AI reads a wedding photographer's site, almost never a restaurant's

Short answer: it depends almost entirely on your trade. When the web-searching assistant recommends a wedding photographer, it had opened that photographer's own site in 53% to 82% of cases. When it recommends a restaurant, it had read the restaurant's site for at most 1 of the 118 restaurants it named. Same assistant, same question, same day.

Say you run a restaurant. An AI assistant recommends you by name to someone asking where to eat in your city. You reasonably assume it has seen your site - your photos, your menu, your opening hours. We looked at the pages the assistant had actually consulted before answering. Your site is almost never among them. It read a city guide, a booking platform, a local news piece - and that is where it got you from.

The trade decides. Where a layer of guides and platforms exists, the assistant reads that layer. Where none exists, it reads you.

What was measured

Waseit's studies ask, in each city and each trade, the question a customer would really type: "what are the best {trade} in {city}?". Only one of the three assistants keeps, inside its answer, the list of pages it consulted to write it: Perplexity, the one that searches the web. So we took its 680 recommendation answers across France, the UK and Brazil, the 10,298 links they carry, and the 5,221 businesses they name. For each named business, one question: was any of those pages on that business's own domain?

Matching a name to a domain is never exact, and the error runs both ways. So we publish two bounds: a floor, where the compacted name must appear inside the domain, and a ceiling, where one distinctive word is enough. We also drop, through a filter that judges no meaning at all, any domain cited across several different cities: a site cited in both Leeds and Bristol is not one Leeds business's own site, it is a directory or a platform. Finally, 10.3% of the extracted "names" are not names - list headers, addresses, criteria - and are removed; the ranking by trade is the same before and after that cleanup.

Why does the trade decide?

Because some trades have a layer of intermediaries sitting above them, and others do not. Ask for the best restaurants in Paris: the assistant reads Time Out, Le Fooding, OpenTable, a what's-on magazine. Eight restaurants named, zero restaurant sites among the sources. In Lyon, in Marseille, in São Paulo, the profile is the same - guides, booking platforms, a review site, sometimes a hotel's blog recommending places to eat. The editorial layer is thick enough that the assistant has no reason to go further.

Now ask for a wedding photographer, a gas engineer, a letting agent: that layer barely exists. There is no Time Out of gas engineers. The assistant falls back on what it can find, which is the businesses' own sites - and it reads them for at least half the names it gives. The same mechanism explains the whole order of the table: the more your sector is covered by guides and platforms, the less your own site gets read.

This is not the first time we have seen it. Back in August, on a far smaller measurement - 157 pages read for 12 answers, and 3 business sites out of 24 - we already wrote that the engine rarely reads the page of the business it recommends. What these 680 answers add is that "rarely" is not a useful average: it is never for some trades, and usually for others.

What it means if you are the business

Start by working out which side of the table you are on, because the job is not the same. If your trade is intermediated - restaurant, hotel, bakery - polishing your own site will not change what the AI says about you: it does not read it. What it reads are the guide and platform pages that mention you, and that is where the description it gives of you is actually decided. If your trade is barely intermediated - photographer, gas engineer, specialist tradesperson - the opposite holds: your site is the source, and what it leaves unsaid, the assistant will not invent in your favour.

Either way, the useful question is not "is my website any good". It is "which pages does the assistant open when someone asks about my trade in my city". There is no single directory to get listed in: 5,774 different sources were cited, and the most cited one accounts for under 4%. The list that matters is the one for your trade, in your city.

And your business - what do the assistants read before talking about it?

That is exactly what the free Waseit audit shows. It asks ChatGPT and Gemini the questions your customers type, in your language and for your market, and returns in about 60 seconds a score out of 100, the competitors named instead of you, the questions where you do not appear - and, when Gemini searched Google, the pages it actually consulted. That list is what tells you where to get listed, guide by guide, platform by platform. Free, no account and no card; an email address only unlocks the competitor names and the action plan. It is a dated snapshot, not a ranking, and we promise no position.

What this does not prove

• The measurement covers Perplexity only, the one assistant whose answers our capture keeps the consulted pages for. For ChatGPT and Gemini the field is missing from every row - that is a gap in our collection, not a fact about them: nothing here says what they read, or whether they read at all.
• Three markets only - France, the UK and Brazil - one phrasing of the question, and a single capture between 31 July and 16 October 2026, through the assistants' programming interfaces, which approximate the consumer apps without reproducing them.
• "Own site read" means a page on that domain was among the sources the answer carries. It does not prove the assistant used it to write its sentence, nor, conversely, that it is unaware of a site absent from that list.
• Matching a business name to a domain name stays approximate in both directions: a business can own a domain unrelated to its name, and a common word can create a false match. Hence the floor and the ceiling, always given together; the true figure sits between them.
• The trade families group three languages, and some rest on a narrow base: 118 names for restaurants, 116 for bakeries. The order of the table is solid; the decimal point on any one row is not.
• We measured what the assistant read, not what gets a business recommended. Nothing here says a site being read improves your position, or that a site left unread damages it.