Two technology policy specialists told a federal appeals court on August 4, 2026 that generative AI cannot discipline Google's search monopoly, because the monopoly paid for the AI in the first place.

The filing landed in the United States Court of Appeals for the District of Columbia Circuit as case No. 26-5023, consolidated with Nos. 26-5047 and 26-5049. It runs 34 pages and 5,183 words, and it supports the Department of Justice and the state plaintiffs. Oral argument has not yet been scheduled.

The brief was submitted by Joel L. Thayer of Thayer, P.L.L.C. in Washington, D.C., acting as counsel of record. The two amici are Thayer himself and Asad Ramzanali, Director of Artificial Intelligence and Technology Policy at the Vanderbilt Policy Accelerator. Both appear in personal capacities. Neither party nor party counsel authored or funded the submission, according to the statement of authorship filed with it.

What separates this filing from the crowded amicus docket is its subject. Amici certify that they are not aware of any other brief addressing how Google has used its unlawful monopoly to extend into emerging artificial-intelligence markets, and why that dynamic bears on the appropriate remedy. Other filers or notice-givers in the appeal include Brave Software, the Washington Legal Foundation, Mozilla Corporation, Apple Inc., Samsung Electronics, a group of former antitrust officials, several sets of law and economics scholars, and OpenAI OpCo, LLC.

The premise under attack

U.S. District Judge Amit P. Mehta found in August 2024 that Google had illegally maintained monopolies in general search services and general search text advertising. He then issued a remedies decision on September 2, 2025 that rejected the Chrome divestiture the Justice Department had demanded and imposed data-sharing and syndication obligations instead. Final judgment followed in December 2025, and Google filed its notice of appeal on January 16, 2026.

Part of Mehta's reasoning was forward-looking. He wrote that generative AI might yet prove a game changer in search, and that tens of millions of people already use chatbots to gather information they previously sought through a search engine. That reasoning is the target.

The brief quotes the liability opinion directly: "Google is a monopolist, and it has acted as one to maintain its monopoly." It also quotes the remedies opinion finding that exclusive default agreements, including the arrangement with Apple, "allowed Google to persistently widen the data moat, ensuring that rivals could not achieve a degree of quality that would pose a threat to Google," and that those agreements "also discouraged investment by existing market actors and new entrants."

The question amici pose is why conduct described in those terms produced remedies they characterize as moderate. The answer Mehta gave was "a healthy dose of judicial humility." Amici argue that humility, properly understood, means applying the law rather than forecasting a technology market.

Monopoly cash, monopoly compute, monopoly data

The first section of the brief reconstructs a spending history.

In 2011 Google launched an internal research team codenamed Project Marvin to build neural networks using deep learning. By the end of 2012 the team had grown to ten people in a unit called Google Brain, and the company had put millions of dollars into compute so the group could run thousands of chips. Neural networks also need data, which Google held in unmatched volume from billions of daily searches and from products it had acquired or launched.

In 2013 Google bought DNNresearch, an academic spinout built on the AlexNet work, for $44 million, and the brief notes the company appeared ready to pay more. The acquisition brought in Geoffrey Hinton, Ilya Sutskever and Alex Krizhevsky. A 2014 magazine account cited in the filing describes the period plainly: "the company began snatching up every deep learning expert it could find." Ian Goodfellow, who had introduced generative adversarial networks, joined that year. So did DeepMind, acquired for $650 million.

Commercial application followed quickly. RankBrain began ordering results for Google Search in 2015 and soon handled 15 percent of queries. In 2017 eight Google employees published the paper introducing the transformer architecture, the design that still underpins modern systems, including the one whose initials end in the letter T. When OpenAI released ChatGPT in November 2022, Google declared an internal code red. Its response included model improvements, but also deeper integration of AI into search results, which the brief says later prompted OpenAI to declare a code red of its own.

The conclusion amici draw is causal rather than incidental: the cash, compute and data that entrenched Google in general search became the essential inputs to its AI position, so a remedy that leaves the search monopoly largely untouched leaves the source of that position intact.

Four layers, all concentrated

The second section treats generative AI as a technical stack of four layers. End users reach applications, which call foundation models, which are trained and served on cloud infrastructure, which runs on chips.

The bottom three layers, according to the brief, are each dominated by between one and three providers.

At the chip layer, recent estimates put Nvidia between 80 percent and above 90 percent of AI chips. Google has begun selling its tensor processing units externally, marketed for AI workloads, after previously reserving them for its own cloud.

