but is really just generated by the language model.”). If a user poses a query that asks for

information outside of its training data, the LLM can generate an incorrect response based on what

31

it views to be the most probabilistic response. Id. at 172:23–173:21 (Durrett) (discussing this

phenomenon); id. at 1036:1-9 (Schechter) (noting that if an LLM “doesn’t have access to the right

tools such as search or, you know, it was just told to give a response even if you don’t know the

answer, it will do its best to give a response and generate a response that it thinks you’d be satisfied

with”).

35. In summary, an LLM “is great at generating content and can generate novel content

but might include mistakes. It is also retrained infrequently, so it won’t have [any] up-to-the-

minute factual information anyway.” Id. at 174:5-8 (Durrett). These limitations can be termed

“factuality” and “recency” issues. See id. at 383:10-14 (Turley).

3. Grounding

36. “Grounding” and retrieval-augmented generation (“RAG”) provide a solution to

problems of factuality and recency. See id. at 168:13–169:17, 170:14–171:12 (Durrett); id. at

2853:6–2854:1, 2933:18-23 (Allan); id. at 3366:19–3367:10 (Collins).

37. “[G]rounding is anchoring the output of a model on factual information or [an]

external database.” Id. at 3336:4-7 (Collins); see also id. at 3511:9-25 (Reid) (describing

grounding as an LLM “ask[ing] for web results . . . to maybe fill in information it doesn’t have, to

correct information”); id. at 3634:11-14 (Reid) (agreeing that “grounding is when an LLM model

uses some class of data, often from the web, in order to improve the accuracy of its response”).

RAG is “[t]he process of accessing additional knowledge through a kind of information retrieval.”

Id. at 168:18-20 (Durrett); id. at 3835:11–3836:13 (Cue) (describing generally how through RAG

techniques a chatbot can produce responses from identified relevant weblinks). RAG is a

grounding technique, and the terms are sometimes used interchangeably. Id. at 170:14–171:12

(Durrett); see also id. at 2853:6–2854:1 (Allan).

32

38. An LLM is grounded by taking the query the user enters and referring to an outside

database such as a search index. See id. at 169:5-17 (Durrett); id. at 3366:24–3367:2 (Collins); id.

at 2853:6–2854:1 (Allan); id. at 3634:11-14 (Reid). The grounding model then surfaces that

information from the outside database and feeds that information into the LLM to generate a

response to the query. Id. at 169:5-17 (Durrett). Through grounding, LLMs can “pull[] in material

that is relevant to the query, or at least believed to be relevant to the query, and then instruct[] the

[GenAI] model to use that information to make references to that, and so that’s where you might

see the equivalent of links in a [GenAI] output.” Id. at 2853:6–2854:1 (Allan).

39. Grounding reduces an LLM’s hallucinations and improves its factuality. Id. at

3370:15–3371:1 (Collins); id. at 3634:8-25 (Reid); id. at 170:14–171:12 (Durrett); see also

PXR0040 at -177 to -178, -202 to -203; PXR0105 at -305. LLMs can receive grounded

information from the web to generate a response. Grounding also allows an LLM to “validate its

responses, fact-check information, or even perform more complex tasks like analyzing sentiment

across the web to tailor its responses.” PXR0040 at -203; see also Rem. Tr. at 169:24–174:22

(Durrett) (explaining that, since LLMs have issues storing information in a lossless way and can

more accurately predict answers based on more common information, grounding allows them to

invoke a search engine and anchor the responses in reality (discussing PXRD003 at 23–26));

see also id. at 3530:16–3631:3 (Reid) (agreeing that GenAI chatbots “need to incorporate aspects

of a search engine” in order to “answer[] a wide range of information needs”).

40. Grounding likewise provides a solution to the recency issue, because the LLM can

pull from and incorporate up-to-date information. See Rem. Tr. at 168:13–169:17 (Durrett)

(explaining that by querying a GSE through grounding, an LLM can provide an answer to a

question not in its training data); id. at 4155:4-7 (Hitt) (stating that “one of the applications of

33

grounding” is “to bring in facts . . . that are real time and could not have been . . . in the training

dataset”).

41. Grounding provides LLMs an opportunity to integrate search capabilities into the

models as it permits LLMs to access content beyond its training data, such as web pages in a search

index. See id. at 3835:11–3836:13 (Cue); see also id. at 640:1–641:20 (Hsiao) (describing how

the Gemini model calls upon Search to ground its response with web content and also provides

links to the underlying web content); id. at 3638:23–3640:19 (Reid) (discussing PXR0105);

PXR0105 at -305 (discussing “Search grounding,” which “allows the Gemini model to check

Search results before generating an answer” or “[i]n other words: Gemini model treats Google

Search like a corpus for [RAG]”).

42. Through grounding, GenAI products can translate user prompts into search queries,

send those queries to a search engine, and then incorporate information from the retrieved search

results into their AI-generated responses. See, e.g., Rem. Tr. at 399:21–401:11 (Turley) (describing

how ChatGPT could incorporate search results); id. at 1022:16–1023:8 (Schechter) (explaining

that Microsoft’s chatbot “will write a query,” send that query to Bing, “[a]nd then just as if you

were a human, the results are then returned, and then the chat is able to read those results”); id. at

1025:23–1026:5, 1030:13–1031:10 (Schechter) (further explaining how Microsoft GenAI

products incorporate Bing search results); id. at 3835:21–3836:2 (Cue) (“And what happens is

when you type something in to Search, it uses a search index to retrieve what are the first 10 hits,

let’s call it, that you would typically see that are linked. The LLM now reads all of those 10 links

and goes out and gathers all of the information, this is the RAG concept, . . . and it’s now able to

provide the answer or the results.”).

34

43. Successful grounding in search requires a high-quality search application program

interface, or API. See id. at 392:11–394:16 (Turley) (discussing quality issues with third-party

search providers, which compelled OpenAI to build its own search index); id. at 1022:6–1023:8,

1039:5-20 (Schechter) (explaining that GenAI products treat search results as fact, so the quality

of search results directly impacts the quality of GenAI responses); Pls.’ Cromwell Rem. Dep. at

74:25–75:20 (explaining that “Copilot relies heavily on how Bing is ranking the results of the

search, so the quality of the Bing index . . . will determine the ordering of . . . the citations that are

going into . . . Copilot”). An API is a connection between computers or software programs that

allows them to communicate and interact. Liab. Tr. at 784:10-20 (Kolotouros); id. at 1234:11-18

(Dischler).

44. To ground its Gemini models, Google uses a proprietary technology called

FastSearch. Rem. Tr. at 3509:23–3511:4 (Reid). FastSearch is based on RankEmbed signals—a

set of search ranking signals—and generates abbreviated, ranked web results that a model can use

to produce a grounded response. Id. FastSearch delivers results more quickly than Search because

it retrieves fewer documents, but the resulting quality is lower than Search’s fully ranked web

results. Id.

45. Google does not make FastSearch directly available to third parties through an API.

Id. at 3512:1-5 (Reid). Rather, the technology is integrated into a Google Cloud offering called

Vertex AI, which is available to third parties to ground on Google Search results or other data

sources. Id. at 3339:5–3340:13 (Collins); id. at 3512:1-5 (Reid); id. at 3637:18–3646:12 (Reid)

(discussing PXR0105 at -305 and PXR0153 at -478). Vertex customers do not, however, receive

the FastSearch-ranked web results themselves, only the information from those results. Id. at

3512:18–3513:4 (Reid). Google limits Vertex in this manner to protect its intellectual property.

35

Id. at 3513:6-12 (Reid). Google has multiple Vertex grounding agreements with third parties. Id.

at 3647:5–3648:4 (Reid) (discussing PXR0153 at -488).

46. Google also uses the Vertex service to ground web results for the Gemini app. Id.

at 3640:1–3641:12 (Reid). The Gemini app, however, receives through Vertex access to portions

of other Search features, like the Knowledge Graph, see infra Remedy-Specific Conclusions of

Law [hereinafter RCOL] § III.B.3, that are not available to third parties. Id. at 3648:5–3649:20

(Reid) (discussing PXR0153 at -481).

D. The GenAI Market

1. Participants

47. Google has incorporated AI technologies into its most important products,

including Search. See id. at 2455:16-18 (Pichai) (“[W]e have deeply used AI technologies for well

over a decade across our most important products. Obviously Search.”); id. at 3329:13-16

(Collins) (“Q. At this point in time, Mr. Collins, are there any products at Google of significance

that generative AI has not been integrated into? A. No.”); id. at 3601:4-9 (Reid) (agreeing that

Google has been incorporating AI and LLM technology into Search for years). In February 2023,

Google released its own GenAI chatbot, Bard. See id. at 2498:19-25 (Pichai). Bard has since

transformed into the Gemini app, which is a GenAI chatbot product that relies on Gemini LLM

models to produce results. See id. at 3627:22–3628:12 (Reid); Pancholi Rem. Dep. at 51:14-23;

Rem. Tr. at 625:6-15 (Hsiao). AI Overviews is a GenAI search feature based on a branch of the

Gemini LLM family. N. Fox Rem. Dep. at 36:9-14, 36:17-19, 36:21-23.

48. Anthropic is an American AI technology company that builds and pre-trains its

own GenAI foundational model. Rem. Tr. at 794:1-6 (Shevelenko); id. at 3334:12-21 (Collins).

Its consumer-facing chatbot application is known as Claude. Id. at 4044:25–4045:15 (Hitt); N. Fox

36

Rem. Dep. at 275:11-18. Google has invested in Anthropic. Rem. Tr. at 3641:12–3642:5 (Reid);

Br. of Anthropic PBC et al. as Amici Curiae, ECF No. 1279-1, at 6–8.

49. DeepSeek is a Chinese technology company that released its eponymous chatbot

to much fanfare and explosive growth in December 2024. Rem. Tr. at 2459:24–2460:2 (Pichai)

(“Last December, DeepSeek launched from China within a month, you know. Literally, I mean,

there have been tens of millions of users downloading them on Android and iOS and using them.”);

id. at 4038:23-25 (Hitt) (“You see entrants like Grok or DeepSeek, that may not have existed six

months ago, are now able to reach the level of performance to wind up in the top ten of these

models.”); id. at 685:4-23 (Hsiao) (“It’s explosive growth. There’s new entrants . . . . You know,

Grok, DeepSeek, all sort of new emerging models that are really, really strong.” (discussing

RDXD-04.007 to .008 and RDX0048)). DeepSeek has open-sourced a pre-trained foundation

model, which at least one company, Perplexity, has used to post-train and develop its GenAI

offerings. Id. at 798:22–799:10 (Shevelenko).

