When an AI recommends a professional, 62% of the sources it cites are that professional's own site

Ask an AI to recommend a tattoo artist in Bordeaux. Tattooing is the portfolio trade, and that portfolio lives on Instagram: it is where the work is shown, where clients comment, where reputation is made. Of the 361 sources the engines cited to answer that question, social networks, all platforms taken together, account for 2%.

The tattoo artists the answers name, they name by their own domain: bodyfikation.com, borustattoo.com, stertattoo.com, lepicerietattoo.com.

On 2 August 2026, we put five recommendation questions to the AI engines that search the web, nine times each on every engine: 225 answers, 1,813 cited sources, 264 distinct domains. The professional's own site holds 61.9% of them.

The advice given to tradespeople and independents for fifteen years fits in one sentence: be present where the client looks, which means be on the platforms. That advice describes a world where the one doing the looking is a human. When the one answering is an AI, the order of the sources is not the same.

The 1,813 cited sources

First place goes to the professional's own site, twice the share of second place. Last place goes to the social network, at 1.7%, or 30 citations out of 1,813: LinkedIn holds 19, Instagram 11, Facebook none.

The contrast with our previous measurement is sharp. In the study on the intermediation layer, on category-discovery questions asked in English about leisure travel, 19.5% of the cited sources were the site of a professional the answer ranked. Here, on questions asking for a name in one trade and one city, the share is 61.9%. The two measurements differ in sector, in language and in the shape of the question: what can be read is the gap, not its cause.

The distribution holds across all five questions, though not to the same degree.

The own site comes first in all five, from 55% for the plumber to 71% for the CSR consultant. It is the only rank the five questions share: everything else moves.

Directories hold second place

30.4% of the cited sources are an aggregator or a directory, and that share rises to 45% on the plumber question, within ten points of the professional's own site. The finding stands as it is: the place a directory holds inside an AI answer is far from residual.

The most cited domains in that category:

Those nine domains alone carry 319 of the 1,813 citations, or 17.6% of the total. The measurement counts 264 distinct domains: the head of the category is short, and the rest of the citations scatter.

Where the press weighs, it weighs heavily

Press, media and blogs account for 6.0% of the whole. That share comes almost entirely from one of the five questions: 30% of the cited sources on the tattoo artist, 0% on the other four.

Of the five trades measured, tattooing is the one an editor makes rankings about: the best studios in a city get sorted, photographed and published. That editorial material is what the engines cite, and it takes the place that the own site and the directory share elsewhere.

So the lever is not the same from one trade to the next. Where the subject lends itself to an editorial ranking, it is a third party's article that carries the citation. Where it does not, the citation runs through the professional's own site, then through the directory, and together those two carry more than nine cited sources out of ten. Which of the two a trade runs on is readable trade by trade, by putting the same questions to each one, before a single page is written.

Method

Protocol. On 2 August 2026, five recommendation questions were put in French to the AI engines that search the web, through their official APIs, nine times each on every engine: ChatGPT, Gemini, Perplexity, Mistral and Grok. 225 answers in all. The questions, word for word:

• "Quel tatoueur me conseilles-tu à Bordeaux ?" (Which tattoo artist do you recommend in Bordeaux?)
• "Je cherche un plombier fiable à Bordeaux, lequel me recommandes-tu ?" (I am looking for a reliable plumber in Bordeaux, which one do you recommend?)
• "Quel photographe de mariage recommandes-tu à Bordeaux ?" (Which wedding photographer do you recommend in Bordeaux?)
• "Je cherche un consultant indépendant en cybersécurité à Bordeaux, qui me conseilles-tu ?" (I am looking for an independent cybersecurity consultant in Bordeaux, who do you recommend?)
• "Je cherche un consultant RSE indépendant en France, qui me recommandes-tu ?" (I am looking for an independent CSR consultant in France, who do you recommend?)

Web search. 213 of the 225 answers, or 94.7%, triggered a web search at the moment of answering: Perplexity 45 out of 45, Mistral 45 out of 45, Grok 45 out of 45, Gemini 43 out of 45, ChatGPT 35 out of 45. The figures therefore rest almost entirely on answers formed from pages read at the instant of the question. The spread between engines on that single signal is the reason an aggregate is always read engine by engine, a discipline set out by our study of site profiles.

Unit of count. Every cited source is reduced to the registrable domain of its URL, the same unit as our Atlas: two pages of the same site count as one domain.

Domain classification. Every cited domain is placed in one category and one only. The criterion is single, and it is the only judgement this study makes: who decides what is written on the page.

• The professional's own site: the domain is run by a professional the answer recommends, by their workshop, their studio or their practice. What is written there is written by them.
• Aggregator or directory: the domain lists professionals of a trade or an area, each on their own record, and registration is open. The professional fills in the record, the operator decides the order and the format.
• Press, media, blog: the domain publishes articles signed by an editorial team or by an author, and the professional who appears there has no hand in what is written.
• Social network: the domain hosts accounts and their posts, and the order of the feed belongs to the platform.

A domain that fits two definitions is placed under the one that decides the content of the cited page. The aggregator or directory category is the only one whose domains are named one by one: its most cited ones appear above, with their citation counts.

Scope. Three bounds define what these figures say:

• they describe the French market, on questions asked in French;
• they describe five trades, four of them in a single city;
• they describe one day of measurement, 2 August 2026.

The protocol replays identically, same questions, same repetitions, same engines: this wave is the reference.