The same question, 24 hours later: an engine that named nobody now names four practices

Yesterday we measured the disagreement between the three assistants, at the same minute. That left the question a business owner actually asks: if I run the test again tomorrow, will it say the same thing?

This morning we replayed yesterday's six questions. Not rewritten: read back from the previous day's file and sent again as they were, to the same three engines. The only variable is time.

The one case that answers on its own

Yesterday, asked for the best dentists in Nice, Gemini named no practice at all. It answered with booking platforms and selection criteria. Today, on the same question, it names four: Cabinet Avident, Arenas Dentistes, Dentego, Clinadent.

Nothing changed at those practices in twenty-four hours. Nothing changed in the question. A Nice dentist who tested yesterday would have walked away convinced the AI names nobody in their trade; the same test today shows them four competitors.

A single test does not measure your visibility. It measures your visibility on that day.

How many names survive one night

We compared the lists name by name, by hand, merging spelling variants (“Victor Plomberie” and “Victor Plombier”, “Caroline Meunier Avocat” and “Cabinet Caroline Meunier”). Across the six cities:

Stability varies enormously by trade. For dentists in Nice, Perplexity kept seven names out of seven and only added one: an almost frozen list. For garages in Nantes, the same engine kept one name out of seven. Same engine, same day, two trades: a stable list on one side, a draw on the other.

When ChatGPT does name, it does not name local businesses

One detail is worth noting. On garages in Nantes, ChatGPT broke its silence today - but the names it gives are Bosch Car Service, Norauto, Midas, Feu Vert, Speedy. National chains, not Nantes garages.

That is consistent with what we measured yesterday: this engine does not recommend a local business, it points to categories and networks. For an independent garage in Nantes, being “absent from ChatGPT” therefore means very little - nobody like them is in there.

What it changes for measurement

This instability is not a flaw in our tools, it is the material itself. Engines that search the web re-read pages that change, and their answers follow. That is exactly why we measure several times rather than once, and why we smooth no variance: the series is the information, not noise to be removed.

And it is why we have never promised anyone a place in ChatGPT. On an engine where a third to a half of the names change overnight, promising a ranking would be promising the result of a draw.

What to do with this

• Conclude nothing from a single test. Neither “I am visible” nor “the AI names nobody”: you only know how it stood that morning.
• Look at the trend, not the point. A name recurring three times across five measurements beats a name that once appeared first.
• Tell stable trades from volatile ones. For dentists a list holds; for garages it recomposes. The useful measurement frequency is not the same.
• Do not judge your presence by ChatGPT alone: across our six cities it named no local business, two days running.

The limits: six questions, six cities, two days, one model per engine. Two points do not make a curve, and we do not claim to measure a trend with two measurements. What they do establish is solid: today's answer is not yesterday's, on strictly identical questions.