50. Meta is an American technology company that builds and pre-trains its own GenAI

models in the “Llama” family. Id. at 3334:16-18 (Collins) (“Meta produces a competitive model

called Llama that . . . recently surpassed a billion downloads . . . .”); see also id. at 794:1-6

(Shevelenko). The Llama models are trained on data derived from open sources. See id. at 165:10-

15, 167:1-5 (Durrett). Meta’s standalone chatbot offering is Meta AI. See id. at 4044:18–4045:19

(Hitt) (discussing RDXD-32.020); see also id. at 2461:24–2462:2 (Pichai). Meta integrates its

GenAI into its proprietary platforms, such as Facebook, Instagram, and WhatsApp. Id. at 506:23–

508:12 (Turley); RDX0091 at -016; RDX0151 at -007; RDX0355 at -029; Rem. Tr. at 4045:22–

4046:24 (Hitt) (“Meta, who primarily distributes their GenAI products through their other

properties like Instagram and WhatsApp and Facebook has also had success in getting these

37

products into consumers’ hands.”). Google has partnered with Meta to provide a Search API for

Meta to ground its LLMs. See Pancholi Rem. Dep. at 69:8–70:11; N. Fox Rem. Dep. at 278:7-16;

see also PXR0101.

51. Microsoft Corporation is an American technology company whose offerings

include GenAI products in addition to its Windows operating system, Edge browser, and Bing

GSE. See Google, 747 F. Supp. 3d at 36 ¶ 10; Rem. Tr. at 1020:23–1021:59, 1021:19-23

(Schechter) (describing “New Bing,” which incorporates LLM technology into Bing’s existing

search technology); id. at 1024:16–1025:5 (Schechter) (describing Copilot Answers, which uses

“LLM technology to . . . summarize the results that are on the” Bing SERP); id. at 1025:6-22

(Schechter) (describing Bing’s “Copilot Search” feature, which uses an LLM to present search

results in a magazine format with a combination of images, text, and links); id. at 1025:23–1026:5

(Schechter) (describing Microsoft’s consumer AI application “Copilot”). Microsoft has integrated

Copilot into Edge and Bing, both as a vertical and through Copilot Answers, which is Microsoft’s

AI-powered search feature analogous to Google’s AI Overviews. Rem. Tr. at 1024:5–1025:22,

1076:17-19 (Schechter). Microsoft licenses some AI models from OpenAI for Bing and Copilot

that it fine-tunes when needed but uses other AI models as well. See id. at 1038:17–1039:2,

1044:1-9, 1081:9-17 (Schechter). Microsoft grounds some of the queries made through its Copilot

products on Bing search results. See id. at 1026:6-11, 1030:13–1032:17 (Schechter).

52. OpenAI is an American AI company that offers both a user-facing AI chatbot,

known as ChatGPT, as well as a developer API for building AI applications on top of OpenAI’s

models. Id. at 373:2-10, 374:10–375:21 (Turley). OpenAI pre-trains its own foundation model.

Id. at 794:1-6 (Shevelenko). ChatGPT offers a free version of its product and a paid subscription

version that gives the user access to more sophisticated AI models. Id. at 376:14–379:3 (Turley).

38

OpenAI previously sought out a partnership with Google for grounding, but Google declined. Id.

at 413:1–416:3 (Turley) (discussing PXR0181). ChatGPT has proven to be an immensely popular

and fast-growing product, with over 100 million daily active users as of the end of 2024. See id.

at 491:18–494:8 (Turley) (discussing PXR0072); id. at 512:22–513:4 (Turley) (agreeing that,

among other strengths, OpenAI has “one of the fastest-growing products of all time” (discussing

RDX0355 at -005)); id. at 3827:13–3828:11 (Cue) (discussing the growth of ChatGPT). In

December 2024, Apple began integrating ChatGPT and OpenAI technology into its devices to

power Apple Intelligence, Apple’s AI experience integrated into its devices, in exchange for

revenue share payments. Id. at 3832:4–3833:11, 3838:17–3839:18 (Cue); id. at 468:1-18, 468:23–

469:7, 499:12-25 (Turley); id. at 500:1-24 (Turley) (discussing RDX0148 at -001).

53. Perplexity is an American AI company that seeks to provide users an “answer

machine.” Id. at 694:8-21 (Shevelenko) (describing Perplexity’s product as “an answer machine”

where users first “ask questions in natural language”; then Perplexity “run[s] a search, identify[ing]

high-quality sources from across the Internet that are relevant to that query,” “retrieve[s] those

snippets,” “rank[s] them,” and “use[s] different large language models to synthesize the

information in a way that’s responsive to the original query”; and finally the user receives “an

answer in text form with citations showing where the information comes from”). Perplexity does

not train its own foundational model but rather relies on foundational models pre-trained by other

GenAI companies (like Meta and DeepSeek) and post-trains them from there. Id. at 794:1-12,

798:16–799:10 (Shevelenko). Perplexity’s consumer-facing products include its website

(Perplexity.ai), a web app, mobile apps, and a recently launched web browser that integrates its

GenAI. Id. at 694:22–695:5, 748:21–749:7 (Shevelenko). Though most of Perplexity’s consumers

use the company’s free chatbot through which it earns some revenue through ads placed below

39

query responses, it also offers a paid subscription version. Id. at 703:8–705:5 (Shevelenko). Like

other GenAI companies, Perplexity’s valuation has surged over the past couple of years. Id. at

731:24–733:12 (Shevelenko) (discussing Perplexity’s $1 billion valuation in spring 2024 and

$9 billion valuation in December 2024).

54. xAI is an American AI company associated with X (formerly Twitter) that recently

released its Grok GenAI product. Id. at 2461:14-25, 2478:16-24 (Pichai); id. at 3334:12-24

(Collins); id. at 4045:4-15, 4048:1-10 (Hitt). Grok is trained on xAI’s own foundation model. Id.

at 794:1-6, 798:22-25 (Shevelenko). xAI integrates Grok into X. Id. at 4050:21-22 (Hitt).

55. Other companies in the GenAI space include DuckDuckGo, which offers a

“Duck.ai” chat service where users can interact with third-party GenAI offerings like ChatGPT,

Claude, or Meta AI. See id. at 1005:13–1007:23 (Weinberg).

2. Competition Among GenAI Companies

56. The GenAI space is highly competitive. See id. at 503:25–504:4 (Turley) (Q. And

let’s talk about the [GenAI] space . . . . You consider that space to be very competitive; correct?

A. Yes, absolutely.”); id. at 3335:19-23 (Collins) (“[Q.] How would you describe the current level

of competition with respect to foundation models as compared to the course of competition over

the years that you’ve seen? A. [It] is the most competitive market I’ve ever worked in.”); id. at

685:4-8 (Hsiao) (“Q. How would you describe the competitive space that the Gemini app

occupies? A. I would say I don’t think I’ve seen a more fierce competition ever in my 20-some

years of working in technology.”).

57. There have been numerous new market entrants. See id. at 685:9-13 (Hsiao) (“It’s

explosive growth. There’s new entrants. . . . You know, Grok, DeepSeek, all sort of new emerging

models that are really, really strong.” (discussing RDXD-04.007 to .008)); id. at 4038:22–4039:4

40

(Hitt) (“You see entrants like Grok or DeepSeek, that may not have existed six months ago, are

now able to reach the level of performance to wind up in the top ten of these models.”); id. at

2459:21-23 (Pichai) (“You have seen over the last few months as many people have launched

chatbots. Very quickly, these chatbots reach tens of millions of users.”).

58. GenAI firms have access to a lot of capital. See id. at 4048:15–4049:20 (Hitt)

(“[A] lot of firms have demonstrated they’ve been able to raise resources, you know, resources

that might be substantial. . . . So in terms of capital, I think that seems to be, at the moment,

plentiful.”); see, e.g., id. at 498:8-14 (Turley) (agreeing that OpenAI recently announced that it has

raised an additional $40 billion in capital at a valuation of $300 billion); Liab. Tr. at 3599:10–

3600:21 (Nadella) (agreeing that, as of the liability phase, Microsoft had invested over $13 billion

in OpenAI).

59. There is constant jockeying for a lead in quality among GenAI products and

models. See Rem. Tr. at 3334:25–3335:9 (Collins) (“The top labs and companies are always

leapfrogging each other and kind of jockeying for the top position. And then the models also

compete on individual capability . . . . [T]he model competition happens both in terms of the best

kind of overall model as well as the best model for specific-use cases.”); id. at 687:15–688:2

(Hsiao) (“[N]o model ever demonstrates a significant lead over the others. If it is better, it’s sort

of a little better, and then another provider will come and make their model like a little bit better

than the previously best one. So it’s very—it’s very tight.”); id. at 4036:22–4039:4 (Hitt) (opining,

in sum, that this is a “very active space, constantly changing position” with “[n]ewer models

tend[ing] to outperform others” and “different models perform[ing] differently”). Today, Google’s

models do not have a distinct advantage over others in factuality or other technical benchmarks.

Id. at 4037:9–4040:13 (Hitt) (discussing RDXD-32.015 to .016); cf. id. at 212:24–214:23, 216:9–

41

218:17 (Durrett) (declining to opine on whether Google’s models are superior to the models

produced by its competitors).

60. A variety of GenAI products have achieved widespread usage. For instance,

OpenAI calculated its share of the U.S. market as of December 2024 to be approximately 85%,

with Claude at 3%, Gemini at 7%, and Perplexity and Copilot making up the remainder. Id. at

485:22–490:14 (Turley) (discussing RDX0355 at -029); see also id. at 512:22–513:4 (Turley)

(OpenAI has “one of the fastest-growing products of all time.”). Google estimated that as of March

28, 2025, its Gemini app had roughly 140 million daily queries, with ChatGPT at 1.2 billion,

MetaAI at over 200 million, Grok at 75 million, DeepSeek at 50 million, and Perplexity at

30 million. See id. at 684:9-22 (Hsiao) (discussing RDXD-04.008).