At the cloud layer, Amazon, Microsoft and Google together account for roughly two-thirds of global revenues, with Google at 15 percent. The brief cites Alphabet's own first-quarter 2026 release, in which the company highlighted cloud growth of 63 percent to $20 billion for the quarter, driven largely by AI. PPC Land reported at the time that the same quarter delivered $109.9 billion in consolidated revenue while Google Network advertising fell 4 percent to $6.97 billion.

At the foundation model layer, the brief relies on a Menlo Ventures estimate published December 9, 2025 finding that three companies, Anthropic, OpenAI and Google, together took 88 percent of API revenues, with Google alone at 21 percent as of the end of December 2025. It also reproduces a finding from the United Kingdom's Competition and Markets Authority, which wrote that the growing presence across the foundation model value chain of a small number of incumbent technology firms "could profoundly shape FM-related markets to the detriment of fair, open and effective competition, ultimately harming businesses and consumers, for example by reducing choice and quality, and by raising prices."

The distribution layer is where marketers live

The application layer is where the brief becomes directly relevant to media buying.

google.com was the most visited web address in the world as of July 1, 2026. Since 2024, Google has placed AI Overviews above search results, powered by Gemini. Citing Similarweb, the brief states that 43.1 percent of Google searches included AI Overviews as of May 2026, which it translates into more than four in ten queries beginning inside Google's AI product suite. The same Similarweb dataset put AI Mode at an estimated 279 million visits in May 2026.

That 43.1 percent figure warrants a caveat the brief does not supply. Measurements of AI Overview frequency diverge sharply. Adthena's June 2026 data recorded the feature in 18 percent of US searches and 23 percent of UK searches, a gap PPC Land documented alongside Similarweb's higher reading, with no published reconciliation between the two methodologies. No industry-standard measurement of AI Overview prevalence currently exists.

What is less contested is the downstream effect. Ahrefs research published February 4, 2026 found AI Overviews correlated with a 58 percent reduction in click-through rates for top-ranking pages, up from 34.5 percent in April 2025. A randomised field experiment covering 1,065 desktop Chrome users, published in April 2026, produced causal evidence that outbound organic clicks fell 39.8 percent when an AI Overview appeared, with sponsored clicks flat.

The brief adds a further point about data. Interactions on AI Overviews and AI Mode generate material that widens Google's data advantage beyond what general search alone produced. YouTube, the second most visited address on the web, now carries an Ask YouTube feature powered by the same models.

Vertical integration, described exhaustively

The filing then walks the full stack across Google's product portfolio, and the length of the walk is the argument. A user might reach AI features in Search, Gmail, Maps, YouTube or Drive, through Chrome, installed via the Play Store, on a Pixel device running Android or a Chromebook running ChromeOS, connected through Google Fi or a Nest router, across backbone and subsea infrastructure the company operates, hitting a Gemini model hosted on Google Cloud, in a Google data centre powered by electricity procured through Google Energy, on servers containing Google TPU chips. Parts of the stack Google does not own may be financed by GV or CapitalG.

Financing is treated as a competitive fact rather than a footnote. Google is among the largest investors in and compute suppliers to Anthropic, and OpenAI has begun using Google Cloud among other providers. The brief argues this blunts the market pressure that might otherwise discipline Google: if Gemini wins, Google benefits, and if a rival wins, Google still sells the compute. Because generative AI is not generally profitable yet, amici write, those investments are themselves funded in large part by search monopoly revenues.

The Apple problem

DOJ had asked the district court to force divestiture of Chrome and potentially Android, to compel search data sharing, to bar self-preferencing, and to stop Google paying Apple for default placement in Safari.

Mehta declined the structural relief. The brief argues the court went further and left a door open: Google may continue paying roughly $20 billion a year for Safari default status so long as the arrangement is not totally exclusive, a condition amici describe as easily met.

Then came January 12, 2026, when the two companies published a joint statement confirming that Gemini would power Siri. PPC Land covered the announcement and the market-power questions it raised at the time, and Apple confirmed the arrangement at WWDC26 on June 8, 2026. For amici, that sequence is the point: the same tactic, with the same partner, applied to a new market, and now blessed rather than blocked.

Adjacent markets are not a defence

The third section is doctrinal. Amici argue that no antitrust decision has ever held that a monopolist entering a distinct, emerging market where it lacks equivalent power absolves liability where it holds power.

