France's competition authority: OpenAI, Google, and Anthropic control 84% of the AI agent market

France's Autorité de la concurrence published a formal opinion on July 17 concluding that OpenAI, Google, and Anthropic together control more than 84% of the global AI agent market, and warning that the shift from chatbots to autonomous agents risks concentrating the digital economy around a handful of vertically integrated firms unless regulators move on data access, interoperability, and default placement before the market hardens. Opinion No. 26-A-05 is the product of an inquiry the authority opened on January 8, 2026, running to more than 3,700 pages with annexes once its evidence and roughly 40 public-consultation contributions are included.

The authority built its own test

Rather than rely solely on submissions from interested parties, the Autorité ran an original empirical study between May 20 and 30, 2026: it built its own agents on top of the ChatGPT and Gemini APIs and posed 550 shopping-related queries — 350 expressing purchase intent across electronics, appliances, hotel bookings, and transport, plus 200 follow-ups — logging exactly which websites each agent visited during research versus which it actually cited in its answer. The two agents barely overlapped in sourcing behavior: ChatGPT visited Reddit for 87.4% of discovery questions but cited it in just 1.0% of final answers, while Gemini leaned on YouTube and the price-comparison site Idealo, which it cited in 87.0% of purchase-completion answers.

That gap between what an agent reads and what it credits is the empirical spine of the authority's disintermediation concern: if an agent's answer satisfies the user without sending them to the underlying site, the traffic that used to flow to publishers and retailers from a search result never arrives. The authority's own estimate puts current traffic redirection from agent answers below 5%, but projects it could reach 25% by 2030 — a shift big enough to reshape how content and commerce sites make money from AI-mediated discovery.

The risks named

Beyond the three-firm concentration figure, the opinion names several specific mechanisms rather than concentration in the abstract. Self-preferencing: vertically integrated firms can embed their own agent by default — Copilot inside Microsoft 365, Gemini inside Android, Meta AI inside WhatsApp — giving them a distribution advantage no standalone competitor can match. Algorithmic collusion: the possibility that AI agents transacting with each other in commerce scenarios could coordinate pricing or behavior without explicit human intent. Protocol fragmentation: Google's Universal Commerce Protocol and OpenAI's Agentic Commerce Protocol remain non-interoperable, meaning a merchant or developer effectively has to pick a side rather than build once. Lock-in through data: the authority notes that inference and orchestration costs have scaled from roughly 4 cents per query in 2023 to about $1.20 per orchestrated task in 2026, raising the capital bar for any new entrant trying to compete on infrastructure rather than just a thinner wrapper.

Easy to enter, hard to grow

The opinion draws a sharper distinction than 'three companies dominate': it separates the barrier to *entering* the AI agent market from the barrier to *expanding* inside it, and finds the two pulling in opposite directions. Entry barriers are comparatively low — a developer can build an agent on top of any of several available foundation models without training one from scratch, which is why the market includes specialist entrants like Mistral AI, Perplexity AI, and xAI alongside vertically integrated giants such as Amazon, Microsoft, Meta, and Nvidia. Expansion is the harder problem: user acquisition, data availability, the technical cost of migrating a personalized agent relationship, and rising inference costs all favor whoever already has scale. One respondent to the authority's public consultation put it plainly regarding Mistral, France's own frontier-model champion: it 'does not benefit from the same investment capacity' as its US rivals — a domestic vulnerability the French regulator is, not coincidentally, the one flagging loudest.

What it recommends — and what it can't force

The authority's six recommendations are notable mostly for how little new machinery they ask for: apply existing frameworks — competition law, the EU AI Act, the Digital Markets Act — 'rapidly and fully' rather than waiting for bespoke AI-agent legislation; monitor equity stakes major platforms take in agent developers; guarantee fair access to distribution for Model-as-a-Service platforms; improve user awareness of default settings and switching costs; establish interoperability and data-portability standards so a user can leave one agent for another without losing functionality; and keep standards governance open rather than letting it fall under the control of the dominant operators.

None of this is binding. An *avis* from the Autorité de la concurrence is advisory — it shapes how the European Commission, national regulators, and courts read the market, and it builds the record for future enforcement, but it doesn't itself fine anyone or force a protocol change. The opinion explicitly frames its own recommendations as things regulators and legislators should now do, not things it has done. That gap between diagnosis and remedy is the real story: France's competition authority has produced one of the most detailed empirical pictures yet of how concentrated the agent layer already is, at a moment when the tools to actually loosen that concentration — interoperability mandates, data portability rules — mostly don't exist yet anywhere in the world.