61. Rival GenAI products have had some success in obtaining distribution with OEMs

and other companies. OpenAI, for instance, has partnered with Apple, T-Mobile, Yahoo,

DuckDuckGo, and Microsoft. See id. at 468:1-18, 468:23–469:7, 499:12-25 (Turley) (describing

OpenAI’s partnership with Apple); id. at 3838:21–3839:18 (Cue) (discussing Apple’s evaluation

of GenAI products and how it chose OpenAI); Giard Rem. Dep. at 51:18–52:5 (discussing T-

Mobile’s partnership with OpenAI); Rem. Tr. at 1278:1-6 (Provost) (stating that Yahoo “work[s]

with OpenAI . . . on AI summarization”); id. at 1005:22–1006:1 (Weinberg) (discussing Duck.ai’s

incorporation of models developed by OpenAI, Anthropic, or Meta); id. at 1044:1-9 (Schechter)

(confirming that Microsoft licenses LLMs from OpenAI to incorporate into Bing). Perplexity has

a distribution deal with Motorola under which Motorola will preload Perplexity’s application onto

new smartphones, although the agreement is not exclusive, the application will not be on the home

screen, and it will not be available via a wake word. See Rem. Tr. at 717:18–719:18 (Shevelenko);

see also id. at 3905:3-19 (Samat) (stating that Motorola will have its own assistant called “MotoAI,

42

which includes Perplexity and CoPilot accessible on the device, invoked by a dedicated hardware

button.”). Perplexity continues to negotiate with other OEMs and browser developers. See id. at

714:14-17, 749:17-23 (Shevelenko). Motorola has also agreed to partner with Microsoft’s Copilot.

See id. at 363:18-22 (Fitzgerald).

62. Google has entered into distribution and promotion agreements for the Gemini app.

See RDX0432 (Google–Samsung Gemini Commercial Agreement); RDX0423 (Google–Motorola

Google One AI Premium OEM Promotion Agreement); RDX0428 (Google–Lenovo Marketing

Agreement).

3. GenAI’s Impact on GSE Usage

63. GenAI products may be having some impact on GSE usage. See Rem. Tr. at

3818:20–3819:4, 3827:23–3828:11, 3846:25–3847:3 (Cue) (testifying that the volume of Google

Search queries in Apple’s Safari web browser declined for the first time in 22 years perhaps due to

the emergence of GenAI chatbots). But GenAI products have not eliminated the need for GSEs.

PXR0176 at -123 (“ChatGPT already expanded what is possible for parts of Search, but users don’t

yet use ChatGPT for the full range of Search needs.”); Rem. Tr. at 648:2–649:21 (Hsiao) (testifying

that Google tracks so-called “cannibalization” of Google Search by GenAI chatbots and the

Gemini app is not diverting queries from Google Search to a significant degree today); id. at

3846:25–3847:17 (Cue) (attributing the recent decline in Safari’s search volume to increasing

usage of GenAI apps but recognizing these apps must improve to compete with Google Search);

id. at 21:2-5 (Opening Arg.) (Plaintiffs’ counsel acknowledging that general search and GenAI

“are different but overlapping products” and that GenAI “is not a replacement for [s]earch today);

see also Google, 747 F. Supp. 3d at 53 ¶¶ 111, 114 (finding during the liability phase that “AI

43

technologies have the potential to transform search” but “AI has not supplanted the traditional

ingredients that define general search”).

64. AI Overviews has potentially strengthened Google’s position in the GSE market.

Since its introduction, Google Search queries in the United States have increased 1.5 to 2%.

Rem Tr. at 3616:6–3617:4 (Reid) (conceding this); see also PXR0033 at -244 (“With our AI

advancements across Search, including . . . AI Overviews . . . we’ve already increased Search

usage by more than hundreds of millions of queries a month in the U.S. alone.”). At present, more

than % of Search queries trigger an AI Overviews response. N. Fox Rem. Dep. at 188:23–

189:15 (as of October 2024, approximately % of U.S. Search queries trigger an AI Overviews

response (discussing PXR0033 at -244)); Parakh Rem. Dep. at 76:22–77:2 (stating roughly “ ,

percent” of U.S. queries triggered an AI Overviews response); see also Rem. Tr. at 2490:19-21

(Pichai) (agreeing that 1.5 billion Search users interact with AI through AI Overviews); Parakh

Rem. Dep. at 64:6-14 (AI Overviews “serve[s] answers at scale to billions of people.”).

65. Certain types of queries with commonplace usage in GSEs are not within the

current use cases of GenAI products. This includes navigational queries. Compare Rem. Tr. at

702:13–703:1 (Shevelenko) (“[N]avigational queries are not a core use case of Perplexity.”), with

Google, 747 F. Supp. 3d at 40 ¶ 39 (“Google’s top five queries by query volume are navigational

queries and nearly 12% of all Google queries are navigational queries.” (internal citations

omitted)). Further, commercial queries are not, at present, a common use case for GenAI

applications and thus far have not cannibalized commercial queries on GSEs. Rem. Tr. at 657:5–

662:3 (Hsiao) (discussing the absence of cannibalization today but believing future cannibalization

may occur based on the quality of a model, its pre-training data, and its grounding capabilities);

44

id. at 3837:12–3838:15 (Cue) (AI chatbots have not mastered commercial queries, but chatbot

developers are increasingly introducing commercial concepts).

66. Even so, “there’s nothing which is fundamentally different between commercial

and non-commercial queries.” Id. at 2460:16–2461:11 (Pichai). GenAI companies hope to attract

more commercial queries, which could be monetized with ads. See, e.g., id. at 658:13–659:15

(Hsiao) (“[T]he providers will try to build great shopping experiences or commercial

experiences.”); id. at 2460:16–2461:11 (Pichai) (referencing ChatGPT’s launch of a shopping

experience to support the prediction that commercial use of chatbots will increase); id. at 662:5–

663:21 (Hsiao) (discussing Google’s desire to kick Gemini to Search with respect to commercial

queries to monetize on the ads served on Search); see also id. at 256:11-22 (Fitzgerald) (Gemini

does not currently serve ads or receive ad revenue with its query responses). The expectation is

that GenAI products will come to serve these needs. See id. at 2460:19–2461:11, 2492:24–2494:4

(Pichai) (predicting that GenAI products will “undoubtedly” be used to answer more commercial

queries, citing ChatGPT’s launching a shopping experience as an example, and that the Gemini

app “will overall expand [Google’s] ability to serve users’ information needs”); id. at 3837:21–

3838:15 (Cue) (observing that “you’re already starting to see” GenAI chatbots answer commercial

queries and launch shopping experiences and that “it’s clear that [companies] have to go in those

directions, and they will,” because that’s “where there’s financial opportunity”); id. at 3543:19–

3544:9 (Reid) (“[W]e have signs that [GenAI chatbots] want to get more and more in [the] business

[of commercial searches].”); see also id. at 3544:1-9 (Reid) (shopping features in Perplexity);

accord id. at 659:16–661:13 (Hsiao) (explaining that a GenAI chatbot would likely need to be

grounded with a search index or with retailers’ databases in order to answer commercial queries,

45

so that it “know[s] actively what are the current products available for sale and what are their

prices”).

II. SEARCH ACCESS POINTS

A. Access to Search Through GenAI Products

67. GenAI products have the potential to provide access points to GSEs. Smutny Rem.

Dep. at 23:8-13 (identifying GenAI products such as Copilot that are “grounded in search” as

“potential future search entry points”). That prospect, however, has not yet fully materialized

given GenAI’s myriad other use cases. Pls.’ Cromwell Rem. Dep. at 40:15-17 (“Q. Mr. Cromwell,

is Copilot’s purpose to drive traffic to Bing? A. No.”); see also Rem. Tr. at 672:2–674:1 (Hsiao)

(describing the other use cases for Gemini and the fact that content on the Gemini app is often

entirely newly generated); cf. id. at 3546:13–3547:3 (Reid) (stating that she did not believe the

Gemini app to be a search access point because it “doesn’t really cause a user to build up this

mental model, [to] go to the Gemini app in order to get to Search”).

68. Google has contemplated developing GenAI products to serve as search access

points as a business strategy. See PXR0113 at -844 (Google presentation recognizing that “[t]he

Search landscape is changing,” pointing to “[n]ew AI access points” such as “chatbots LLM” and

stating that “[t]o continue to drive Search growth & diversification in this new environment, we

must introduce new user access points” including by “embedd[ing] [a] Search experience” in

“Browsers/Search Engines (alongside AI chatbots) and Mobile Apps (AI/non-AI)”); Rem. Tr. at

3608:21–3609:1 (Reid) (agreeing that Google is exploring and testing new Search entry points);

id. at 662:5–665:8 (Hsiao) (discussing commercial queries and a desire for Gemini to kick back to

Search as it relates to these queries for monetization purposes).

46

69. Within the Gemini app, as of the evidentiary hearing, users can access Search via

the “related search” function, which is accessible via a “G” icon that appears in response to certain

queries. See Rem. Tr. at 641:21–647:20 (Hsiao) (explaining what the “G” button in the Gemini

app does, including corroborating answers and providing search-related topics). After the Gemini

app generates a response to a query, one typically grounded in Search, a user can click the “G”

icon and will be presented with “Search related topics.” See id. (Hsiao) (discussing PXRD005 at

7–10). Should the user click one of those Search related topics, the user will “get a screen” that is

“sort of a traditional Google Search results page.” Id. at 647:15-20 (Hsiao) (discussing PXRD005

at 7–10).

70. This functionality was built “as an early feature that was really around trying to

help with hallucinations” but has not always been present and is “very rarely clicked.” Id. at

665:3–6 (Hsiao); id. at 642:2-5 (Hsiao) (stating that Google had previously removed the

“G” button).

