France's competition authority has concluded that OpenAI, Google and Anthropic together hold more than 84 percent of the global AI agent market, and warned that the shift from chatbots to autonomous "agents" could concentrate the digital economy around a small number of vertically integrated firms unless regulators act quickly on data access, interoperability and default placement.

The Autorité de la concurrence published Opinion No. 26-A-05 on 17 July 2026, the product of an inquiry opened on 8 January 2026 into the competitive functioning of the artificial intelligence agents sector. The document runs to more than 3,700 pages once its annexes are included, and it does something France's regulator had not previously attempted at this scale: it built its own AI agents, asked them 550 shopping-related questions, and logged exactly which websites they visited and cited in response.

The opinion does not accuse any company of breaking the law. It is advisory, issued under Article L. 462-4 of the French Commercial Code, and states explicitly that its purpose is not to assess conduct against Articles 101 and 102 of the Treaty on the Functioning of the European Union. What it does instead is describe how a market moves from a set of competing chatbots toward something closer to a gateway, and where that gateway could narrow.

What the Authority found

According to the opinion, generative AI tools have moved well past their original role as simple conversational agents. The report distinguishes this earlier category from what it calls "AI agents": systems capable of reasoning, planning and coordinating other agents with a degree of autonomy, rather than merely answering a single question. The European Commission's own definition, quoted in the opinion, describes these as software applications that "perceive and interact with a virtual environment" and "operate autonomously, that is to say, without being directly controlled by a human."

Three companies dominate the resulting space. Citing data from Sensor Tower, the Authority states that OpenAI, Google and Anthropic together accounted for over 84 percent of the global AI agent sector in May 2026. Citing Médiamétrie, the opinion notes ChatGPT is the leading AI chatbot in France by unique visitors, reaching 21.6 million in September 2025. OpenAI itself reported more than 900 million weekly active users and over 50 million subscribers worldwide as of March 2026.

Around that scale sit vertically integrated companies - Amazon, Microsoft and Nvidia among them - plus specialist firms such as Mistral AI, Perplexity AI and xAI. Mistral, a French company, is repeatedly cited as the sole European alternative of scale; one respondent noted that Mistral "does not benefit from the same investment capacity, infrastructure and large-scale adoption as the global US players."

Why barriers to entry differ from barriers to growth

One of the opinion's more granular arguments is that building an AI agent has become comparatively easy, while growing one to meaningful scale has not. Access to third-party generative models through an API, or through open-weight releases, has lowered the technical barrier to entering the market, but the Authority states plainly that this ease of entry does not translate into an ease of expansion.

Three factors explain the gap, according to the opinion. Reaching users is difficult when established platforms enjoy default integration into ecosystems many people already use - Copilot inside Office 365, Gemini inside Android, Meta AI inside WhatsApp. Access to high-quality data compounds an advantage for incumbents, who combine proprietary usage data with partnerships with outside sources such as news publishers, and who benefit further if they already operate a search engine capable of ranking relevant documents. And the costs of running an AI agent at scale, what the industry calls inference, now can dwarf the cost of training the underlying model, because those costs recur and grow with every use.

The opinion cites a specific figure for that cost growth: an EY study referenced in the report estimates that the average cost of "orchestration" by an AI agent - the organisation of a complex, multi-step task - reached 1.2 US dollars in 2026, up from 4 US cents for a single response in 2023. Google, separately, told the market in May 2026 that it was processing more than 3.2 quadrillion tokens per month, seven times more than a year earlier.

The experiment: how ChatGPT and Gemini answer shopping questions differently

The most concrete section of the opinion is also its most unusual: an experiment the Authority designed and ran itself. Between 20 and 30 May 2026, using the ChatGPT and Gemini APIs with search enabled, the Authority posed 350 questions expressing an intention to purchase something, plus 200 follow-up questions about completing that purchase, across four categories: electronics, household appliances, hotel bookings, and transport bookings. It then logged which websites each agent visited and which it actually cited. The results, in detailed tables within Annex 2, show two agents drawing on almost entirely different parts of the internet to answer comparable questions.

