“Cheap”, “urgent”, “the best”: five ways to ask, five different answers

Your customers do not all ask the same question. One asks for “the best plumber in Lyon”. Another, worried about the bill, asks for “a cheap plumber”. A third has water coming through the ceiling and asks “who can come right now”. Three customers, same trade, same city, same day. Do they get the same answer?

We measured it this morning, 11 August 2026. One assistant - Gemini, with web search - one city, one trade, and one thing changing: how the question is phrased. Five intents: the best, cheap, urgent, best reviews, a careful project. Then the same for dentists in Montpellier.

The result, checked name by name: no business appears across all five phrasings. Not in either city.

The question decides who gets named - and sometimes, whether anyone does

The striking part is not that the lists differ: it is that on some phrasings, the assistant names nobody. Asked “which cheap plumber in Lyon?”, it cites no plumber at all. It explains how to judge a price - insist on a quote, be wary of leaflets, check the registration - and points to platforms. Same on the dental side: “cheap” produces no practice, only the notion of a state-tariff practitioner.

On the price question, no business wins - because the AI refuses to name anyone.

Conversely, “urgent” produces an entirely separate list: in Lyon, names that appear on no other phrasing, phone numbers included. In Montpellier, urgency surfaces dental centres that “the best dentist” never cited - that latter question having named almost no practice at all, listing selection criteria and a directory instead.

What survives across questions is not a business

One detail reframes the whole measurement. The names that recur most from one phrasing to the next are not plumbers or dentists: they are platforms. In Lyon, one review directory appears on four intents out of five, another on three. In Montpellier, the appointment platform appears on three. Actual businesses cap out lower: the best-placed Lyon plumber appears on three phrasings, and no Montpellier practice exceeds two.

It is the same lesson as our three national studies, seen from another angle: what travels across questions are the intermediaries. Being present on the platform the assistant cites whatever the phrasing puts you in the answer through the back door - often more reliably than hoping to be named directly.

Why a rephrasing changes so much

This is not a whim of the model. Modern answer engines do not process your question as written: they break it into several sub-questions run in parallel, then synthesise. Google's AI Mode is reported to fire nine to eleven of them for a complex question, ChatGPT two to three, according to analyses published this year. Change one word - “cheap” instead of “the best” - and you change the whole cluster of sub-questions, therefore the pages read, therefore the names cited.

An honesty note, because it nearly cost us a false figure. Our first automated count claimed 85 to 94% of businesses appeared on a single intent only. Reading the answers one by one, that number collapsed: the counter treated “Plombier 7” and “PLOMBIER 7” as two businesses, and caught fragments of advice (“quote required”) as names. What remains - no business across all five phrasings, nobody named on price - was verified by hand. We would rather publish the verified result than the spectacular one.

What should you do tomorrow morning?

• Stop testing a single question. “Best X in Y” is the least representative phrasing: it is the one where assistants most often answer with criteria rather than names.
• Look at what you are missing per intent. If you do emergency work, the question that matters is “urgent”, not “the best” - and it has its own set of winners.
• Be present on the platforms that do travel across every phrasing: in our Lyon measurements, one review directory appeared four times out of five. Being listed there puts you in the answer even when you are not named.
• Do not expect to win the price question: across both cities, no professional was named on it. Spend your energy on the phrasings where the AI actually cites businesses.

Two cities, two trades, ten questions: that is a sounding, not a census, and assistants' answers move from day to day - we measured it this week, only one name in six survives forty-eight hours. But the structure is clear: your visibility is not a single number, it is a map. You can be first on one question and absent from the next, without anything changing on your side - except how the customer phrased their need.