When it comes to Google’s leveraging its power from search into AI, antitrust scholars have rightly focused their attention on the harms to publishers: lost traffic, lost licensing revenue, the gradual starvation of the outlets that actually produce the content search engines crawl.
But the same anticompetitive scheme has a second set of victims that are less sympathetic, but equally harmed: Google’s AI competitors. OpenAI, Anthropic, Meta, Grok, and every other large language model (LLM) developer is forced to compete for training data on a tilted and tied playing field Google doesn’t have to—thanks to its ability to leverage a monopoly a federal court already found illegal.
Google is a monopolist seller of general search. For its search monopoly, Google needs one input above all others: publisher content delivered through its web crawler. That makes Google not only a monopolist seller of search but also the monopsonist buyer of crawl access: the dominant, essentially unavoidable purchaser of the one thing every publisher has to sell if it wants to be found online by the entire population of English speaking internet users.
In 2023, Google introduced Google-Extended, a robots.txt directive that instructs how search engines interact with a website. It ostensibly lets publishers opt their content out of AI training. On paper it looked like an unbundling of services to give its dependents more choice: keep the search visibility, withhold the AI training data. In practice, it was not.
Evidence from the search antitrust trial showed Google kept using content from publishers who had opted out via Google-Extended to power AI Overviews. Internally, the company treated the opt-out as applying only to DeepMind products, not to AI Overviews. Despite its explicit name, Google classified AI Overviews as “search org” instead of an AI feature, a distinction it never clarified explicitly with publishers. Internal documents also showed that Google considered, and rejected, more genuine unbundling options: opt-in regimes, content-level controls, letting publishers allow training but not grounding or the reverse.
The result is that publishers cannot actually separate “let Google index me for search” from “let Google train its AI models on my content.” This is a classic case of monopoly leveraging: monopoly power in one market (search) begetting market power in an adjacent market (AI). Both run through the same crawler, and Google alone decides whether to honor the distinction. A publisher who wants out of the AI-training relationship has one real option: block Googlebot entirely and vanish from Google Search. For most publishers that option is close to existential, given how dependent they remain on Google referral traffic. In June, the UK Competition and Markets Authority, troubled by Google’s behavior towards publishers, intervened.
A lopsided playing field
This tying does not just extract value from publishers over whom Google wields monopsony power. It structurally advantages Google’s models over every AI rival built by a company that does not also run the dominant search engine. As Madhavi Singh has noted in these pages, government enforcement is one pathway to addressing this competition problem. But the choice is not limited to the government or a class of economically powerless and dispersed small publishers.
When GPTBot arrives, the publisher can block it freely. The only cost is OpenAI’s ability to train on that content. Reuters Institute survey data confirm the pattern: the large majority of major news publishers block OpenAI’s crawlers, a far higher share than those blocking Google’s AI-specific crawler. Refusing OpenAI carries no search-visibility risk, whereas refusing Google does. ClaudeBot faces the same free rejection. So does Perplexity’s crawler, Meta’s, and everyone else’s. Every LLM developer except Google’s adjudicated adjacent multi-trillion dollar search monopoly actually must negotiate for training data the way a normal buyer deals with a normal seller who has other options and real leverage–sometimes through licensing deals and sometimes through lawsuits like New York Times v. OpenAI.
Google faces none of that supplier cost or discipline. It gets the same content without negotiating, because refusing Google means forfeiting search traffic, not merely AI training participation. This is exactly the kind of monopoly leveraging Section 2 of the Sherman Act is meant to stop: a firm using power entrenched in one market to gain an advantage in a distinct, adjacent market. As the Supreme Court said in 1948, “If monopoly power can be used to beget monopoly, the [Sherman] Act becomes a feeble instrument indeed.” A market in which one player acquires its key input at zero marginal cost through leverage over an unrelated monopoly, while every other player pays market rates or gets blocked outright, is not a fair or competitive market. It is a lopsided market with a dominant firm operating under a standing subsidy extracted from captive counterparties.
The scale of the advantage is not trivial. Google’s own internal reports indicated that fully honoring publisher opt-outs would have required excluding roughly half its training corpus, on the order of 80 billion of 160 billion tokens. That is the volume of data Google’s rivals are, in principle, being asked to go acquire the hard way, one license or lawsuit at a time, while Google simply keeps it by declining to build a working opt-out. An advantage of that size, obtained through conduct a monopsonist could get away with only because of its unrelated search monopoly, is not something OpenAI or Anthropic can simply innovate around.
AI rivals should step up
The natural next step is to apply the tying and monopoly-leveraging theories already developed with respect to publishers to their second audience: rival AI developers locked out of competing for the same input on equal terms. Courts have long recognized that the “vice of tying arrangements lies in the use of economic power in one market to restrict competition on the merits in another, regardless of the source from which the power is derived and whether the power takes the form of a monopoly or not.” Restricting competition on the merits in the market for AI models is what happens when one competitor’s data pipeline is compelled and free while every other competitor’s pipeline is negotiated and priced. The government should act but if it does not, well capitalized private technology companies can initiate antitrust litigation to stop Google’s conduct, and competing AI labs would be formidable plaintiffs themselves.
The remedy sought should match the structural nature of the harm. At a minimum, Google should be required to implement genuinely independent controls that let publishers permit search indexing while refusing AI training, with enforceable prohibitions on retaliation against publishers that opt out. But if those behavioral remedies prove not to reliably prevent Google from exploiting its search monopoly to obtain an input advantage in AI, courts should consider structural relief.
Structural separation would be a significant remedy, requiring proof that narrower measures cannot adequately restore competition, not an automatic consequence of finding liability. But the point of antitrust is not merely to compensate publishers for extracted value. It is to preserve competitive markets. If Google cannot be trusted to maintain a meaningful boundary between its monopoly search service and its AI ambitions, that boundary may ultimately need to be imposed by a structural remedy.
Shishene Jing is a Fellow at the Thurman Arnold Project at the Yale School of Management and a former antitrust attorney with the Federal Trade Commission Bureau of Competition Technology Enforcement Unit, where she worked on an investigation into the artificial intelligence industry.