What your customers see when they ask in their own language

There is a measurement problem sitting underneath most AI visibility work, and it is easy to miss because the dashboard looks fine.

The dashboard is in English. The prompts behind it were written in English. The sources the assistant leaned on when it answered were, for the most part, English. And the brand being measured may well be visible in English, because English is the language its own marketing site is written in.

None of that tells you what a customer in Brno or Braga sees.

An assistant answering in German is not translating an English answer

This is the part that gets skipped, so it is worth being precise about.

When an assistant answers a question in German, it is not answering the English question and translating the result. It is answering a German question, drawing on German-language material, and producing a German answer. The retrieval is different. The sources are different. The competitor set that comes back is frequently different.

That means the whole chain diverges:

• Different corpus. German trade press, German directories, German forums, German review platforms. Your English case studies are not in it.
• Different sources cited. An answer in Portuguese tends to cite Portuguese-language pages. If nothing about you exists in Portuguese, there is nothing to cite.
• Different competitors. The strong local player who has no English presence at all does not appear in your English measurement. They appear in the answer your customer gets.
• Different phrasing. Which changes the answer, and is the subject of the next section.

So a score can be high in English and near zero in Czech, and both numbers can be correct. They are measurements of two different markets that happen to share a brand name.

Translating the prompt is not asking the question

The usual attempt at multilingual coverage is to take the English prompt set and machine-translate it. This is better than nothing and worse than it looks.

A translated prompt is a grammatically correct sentence that no native speaker would type.

• German compounds a great deal of what English spreads across several words, and the compound is what people search for. The literal translation of an English noun phrase is often not the word in use.
• Czech declines nouns, so the form of a city or category name in a question is not the form in your keyword list.
• Portuguese written for Portugal and Portuguese written for Brazil differ in vocabulary and register, and a customer in Lisbon notices immediately which one they are reading.
• Formality is a real variable. Business questions in German and French are frequently phrased formally; a translated prompt that lands informally is asking a different question of a model that is sensitive to register.
• Some categories are simply named differently. The local term for a trade, a professional service or a regulated activity is often not a translation of the English term but a separate concept with its own boundaries.

The result is a prompt set that measures how the assistant responds to slightly foreign-sounding questions. That is a real thing to measure. It is not the thing the buyer wanted.

Language is not the same as market

There is a second confusion layered on top, and it cuts the other way.

Ten languages do not mean ten markets. German is the working language of Germany, Austria and Switzerland, and those are three different buying environments — different regulators, different tax wording, different currency in one of them, different local competitors, different sources an assistant will reach for.

Handled properly, this is not a problem, it is a distinction:

• The article is written once per language. There is no Austrian edition of an argument.
• The market work — the prompts, the local sources, the competitor set, the commercial context — is per country.

Tools that collapse these two into one axis get it wrong in both directions. They either treat German as one market and miss Switzerland entirely, or they treat every country as needing its own translation and spend the budget three times for one piece of prose.

How the gap shows up in practice

The failure is quiet, which is what makes it expensive. Nothing breaks. There is no error message. The English number goes up and to the right, and the pipeline in the non-English markets stays flat, and the two facts never get connected because they live in different reports.

Typical shapes it takes:

• A brand is recommended by name for its category in English and is absent from the same question asked in the local language.
• A brand appears in the local language but is described generically, because the only material available in that language is a thin translated summary page.
• A brand appears with out-of-date or wrong details, because the source the assistant found in that language is a third-party directory nobody has updated.
• A brand loses to a smaller local competitor who publishes properly in the language and is therefore the only citable option.

The last one is the most common and the most annoying, because it is not a technology problem. It is a publishing problem with a technology symptom.

How to check it yourself, without buying anything

You do not need a tool to find out whether you have this problem. You need someone who speaks the language.

1. Write down the five questions a customer would actually type in that market. Have a native speaker write them, not a translator working from your English list.
2. Ask them of the assistants your customers use, in that language.
3. Read what comes back. Are you in it? Named, or described vaguely? Accurate?
4. Look at what the answer cited. Are those sources in the local language? Do you exist in any of them?
5. Note who else is in the answer. If a competitor is there and you are not, look at what they publish in that language.

That exercise takes an afternoon and it will tell you more than a dashboard that has only ever asked in English.

The honest limits of this argument

Two things are worth saying against the case above, because they are true.

First, some categories genuinely are English-first. If you sell developer tooling to an audience that reads English documentation by preference, an English measurement may be close enough to the truth. The question is whether that is a finding or an assumption.

Second, multilingual measurement is harder to do well and easier to do badly. A tool that asks bad questions in ten languages is not better than one that asks good questions in one. The thing worth paying for is not the number of languages on the box. It is whether the questions in each language were written by someone who would ask them.

What we take from it

Being visible in English is worth exactly as much as your English-speaking demand, and no more.

If your customers ask in their own language, that is the language the measurement has to be taken in — with prompts a native would use, sources in that language, and a competitor set drawn from that market rather than translated into it.

Anything else is a report about a market you do not sell in.