Every economic era has a defining transaction. The industrial age’s mills and mines extracted value from labor and land; the financial age’s traders and funds extracted it from risk and leverage. The platforms of the emerging age extract it from language itself—from the centuries of writing, reporting, and argument that make up the open web—and return remarkably little to those who produced it.
Google started doing exactly this decades before AI was even a consideration, by attaching commercial value to keywords and linking them to the web pages that matched them—what Frederic Kaplan has called “linguistic capitalism.” Generative AI has now intensified this logic. It ingests the works built up across the open web and answers questions by recombining them—without royalty, without consent in any meaningful sense, and increasingly without even the courtesy of a click.
Google’s AI Overview is one instance of this pattern. It sits atop Google’s search results page and answers the user’s query using the publisher’s content, yet returns nothing resembling the traffic relationship that once made producing that content economically rational. Because for two decades, the implicit bargain of the open web was simple. Publishers—broadly understood as anyone producing content online—create, Google indexes and ranks it, and in exchange for appearing in the index, publishers receive a share of user attention in the form of clicks. It was never an equal bargain; Google held the leverage throughout and cemented it with each product improvement, capturing the advertising value of search while sending only traffic, not money, downstream. But it was a bargain, and it sustained an entire economic ecosystem (digital journalism chief among it) on a thin and precarious margin. Google’s AI Overview doesn’t renegotiate that bargain. It ends it, unilaterally, while keeping the publisher inside the index that makes the rest of Google’s business model work. This appropriation of value inflicts an unusually tractable harm on publishers. The conduct has a clear mechanism, a measurable effect, and a remedy logic that competition lawyers and economists will recognize: restore the conditions of exchange that the dominant firm has unilaterally withdrawn.
But the response on the ground looks nothing like that logic. Even the most advanced democracies in the world have so far managed only partial, halting responses to this extraction. The European Commission has opened a formal investigation into whether Google’s use of publisher content for AI Overview breaches competition law and has committed to no remedy yet. In the United States, a federal court found Google to hold an illegal monopoly in search, yet declined to impose any remedy on Gemini or AI Overview, treating generative AI chiefly as a threat to Google’s dominance rather than an extension of it. A regional court in Munich has held Google directly liable for the false statements its AI Overview generates—but that ruling is about accuracy, not about extraction where the output happens to be true. Only the UK’s Competition and Markets Authority (CMA) has gone as far as imposing a binding requirement addressed squarely at AI Overview. This essay examines that requirement in some detail, and explains that for all its apparent specificity, it is substantially cosmetic.
CMA’s AI Overview Remedy
In October of last year, the CMA designated Google as holding strategic market status in general search—the threshold finding that gives the CMA power to impose conduct requirements under the new digital markets regime. In late January of this year, the CMA published a proposed conduct requirement for consultation, addressing Google’s use of publisher content in AI Overview. In June, the agency imposed that requirement.
Three concerns animated the conduct requirement: (1) insufficient publisher choice over the use of their content in Google’s generative AI; (2) lack of transparency about how that content is used; and (3) ineffective attribution. These are real concerns. But they are the vocabulary of procedural fairness—they assume a relationship between Google and publishers that is basically sound but poorly governed, in need of better disclosure and clearer consent rather than structural correction. A conduct requirement built on choice, transparency, and attribution merely addresses the governance of extraction; it does not address the extraction itself. In the sections below, I examine these obligations one by one and show why.
The first obligation imposed on Google addresses the concern the CMA called insufficient publisher choice; Google must now provide publishers with effective controls to withhold their content from being used in training and “grounding”—that is, answering queries based on current, verifiable data rather than only what it learned during training. Within search, the controls must operate at directory level and at page level, and Google is barred from circumventing a publisher’s choice by acquiring the same content through other sources. Yet the obligation does not address what happens to publishers who do not withhold their content—which is to say, most publishers, most of the time, because their traffic depends heavily on their place on Google search and they simply can’t risk losing it. So, choice architecture, however well built, governs only the exit. It says nothing about what occurs for everyone who stays.
