The AI reopened the same pages, then named different businesses

Short answer: what changes is not its reading, it is its answer. When the same need is reworded, the web-searching assistant reopens 72.8% of the very pages it had just read. Yet of the 5,221 businesses it had named, only 1,939 come back - 37.2%, or 56.3% if names are matched loosely. The sources hold; the list does not.

You ask an assistant for the best plumbers in your town. It gives you eight names. Someone else puts the same need differently, "list the most recommended ones and explain why", and the list is no longer the same. The explanation that comes to mind is that it went looking elsewhere. We checked, page by page, across 682 pairs of answers: it had reopened almost exactly the same ones. What moves is not what it reads. It is what it takes from it.

It does not change what it reads. It changes what it answers.

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}?" - then the same need reworded: "list the most recommended {trade} in {city} and explain why". Only one of the three assistants keeps, inside its answer, the list of pages it consulted: Perplexity, the one that searches the web. So we isolated its pairs of answers: same market, same city, same trade, the two wordings. 682 pairs, 39 cities, 58 trade-and-market cells, across France, the UK and Brazil. For each pair, two comparisons and nothing else: how many of the pages read the first time are read again the second, and how many of the businesses named the first time are named again.

An overlap means nothing until you know what chance would give. So we paired, the same way, answers with no relation to each other: same country and same trade but two different cities, then same city and two different trades. Shared exact pages fall to 0.7% there, and shared names to 0.1% - in other words, the 72.8% measured on the real pairs is not a background effect. Two further controls: names are matched under two bounds, strict equality and a single shared distinctive token, and we conclude only on what holds under both; and because the page list grows a little from one wording to the other (13.7 domains against 14.8), we redid the count on the pairs whose two lists are exactly the same size. The gap does not narrow, it widens: 82.7% of domains against 38.8% of names.

Why does the list move if the reading does not?

Because these are two different operations, and only one of them is stable. Finding the pages that match "best plumbers in Leeds" is a retrieval problem: the same words bring back the same pages, and that is what the figure shows. Picking eight names from inside those pages is a writing problem, and it starts from scratch in every answer. The number of names does not move either: 7.7 on average each way. It is not that it says more or less; it says just as much, and not the same.

An ordinary pair from the corpus shows it plainly. For one trade in one French city, the assistant cited ten pages the first time and seven the second - and all seven were among the ten. It had the same material in front of it. Of the eight businesses named, three or four, depending on how closely names are matched, were not in the first list. Same reading, different outcome. The same contrast holds in every one of the 49 trade-and-market cells with enough pairs to count, from British hairdressers to French heating engineers, without a single exception.

What it means if you are the business

This is good news and bad news, and they have to be taken together. The good news: the pages where your visibility is decided are not drawn at random for every question. For your trade in your town, the assistant comes back to roughly the same sites, which means a handful of pages matter and are worth the work - bearing in mind only that there is no single directory to be listed in: those pages are the ones for your trade in your town, and they differ from one trade to the next.

The bad news: being on them does not put you in the answer. Once the page is read, the choice of names begins again, and the list changes as soon as the question is put another way. So there is no position to hold, and nobody can sell you one. What can be worked on sits upstream: being present, legible and unambiguous on the pages that keep coming back, and on your own site, of which the assistant most often opens only the homepage. What can be measured sits downstream, and has to be measured again, because one reading says nothing about the next.

And your business - what does the assistant take from it?

If the reading is steady and the answer is not, the only question worth asking becomes: today, on that question, are you in the answer or among the pages that were passed over? That is exactly what Waseit's free audit shows. It puts to ChatGPT and Gemini the questions your customers actually type, in your language and on your market, and returns in about 60 seconds a score out of 100, the competitors named instead of you question by question, the questions where you do not appear, and - when Gemini searched Google - the pages it actually consulted, which is to say where it would be worth being listed. Free, no account, no card; an email address only unlocks the competitors' names and the action plan. The audit page says it already, and this measurement confirms it on another axis: it is a dated snapshot, not a ranking, and we promise no position.

What this does not prove

• The measurement covers Perplexity alone, the only assistant whose capture keeps the pages consulted. For ChatGPT and Gemini that field is absent from every row: that is a collection shortcoming on our side, and nothing here says whether their reading is steady, or even whether they read.
• Three markets only - France, the UK, Brazil -, 39 cities, a single capture between 31 July and 16 October 2026, through the assistants' programming interfaces, which approximate the consumer apps without reproducing them.
• Two wordings are not every wording. A very different request, "a cheap plumber open this evening", was not tested here and would probably bring back other pages.
• We compare the pages the answer states it consulted. A page coming back does not prove it weighed on the choice of names, nor that no other source was used without being cited.
• The steadiness of the pages may come from the search engine queried underneath rather than from the assistant: two close wordings probably produce the same web query. We measure the outcome, not the mechanism.
• Our name extraction keeps at most eight businesses per answer, and sometimes takes a subheading for a name. That is why both matching bounds are published together, and why the finding is about the gap between pages and names, not about the exact value of 37.2%.
• Nothing here says a better-written page gets you named. We measured what is read again and what is named again, not what decides.