To make the contradiction concrete, the brief borrows an analogy from the Federal Trade Commission's post-trial memorandum in its Meta case, filed July 3, 2025:

"[I]f the only supermarket in a town starts selling pet food, the supermarket would find itself in newfound competition with Petco and PetSmart. But those competitors would not alter the supermarket's dominance, because consumers cannot patron pet-store retailers to accomplish the purpose of buying groceries for the week."

The brief also revisits Foundem, the vertical search product built by Adam and Shivaun Raff. Reporting cited in the filing described a service precise enough to "figure out which websites charged hidden shipping fees and which offered truly good deals." Once Google identified it as a threat, the product was buried as far as 170 pages down, even for searches on the company name, while MSN Search and Yahoo ranked it highly.

Amici close by asking the court to reject the premise that generative AI might create competitive pressure sufficient to diminish Google's search monopolisation, and to ensure the remedy is strong enough to stop Google using what the district court itself called the same anticompetitive playbook for its generative AI products that it used for Search. In a post on LinkedIn describing the filing, Ramzanali put it more compactly: "The humblest judicial posture is to apply the law, not to forecast where AI is heading."

Why this matters to advertisers and publishers

The appeal will determine whether the six-year behavioural regime that took effect on February 3, 2026 survives, is widened, or disappears. Google asked the D.C. Circuit on May 22, 2026 to reverse the judgment outright, arguing the trial judge applied the wrong legal standard at nearly every stage. The government cross-appealed. This brief pushes in the opposite direction, asking the court to vacate the remedy as too weak rather than too strong.

For media buyers, the practical stakes sit in the syndication and data-sharing provisions. Those provisions were designed to let rival search products and AI answer engines reach competitive quality, which over time would determine whether search advertising budgets have anywhere else to go. The Glue data system requirement and the one-time index snapshot for qualified competitors are the mechanics of that theory, and a legal paper covered by PPC Land has already argued that implementing them fairly is harder than the order acknowledges.

For publishers, the connection is more direct. If the appeal narrows the remedy, the distribution advantage that produces AI Overviews on more than four in ten queries persists without structural constraint. DuckDuckGo's chief executive called the original remedies insufficient within days of the September 2025 ruling, and the traffic data published since has not weakened that critique.

There is also a jurisdictional echo. A Munich court ruled on May 28, 2026 that AI Overviews forfeit host privilegebecause they speak in their own words rather than reproducing sources. Different legal question, same underlying fact pattern: the summary layer is a product, not a directory, and courts on two continents are now being asked what follows from that.

Timeline

Summary

Who: Joel L. Thayer, president of the Digital Progress Institute and a senior fellow at both the America First Policy Institute and the Vanderbilt Policy Accelerator, and Asad Ramzanali, Director of Artificial Intelligence and Technology Policy at the Vanderbilt Policy Accelerator and formerly chief of staff at the White House Office of Science and Technology Policy. Thayer serves as counsel of record. Both file in personal capacities, supporting the United States and the state plaintiffs against Google LLC.

What: A 34-page, 5,183-word amicus curiae brief arguing that the district court erred in treating generative AI as a potential competitive check on Google's general search monopoly. The brief makes three claims: that Google's search monopoly financed and seeded its AI position through investments including the $650 million DeepMind acquisition; that generative AI is already concentrated, with three firms holding 88 percent of foundation model API revenues and Google alone at 21 percent as of December 2025; and that a monopolist's entry into an adjacent market has never functioned as a defence. It asks the court to vacate the remedy as inadequate.

When: Filed August 4, 2026, as document 2186578. Oral argument has not been scheduled. The underlying liability ruling dates to August 2024, the remedies decision to September 2, 2025, and the final judgment to December 5, 2025.

Where: The United States Court of Appeals for the District of Columbia Circuit, case No. 26-5023, consolidated with Nos. 26-5047 and 26-5049, on appeal from the U.S. District Court for the District of Columbia in Nos. 20-cv-3010 and 20-cv-3715.

Why: The district court limited its remedy in part because it expected generative AI to introduce competitive pressure into search. Amici argue that expectation is unsupported by market structure, that Google is the most vertically integrated participant across chips, cloud, foundation models and applications, and that the arrangement placing Gemini inside Siri repeats in AI the exact distribution tactic that produced the search monopoly. The outcome determines whether search advertising and publisher traffic remain governed by a single vertically integrated stack.