AI and Search / Package of Measures Proposed by the UK CMA for Google Search / Reflections on the Intersection of AI Overviews and Digital Market Competition Regulation

Key Points

On January 28, 2026, the UK Competition and Markets Authority (CMA) proposed a package of measures for Google's search service regarding the operation of AI features, including AI Overviews. (The deadline for public comments is February 25, 2026.)

The core of the proposal is summarized in four points: the introduction of a mechanism ensuring that publishers' search rankings are not lowered if they opt out of using their content for AI features, the obligation to prove the fairness of rankings that include AI-generated results, the legal mandate for a search engine choice screen on Android and Chrome, and the assurance of search data portability.

This proposal serves as a precedent showing that AI governance can be constructed not only under the label of AI regulation but also within the framework of competition policy. It has high compatibility with the regulation of abuse of superior bargaining position under Japan's Antimonopoly Act and may influence the future practices of the Japan Fair Trade Commission.

0 Introduction

Search has long been the gateway to the internet. However, with the integration of generative AI, search is becoming both the gateway and the answer itself. What happens when the gateway becomes the answer? Value such as distribution, advertising, attention, trust, and content may become even more concentrated in the platforms that generate the answers.

In this context, the UK's competition authority, the CMA, has proposed a package of measures for Google's search service regarding the operation of AI features, including AI Overviews, and has begun soliciting public comments. The announcement was made on January 28, 2026, and the deadline is February 25 of the same year (Note 1). This article uses the content of this package to organize some reflections on the current situation where AI governance can no longer be discussed solely under the banner of AI regulation.

1 Location of the Problem

Let's confirm the current state of the UK search market with figures. According to the CMA's announcement, over 90% of general search queries in the UK depend on Google Search (Note 1). More than 200,000 companies in the UK spent over 10 billion pounds on Google search advertising in the previous year (Note 1). These figures clearly indicate that search is not just a convenient feature but a fundamental infrastructure for the economy and society.

AI features are now being combined with this. When generative AI-like UIs such as AI Overviews and AI Mode are permanently integrated into search, not only the ranking of search results but also which information is integrated as an answer and which sources are referenced and how, are incorporated into the platform's design domain. In addition to typical issues that traditional competition law has dealt with—such as superior bargaining position, self-preferencing, exclusionary conduct, and tying—issues closer to AI governance, such as content usage, compensation and bargaining power, source attribution and verifiability in generative AI output, default settings and choice design, and data transfer and interoperability, are appearing on the main stage of competition policy.

Will these be handled in the name of AI safety or AI ethics, or in the name of market structure, bargaining power, and choice? The CMA's recent move strongly suggests the latter.

What has become apparent with the rise of generative AI is the prisoner's dilemma situation for publishers. AI Overviews, which Google is promoting, are highly convenient for users because they can obtain information with zero clicks. However, for publishers who are the sources of information, this carries the risk of reduced traffic to their own sites and damage to advertising revenue. It is rational for publishers to want to refuse the use of their data to prevent their content from being free-ridden for AI learning and answer generation.

However, in the traditional ecosystem, blocking crawlers to refuse use for AI simultaneously meant disappearing from Google Search's index. If search traffic is cut off, survival as a web media outlet is nearly impossible. Publishers are forced into an ultimate choice between providing content to AI or disappearing from the web's main stage, and they have effectively had no bargaining power. This state can be evaluated as a kind of tying by a dominant operator, and it has been a factor that significantly undermines the bargaining power of publishers.

The CMA's intervention this time is intended to prevent this search market dominance from unfairly spilling over into the generative AI market and to correct the bargaining power between publishers and the platform.

2 Institutional Background

In the UK, the Digital Markets, Competition and Consumers Act 2024 (hereinafter the DMCC Act) established new competition regulations for digital markets and gave the CMA a new framework of powers (Note 3). According to the CMA's explanation, this system designates Strategic Market Status (SMS) for specific digital activities for very large tech companies and then designs proportional and targeted conduct regulations in light of the objectives of fairness, open choices, and trust and transparency (Note 4).

On October 10, 2025, the CMA designated Google as having SMS regarding general search and search advertising services (Note 5). What is noteworthy here is how AI features fall within the scope of the SMS designation. The CMA explicitly states that while Google's Gemini AI assistant was at least outside the scope of the SMS designation at the time, AI-based search features such as AI Overviews and AI Mode are within the scope of the designation (Note 5). A line is drawn where general-purpose AI assistants are excluded for the time being, while generative AI integrated into search is brought within the scope of competition regulation.

On January 28, 2026, based on this SMS designation, the CMA presented its first proposed conduct regulation regarding search services and began soliciting public comments (Note 1 and Note 2). The CMA itself positions this as a milestone for the first conduct regulation under the UK's digital market competition system (Note 1).

3 Framework of the Proposed Package

The proposal published by the CMA is organized into four areas (Note 1 and Note 2). The content and practical implications of each area are summarized below.