The second obligation addresses transparency. Google must publish information explaining how search content is used for training and grounding; it must also provide publishers with clear and detailed metrics on user engagement with their content in search generative AI features. On its own terms, this requirement is more substantial than it first appears. Engagement metrics, in particular, give publishers something they could not previously obtain: visibility into how their content is performing within a system they have no other means of observing. The difficulty is not that this information is worthless, but that it is only information. Knowing precisely how one’s content is used, and how poorly it performs once used, does not change the terms on which it is used. A publisher equipped with detailed metrics is in exactly the same bargaining position as one without them—better informed about their own disadvantage, but not less disadvantaged.
The attribution requirement exposes a deeper confusion in the CMA’s framework. Google is required to take reasonable steps to ensure that search content is clearly and accurately attributed, and that users have a clear means of accessing it. But AI Overview synthesizes across multiple sources. Therefore, the relationship between its output and any individual input is probabilistic and lossy by design, which makes clear and accurate attribution of a synthesis simply an architectural impossibility. The requirement also conflates two distinct goals: attribution that allows users to verify what they are reading, and attribution that allows publishers to sustain brand value. A citation link may satisfy the first in form while delivering nothing on the second in substance. It does not generate a click. It does not restore the traffic relationship. It does not compensate the publisher for the use of their content in grounding or training. What the CMA calls attribution is, in practice, a footnote on extraction.
A fourth obligation sits alongside these three: Google may not retaliate against publishers who use the controls mentioned above—for instance, by downranking their content relative to publishers who remain opted in. This requirement is a necessary safeguard, but only in the narrow sense that any opt-out regime needs one. It protects the choice already shown to matter little. Therefore, the CMA’s remedies fall short as they treat Google’s extraction as a contracting problem between two parties of roughly comparable power, when it is closer to a sovereign’s decision to stop paying tribute it once paid.
Locating the harm in competition law
A number of publishers—including Penske Media and The Hollywood Reporter—have sued Google in the United States, arguing that its conduct violates Sherman Act Section 2 by illegally exploiting its position as a dominant buyer of publisher content and data. In a recent paper in Penn Law Review, Singh and Scott-Morton support this claim. Their argument runs as follows.
Google is not just a dominant seller in the output market for general search; it is also the dominant buyer in the input market for publisher content that search depends on because, without access to search, most publishers have nowhere else to take their content. The authors explain that “Google’s monopoly power in general search and its monopsony power over publisher data are two sides of the same coin.”
Google requires publishers to accept the use of their content for AI training and grounding as a condition of staying visible in search. Because appearing in Google Search is essential for most publishers to be found at all, they have no real choice but to accept this condition. Singh and Scott-Morton argue this conduct amounts to illegal tying—bundling one product or service (access to search users) to the purchase of another (use of content for AI training and grounding). The basic competition concern with tying is that it lets a monopolist/or monopsonist use its power in one market to take over a second market.
To prove tying, two conditions must be shown. First, that there are two genuinely separate products or services. Second, that the seller only supplies one on the condition that the buyer also takes the other. Singh and Scott-Morton argue both elements are present here. On the first point, supplying data for search indexing and supplying data for AI training/grounding are clearly different services. Courts decide whether two things count as separate products by asking whether there is enough independent demand for each that a firm could profitably offer them apart. This proof is achieved by studying whether customers have asked for them separately, whether competitors already sell them separately, and how the company itself has behaved in the past. It is worth mentioning that a federal court in the Eastern District of Virginia recently ruled, in the Google Ad Tech case, that Google’s bundling of its publisher ad server (DFP) with its ad exchange (AdX) was per se illegal—meaning the court did not need separate proof of anticompetitive effects to reach that conclusion.
The tying theory of harm is, in my view, the strongest available account of Google’s conduct, and I agree the case fits. The theory does not need to establish Google’s dominance afresh—it borrows Google’s already-established monopoly in search and uses it to explain why publishers cannot meaningfully refuse the AI-related conditions attached to indexing. The main vulnerability, however, is the claim that search indexing and AI training/grounding are separate products/services. Not because they are not separate, but because it will face the market-definition trap that haunts so much of platform antitrust litigation: courts insist on a clean, provable boundary between two markets/ before they will call a bundling arrangement a tie, and Google will argue that indexing and AI ingestion are simply one integrated pipeline rather than two products awkwardly forced together. None of this means the underlying conduct is innocent. It means that antitrust adjudication, as currently practiced, sets a burden of proof that is difficult to meet even where the harm is real and visible—and that difficulty tends to favor the party with the resources and the incentive to contest every definitional question for as long as possible. Google might also respond by claiming that using publisher data to train and ground its AI systems benefits consumers, because it improves search quality and features like AI Overviews and AI Mode make for a better user experience. Singh and Scott-Morton reject this defense outright, citing several precedents in which it was ruled that competition law harm in one market cannot be excused by pointing to benefits in another market.