ChatGPT, according to the Authority's findings, consulted Reddit for 87.4 percent of discoverability questions, by far its most frequently visited source - yet Reddit was cited directly in only 1.0 percent of those responses. Wikipedia followed a similar pattern: consulted in 32.0 percent of ChatGPT's discoverability answers but cited in just 3.5 percent. Gemini diverged sharply: YouTube was its most-visited source at 19.4 percent of discoverability queries, and for completing a purchase, Gemini relied heavily on Idealo (87.0 percent) and Cdiscount (62.5 percent), both far ahead of anything ChatGPT consulted at the equivalent stage.

The opinion also documents a structural difference in how the two agents present their work. ChatGPT's responses include citations embedded directly in the text alongside separate metadata on sources consulted. Gemini, by contrast, returns a plain-text answer with no inline citations; only its "GroundingChunks" metadata records what it consulted. ChatGPT typically issues four internet queries per response, while Gemini averages 2.8 at the discovery stage but jumps to more than 6 during purchase completion. A further quirk: ChatGPT frequently translates French-language questions into English-language search queries, which the Authority says explains the higher proportion of English-language sources in its answers to French users.

Taken together, across both agents, the ten most frequently visited sites accounted for 24 percent of all visits, but only 22 percent of ChatGPT's citations and 20 percent of Gemini's - meaning the sites an agent visits most are not necessarily the sites it tells the user it used. The Authority frames this as illustrative rather than conclusive: the tests, it states, "are not statistically significant" and "cannot be regarded as definitive findings," reflecting a snapshot of May 2026 rather than a permanent picture of conversational commerce.

The protocols racing to define agent-based commerce

Behind the experiment sits a wider argument about the plumbing of agent-based commerce - the automation of some or all of a purchase through an AI agent rather than a person clicking through a website. The opinion identifies the Universal Commerce Protocol (UCP) and the Agentic Commerce Protocol (ACP) as the two protocols currently structuring the market, describing their architectures as meaningfully different from one another.

UCP, launched by Google in early January 2026 alongside a coalition that includes Shopify, Walmart, Visa and Stripe, is built around merchants declaring their capabilities through a file called "well-known/ucp." ACP, developed by OpenAI with Stripe and launched in September 2025 under the Apache 2.0 licence, requires merchants to apply to join a programme and submit product catalogues before setting up dedicated APIs. The opinion also references adjacent protocols, including Agent2Agent and the Agent Payment Protocol, both developed by Google, alongside cryptocurrency-oriented payment protocols such as x402 from Coinbase.

The Authority's concern is not that these protocols exist, but that whoever controls their governance controls a meaningful chokepoint in how commerce flows through AI agents. Recommendation No. 6 calls for standards in the sector to be "developed and maintained in a transparent, open and collaborative manner, in order to avoid any situation of excessive control or influence by a dominant player."

That concern tracks a pattern PPC Land has documented since UCP's launch: a tracking dashboard published by Originality.ai in May 2026 found only 26 public websites out of more than 3 million scanned had implemented the protocol, four months after its debut at the National Retail Federation's conference. Since then, Amazon, Meta, Microsoft, Salesforce and Stripe have joined the UCP Tech Council that steers the standard, broadening governance beyond Google and its original retail partners - though ACP remains active in parallel, and Walmart's early experience with ACP-based checkout inside ChatGPT reportedly converted three times worse than ordinary click-out links.

Disintermediation and the publisher traffic problem

A recurring theme in the opinion is what it calls "disintermediation": the risk that a single AI-generated answer satisfies a query so completely the person never visits the website that supplied the underlying information, undermining the revenue that site depends on. The Authority states that traffic redirected from AI agents to e-commerce websites in France sits below 5 percent, but forecasts this could reach nearly 25 percent by 2030.