71. Today, the Gemini app drives little traffic to Google Search. Id. at 665:7-8 (Hsiao)

(“I would say Gemini does not drive much, if any, meaningful traffic to Search today.”); id. at

4059:24–4061:8 (Hitt) (stating that “[o]nly a very small fraction” of queries yield the “G” button

and lead the user to “actually even consider going to Search, and then only a small fraction of those

actually wind up in Search” (discussing RDXD-32.030)); see also id. at 664:16–665:2 (Hsiao)

(“I actually think if the user has to search after using Gemini, it’s a failure of the product. Like,

the point of Gemini is to help you answer the question and help you get things done with the AI.

If you then have to use another tool, you know, that’s not our desired outcome . . . .”).

47

B. Circle to Search

72. “Circle to Search” is a recently launched search access point by Google for Android

devices. It allows a user to execute a query by simply drawing a circle on a device’s display screen.

See id. at 243:10–244:14 (Fitzgerald); id. at 2494:5-16 (Pichai); id. at 3553:13–3555:16 (Reid); id.

at 3907:17–3908:22 (Samat). Circle to Search returns a SERP in response to the “circled” query.

Id. at 244:9-14 (Fitzgerald); 2494:8-16 (Pichai); 3554:16-24 (Reid); see also id. at 3602:21-23

(Reid) (agreeing that a user “can access Search through Circle to Search”).

73. An OEM must modify its user interface to allow Circle to Search to operate, but

the search functionality itself is built on the open-source Android platform. Id. at 3908:25–

3910:24, 3979:19-24 (Samat). An OEM can select the default search engine that will answer the

query, which may be Google or some other search product that has the capacity to execute a

“circle” query. See id. at 3908:25–3910:24 (Samat) (confirming that Circle to Search can be used

with a search application other than Google and discussing Perplexity’s visual search offering).

C. Google Lens

74. “Google Lens” is a recent visual search capability built into Chrome. See id.

at 1639:10–1640:7 (Tabriz) (stating that Google Lens “uses AI” and is “built by the [S]earch team”

and agreeing that “Google Lens and Circle to Search” can be viewed as “search features or search

products”). “[I]nstead of searching using text,” a user “take[s] an image and search[es] based on

that image” with Search “providing results based on what that image is.” Id. at 3823:11-21 (Cue)

(“So if I were to search [using Google Lens], for example, I could take an image of this water, it’s

called ‘Deer Park,’ and it would search for Deer Park Water and give the results.”); see also id. at

1657:3-14 (Tabriz).

48

75. Google Lens’s current integration in Chrome only works if Google Search is set as

the default GSE in the browser. Id. at 1656:22–1659:9 (Tabriz) (describing how visual search

works with and without Google Search as the default GSE).

76. Apple and Google entered into an agreement to bring Google Lens functionality to

Apple devices in August 2024, with Google paying a revenue share on ads clicked on Google

Lens-generated SERPs. Id. at 3823:22–3825:2 (Cue).

III. GOOGLE’S DISTRIBUTION CONTRACTS: AMENDMENTS & WAIVERS

77. Since the conclusion of the liability trial, Google has entered into several additional

agreements, amended agreements, or extended agreements to distribute its Search and GenAI

products. See Google, 747 F. Supp. 3d at 89–106 ¶¶ 289–396 (chronicling the at-issue browser

agreements, MADAs and RSAs); PXR0567 (June 2024 amendment to Google–Samsung Google

Mobile RSA); PXR0515 (December 2024 amendment to the Google–AT&T Google Mobile RSA);

PXR0571 (January 2025 Google–Samsung Gemini Commercial Agreement); PXR0535

(June 2024 Google–Motorola Google One AI Premium OEM Promotion Agreement); PXR0541

(February 2025 amendment to Google–Motorola/Lenovo Google Mobile Incentive Agreement);

PXR0543 (February 2025 Google–Motorola Gemini Fund Agreement); PXR0370 (March 2024

Google–Mozilla contract extension); PXR0610 (March 2024 Google–T-Mobile RSA notice of

renewal).

A. OEMs

1. Samsung

78. On July 1, 2024, Google and Samsung entered into an agreement amending the

Google Mobile RSA. See Rem. Tr. at 237:17–240:14 (Fitzgerald); see PXR0567; RDX0424. This

amendment merely extended the provisions of the original RSA through March 31, 2025.

49

See PXR0567; see also Google, 747 F. Supp. 3d at 103 ¶¶ 380–383 (discussing the Google–

Samsung RSA).

79. On April 19, 2025, mere days before the start of the remedial hearing, Google and

Samsung entered into an RSA lasting until September 30, 2025, with a retroactive effective date

of April 1, 2025. Rem. Tr. at 317:8–319:11 (Fitzgerald); see also PXR0608. This agreement

covers the United States only. See Rem. Tr. at 347:9-11 (Fitzgerald) (discussing PXR0608 and

RDXD-02.002); PXR0608 at -046 § 1.69. It provides for revenue share payments on an access-

point-by-access-point and device-by-device basis for both new devices and installed-base devices

(i.e., devices already in circulation). See Rem. Tr. at 347:12–348:8 (Fitzgerald) (discussing

RDXD-02.002); PXR0608 at -044 § 1.49. This means that “Samsung has the choice device by

device whether it wants to meet . . . the requirements that would open them up for promotional

payment” and can choose which “access points” are set to Google Search. See Rem. Tr. at 347:12–

348:8 (Fitzgerald) (discussing RDXD-02.002). Under this RSA, access points include the Google

Search widget on the device’s default home screen and the Chrome browser in the “hot seat,” the

application dock at the bottom of the device’s screen. Id. at 347:23–348:4 (Fitzgerald); see also

PXR0608 at -064 to -065 (Attach. C-1). Revenue share payments depend on device category, can

apply to new and installed-base devices, and range between % to %. See PXR0608 at -060 to

-062 (Attach. A).

80. Under the April 2025 RSA, Samsung is not required to exclusively distribute any

Google product or service, and the agreement explicitly states that Samsung may work with

another GSE, assistant, or GenAI service. See Rem. Tr. at 348:9–349:14 (Fitzgerald) (discussing

PXR0608 at -049 § 9.1, and RDXD-02.002). Further, the RSA contains no Gemini promotion

requirements, which is covered by a separate agreement. Id. at 349:12-18 (Fitzgerald).

50

81. That separate agreement is the Gemini Commercial Agreement, a worldwide

agreement between Google and Samsung governing GenAI services. See id. at 349:15-21

(Fitzgerald); see also PXR0571; RDX0432. The agreement went into effect on January 1, 2025,

and continues through December 31, 2027, with an automatic one-year renewal through December

31, 2028. See Rem. Tr. at 245:14–246:4 (Fitzgerald); PXR0571 at -363.

82. Pursuant to the Gemini Commercial Agreement, Google agrees to make revenue

share payments from Samsung’s “implementation of a certain Gemini experience on its Android

phones and tablets” for consumers. PXR0571 at -364 (“Background” section); see also Rem. Tr.

at 246:7–247:1 (Fitzgerald). Samsung only receives payments on “Gemini Qualified Devices,”

which must meet certain preinstallation, placement, default, and usage requirements. Rem. Tr. at

247:7–249:4, 251:4-12, 253:15–256:10 (Fitzgerald); PXR0571 at -368 to -369 § 2.3; id. at -384 to

-389 (Attach. A, A-1, A-2, and B).

83. To qualify as a Gemini Qualified Device, a new Samsung device must be

preinstalled with the Gemini app placed in the App Tray, either in alphabetical order or next to the

Play Store Icon. See Rem. Tr. at 255:9-18 (Fitzgerald); see also PXR0571 at -389 (Attach. B). As

for installed-base devices, the Gemini app must be installed at the end of the Samsung App Tray.

Rem. Tr. at 255:9-20 (Fitzgerald); see also PXR0571 at -389 (Attach. B); Rem. Tr. at 256:5-8

(Fitzgerald) (confirming that installed-base devices are devices already in circulation). Further,

Samsung must set the long-press side key and the “hot word” to invoke the Gemini app by default.

See Rem. Tr. at 254:18–255:5, 255:21–256:4 (Fitzgerald); id. at 255:24–256:4 (Fitzgerald)

(discussing how invocation of Gemini via “Hey Google” or “Hey Gemini” is required to receive a

revenue payment); PXR0571 at -389 (Attach. B). Google may choose other Gemini entry-point

offerings during device setup, and Samsung must install AICore and Gemini Nano if the device

51

meets applicable technical requirements. Rem. Tr. at 247:22–248:23 (Fitzgerald); id. at 248:24–

250:8 (Fitzgerald) (explaining that AICore is a way to run an on-device LLM and Gemini Nano is

the on-device LLM); PXR0571 at -368 to -369 § 2.3(c)(i), (d).

84. Under the Gemini Commercial Agreement, Google makes fixed payments, bounty

payments, and revenue share payments to Samsung. See Rem. Tr. at 250:11–253:14 (Fitzgerald);

PXR0571 at -384 to -385 (Attach. A).

85. Samsung can receive fixed payments that total up to $ million per month if the

Gemini Qualified Devices meet certain “Key Performance Indicators” (“KPIs”) based on usage.

See Rem. Tr. at 250:22–251:12 (Fitzgerald); PXR0571 at -384 (Attach. A), -386 to -387 (Attach.

A-1). These payments can vary based on whether KPIs are met. See Rem. Tr. at 251:4-9

(Fitzgerald); PXR0571 at -384, (Attach. A); see also Kim Rem Dep. at 173:7-9, 173:12 (agreeing

that Samsung receives more money when it preinstalls Gemini on more devices).

86. As for bounty payments, Samsung receives up to $ per device for activation

of Gemini Qualified Devices in the United States. See Rem. Tr. at 251:14–252:9 (Fitzgerald);

PXR0571 at -384 (Attach. A), -388 (Attach. A-2). The amount of the bounty payment depends on

the geographic location of the device and whether the device is new or part of the installed base.

See PXR0571 at -388 (Attach. A-2).