It is worth noting, too, that in the EU’s Google Shopping case, Google’s “product improvement” defense was rejected as a justification for anticompetitive conduct. Proving harm is somewhat easier under EU and UK competition laws, because both jurisdictions recognize exploitative abuse as well as exclusionary abuse, unlike the U.S. system, which deals only with the latter. Under this broader framework, imposing unfair conditions on a trading partner can itself constitute an abuse of dominance—without any need to show that a rival was excluded. This lens is precisely the framing the European Commission used in its press release announcing the Google AI investigation. There is another, perhaps more novel theory of harm available in the EU, with some refinement: self-preferencing—the same conduct for which Google was fined in Google Shopping, where it placed its own comparison-shopping service more prominently than rival services on its search results page. As Davies and Cohen point out, by answering the user’s query directly within AI Overview, rather than directing them to a publisher’s site, Google has placed itself in direct competition with publishers for traffic—a dynamic in which Google favoured its own service. The conduct in AI Overview may not fit squarely within Google Shopping, but several elements of the Commission’s and the EU courts’ reasoning there could be drawn on to build a self-preferencing case against AI Overview.
That’s the economic harm. But the harm does not stop at the publisher’s balance sheet. If the smaller, less diversified outlets are the ones cut off from the traffic that AI Overview withholds, then over time they are also the ones that close. What survives is a press that is smaller, more concentrated, and less plural—and a less plural press is, in the long run, a democracy with fewer independent eyes on power.
There is a narrower version of this problem too, on the epistemic side: AI Overview gets things wrong, sometimes by misreading a source, sometimes by inventing a connection that was never there. The more it becomes the public’s first and only answer, the more those errors travel uncorrected. Many would say these are not competition law problems. Maybe, not entirely. Yet competition law, and the agencies that enforce it, can do a great deal here. They simply choose not to. The CMA is just one example of this.
What, then, should be done?
It would be naive to suggest that any single remedy can resolve the extraction problem outlined above, because the problem is not a defect to be patched but a feature of the underlying economic structure. Google’s dominance in search is what allows it to extract publisher content on its own terms in the first place; no opt-out, attribution standard, or transparency obligation changes that underlying architecture, and any remedy confined to AI Overview alone will leave the same dominance free to express itself through the next product. The right level of analysis is structural, not behavioral—the question is not how Google should be made to treat publishers more fairly within its search business, but why a single firm is permitted to control the gateway through which the public finds information at all. A behavioral fix, however well designed, is just a settlement with that structure.
That said, structural reform of this kind is slow, contested, and uncertain, while publishers are disappearing now. A remedy that gives them some comfort in the meantime has value, even if it cannot be the whole answer. It is worth engaging seriously, then, with what a better-designed remedy might look like. Singh and Scott-Morton, writing from a tying/monopsony framing, consider several alternatives: a clear ban on Google tying search indexing to AI training and grounding; an opt-in regime—requiring publishers’ content to be excluded from AI use unless they actively agree to it, rather than included unless they object along with granular publisher choice; permitting collective bargaining among publishers; and judicial or regulatory rate-setting for the use of publisher content. Each addresses a piece of what the CMA’s conduct requirement does not. Opt-in reverses the default the CMA left in place, so that the burden of inertia falls on Google rather than on the publisher. Collective bargaining responds directly to the disparity in leverage between Google and any individual publisher, allowing smaller outlets to negotiate as a bloc rather than alone. Rate-setting confronts the question the CMA’s remedy avoids entirely: what publisher content is actually worth, and who should pay for it. None of these is without difficulty. But each, unlike the CMA’s remedy, at least begins from the right question.
Dilan Alma is Associate General Counsel at Pogust Goodhead, a London-based claimant group litigation firm. She holds an LL.M.in Competition, Innovation, and Trade Law from The London School of Economics (LSE).