This concern echoes findings that have accumulated across PPC Land's coverage of AI Overviews and publisher trafficover the past year. A randomized field experiment involving more than a thousand desktop Chrome users, published in April 2026 and last revised in June, found that Google's AI Overviews cut outbound organic clicks by 39.8 percent and increased zero-click searches by 34.5 percent when the feature appeared. Ahrefs data cited across that reporting shows the click-through-rate impact for top-ranked pages growing from 34.5 percent in April 2025 to 58 percent by February 2026, and Chartbeat figures show small publishers losing an average of 60 percent of their search referral traffic over two years. The French opinion frames this dynamic as a structural risk to the "diversity of supply" across the wider digital economy, not merely a search-engine issue.

WhatsApp and the interoperability question

The opinion's discussion of platform integration draws on an ongoing case PPC Land has tracked closely: the European Commission's antitrust action against Meta over WhatsApp, cited in the report as an example of how major platforms bundle their own AI agents into services people already use by default. Meta restricted third-party AI chatbots from the WhatsApp Business API in January 2026, leaving Meta AI as the sole assistant on a platform serving billions of users in the European Economic Area. The Commission escalated in April 2026, finding that a replacement fee structure produced the same practical outcome as the original restriction, and by June 2026 it had imposed binding interim measures ordering Meta to restore access while its investigation continues. Teresa Ribera, the Commission's Executive Vice-President for Clean, Just and Competitive Transition, said at the time that "in rapidly evolving markets, competition can be lost long before a final decision is adopted." The French opinion treats this as illustrative of a broader pattern it wants regulators to watch: Recommendation No. 2 calls for particular vigilance regarding "the effective ability of users to select and use AI agents that compete with those integrated by default within the ecosystem of major digital companies."

Monetisation: advertising, subscriptions, and one notable holdout

The opinion devotes a section to how AI agent providers make money: a freemium structure for individual consumers, free basic access followed by paid subscription tiers, alongside pay-as-you-go billing for API access priced per token. Individual subscriptions can run as high as 100 US dollars, roughly 88 euros, per month for products such as ChatGPT Max or Claude Pro.

Advertising within AI agents is only beginning to emerge, and the opinion notes a divergence in strategy among leading providers. Trials of advertising within ChatGPT began in the United States in January 2026, matching PPC Land's reporting of OpenAI's 16 January 2026 announcement. Perplexity abandoned its own advertising tests in early 2026 "for fear of losing credibility and the trust of its users." Anthropic went further, announcing in February 2026 its intention to keep Claude ad-free entirely, "in order to ensure that its responses are guided by the principle of helping users" - a position reflected in the company's own published essay. One consultation respondent added a caveat: "the integration of advertising into conversational agents is still at a very early stage and is set to evolve significantly in the future."

What the Authority is recommending

The opinion closes with three sets of recommendations rather than binding rules. The first calls for the existing regulatory toolkit - the AI Act, the DMA, the DSA, GDPR and the Data Act - to be applied "rapidly and fully," with new rules prioritised at EU level and soft-law instruments favoured over hard legislation. The second calls for vigilance around three areas: equity investments and partnership agreements between major digital companies and publishers of competing AI agents; the practical ability of users to switch away from an AI agent pre-integrated into a major platform's ecosystem; and the procedures that determine how an AI agent ranks or recommends the services it presents to a user.

A third addresses distribution channels directly, calling on competition authorities to watch "Model as a Service" (MaaS) platforms - the cloud-hosted marketplaces through which businesses access third-party AI models - and suggesting the Commission consider designating such marketplaces as essential platform services under its ongoing DMA review. The opinion also calls on industry to guarantee that users can migrate between AI agents "without any significant loss of information or functionality," and for technical standards to be developed openly rather than under a single dominant player's control.

