Asked the same need twice, the assistant changed most of its list

Short answer: ask ChatGPT for the best plumbers in your city, then ask it for the most recommended plumbers in your city. Same need, different list. Across 682 pairs of answers it kept only 16 to 29 % of the businesses it had just named, and in a quarter to a half of the pairs it kept none of them at all.

You want to know whether the AI recommends you. The question assumes there is a list, and a place in that list. Our readings say otherwise: there is an answer, and it is attached to the exact sentence someone typed. Change the sentence without changing the need, and the same assistant, on the same day, for the same city and the same trade, answers largely with other businesses. One assistant stood apart, and it is the one that went and searched the web.

A place in the list is not a position. It is the answer to one sentence.

What was measured

Waseit's eight studies ask, in every city and every trade, two phrasings of the same need: "what are the best {trade} in {city}?" and then "list the most recommended {trade} in {city} and explain why". Until now we compared assistants with each other. This time we compare each assistant WITH ITSELF, across its own two answers. We kept the pairs where both answers name at least one business: 3,729 pairs, 8 countries, 89 cities, 148 country-trade pairs, captured between 31 July and 16 October 2026, each time in the language of the country. To compare the three assistants with each other we restricted the count to France, the United Kingdom and Brazil, the only countries where all three answered: 2,044 pairs, roughly 681 per assistant.

Matching business names has no obvious rule: "Poilane" and "Boulangerie Poilane" are the same bakery, "Table de l'Ours" and "Table de l'Olivier" are not. So we give two bounds rather than a single figure. The STRICT bound matches only names that are identical once normalised; it understates what was kept. The BROAD bound matches as soon as one distinctive word is shared, after setting aside trade words and city names; it over-matches on purpose. The truth sits between them, and we publish only what holds under both.

What survives the question being rephrased?

Most of it goes. ChatGPT keeps 16 to 29 % of the businesses in its first answer, Gemini 12 to 28 %. Put plainly, around three quarters of the list changes when the request changes wording without changing meaning. And often nothing at all remains: one rephrasing in four for ChatGPT under broad matching, one in two under strict matching. The business placed on top holds up even less: ChatGPT keeps it there in 11 to 19 % of pairs, Gemini in 6 to 11 %.

One assistant stands clearly apart, and it is the same one under both matching rules: Perplexity keeps 37 to 50 % of its list and puts the same business back on top three times more often than Gemini. It is also the only one of the three that fetched web pages before answering; ChatGPT and Gemini were queried through their programming interface, with no search. Between ChatGPT and Gemini, by contrast, the gap does not survive the change of matching rule (29 % against 28 % under the broad one): we do not separate them.

What it means if you are the business

A single check proves nothing. If you type your trade and your city into ChatGPT once and your name comes out, you have learned something true and very narrow: that sentence, that day, surfaced you. The neighbouring sentence, which means the same thing to your customer, may leave you out. That is why a position is not what gets measured: a frequency is, across several ways of asking. It also explains the competitor who swears he "always comes out first": he remembers the wording that surfaces him. We showed earlier that two assistants rarely answer the same thing; the same assistant, asked twice, does not answer itself either.

And your business - what do the assistants say?

That is exactly what Waseit's free audit does: it does not ask one question, it asks eight, the ones your customers really type, in their language and for your city, to ChatGPT and to Gemini, plus a notoriety test that separates "the AI does not know you" from "the AI knows you and does not recommend you". In about 60 seconds you see a score out of 100, the competitors cited instead of you, the questions where you do not appear, and which assistant actually searched the web rather than answering from memory. Free, with 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: this article is precisely about how much the answers move, and we promise no position.

What this does not prove

• Matching business names has no single rule, which is why we give two bounds. The broad one over-matches on purpose: it joins "Table de l'Ours" and "Table de l'Olivier" because they share a word. The strict one separates "Poilane" from "Boulangerie Poilane". The real figure lies between them, and we kept only the findings that are true under both.
• Under broad matching, ChatGPT and Gemini are within a point of each other: we do not separate them, and this article does not say which of the two is steadier. Only Perplexity's gap with the other two holds under both bounds.
• Our stored answers keep only the first eight businesses named. So we are comparing short lists, not everything the assistant could have cited: two answers might share a ninth name we never see.
• Two phrasings only, each asked once, on a single capture between 31 July and 16 October 2026. This is not a measurement of what changes from one day to the next, which is a different question.
• Perplexity answered only in France, the United Kingdom and Brazil in this capture, so the comparison between the three assistants is limited to those countries. And the fact that it searched the web goes together with its steadiness, without our having shown that it is the cause.
• We counted changes of names, not mistakes. Nothing here says one of the two lists is better than the other, or that it matches the businesses actually serving the city. Automatic name extraction can also take a criterion or a platform for a business, as our earlier editions measured.