The same question in another language: 72% of the businesses named appear on only one side
Measured on 15 August 2026 across six cities, in French, Portuguese and English. Re-asking in the same language returns 61% of the names; asking in another language, 20%.
In this article
This morning we asked ChatGPT for the best accountants in Belo Horizonte. In Portuguese it named nobody - it answered with questions of its own: for an individual or a company? what kind of service? At the same minute, the same question in English produced an immediate list: PwC Brasil, Deloitte Brasil, EY Brasil, KPMG Brasil, Grant Thornton Brasil.
Two answers, one city, one engine, one instant. Which one describes accountancy in Belo Horizonte? And more to the point: if you run a firm there, which of the two are you looking for yourself in?
Why a difference on its own proves nothing
Yesterday we published a result that makes today's harder: the same question, to the same engine, twenty-four hours later, does not return the same list. An engine already disagrees with itself. Noticing that French and English diverge therefore says nothing until you know how far an engine diverges from itself.
That control is what this measurement is really about. Every question was asked twice in each language, back to back. The two repeats give the noise floor. The comparison across languages only means something above that floor.
Without that control, we would have published noise and called it a language effect.
What we measured, on 15 August 2026
Six trade-and-city pairs, matched across two markets: plumbers, accountants and bakeries in Lyon, Nantes and Bordeaux; the same three trades in São Paulo, Belo Horizonte and Curitiba. Each question was asked in the country's language and in English, twice each, to all three assistants - 72 answers. The questions are the ones a paid audit asks, not questions written for the occasion.
- Same language, question re-asked: 61% of names come back. Six in ten. That is the noise floor.
- Language changed: 20%. Two in ten.
- A 41-point gap. Crossing the language moves you far further than re-asking does.
- Of 123 businesses named in total, 35 appear in both languages - and 88 in only one, or 72%.
Which engines does this affect?
The headline number hides the important part: the effect is not shared equally. Perplexity keeps 69% of its names when the question is re-asked, and 5% when the language changes. Gemini goes from 45% to 25%. With ChatGPT the language gap is not distinguishable from noise - but that engine names so few local businesses that the calculation rests on a handful of cases, so we flag it rather than count it.
That split matches what we have been measuring for ten days: the effect belongs to the engines that go and search the web. An engine answering from memory answers much the same in either language. An engine that searches does not search the same place depending on the language - and so brings back something else.
The clearest case: plumbers in Lyon
Asked in French for the best plumbers in Lyon, Perplexity names Plomberie Roche Lyon and Mb Plomberie. Asked in English, at the same minute: AS DEPANN, Abatir, APAMS Plomberie, Plomberie Paul.
These are not international firms displacing local ones. They are Lyon plumbers on both sides - just not the same ones. Meanwhile, re-asking in French returns two thirds of them. Same engine, same city, same trade, same instant: two disjoint lists of real plumbers.
Has anyone else seen this?
Yes, and it deserves saying. On 12 February 2026 the vendor Peec AI published an analysis of its own data - over 10 million prompts and 20 million background searches triggered by ChatGPT: 43% of those searches run on the English-language web while the original question was asked in another language. Turkish switches to English most often (94%), Spanish least (66%), and no non-English language falls below 60%. Take it for what it is: a vendor's statement about its own data, with no published methodology or dataset, so unverifiable from outside. But it describes exactly the mechanism our numbers make visible on the results side.
BrandGEO Global published a Paris test on 1 August 2026, collected on 10 July: in wealth management the French-language answers named independent French boutiques, the English-language ones major international private banks; one firm appeared in three of the four French answers and none of the four English ones. The denominator, as the source gives it: four categories, one question per language, four engines returning usable data that day, ChatGPT failing on every Paris prompt. That is a one-off signal, not a series - which is exactly the gap we set out to close with a control.
Finally, a model maker says it plainly. The Claude Opus 5 system card, published by Anthropic on 24 July 2026, states: « Claude is multilingual, typically responding in the same language as the user's input. Output quality varies by language. » Output quality varies by language. The sentence says nothing about which businesses get named, and we do not stretch it further.
Why this matters now
On 22 July 2026 Google announced on its France blog, signed by Sébastien Missoffe, the launch of AI Overviews and AI Mode in Search in France, rolling out « dès aujourd'hui sur mobile et ordinateur, ainsi que dans l'application Google sur Android et iOS ». The French generative surface is three weeks old. If you serve a non-English market, the question of which language you measure in stopped being theoretical this summer.
One detail we checked before concluding
Two of our six native-language questions were grammatically wrong: our templates agree in the masculine, which produces « les meilleurs boulangeries » and « os melhores padarias ». The file holding those templates asserts that a badly written question gets a worse answer from every assistant. If that is true, those two cases were measuring a grammar slip, not a language.
So we re-asked those two questions with correct agreement and nothing else changed: a 4.4-point difference, against 48% overlap when the same question is simply re-asked. That is noise. The belief is not confirmed, the two cases are not disqualified - and we did not change the templates, because a mistake that grates on reading is not sufficient reason to move every customer's measurement.
What to do with this
Measure in your customers' language.
A tool that queries the assistants in English to judge your visibility in France or Brazil is not measuring your market - it is measuring a different one, where your competitors are not the same.
Conclude nothing from an English-only test.
Neither « I am visible » nor « nobody in my trade gets named ». All you know is what English says about your city.
If some of your customers search in English, that is two grounds, not one.
In Nantes the English question surfaces PwC Nantes, Groupe Secob and Fiteco; the French one surfaces Cerfrance, Capeos and One Ace. A firm can own one and be absent from the other.
None of these numbers promises a ranking.
On material where re-asking the same question already changes four names in ten, promising a position would be promising the result of a draw. We do not do that.
The limits, as always
Six trade-and-city pairs, a single morning, three assistants: this is a poll, not a census. We compare the first eight names in each answer, which is what our extractor keeps. The lists were read by hand, and the count follows through: of 334 strings extracted, 154 were not business names at all - section headings, selection criteria, advice - and 23 were directories. That left 157 names, 4 of which were dropped on a second reading. The 153 surviving spellings resolve to 123 distinct businesses once variants of the same establishment are merged - without which « Aquarius » and « Panificadora Aquarius » would have counted as two Curitiba bakeries.
That cleanup does not flatter our conclusion; it strengthens it. Before filtering, the noise floor sat at 32% and the gap at 21 points. Removing the noise is what widened the gap - noise, after all, speaks no language.