87. As for revenue share payments, Samsung receives % of “Net Gemini Ad

Revenue” and “Net Gemini Subscription Revenue” per month. See Rem. Tr. at 252:10–253:14

(Fitzgerald); see PXR0571 at -384 to -385 (Attach. A). “Net Gemini Ad Revenue” is defined as

% of all ad revenue generated on Gemini Qualified Devices. See PXR0571 at -364 to -365

§§ 1.9, 1.19; id. at -367 § 1.39. “Net Gemini Subscription Revenue” is defined as % of all

subscription revenue generated on Gemini Qualified Devices. See id. at -364 to -365 § 1.9; id. at

52

-367 §§ 1.40, 1.50; see also Rem. Tr. at 256:11-18 (Fitzgerald) (stating that Google receives

subscription revenue on the advanced version of Gemini).

88. The Gemini Commercial Agreement does not limit Samsung’s ability to work with

GenAI services other than Gemini. See Rem. Tr. at 350:21–351:12 (Fitzgerald); PXR0571 at -371

§ 4.7. Further, other device entry points remain available to other GenAI services, including

additional side-key buttons, hot words, or placement in the hot seat or on the default home screen.

Rem. Tr. at 351:21–353:13 (Fitzgerald). Finally, the Gemini Commercial Agreement provides a

mechanism for the parties to exit annually. Id. at 350:4-20 (Fitzgerald); PXR0571 at -375 to -376

§ 8.4(a); see also RDXD-02.002.

89. As of the evidentiary hearing, the Samsung-Google MADA has lapsed in the United

States. See Rem. Tr. at 353:15–354:15 (Fitzgerald) (stating that Samsung has no obligation to put

either the Google Search widget or Chrome on devices in the United States to receive revenue

payments under the RSA); see also Google, 747 F. Supp. 3d at 97–99 ¶¶ 348–358 (discussing the

Samsung-Google MADA).

2. Motorola

90. Google has entered additional agreements with Motorola and its parent company,

Lenovo. See Rem. Tr. at 258:4–261:16 (Fitzgerald) (discussing PXR0535); id. at 354:16–359:2

(Fitzgerald) (discussing RDXD-02.003); see also PXR0541 (February 2025 amendment to

Google–Motorola/Lenovo Google Mobile Incentive Agreement); RDX0442 (same); PXR0535

(June 2024 Google–Motorola Google One AI Premium OEM Promotion Agreement); PXR0537

(June 2024 Google–Lenovo Marketing Agreement); PXR0543 (February 2025 Google–Motorola

Gemini Fund Agreement).

53

91. In June 2024, Google and Motorola signed a marketing agreement, whereby

Motorola received $ million from Google to promote Google services, including Gemini, in the

United States and elsewhere with respect to Motorola’s Razr and Edge devices. Laflamme Rem.

Dep. at 43:11–45:2, 45:4-10, 45:13–47:1, 47:5–48:5; PXR0537 at -280 to -281 (Attach. A). During

the term of the agreement, which lasted from June 1, 2024, to December 31, 2024, Motorola was

restricted from using the funds to market non-Google assistive services, including other GenAI

assistive services. See Laflamme Rem. Dep. at 46:21–47:1, 47:5–48:5; PXR0537 at -280 (Attach.

A). Motorola viewed the additional funding received from Google as bringing it closer to where

Motorola sought to be with revenue share payments. Laflamme Rem. Dep. at 49:16-20, 50:1-6.

92. Also in June 2024, Google and Motorola entered into an agreement to offer

“AI Premium,” with an end date of September 30, 2025. Rem. Tr. at 259:14-16 (Fitzgerald);

PXR0535 at -164. Under this agreement, Motorola is required to preinstall the Gemini app and

Google One, which is a cloud storage solution, on the default home screen of four devices.

See Rem. Tr. at. 259:22–261:14, 357:13–358:15 (Fitzgerald); PXR0535 at -067 to -068 § 2.2.

Google agreed to pay Motorola a bounty for Gemini app subscriptions pursuant to this agreement.

PXR0535 at -069 to -070 § 4.1. The agreement is also non-exclusive, meaning that it does not

restrict or limit Motorola from preloading or determining the placement of third-party apps beyond

the placement and setup requirements listed. Rem. Tr. at 358:19–359:2 (Fitzgerald); PXR0535 at

-068 § 2.3.

93. In February 2025, Google and Motorola signed a Gemini funding agreement lasting

from January 1, 2025, to December 31, 2025. See Laflamme Rem. Dep. at 71:6–72:1; PXR0543

at -653, -667. This agreement allocates $ million for Gemini-related marketing activities.

See Laflamme Rem. Dep. at 72:9–73:3; PXR0543 at -656 § 4.1; id. at -668 (Attach. A).

54

94. Around the same time, Google and Motorola also amended their RSA with an

effective date of January 1, 2025. See Google, 747 F. Supp. 3d at 103 ¶¶ 380, 384 (discussing

Motorola’s RSA with Google); PXR0541; RDX0442; Laflamme Rem. Dep. at 34:13-16.

95. In amending the RSA, Motorola sought to both increase its revenue share payments

and gain flexibility in partnering with other companies. Laflamme Rem. Dep. at 36:9–38:13. The

amended RSA increases some of the revenue share payments, provides for revenue share payments

on an access-point-by-access-point and device-by-device basis for devices sold in the United

States, removes Search widget and Chrome placement obligations for devices sold in the United

States, and provides no restrictions on preloading alternative search or GenAI services, meaning

that Motorola could still receive payments under the RSA while having other GSE or GenAI

providers on the same device. See Rem. Tr. at 355:18–357:2 (Fitzgerald); PXR0541 at -196 to -

197 §§ 2.3, 2.9, 2.10; id. at -199 to -200 §§ 2.17, 2.18; see also Laflamme Rem. Dep. at 56:18–

57:2, 57:5–58:4. That said, Motorola receives less revenue share if it does not fulfill certain

placement requirements, such as Chrome as the default browser in the application dock or the

Google Search widget on the default home screen. See Laflamme Rem. Dep. at 69:15–70:11,

70:14-18; PXR0541 at -199 to -200 § 2.18.

B. Carriers

96. Google has also extended RSAs with U.S. carriers, including AT&T, Verizon, and

T-Mobile. See Ezell Rem. Dep. at 24:20–25:6 (stating that AT&T entered into a short-term

extension from December 2024 to September 2025); PXR0515 (December 2024 amendment to

Google–AT&T Mobile RSA); RDX0438 (same); PXR0597 (January 2025 amendment to Google–

Verizon RSA); RDX0440 (same); PXR0610 (March 2024 Google–T-Mobile RSA notice of

renewal).

55

97. The December 2024 AT&T amendment provides for revenue share payments on an

access-point-by-access-point and device-by-device basis. See Rem. Tr. at 361:1–7 (Fitzgerald);

Ezell Rem. Dep. at 27:5–28:15 (“[T]he biggest change in the new extension is that . . . [AT&T]

can elect to work with Google on a particular search entry point, but that doesn’t obligate [it] to

work with Google on another search entry point.”); PXR0515 at -119 to -120 § 2.5, -123

(Attach. B). Further, the agreement places no restrictions on promoting alternative search services.

See Rem. Tr. at 361:1-4 (Fitzgerald); PXR0515 at -120 § 2.8. To receive a revenue share payment

for a particular device, AT&T must satisfy certain requirements, such as placement obligations or

Chrome as a default browser. Ezell Rem. Dep. at 28:16–29:1; see also PXR0515 at -123 to -124

(Attach. B).

98. Verizon and Google amended its RSA in January 2025, effective from January 1,

2025, through September 30, 2025. See PXR0597 at -320, -323; Rem. Tr. at 359:8-13 (Fitzgerald).

Like AT&T’s amended RSA, Verizon’s amendment provides for revenue share payments on an

access-point-by-access-point and device-by-device basis and no restrictions on alternative search

services. See Rem. Tr. at 359:14-20 (Fitzgerald); PXR0597 at -321 to -322 §§ 2.4, 2.5, 2.9; id. at

-324 (Attach. B).

99. T-Mobile had previously exercised an option to extend its RSA for an additional

year through June 2025. See PXR0610 (March 2024 email exercising extension option); see also

Google, 747 F. Supp. 3d at 102–03 ¶¶ 378–379 (describing the T-Mobile RSA). By the time of the

evidentiary hearing, Google and T-Mobile were in discussions regarding a new agreement.

See Rem. Tr. at 361:23–362:10 (Fitzgerald). Consistent with the Verizon and AT&T agreements,

Google proposed that the revenue share payments be access point by access point and device by

device and that there be no restrictions on alternative search, assistant, or GenAI services. Id. at

56

362:11-19 (Fitzgerald) (discussing RDXD-02.006). Further, Google proposed that there need not

be any Assistant preinstallation requirements. Id. at 362:11-21 (Fitzgerald) (discussing RDXD-

02.006). Under Google’s proposal, T-Mobile could configure its device to preinstall alternatives

to Google Search and Gemini and still receive revenue payments. Id. at 363:1-5 (Fitzgerald).

C. Mozilla

100. Mozilla exercised an option under its existing RSA with Google to extend it for an

additional year, with an expiration date of December 1, 2026. See id. at 3166:10–3167:4

(Mullheim); PXR0370 (March 13, 2025, email); see also Google, 747 F. Supp. 3d at 96 ¶¶ 334–

336 (discussing the Google–Mozilla RSA whereby Mozilla earns revenue share payments in

exchange for Google’s default search placement on the Firefox browser).

D. Waiver Letters

101. Right before the evidentiary hearing, Google sent letters waiving certain

obligations contained in its agreements with Motorola, AT&T, and Verizon. See Rem. Tr. at 357:3-

12, 359:21–360:3, 361:9-17, 369:3-10 (Fitzgerald); PXR0607 (April 17, 2025, waiver letter to

Motorola); PXR0606 (April 17, 2025, waiver letter to AT&T); PXR0609 (April 17, 2025, waiver

letter to Verizon).

102. The Motorola letter waived Google Assistant requirements in Motorola’s RSA and

clarified that there are no restrictions in the RSA on alternative assistive services. See Rem. Tr. at

357:3-12 (Fitzgerald); PXR0607; see also Laflamme Rem. Dep. at 64:5–65:19 (Motorola wanted

the February 2025 extension to resolve ambiguity around whether partnership with other GenAI

products would violate the RSA).

