Why AI Visibility Breaks When You Change Language

Ask ChatGPT, in English, which companies make industrial dosing pumps for food production. Write down the names. Now ask the same question in German. You will not get a translation of the first list. You will get a different list, with different companies in it, and there is a fair chance your brand is in one and not the other.

People find this surprising. It is not a glitch, and it is not a translation problem you fix with a language switcher. It is a direct consequence of how these models learn, and if you export, it is probably the largest hole in your marketing measurement right now.

There is no global answer about your brand

It is tempting to imagine a model holding one settled opinion of your company, stored in its weights, which it then renders into whatever language you asked in. That is not what happens. What the model holds is the residue of everything it read, and what it read about you in German is not what it read about you in English.

Consider what the corpus for a language contains: trade press, distributor pages, forum threads where someone asks which supplier to use, review sites, comparison articles written by people nobody in the comparison paid. All of it written by people in a specific market, and dense in some languages, nearly empty in others. So when a model answers in Spanish, it leans on what the Spanish-speaking web says. Retrieval-based systems make this literal: they fetch Spanish-language pages and summarize them. If your presence in Spanish is a translated homepage and nothing else, the model is working from a corpus in which you barely exist. It will still answer confidently. It will simply name somebody else.

A caveat, because this gets oversold: knowledge does cross languages to some degree. Large models share internal representations, and a brand everybody writes about everywhere survives the jump intact. Nobody needs a German corpus to know what Bosch makes. The transfer is weak exactly where most exporters live: a company well known at home and in its industry, whose whole third-party footprint is a few dozen pages in one or two languages. For them, crossing a language boundary is closer to starting over than to being translated.

Translating your website moves less than you think

Your own site is one voice, and it has the weakest claim to objectivity in the whole corpus: everything on it is what you say about yourself. Models have read a great many corporate about-pages and treat them accordingly. What establishes that you exist in a language is what other people wrote about you in that language, in places where they had no reason to flatter you. Translation is the obvious move, and the one that changes least.

So the question is not "is our site available in German". It is: who, writing in German, has ever mentioned us?

You can check that this afternoon. Search your brand name with results filtered to the target language, excluding your own domain: "yourbrand" -site:yourdomain.com. Ignore the result counter and read what comes back. If it is your distributors reprinting your press release, that is your voice again wearing a different hat. If it is nothing at all, you now know what the model had to work with when it decided not to mention you.

The market where you are most invisible is the one you never check

Here is the structural reason this stays hidden. Marketing teams at exporting companies test their visibility in one of two languages: the language of head office, or English. Those are the languages the people doing the testing read. Nobody types a procurement question into ChatGPT in Polish, because nobody at head office thinks in Polish.

The blind spot, then, is not random. It is exactly the set of markets where you have revenue but no local press and no local community talking about you: by construction, the markets where an AI answer is least likely to include you, and the ones you are least likely to test. Your reporting says visibility is fine. Your reporting is reading the two languages where you were never the problem.

The asymmetry cuts both ways

This is not only a threat. In an export market, the local competitors a model recommends in your place are frequently much smaller than you: less revenue, a narrower product line. What they have is density in that language. A trade magazine profiled them, an association lists them, a forum argues about their products. The model was fed a market in which they are the obvious answer, because in the text it read, they are close to the only answer.

The gap between your real position in a country and your presence in the machine's answers about that country is the most actionable thing an exporter can measure. It is far more useful than one global visibility number, which averages away the exact information you need. Imagine a manufacturer with solid sales in France and no local content: the question worth asking is not its home-market score, it is why French buyers who ask a model get told about someone else while the sales figures say those buyers keep choosing it anyway. That discrepancy is a warning about the buyers who have not called yet.

How to measure it without fooling yourself

Three rules, and the second is where most attempts go wrong.

Ask what a buyer asks. Testing your brand name tells you nothing: hand a model a name and it will find something to say. The questions that matter come from someone who does not yet know you exist. Who supplies this component in Europe. What are the options at this company size. Which vendors meet this certification.

Reformulate for each market, do not translate. This is the step everyone skips. A literal translation of an English prompt produces a sentence no local buyer would type. Markets frame their categories differently: the product goes by a different name, the qualifier that matters is a local standard rather than a global one, the regulator a German buyer cites is not the one an Italian buyer cites. A prompt set pushed through a translation engine measures a question nobody asks, and hands you a clean, wrong answer.

Repeat, then compare. These models are probabilistic: ask twice, get two answers. A single result in a single language is an anecdote. What you want is the same reformulated prompt set, run across the platforms people use, in every language you sell in, repeated over time. The delta between markets is the number that matters, not the score in any one of them.

Where a tool becomes necessary

You can do all of this by hand, once, for one language. You cannot do it across dozens of prompts, four AI platforms and six markets, repeatedly, and still have a job. That is where it has to be automated, and it is what PSentry was built for: an AI Visibility Score computed per language and per market rather than globally, the local competitors named in your place in each one, and an export gap analysis that isolates the markets where you sell but the machines do not mention you.

The same caveats hold in every market, not just the ones you happen to check most. Twice a month, on a fixed calendar rather than the moment something changes, is how often a market gets read, because a language's standing inside these models does not move fast enough to reward watching it continuously. And in no market does the tool put a hand on what a model says, or guarantee that a gap you found closes just because you found it. It gives you a reading of where you stand in each market, which most exporters do not have in any form today.

The short version

Your AI visibility is not one number. It is one number per language, and those numbers can differ wildly. The web talks about you in some languages and is silent in others, the machines read exactly that, and they answer accordingly. Translating your own pages does not fix it, because you were never the source that counted. And the market where the gap is widest is, almost by definition, the one nobody has thought to check.

Frequently Asked Questions

If I publish German content on my site, will the German answers change?

It helps, but on its own it is usually not enough. Your site is one source, and a model weighs it as what it is: a company describing itself. Presence in a language is built mostly out of third parties, so the question that matters is whether anyone writing in German has a reason to mention you.

Does a language behave the same across the countries that share it?

Roughly, and not entirely. A model tends to draw on the same German-language sources whether the buyer is in Germany, Austria or Switzerland. But which suppliers and regulations get named can differ, because the model reflects the text it read. Treat a language as a starting unit, not as proof that every country using it looks the same.

Should I test English in markets where English is not the local language?

Yes. Technical and industrial buyers often research in English wherever they sit. English is not a substitute for the local language, but it is not irrelevant either. It is one more language in the set, not the set itself.

My prompts are already translated. Why is that not enough?

Because a translated prompt is your question in their words, not their question. Buyers elsewhere name the category differently and frame the comparison on different terms. If the prompt is not one a local buyer would plausibly type, the answer you measure is not the answer they see.