Regarding publisher management, the CMA states that content providers, including news outlets, should be given greater choice and transparency regarding the use of their content in AI Overviews and similar features. Specifically, it calls for Google to allow publishers to opt out of having their content used to power AI features like AI Overviews or to train AI models outside of Google Search, and further demands measures that effectively guarantee appropriate attribution in AI results (Note 1, Note 2). Most importantly, it prohibits Google from lowering the ranking of such publishers in regular search results as a consequence of choosing to opt out. This can be evaluated as a measure that legally decouples search indexing from AI training data, guaranteeing publishers an intermediate option to remain in search while opting out of AI usage.

Regarding fair ranking, the CMA demands that search result rankings be fair and transparent, and requires the establishment of effective processes for raising issues and conducting investigations. The explicit inclusion of AI Overviews and AI Mode reflects the changing times (Note 1). Specifically, it obligates Google to demonstrate fairness between its own services and those of competitors. This aims to prevent self-preferencing, and transparency in algorithms is required. The CMA imposes a burden of proof on Google to demonstrate fairness, including the development of complaint-handling processes (Note 2).

Regarding choice screens, the CMA proposes legally mandating default choice screens on Android devices and introducing them in the Chrome browser as well, to make it easier for users to switch search services (Note 1, Note 2). In the words of the CMA, this is an intervention to support active and informed choices (Note 1). This approach is similar to measures already introduced under the European Digital Markets Act, aiming to break the maintenance of monopoly power that relies on status quo bias.

Regarding data portability, this is a measure to enable individuals and businesses to make better use of Google search data. The CMA blog proposes placing Google's existing data portability tools on a legal foundation (Note 2). Search data can densely contain information about a person's interests, health, economic status, and ideological leanings. Therefore, data portability is a means of promoting competition—such as reducing switching costs and fostering innovation—while simultaneously being a high-risk area for privacy and security.

4 Random Thoughts

(1) Those who control the entrance to AI also hold the exit

The integration of generative AI expands the competition problem of search from a ranking issue to an answer issue. Ranking unfairness is a problem regarding exposure in search results. However, the design of answers reshapes not only exposure but the very structure of information distribution—whose information is summarized and integrated, how it is done, and how it is referenced.

The fact that the CMA's proposal places publisher management at the forefront is likely due to this structural change. When search becomes an answer, the value of content can no longer be measured by clicks alone. Value as material for answers, value as training data, and value as a reference point for building trust begin to overlap. Discussions on the fairness of transaction terms and bargaining power, which competition policy has handled, are flowing into the design theory of AI Overviews.

The fact that the CMA blog explicitly mentions the long-term sustainability of reliable information and the verifiability of sources is also suggestive (Note 2). Rather than just protecting the content industry, the perspective of information infrastructure is coming to the fore—that if a reliable information environment collapses, both the market and democratic society will suffer.

(2) Fair ranking has a different face from AI explainability

When people talk about AI explainability, they often think of internal model explanations or technical transparency. However, the CMA's fair ranking is not the same thing. The focus is on the fairness of commercial transactions, whether businesses are unfairly disadvantaged, the development of complaint and investigation processes, whether it is verifiable when disputes arise, and explanation to users—whether they can at least understand what is happening (Note 1).

What is being questioned here is governance transparency rather than model transparency. When generative AI is mixed in, the rationality of ranking is dispersed not only across search algorithms but also across generation logic, source selection, citation methods, and UI presentation. As a result, the location of responsibility is also dispersed. If this dispersion is left unchecked, distrust and conflict—where people are demoted or buried for reasons they do not understand—will become normalized. The CMA's demand for an effective process for raising issues and conducting investigations appears to be an institutional safeguard to avoid this normalization of distrust.

What is practically troublesome is how to define and audit fairness for the behavior of an LLM that generates tokens probabilistically. With traditional search algorithms, a certain level of explainability could be guaranteed by auditing indicators like PageRank or the weighting of specific parameters. However, in the case of AI Overviews using RAG, which sources the AI cites and which it ignores depends on context, prompts, and the probabilistic fluctuations of the model. To fulfill accountability beyond Google simply saying 'the model decided so' in response to a publisher's question about why their article was not cited, a difficult technical challenge—the visualization of the inference process—must be cleared. What the CMA is seeking is likely not full disclosure of the algorithm, but the explanation of fairness and the assurance of verifiability; however, where to draw the line is a design problem.

(3) Choice screens are a minimum weapon in digital regulation

Choice screens are neither flashy structural separation nor intense prohibition. But in a world where the entrance has become an answer, the escape route of switching if you don't like the exit is actually the most important thing. The reason the CMA specifically names Android and Chrome is likely because those are the primary access points to search (Note 1, Note 2).

However, choice screens are not a panacea. Humans gravitate toward defaults. Moreover, the more convenient generative AI is, the more they will gravitate toward it. Unless choice screens are designed to the extent that switching is not only theoretically possible but also psychologically and practically feasible, they are likely to become empty. The fact that the CMA blog includes the ability to change defaults quickly and easily at any time suggests they are aware of this point (Note 2).

