2026-08-06 · 2 min read
What AIs actually read before recommending a brand
Across a 360-measurement audit, the most-cited sources were neither the brands' own sites nor the top Google results. Here is what we found.
When an AI assistant recommends a supplier, it does not pull that name out of nowhere. It read pages. Which ones?
We have the answer for a real case: a French reseller of rare spare parts for classic cars. Forty questions, three assistants, three passes — 360 answers analysed, and for each one, the list of domains actually consulted.
The ranking is counter-intuitive
The most-cited domains, across all answers:
| Rank | Type of source | Citations |
|---|---|---|
| 1 | A general-purpose forum (Reddit) | 41 |
| 2 | A brand enthusiasts' forum | 36 |
| 3 | An artisan marketplace | 36 |
| 4 | A competing specialist reseller | 34 |
| 5 | The client's own site | 31 |
Three immediate lessons.
The top result is not a shop. It is a forum where strangers have been answering each other for ten years. No SEO strategy put it there; the AI goes because that is where the information genuinely lives.
The client comes fifth on home ground. In his own market, with a decent site and real stock, he is read less often than three spaces he does not control and one direct competitor.
The long tail is community-driven. Further down the ranking: a German forum, a YouTube channel, several European niche shops. Almost no traditional media at all.
The visibility numbers
Across the 360 answers:
- the client's site was cited as a source in 8.6% of cases;
- his brand was named in the text in 10.3% of cases.
Roughly one answer in ten. Put another way: nine out of ten buyers who ask an AI never hear about him — despite his being an established player in his sector.
The spread between the three assistants is instructive: from 5% to 12% recommendation depending on the model. A measurement on a single assistant would have produced a false diagnosis, in one direction or the other.
What to take from it
The natural reflex on seeing these numbers is to want to "redo the site". That is almost always the wrong first move.
If AIs primarily read two forums, a marketplace and a competitor, the useful question is not "how do I improve my tags" but "how do I exist honestly in those four places": answer where the questions are being asked, get listed on the marketplaces that matter, get the communities that hold authority in your niche to write about you.
The site remains necessary — it is what lets the AI verify that you exist, what you sell and to whom. But it is rarely the missing lever.
Your own numbers
The figures above apply to a niche, highly community-driven market. In travel or software the ranking of sources is completely different — comparison sites, trade media, directories.
Which is exactly why it is worth measuring on your market rather than reasoning by analogy.