103. The AT&T and Verizon letters waived restrictions on alternative assistive and

GenAI services and changed Google Assistant to an “à la carte” access point, meaning that the

57

carriers could decide whether to have Assistant and receive revenue share payments, rather than

be subject to an all-or-nothing requirement. See Rem. Tr. at 359:21–360:18, 361:9-17 (Fitzgerald);

see also PXR0606; PXR0609. These waiver letters made clear that partners could preinstall

alternative search or GenAI products and still receive revenue share payments. See Rem. Tr. at

360:1-6, 361:18-21 (Fitzgerald).

CONCLUSIONS OF LAW

In this section, the court sets forth conclusions of law that will frame its later evaluation of

the parties’ proposed remedies. The court here covers: (1) the general legal principles of antitrust

remedies; (2) the sufficiency of the liability-phase factual findings to support the proposed

remedies; (3) the “fruits” of Google’s exclusionary conduct; and (4) the propriety of including

GenAI firms and products within the scope of remedies.

I. LEGAL FRAMEWORK

A. General Principles

It is the duty of the district court, upon finding a violation of the antitrust laws, to redress

the violation and restore competition. See United States v. U.S. Gypsum Co., 340 U.S. 76, 88

(1950); Ford Motor Co. v. United States, 405 U.S. 562, 573 (1972). The remedy in a Section 2

enforcement action “must seek” to “unfetter a market from anticompetitive conduct,” “deny to the

defendant the fruits of its statutory violation, and ensure that there remain no practices likely to

result in monopolization in the future.” United States v. Microsoft Corp., 253 F.3d 34, 103

(D.C. Cir. 2001) [hereinafter Microsoft III] (en banc) (internal quotation marks and citations

omitted).3

3

The D.C. Circuit, quoting Supreme Court precedent, has identified a fourth remedial objective: “terminate the illegal

monopoly.” Microsoft III, 253 F.3d at 103 (quoting United States v. United Shoe Mach. Corp., 391 U.S. 244, 250

(1968)). On remand, the district court deemed this an improper objective in that case. New York v. Microsoft Corp.,

58

An appropriate remedy “restor[es] conditions in which the competitive process is revived,”

Massachusetts v. Microsoft Corp., 373 F.3d 1199, 1231 (D.C. Cir. 2004), and is “tailored to fit the

wrong creating the occasion for the remedy,” Microsoft III, 253 F.3d at 107; see Ford, 405 U.S. at

575 (emphasizing that the relief ordered “must fit the exigencies of the particular case” (internal

quotation marks and citation omitted)); accord 3 PHILLIP E. AREEDA & HERBERT HOVENKAMP,

ANTITRUST LAW ¶ 650a(2)(D) (5th ed. & Supp. 2025) [hereinafter AREEDA & HOVENKAMP]

(“[T]he relief must be reasonably tailored to prevent a recurrence of the challenged practices and,

to the extent practicable, to restore competitive conditions to the dominated market.”). The district

court is vested with broad discretion in crafting the remedial decree. Ford, 405 U.S. at 573

(collecting cases); Microsoft III, 253 F.3d at 105 (“[A] district court is afforded broad discretion

to enter that relief it calculates will best remedy the conduct it has found to be unlawful.”).

The ordinary starting point is an injunction terminating the anticompetitive conduct.

See Ford, 405 U.S. at 575 (stating that the relief ordered should “assure the public freedom from”

continuation of the exclusionary acts (citation omitted)); see also Microsoft III, 253 F.3d at 106.

But “relief, to be effective,” must often “go beyond the narrow limits of the proven violation.”

Gypsum, 340 U.S. at 90; see New York v. Microsoft Corp., 224 F. Supp. 2d 76, 148 (D.D.C. 2002)

[hereinafter New York I] (recognizing that “equitable relief beyond a mere injunction against

repetition of the act is generally appropriate” where “a monopolist has consummated an

exclusionary act” (quoting 3 PHILLIP E. AREEDA & HERBERT HOVENKAMP, ANTITRUST LAW ¶ 653c

(2d ed. 2000))), aff’d, Massachusetts, 373 F.3d 1199. Accordingly, “the district court is

‘empowered to fashion appropriate restraints on [the defendant’s] future activities both to avoid a

224 F. Supp. 2d 76, 100–01 (D.D.C. 2002). This court, however, need not decide this issue, because there are

independent reasons that remedies designed to eliminate the defendant’s monopoly—i.e., structural remedies—are

inappropriate in this case. Moreover, Plaintiffs are not advocating for monopoly termination as a rationale to support

their remedies. See Pls.’ Reply at 3; Rem. Tr. at 4756:11–4757:1 (Closing Arg.).

59

recurrence of the violation and to eliminate its consequences.’” Massachusetts, 373 F.3d at 1216

(quoting Nat’l Soc’y of Pro. Eng’rs v. United States, 435 U.S. 679, 697–98 (1978) [hereinafter

NSPE]); see United States v. Bausch & Lomb Optical Co., 321 U.S. 707, 724 (1944) (“Equity has

power to eradicate the evils of a condemned scheme by prohibition of the use of admittedly valid

parts of an invalid whole.”); accord 3 AREEDA & HOVENKAMP ¶ 653f (“[E]quitable relief properly

goes beyond merely ‘undoing the act’; the proper relief is eradicating all the consequences of the

act and providing deterrence against repetition; and any plausible doubts should be resolved

against the monopolist.”).

Consistent with that authority, the court may prohibit “practices connected with acts

actually found to be illegal,” Gypsum, 340 U.S. at 89, including practices “which are of the same

type or class as unlawful acts,” Zenith Radio Corp. v. Hazeltine Rsch., Inc., 395 U.S. 100, 132

(1969) (citation omitted); see Massachusetts, 373 F.3d at 1233 (concluding that the district court

“did not abuse its discretion by adopting a remedy that denie[d] Microsoft the ability to take the

same or similar actions to limit competition in the future”); accord 3 AREEDA & HOVENKAMP

¶ 653f (“[I]njunctive relief must be tailored with sufficient breadth to ensure that a certain ‘class’

of acts, or acts of a certain type or having a certain effect, not be repeated.”). It may even impose

affirmative obligations on the defendant. See Massachusetts, 373 F.3d at 1215 (approving

“forward-looking” provisions requiring Microsoft to disclose certain APIs and communications

protocols, even though “non-disclosure of this proprietary information had played no role in our

holding Microsoft violated the antitrust laws”).

Behavioral remedies, then, need not be confined to “end[ing] specific illegal practices.”

Int’l Salt Co., 332 U.S. at 401; see New York I, 224 F. Supp. 2d at 107 (acknowledging

“unquestionable legal authority which indicates that the Court may address conduct beyond the

60

precise parameters of that found to violate the antitrust laws”). Nor must they be narrowly directed

at “restor[ing] the status quo ante.” Ford, 405 U.S. at 573 n.8. Rather, they must “represent[] a

reasonable method of eliminating the consequences of the illegal conduct.” NSPE, 435 U.S. at

698 (upholding an injunction “go[ing] beyond a simple proscription against the precise conduct

previously pursued”); see Massachusetts, 373 F.3d at 1216, 1218 (applying NSPE’s “reasonable

method” standard to “forward-looking” mandatory disclosure provisions); see also Bausch &

Lomb, 321 U.S. at 726 (summarizing Supreme Court precedents as “uphold[ing] equity’s authority

to use quite drastic measures to achieve freedom from the influence of the unlawful restraint of

trade,” provided such measures “reasonably tend[] to dissipate the restraints and prevent

evasions”).

Structural relief, such as divestiture or dissolution, is also available to redress a Section 2

violation. The “most drastic, but most effective, of antitrust remedies,” United States v. E. I. du

Pont de Nemours & Co., 366 U.S. 316, 326 (1961), and perhaps “the most important,” id. at 331,

structural measures are “designed to eliminate the monopoly altogether” and therefore “require[]

a clearer indication of a significant causal connection between the conduct and creation or

maintenance of the market power,” Microsoft III, 253 F.3d at 106 (quoting 3 PHILLIP E. AREEDA

& HERBERT HOVENKAMP, ANTITRUST LAW ¶ 653b (1st ed. 1996)). Courts have traditionally

ordered such remedies where the monopoly arose by merger or acquisition, see Microsoft III, 253

F.3d at 105, and where “other measures will not be effective to redress a violation,” E. I. du Pont,

366 U.S. at 327; see also Microsoft III, 253 F.3d at 49 (observing that “[c]onduct remedies may

be unavailing” in cases involving technologically dynamic markets “because innovation to a large

degree has already rendered the anticompetitive conduct obsolete”); accord 3 AREEDA &

HOVENKAMP ¶ 653c1 (explaining that “[t]he rationale for a ‘structural’ remedy is that injunctive

61

relief is inadequate” and “[t]he case for a breakup remedy is strongest with respect to acquired

assets”).

The harshness of a remedy is not reason alone to reject it. See E. I. du Pont, 366 U.S. at

327. On the contrary, those who violate the antitrust laws cannot “avoid an undoing of their

unlawful project on the plea of hardship or inconvenience.” Id. at 326–27 (citation omitted);

see FTC v. Whole Foods Mkt., Inc., 548 F.3d 1028, 1033 (D.C. Cir. 2008) (“Even remedies which

‘entail harsh consequences’ would be appropriate to ameliorate the harm to competition from an

antitrust violation.” (quoting E. I. du Pont, 366 U.S. at 327)). Antitrust actions would be “futile

exercise[s]” indeed “if the Government prove[d] a violation but fail[ed] to secure a remedy

adequate to redress it.” E. I. du Pont, 366 U.S. at 323.

Of course, the district court’s equitable powers are not without limits. As the Supreme

Court has made clear, the ends of equity are ill-served by punishing the monopolist for past

transgressions, Int’l Salt, 332 U.S. at 401; prohibiting “all future violations of the antitrust laws,”

Zenith Radio, 395 U.S. at 133; or providing aid to a particular competitor, Brooke Grp. Ltd. v.