(4) Circumventing copyright issues through competition law

If we delve into this proposal from a legal perspective, it is interesting that the CMA defines search engines and generative AI as separate services and attempts to sever the connection between them. Until now, in copyright law discussions, the focus has been on the pros and cons of the right to refuse crawling for AI training and the technical standards for it. However, no matter how much an opt-out right is recognized under copyright law, if the platform presents contract terms or specifications that say 'if you opt out, you won't be in search,' the right becomes a mere pie in the sky.

The CMA's approach does not operate on the playing field of copyright law, but on that of competition law, prohibiting as an abuse of dominance the linking of access rights to search engines—which have characteristics close to essential facilities—with the provision of data for AI development. With this, publishers gain the card to refuse the provision of training data for AI, or to negotiate for compensation, while maintaining the fruits of traffic from search.

5 Remaining Issues and International Consistency

Can a meaningful choice for publishers truly be established? Even if an opt-out is formally provided, if opting out effectively kills a business, then the choice is as good as non-existent. The CMA's statement that it will provide publishers with bargaining power is based precisely on this point (Note 2). However, how to concretize this bargaining power is difficult. The CMA blog states that regarding publisher relations, it will first observe the effects of the initial measures for 12 months and consider additional measures if necessary (Note 2). While cautious, it is uncertain how the information environment will change during that time.

To what extent will attribution solve the problem of declining clicks? If AI Overviews resolve a user's query on the spot, the motivation for the user to return to the original source is weak, even if the source is displayed. While attribution builds a foundation for trust, it does not necessarily guarantee the sustainability of the content industry. Since the CMA uses the term 'fairer trading' (Note 1), there remains room to delve into discussions of compensation and terms in the future.

The trade-off between transparency in ranking and the potential for abuse cannot be overlooked. The more transparent rankings become, the more SEO manipulation will evolve. Especially when AI-generated results are included, transparency carries the risk of spreading 'how-to' strategies for getting cited across the market.

International consistency is also a challenge. The CMA has indicated that it will consider measures in the UK while referring to trends in other jurisdictions (Note 2). Search crosses borders, and AI crosses borders as well. If interventions by each jurisdiction contradict each other, it is a burden for companies, but it could create loopholes for users, competitors, and publishers. If a certain degree of convergence occurs, standard governance for search and generative AI could be formed globally. The CMA's current proposal is worth viewing as one candidate for that convergence.

6 Implications for Japanese Law

Turning our eyes to the situation in Japan, the Act on Promotion of Competition for Specified Software Used on Smartphones, which was enacted in 2025, mainly targets app stores, OSs, and browsers, and does not directly regulate search engine algorithms or AI training data itself.

Furthermore, Article 30-4 of the Copyright Act is a provision that broadly permits the use of copyrighted works for AI training, which is rare even globally, and in principle, AI training is possible without the publisher's permission. Therefore, it seems difficult for publishers to counter Google using only the approach of the Copyright Act.

However, the CMA's approach uses the framework of competition law rather than copyright law, and it has high affinity with discussions on the abuse of superior bargaining position and restrictive trade practices under Japan's Antimonopoly Act. If Google were to take action in Japan such as lowering search rankings if AI training is refused, it could potentially become an issue under the Antimonopoly Act, even if it is legal under the Copyright Act. The UK CMA's current proposal will likely influence the future practice of the Japan Fair Trade Commission as a case that has drawn a clear line on the data collection authority of platformers in the intersection of copyright and competition policy.

A 'London Effect' originating from the UK, rather than the EU's 'Brussels Effect,' is also expected. For Google, building an opt-out function or data portability mechanism solely for the UK market and not implementing it in other countries is costly both technically and operationally. If Google deploys the publisher control dashboard developed in response to the CMA's request globally, Japanese publishers will also benefit from it. Conversely, there is a possibility that Google will restrict the provision of services in the UK, and the outcome of that maneuvering is unpredictable.

7 In Conclusion

When we think of AI governance, we tend to think of obligations for model developers, safety evaluations, risk management, and transparency. However, the main routes through which generative AI affects society are often distribution, circulation, and UI. Search is a prime example of this. Now that the entrance has become the answer, competition policy is no longer on the sidelines of the market. It is becoming a player that shapes the social implementation of AI itself.

The value of the CMA's proposal lies in the fact that it has bundled issues that support the information infrastructure of the AI era—publisher choice and attribution, fairness in ranking including AI, design of choice, and data portability—using the language of competition policy. Even if the regulation is labeled as a digital markets competition regime, its content is the governance of AI becoming society's answer. This is where the fascination and complexity of AI and law are condensed.

As search by AI and search for AI merge, how can we balance the fair use of search data with the maintenance of the publisher ecosystem? The UK's experiment is the front line of future global AI regulation. Legal researchers and practitioners should focus not only on model performance and ethical guidelines, but also on the direction of such gritty regulation of trade terms. This is because what determines the shape of the information we encounter daily is ultimately not the superiority of technology, but the dynamics of law and regulation.

(Magazine) 'AI and Law - Reflections'

*Please refer to the following for the table of contents

Note General Terms of Service Article 3, Paragraph 2, First Sentence
3.2 The copyright of digital content produced by creators belongs to the creator.