France Just Audited What AI Shopping Agents Actually See. Most Stores Would Not Like Their Result.
France's competition regulator asked ChatGPT and Gemini 550 shopping questions and logged every site the AI visited and cited. Here is what the numbers mean for your store's visibility, and what to do about it.
On July 17, France's competition authority (the Autorité de la concurrence) published Opinion 26-A-05 on AI agents, the assistants that can research a product, compare options, and soon complete the purchase for you. (Official press release, detailed coverage at ppc.land.)
Buried in the legal language is something remarkable. The regulator did not just hold hearings. It ran an experiment.
The regulator ran the audit most merchants never run
Between May 20 and 30, the authority's team posed 550 shopping questions to ChatGPT and Gemini with search turned on: 350 questions expressing an intent to buy something, plus 200 follow-ups about completing the purchase, across electronics, household appliances, hotel bookings, and transport bookings. Then they logged exactly which websites each AI visited while researching, and which ones it actually cited in its answer. (ppc.land)
In other words: a national regulator just ran an AI visibility audit of the shopping journey.
The headline finding is a gap most store owners have never seen measured. ChatGPT consulted Reddit in 87.4% of product-discovery questions, making it the AI's most-visited research stop. But Reddit showed up as a cited source in only 1.0% of the answers. Gemini behaved differently: it leaned on YouTube for discovery (19.4% of visits) and on price-comparison sites such as Idealo for purchase completion (87.0%).
Read that again. The sites an AI reads and the sites an AI shows you are two different lists. And neither list is visible to the merchants on (or missing from) them.
Why this matters for your store now, not in 2030
Today, AI agents send less than 5% of traffic directly to e-commerce sites. The forecasts the regulator cites put that at 20 to 25% by 2030. Both figures are read on the French market, so treat them as the shape of the trend rather than a number for your own country: the opinion itself states the current one as "less than 5% in France", and the 2030 range is an estimate from some respondents to the public consultation rather than a measurement. (Press release, Opinion 26-A-05, English text)
That is not a niche channel. That is a fifth to a quarter of your future customers asking an AI "what should I buy?" and the AI deciding which stores exist.
The pressure is already visible on the search side. In a randomized field experiment on 1,065 desktop Chrome users, Saharsh Agarwal of the Indian School of Business and Ananya Sen of Carnegie Mellon found that showing a Google AI Overview cut outbound organic clicks by 39.8% and raised zero-click searches by 34.5%. That study is separate from the French opinion, not one of its findings. Buyers increasingly get their answer without ever reaching your site.
And the layer making these decisions is concentrated. The opinion reports that OpenAI, Google, and Anthropic together hold more than 84% of the global AI agent market (Sensor Tower data, May 2026). A handful of gatekeepers, each with its own habits about what to read and what to cite.
The regulator named the risks plainly: merchants getting disintermediated, agents ranking or recommending in ways nobody outside can inspect, buyers losing the ability to make an informed choice, and stores losing the customer behavior data they rely on today.
What a store owner should actually do
The French opinion's recommendations are aimed at regulators and platforms: apply existing rules, watch the big players' investments and default deals, make it possible to switch agents without losing your data, and build open standards for agent-driven checkout. Useful, but slow. (Press release)
Here is what you can do without waiting for Brussels or Paris:
- Measure your AI visibility the way the regulator did. Ask the engines the questions your buyers ask. Which stores get named for your product category? Are you cited, or just silently read? You cannot fix a surface you have never looked at. This is exactly what our ForIntel Search & AI Visibility reports do for you: we run the audit across the AI answer surfaces for your category and hand back which engines quote whom, where you are cited versus silently read, and the wedge queries to fix first.
- Check the sources the agents actually read. The experiment showed the research path runs through community and comparison sites (Reddit, YouTube, price comparators). If your product has no footprint where the agents do their reading, you are invisible one step before the answer is even written.
- Treat your Google rank as one signal, not the scoreboard. The click-through data says the answer box is eating the clicks. Ranking #1 under an AI Overview that never mentions you is a quiet loss that a rank tracker will never flag.
- Watch the checkout protocols, but do not panic. Agent-driven checkout is real (OpenAI's protocol launched September 2025, and Google announced its Universal Commerce Protocol in January 2026 with Shopify, Target and Etsy), but merchant implementation is early. Originality.ai keeps a running crawl of how many sites have adopted Google's version, and the count is the number to check before you budget for it, because any figure quoted here would be out of date by the time you read it. (Originality.ai UCP adoption tracker) Visibility is the near-term battle. Checkout plumbing is the later one.
The bottom line
When a competition regulator asks AI agents 550 shopping questions and logs every site visited and cited, the message to merchants is simple: this surface is now important enough to audit, and it is auditable. The stores that measure their AI visibility in 2026 will not be the ones surprised by where their traffic went in 2030.
If you want to see what the AI engines say when a buyer asks about your category, that is precisely what our ForIntel Search & AI Visibility reports measure, and deliver as a report you can act on. Start with the question the French regulator started with: what do the agents actually see?