Brown & Williamson Tobacco Corp., 509 U.S. 209, 224 (1993). Further, the “[m]ere existence of

an exclusionary act does not itself justify full feasible relief against the monopolist to create

maximum competition.” Microsoft III, 253 F.3d at 106 (quoting 3 AREEDA & HOVENKAMP,

ANTITRUST LAW ¶ 650a (1st ed. 1996)). And the court must be sensitive to remedies that risk

substantially stifling technological innovation or impairing consumer welfare. See Massachusetts,

373 F.3d at 1219 (affirming the district court’s rejection of a remedy that would work a

“substantial” effect upon Microsoft’s incentive to innovate and thereby harm consumers); see also

United States v. Am. Tobacco Co., 221 U.S. 106, 185 (1911) (counseling that antitrust violations

should be remedied with “as little injury as possible to the interest of the general public” and with

62

“proper regard” for relevant private interests); New York I, 224 F. Supp. 2d at 100 (“Equitable

relief in an antitrust case should not ‘embody harsh measures when less severe ones will do . . . .’”

(quoting 2 AREEDA & HOVENKAMP, ANTITRUST LAW ¶ 325a (2d ed. 2000))).

Ultimately, when crafting a remedial decree, judges must “be mindful . . . of their

limitations” and approach the task with “a healthy dose of judicial humility.” Nat’l Collegiate

Athletic Ass’n v. Alston, 594 U.S. 69, 106–07 (2021). After all, judges “are neither economic nor

industry experts,” and “[a]n antitrust court is unlikely to be an effective day-to-day enforcer of a

detailed decree.” Id. at 102 (internal quotation marks and citation omitted). Thus, “[w]hen it

comes to fashioning an antitrust remedy, . . . caution is key.” Id. at 106.

B. Causation

Causation plays an important role in calibrating the scope of a remedial decree. The relief

granted should not “exceed[] evidence of a causal connection” between the defendant’s

anticompetitive behavior and its dominant position in the relevant market. Massachusetts,

373 F.3d at 1234.

To make that determination in this case, the court must resolve two threshold issues: (1) the

strength of the causal connection required to impose certain remedies and (2) the types of remedies

to which that standard applies. Both sides agree that an order enjoining Google’s exclusionary

practices is appropriate where the requisite causal connection is found only through inference.

See Pls.’ Br. at 10–11; Google’s Br. at 12. Turning to the decree’s outer limits, the parties further

agree that structural remedies must satisfy a heightened causation standard—namely, the standard

set forth in Microsoft III: “a clearer indication of a significant causal connection between the

conduct and creation or maintenance of the market power.” Microsoft III, 253 F.3d at 106 (quoting

3 AREEDA & HOVENKAMP, ANTITRUST LAW ¶ 653(b) (1st ed. 1996)); see Pls.’ Br. at 10–12;

63

Google’s Br. at 17, 24–25. The rub, then, is how to evaluate more robust behavioral remedies that

fall between these two poles.

1. The Strength of the Causal Connection

In a redux of its liability-phase defense, Google contends that Microsoft III demands a

showing of but-for causation as to every remedy except the narrowest one—an injunction against

the specific conduct held unlawful. Google’s Br. at 5–7, 17–24; Google’s Reply at 3–5; see also

Google, 747 F. Supp. 3d at 152–55 (discussing Google’s but-for argument at the liability phase).

According to Google, “a clearer indication of a significant causal connection” is synonymous with

a showing of but-for causation, and this court cannot grant any additional relief unless it finds that

Google would not have maintained its monopoly in the relevant markets absent its exclusive

distribution agreements. Google’s Br. at 5–7, 17–24; see also id. at 23 (“The but-for analysis in

this case is more straightforward than in many cases because the exclusionary conduct is limited

to specific terms of identifiable contracts.”). To justify greater relief, Plaintiffs had to show that

Google’s market dominance would have declined or that Google would not have enjoyed market

advantages “but for” the exclusive distribution agreements, which they failed to do. See id. at 21

(arguing that Plaintiffs failed to establish a causal link because they “did not attempt to identify

the market share that purportedly would have shifted to other search engines if Google had entered

only non-exclusive agreements, the amount of data that rivals purportedly would have needed to

improve their quality or monetization, or the scale that rivals supposedly would have obtained if

not for the exclusionary conduct”).

Google’s reliance upon Microsoft III for this proposition is misplaced. An extended

discussion of that case is necessary to understand why.

64

In Microsoft III, the trial court found that Microsoft had violated Section 2 of the Sherman

Act in part because it had closed off distribution channels for nascent Netscape and Java

technologies, which Microsoft perceived as threats to its dominance in the operating systems

market. Microsoft III, 253 F.3d at 68–70, 75–76. On appeal, Microsoft urged the D.C. Circuit

sitting en banc to reject that liability finding because the necessary causal link between Microsoft’s

foreclosure of distribution channels and its maintenance of an operating system monopoly was

missing. Id. at 78. To support that contention, Microsoft pointed to the trial court’s factual finding

that “[t]here is insufficient evidence to find that, absent Microsoft’s actions, Navigator and Java

already would have ignited genuine competition in the market for Intel-compatible PC operating

systems.” Id. (emphasis added).

The court rejected Microsoft’s argument. It explained that no authority stood for the

“proposition that, as to § 2 liability in an equitable enforcement action, plaintiffs must present

direct proof that a defendant’s continued monopoly power is precisely attributable to its

anticompetitive conduct.” Id. at 79. Instead, causation based on inference—that is, whether the

exclusionary conduct “reasonably appears capable of making a significant contribution to

maintaining monopoly power”—was an appropriate standard to establish liability in those cases.

Id. (citations omitted) (cleaned up). Such inference is appropriate, the court said, whether (as in

this case) the conduct was “aimed at producers of established substitutes” or (as in Microsoft III)

directed at “nascent threats.” Id. But in no event did government enforcers need to make a but-

for showing to establish liability. “To require that § 2 liability turn on a plaintiff’s ability or

inability to reconstruct the hypothetical marketplace absent a defendant’s anticompetitive conduct

would only encourage monopolists to take more and earlier anticompetitive action.” Id. The court

also pointed to the practical problem of proof. “Neither plaintiffs nor the court can confidently

65

reconstruct a product’s hypothetical technological development in a world absent the defendant’s

exclusionary conduct.” Id. A more relaxed causation standard for establishing liability was thus

appropriate, as the defendant should be “made to suffer the uncertain consequences of its own

undesirable conduct.” Id. (quoting 3 AREEDA & HOVENKAMP, ANTITRUST LAW ¶ 651c (1st ed.

1996)).4

But the court did not end its discussion of causation there. It reminded that causation is

also relevant for determining antitrust remedies. See id. at 80 (“Microsoft’s concerns over

causation have more purchase in connection with the appropriate remedy issue, i.e., whether the

court should impose a structural remedy or merely enjoin the offensive conduct at issue.”). That

is not something the trial court had appreciated. It had ordered splitting Microsoft into an

applications company and an operating systems company, without either holding an evidentiary

hearing or providing an adequate explanation. Id. at 48, 101–03. These procedural deficiencies

required a remand. Id. at 103. The trial court would have to consider anew whether “it should

impose a structural remedy or merely enjoin the offensive conduct at issue.” Id. at 80. In that

context, the Court of Appeals explained that the standard of proof to sustain a structural remedy

was more demanding than establishing causation by inference. “[S]tructural relief . . . ‘requires a

clearer indication of a significant causal connection between the conduct and creation or

maintenance of the market power.’” Id. at 106 (alteration omitted) (quoting 3 AREEDA &

HOVENKAMP, ANTITRUST LAW ¶ 653b (1st ed. 1996)). The Court of Appeals then observed that

4

AREEDA & HOVENKAMP continues to support the “reasonably appears capable” standard and the underlying

proposition that “the defendant is made to suffer the uncertain consequences of its own undesirable conduct.”

3 AREEDA & HOVENKAMP ¶ 651g. Moreover, this proposition applies at both the liability and remedies stages. See id.

at ¶ 653f (“It is thus proper to invoke again the proposition that the monopolist bears the risk of the uncertain

consequences created by its exclusionary acts. . . . [T]he proper relief is eradicating all the consequences of the act

and providing deterrence against repetition; and any plausible doubts should be resolved against the monopolist.”); id.

at ¶ 650a(2)(B) (“[I]t is always appropriate to deprive the defendant of the continuing benefits of past misbehavior.

In devising the ‘tailored’ remedies for this purpose, reasonable doubts will ordinarily be resolved against the

defendant.”).

66

the trial court “expressly [had] not adopt[ed] the position that Microsoft would have lost its

position in the [operating system] market but for its anticompetitive behavior.” Id. at 107. “If the

court on remand is unconvinced of the causal connection between Microsoft’s exclusionary

conduct and the company’s position in the [operating system] market, it may well conclude that

divestiture is not an appropriate remedy.” Id.

Google reads Microsoft III as effectively requiring a showing of but-for causation for any

remedy greater than “an injunction against continuation of [the unlawful] conduct.” Google’s Br.

at 23–24 (quoting Microsoft III, 253 F.3d at 106). It urges the court to reject Plaintiffs’ remedies

for a failure of proof “because the record does not show that Google ‘would have lost its position

in the relevant markets but for its anticompetitive behavior.’” Id. at 23 (quoting Microsoft III, 253

F.3d at 107) (alterations omitted)). But Google’s interpretation stretches Microsoft III beyond

recognition.

To state the obvious, the Court of Appeals did not use that simple, familiar phrase “but for”

when describing the remedies causation standard. (Nor did the AREEDA & HOVENKAMP treatise

on which it relied, for that matter.) Importantly, the court remanded the matter for further

proceedings, even though it acknowledged “the District Court expressly did not adopt the position

that Microsoft would have lost its position in the [relevant] market but for its anticompetitive

behavior.” Microsoft III, 253 F.3d at 107. In other words, the plaintiffs had not established but-

for causation at the liability phase. If divestiture was dead on arrival for want of such connection,

a remand would have been unnecessary. The Court of Appeals simply could have directed the

trial court to enter a prohibitory injunction. But that is not what it did. Presumably, the structural

remedy, as well as the other “conduct remedies,” see id. at 48, remained available even without

proof of but-for causation.

67

Google’s reading of Microsoft III is further undercut by the D.C. Circuit’s reasoning.

