What the AI reads before naming you: 157 pages for 12 answers, and 3 business websites out of 24

Yesterday we showed that the names surviving across every phrasing of the same question are not businesses, but platforms. Why remained open. Today we go one floor down: instead of looking at the names the AI produces, we look at the pages it read to find them.

This is possible because the grounded search interface returns more than an answer: it also returns the sub-questions the engine asked itself and the list of pages it consulted. So we see the raw material, not just the finished product.

Twelve questions this morning - four trades in four different French cities (plumber in Lyon, dentist in Montpellier, garage in Rennes, hairdresser in Lille), each from three angles: the best, urgency, reviews.

Thirteen pages read per answer - and you are probably not one of them

Raw tally: 12 questions triggered 34 sub-questions, which led to 157 pages being read across 85 distinct domains. That is thirteen pages consulted per answer, from six at the low end to twenty at the high end. To end up citing a handful of names.

The number that matters to a business owner lies elsewhere. Twenty-four domains were read in at least two answers out of twelve - the ones that owe nothing to the accident of a single question. Eighteen are intermediaries: directories, booking platforms, comparison sites, review sites, forums, city blogs. Of the remaining six, one is a hospital and two are national chains. That leaves three sites belonging to an independent local business: a Lyon plumber and two Lille hair salons.

Of the twenty-four sources the AI re-reads from one answer to the next, three are an independent business's own website.

One directory read in half the answers

A single domain stands out: threebestrated.fr, read in 6 answers out of 12. It shows up across all four trades and all four cities - and above all, it is read on the “reviews” question in all four cases, without exception. When a customer asks who has the best reviews, the engine goes to the same aggregator, in Lyon as in Lille, for the dentist as for the garage.

Reddit appears three times out of twelve, the national yellow pages three times, the appointment platform three times. None of those sites belongs to you. All of them can decide whether you appear in the answer.

A correction, on a figure we published yesterday

Yesterday we wrote that Google's AI Mode reportedly fires nine to eleven sub-questions for a complex question, according to analyses published this year. We measured it ourselves: on the interface we query, it is two to four sub-questions, median three. It is not the same surface - the consumer AI Mode and the programming interface do not behave alike - and that is exactly the nuance yesterday's phrasing let slip. The figure we measure is the interface's; the one we quoted comes from another.

One detail among those 34 sub-questions is worth noting: two of them were asked in English for a French question. “top rated hair salons Lille”, no accents, no local framing. So an English-language page can weigh on the answer served to a French customer.

What the market observes, pointing the same way

Measurements published this year converge on one point: there is no universal dominant source. The same site carries weight on one engine and vanishes on another - Reddit accounts for a major share of Google AI Overviews' social citations and a marginal one on Gemini per Tinuiti and Profound's readings, and Semrush watched Reddit's share on ChatGPT fall from 60% to 10% within weeks. A separate analysis of 23,387 citations attributes 17% of citations to directory sites.

Translated for a local business: the question is not “should I be on directories”, but “which ones, for my trade and my city, are actually read - this month”. Our twelve measurements give four different answers depending on city and trade. A generic ranking of “best directories” will not tell you which one decides your case.

What to do with this

• Stop judging your visibility by your own website alone. In our twelve measurements, the cited business's own site was almost never among the sources read: the engine goes through third parties.
• Identify the third parties actually read in YOUR trade and YOUR city, not those of a generic guide - they change from one city to the next, and we measured that across four.
• Treat the “reviews” question separately. That is where one aggregator recurs everywhere: your listing there is an asset, your absence a hole.
• Treat English-language content as part of the setup, even for a French-speaking clientele: the engine sometimes searches in both languages.

The limits, as always: one engine, one country, twelve questions, one morning. Assistants' answers move from day to day - we measured last week that of 81 names cited on 6 August, only 15 were still there on the 8th. What this measurement adds is the layer underneath: the names do not move on their own, the pages read to produce them do.