When you ask an AI assistant what to buy, the assistant also decides which parts of the web get a hearing.
France’s competition regulator made that choice unusually visible in a 104-page opinion on AI agents released Friday. Its broad warning is familiar: assistants that recommend services, remember users and eventually complete purchases could become powerful middlemen. The useful new evidence is a small experiment showing how differently two assistants already assemble a shopping answer.
The test. From May 20 to 30, the authority sent ChatGPT and Gemini 350 product-discovery questions covering electronics, appliances, transport and hotels. It then asked 200 follow-up questions requesting specific prices for electronics and appliances. Both systems were accessed through APIs with web search enabled. The researchers kept not only the final answers but also metadata about the pages each system consulted.
ChatGPT often looked at Reddit without showing it. Reddit appeared in the material consulted for 306 of ChatGPT’s 350 discovery prompts, or 87.4%. Yet it appeared as a citation only twice. Wikipedia was consulted for 32% of those prompts and cited seven times. TechRadar and Tom’s Guide were both consulted and cited more visibly.
Gemini followed a different path. Its most consulted discovery sources were YouTube at 19.4% of prompts, Cdiscount at 18.9%, and French technology site Les Numériques at 14.9%. When the questions moved from “what should I buy?” to “where can I buy it, and at what price?”, Gemini consulted Idealo for 87% of the follow-ups, Cdiscount for 62.5% and Fnac for 49.5%.
These figures are not a league table. The systems expose source metadata differently, and the experiment did not measure which recommendations were best. What it does show is that choosing the assistant also changes the information environment behind the answer. One assistant may lean on community discussion, another on video, retailers or price-comparison sites. Much of that sourcing may remain invisible to the person reading the final recommendation.
That becomes a competition problem when the answer turns into an action. A search page normally shows several links, even if its ranking is imperfect. An assistant compresses that field into a short response. A shopping agent may go further: choose an offer, complete the purchase and keep the user inside its own interface. At that point, source selection is not just an editorial choice. It can redirect demand toward some merchants and away from others.
The authority therefore focuses on ordinary platform questions rather than waiting for fully autonomous shopping. Who is installed by default? Can users choose another assistant? Can they move their history, preferences and connections without starting over? Are recommendations influenced by partnerships, advertising or a company’s own services? Its recommendations call for scrutiny of ranking and defaults, easier switching between agents, and open standards for connecting assistants to outside services.
The limits matter. This was an illustrative API test, not a statistically significant audit of the consumer apps. It captures ten days in May, and the authority says direct product-catalog connections could soon make its web-search method obsolete. The published opinion also says its data and code are public, but the cited footnote currently leads only to the authority’s GitHub profile; I could not find the experiment repository there.
What to watch now is not only whether agents can buy things. Watch what they disclose before the purchase: which sources they consulted, why particular offers appeared, whether a placement was paid, and whether users can take their accumulated context to a competing assistant. The hidden route to the recommendation may matter as much as the recommendation itself.
Source graph: Semble source collection