Why this matters for marketing professionals

For anyone working in paid media, search, or e-commerce, the opinion's practical significance lies less in its legal conclusions - it draws none - and more in what its experiment demonstrates empirically: ChatGPT and Gemini do not treat the same web the same way, and the sources cited to a user may not reflect the sources actually consulted to generate an answer. A brand's visibility strategy built around ranking well in Google Search does not automatically transfer to visibility inside an AI agent's shopping recommendation, since the agent may be pulling from Reddit, YouTube, or a price-comparison site absent from that brand's existing search optimisation efforts.

The opinion's warnings about lock-in and default integration also matter for anyone planning a channel strategy around agentic commerce. If migrating between AI agents remains costly due to the retention of historical, personalised usage data, early integration choices - which protocol to build for, which agent to prioritise - may prove harder to reverse than a typical platform decision, a live question given that UCP and ACP still operate as separate, non-interoperable standards more than a year after launch. The finding that three companies control more than 84 percent of a market functioning as an increasingly important gateway to digital services gives concrete shape to a concern the industry has been circling for some time: the infrastructure connecting consumers to brands is consolidating even as the tools built on top of it multiply.

Timeline

  • 28 June 2024 - The Autorité de la concurrence issues Opinion No. 24-A-05 on the competitive functioning of the generative AI sector, focusing on the upstream part of the value chain.
  • 17 December 2025 - The Authority publishes a study on competition issues relating to the energy and environmental impact of artificial intelligence.
  • 8 January 2026 - The Authority opens an ex officio inquiry into the competitive functioning of the AI agents sector.
  • 29 January 2026 - The Authority launches a public consultation on conversational agents, inviting stakeholder responses through 6 March 2026.
  • 6 March 2026 - The public consultation period closes, having gathered input from around forty stakeholders.
  • 20 to 30 May 2026 - The Authority conducts its own experiment, posing 550 questions to ChatGPT and Gemini via their APIs to analyse cited and consulted sources.
  • May 2026 - Sensor Tower data shows OpenAI, Google and Anthropic together holding more than 84 percent of the global AI agent market.
  • 10 June 2026 - Rapporteurs and other officials are heard at the Authority's sitting on the matter.
  • 17 June 2026 - Prices for the main AI models are observed as part of the opinion's cost analysis.
  • 17 July 2026 - The Autorité de la concurrence publishes Opinion No. 26-A-05 on the competitive functioning of the AI agents sector.

Summary

Who: The Autorité de la concurrence, France's national competition authority, issued the opinion. It names OpenAI, Google and Anthropic as the three companies holding a combined 84 percent share of the global AI agent market, alongside vertically integrated firms including Amazon, Microsoft, Meta and Nvidia, and specialist entrants such as Mistral AI, Perplexity AI and xAI.

What: Opinion No. 26-A-05 concerning the competitive functioning of the artificial intelligence agents sector, an advisory document analysing barriers to entry and expansion, risks tied to platform integration and default placement, and the automation of purchasing through agent-based commerce, accompanied by an original experiment testing how ChatGPT and Gemini cite and consult sources when answering shopping questions.

When: The opinion was published on 17 July 2026, concluding an inquiry the Authority opened on 8 January 2026 and a public consultation that ran from 29 January to 6 March 2026. The Authority's own source-citation experiment was conducted between 20 and 30 May 2026.

Where: The opinion concerns the French market specifically but situates its analysis within the broader European Union regulatory framework, referencing the AI Act, the Digital Markets Act, the Digital Services Act, GDPR and the Data Act, and drawing comparisons with parallel work by competition authorities in Portugal, Hungary, Italy, the European Commission itself, the United States and the United Kingdom.

Why: The Authority concluded that AI agents are increasingly functioning as intermediaries for accessing digital services and completing purchases, and warned that without close regulatory attention to data access, interoperability, default integration and standards governance, the sector could concentrate around a small number of vertically integrated players in ways that reduce consumer choice and disrupt the business models of websites and digital services that depend on referral traffic.