When it comes to analyzing causation, the Court of Appeals explained, plaintiffs and courts alike

face a fundamental “proof problem”: the “inability to reconstruct the hypothetical marketplace

absent a defendant’s anticompetitive conduct.” Id. at 79. The D.C. Circuit is not alone in

recognizing this evidentiary predicament. The Supreme Court has likewise concluded that but-for

causation “would be a standard of proof if not virtually impossible to meet, at least most ill-suited

for ascertainment by courts” in exclusive dealing cases. Standard Oil Co. of Cal. v. United States,

337 U.S. 293, 309–10 (1949). So, too, have leading scholars. Since Microsoft III, Areeda and

Hovenkamp have continued to endorse the “reasonable inference” standard on the ground that

“[m]any exclusionary practices, just like many negligence torts, are one-of-a-kind situations in

which it is impossible to prove that an outcome would have been different absent the violation.”

AREEDA & HOVENKAMP ¶ 657a2. “Once the challenged events have occurred, the alternative

reality can never be re-created.” Id. “For this reason,” the authors submit, “the government suitor

need not show that competition is in fact less than it would be in some alternative universe in

which the challenged conduct had not occurred. It is enough to show that anticompetitive

consequences are a naturally-to-be-expected outcome of the challenged conduct.” Id.

This problem is no less intractable at the remedies stage than it was at the liability stage.

(Google does not say otherwise.) To reconstruct the but-for world would require the court to

determine how Google Search and its rivals would have evolved had Google not secured exclusive

default placement for more than 10 years across the most effective channels of distribution. Such

an exercise is even more challenging in cases, like this one, involving technology markets. As the

D.C. Circuit observed in Microsoft III, even “six years seems like an eternity in the computer

industry”; during that period, “firms, products, and the marketplace are likely to have changed

68

dramatically.” 253 F.3d at 49. If judges could accurately chart the path innovation would have

taken, we would work on Wall Street (or the Las Vegas Strip), not Constitution Avenue. Of course,

judicial decision-making is informed by the opinions of economic experts. But not even Google’s

expert claimed “to be able to precisely specify a but-for world” “in a case like this.” Rem. Tr. at

4215:16-18 (Murphy); see id. at 4215:9-12 (Murphy) (“The idea that I could tell you, and the world

will look just like this, nobody is going to do that. And nobody is going to do that in this case,

nobody is going to do it in most cases.”).

Lest any doubts remain, the decisions implementing Microsoft III’s remedial directives laid

them to rest. On remand, Microsoft pressed a virtually identical argument to the one Google raises

here. Compare New York I, 224 F. Supp. 2d at 147 (Microsoft claiming that the plaintiffs were

“entitled to no more than a simple proscription against the offensive conduct” because there was

no proof that Microsoft’s rivals “would have achieved near ubiquitous distribution,” evolved into

viable platform substitutes, and “thereby eliminated Microsoft’s monopoly” “absent the

12 Microsoft acts found to be anticompetitive”), with Google’s Br. at 17 (Google claiming that the

record evidence does not justify relief beyond a prohibitory injunction because Plaintiffs have not

shown that Google’s rivals “would have achieved [sufficient] scale” to “displace[]” Google’s

monopoly position “in the absence of its search distribution agreements”). The district court

unequivocally rejected Microsoft’s argument. Its position, the district court reasoned,

“demand[ed] of Plaintiffs precisely what the appellate court deemed to be largely unattainable”

and overlooked the fact that the appellate court “was well aware of [the liability-phase] finding”

that but-for causation had not been established, yet “did not indicate that Plaintiffs must overcome

it in order to obtain a remedy exceeding a mere proscription of the illegal conduct.” New York I,

224 F. Supp. 2d at 147–48. On appeal, the D.C. Circuit, again sitting en banc, unanimously

69

“affirm[ed] the district court’s remedial decree in its entirety.” Massachusetts, 373 F.3d at 1204.

At no point did it question the district court’s causation analysis.

Google contorts these decisions to tease out its desired standard. With respect to

New York I, Google trains its focus on the lone instance in which the district court mentioned but-

for causation. It asserts that “the district court rejected a proposed remedy that would have

‘place[d] the Java technology “on equal footing” with Microsoft’s technology’ because ‘[t]here is

no evidence that Java would today possess “equal footing,” in terms of distribution, with

Microsoft, but for Microsoft’s anticompetitive conduct.’” Google’s Br. at 6 (alterations in

original) (emphasis added by Google) (quoting New York I, 224 F. Supp. 2d at 262). That finding,

however, came amidst a discussion of the plaintiffs’ proposal on remand that Microsoft be ordered

to distribute Java, a victim of Microsoft’s exclusionary conduct. New York I, 224 F. Supp. 2d at

260. The district court declared this remedy inappropriate because it amounted to “nothing more

than ‘market engineering.’” Id. at 262 (citation omitted). Far from “facilitating entry,” the district

court explained, mandatory Java distribution would “manipulat[e]” the market, id. at 262, and

“single[] out” a particular competitor by “anoint[ing] [it] with special treatment not accorded to

other competitors in the industry,” id. at 189. Such “favoritism” was clearly antithetical to both

the purpose of the antitrust laws and binding precedent. Id. at 189 (collecting cases); see Brooke

Grp., 509 U.S. at 224 (“It is axiomatic that the antitrust laws were passed for ‘the protection of

competition, not competitors.’” (emphasis in original) (quoting Brown Shoe Co. v. United States,

370 U.S. 294, 320 (1962))). The D.C. Circuit agreed. See Massachusetts, 373 F.3d at 1231

(“There is a real difference. . . between redressing the harm done to competition by providing aid

to a particular competitor and redressing that harm by restoring conditions in which the

competitive process is revived and any number of competitors may flourish (or not) based upon

70

the merits of their offerings.”). Thus, the “significant causal connection” standard played no role

in either court’s rejection of the “Java Distribution” remedy. The finding highlighted by Google

merely reinforced the district court’s conclusion that opening the unlawfully foreclosed channels

of distribution to competition on the merits does not mean filling them with a rival product,

especially where the record did not support such an outcome.

Google’s reading of Massachusetts fares no better. As Google tells it, the D.C. Circuit

squarely rejected the argument that “the district court erred in applying a ‘stringent but-for test’ of

causation in determining whether ‘advantages gained by Microsoft could be considered a fruit of

Microsoft’s illegality.’” Google’s Br. at 6 (quoting Massachusetts, 373 F.3d at 1233). But it was

the appellants (a group of states led by Massachusetts)—not the district or appellate court—that

characterized the district court’s causation standard in this way. See Massachusetts, 373 F.3d at

1233 (“Massachusetts also complains the district court erred in applying a ‘stringent but-for test’

of causation in determining whether ‘advantages gained by Microsoft could be considered a fruit

of Microsoft’s illegality.’”); Br. for Pls.-Appellants, Massachusetts v. Microsoft Corp., 373 F.3d

1199 (D.C. Cir. 2004) (Nos. 02-7155 & 03-5030), 2003 WL 22340413 (“The district court erred

as a matter of law in the causation burden it applied to analyzing the fruits of Microsoft’s illegality,

imposing a stringent but-for test before the advantages gained by Microsoft could be considered a

fruit of Microsoft’s illegality.”). In any event, the D.C. Circuit nowhere discredited the district

court’s approach to analyzing causation or its findings as to the strength of the causal connection

established—a point Google does not contest. See Google’s Br. at 6 (emphasizing that “an en banc

D.C. Circuit unanimously ‘affirmed the district court’s remedial decree in its entirety’” (alteration

and citation omitted)). Against that backdrop, this court is hard-pressed to conclude that

Microsoft III and its progeny require plaintiffs to satisfy such a demanding standard.

71

2. Causation Standard Applicable to Certain Remedies

Having determined that “significant causal connection” is not tantamount to but-for

causation, the court turns to the second threshold issue—the types of remedies to which the former

standard applies. Google accords substantial weight to Microsoft III’s instruction that “the

antitrust defendant’s unlawful behavior should be remedied by ‘an injunction against continuation

of that conduct’” absent evidence of a significant causal connection. Microsoft III, 253 F.3d at

106 (quoting 3 AREEDA & HOVENKAMP, ANTITRUST LAW ¶ 650a (1st ed. 1996)); see Google’s Br.

at 5–7, 23–24. With that language, Google says, the D.C. Circuit drew a sharp line between a

simple proscription of exclusionary practices and all other remedies. Google’s Br. at 5–7;

Google’s Reply at 3–5. Whereas the former may be imposed based on inference—what the

D.C. Circuit characterized as a “rather edentulous” standard of causation, Microsoft III, 253 F.3d

at 79—the latter are justified only when a plaintiff has met the heightened “significant causal

connection” test. Google’s Reply at 3–5.

Here again, Google divorces the quotation from its context. Microsoft III did not address

any remedies beyond “the structural remedy of divestiture” and “an injunction against continuation

of [the defendant’s illegal] conduct.” Microsoft III, 253 F.3d at 105–07 (citation omitted).

New York I did on remand, and it did not adopt the rigid dichotomous framework Google espouses.

By the district court’s own account, the remedial decree in that case “exceed[ed] a mere

proscription of the precise conduct found to be anticompetitive and [was] forward-looking in the

parameters of the relief provided.” New York I, 224 F. Supp. 2d at 193. It even included

“affirmative measures.” Id. at 194. The court adopted such greater measures even though it also

found that the significant causal connection needed to support structural remedies was lacking. Id.

at 185–86 (rejecting proposal to make Internet Explorer open source, which the court viewed as a

72

structural remedy, because “the court is unconvinced that there exists a sufficient relationship

between the liability findings and” the proposed “open-sourcing of IE”), id. at 177–78 (imposing

a reasonable royalty as a condition to receiving Microsoft’s APIs, communications protocols, and

related technical information, because a royalty-free disclosure of Microsoft’s intellectual property

was tantamount to a divestiture, for which the required causal connection was lacking).

Rather than apply the “significant causal connection” test to all but the narrowest

behavioral remedies, as Google urges here, the district court considered the “proportionality

between the strength of the evidence of the causal connection and the severity of the remedy.”

New York I, 224 F. Supp. 2d at 102. Its task was to “assess the strength of the causation evidence

that established liability and tailor the relief accordingly.” Id. The district court thus viewed the

potential remedie

This text is long and has been trimmed here. Open the source document for the complete record.