Translating Your Site Is Not Enough for AI Visibility

You paid for the translation. The site exists in five languages, the product pages read fluently in every one, the language switcher works. Then you ask a model, in German, which companies supply what you supply, and your name is not in the answer. You ask in Spanish, and it is not there either. In English it appears in the first sentence.

The instinct is to suspect the translation. Something must be off in the German copy. Almost always, nothing is off. The translation did exactly what a translation does. That is simply a much smaller thing than what you needed it to do.

A translated site makes you readable, not quoted

Consider what a model draws on when it decides which companies to name. Not your website by itself. Your website is one voice in the corpus, and the one with the most obvious interest in the outcome. Models have read an enormous number of pages in which a company explains why it is the leading provider of something, and they treat them as what they are: self-description. Useful for knowing what you sell. Worthless as evidence that anyone should recommend you.

What carries weight is what other people wrote about you. A trade journalist covering the category. A buyer in a forum thread asking who to go with, and strangers answering. A distributor catalogue. An association membership list. A comparison article by someone with no stake in your revenue.

Translating your site relocates your sentences into German. It does nothing to the German sentences written by other people, because those either exist or they do not, and no amount of work on your own domain conjures them into being. If nobody has written about you in German, the German corpus holds a company that describes itself well and that no one else has mentioned. The model answers from that, confidently, and names somebody else.

Every market runs on its own set of sources

What is easy to miss, if you have only ever measured your home language, is that the sources deciding the answer are not the same everywhere. In your domestic market you know which publications count, which forum the practitioners read, which directory a buyer checks before shortlisting. In an export market you have no map. There is a trade magazine half the industry reads and you have never heard of it. There is a national association whose member list gets reproduced across dozens of other sites, and you are not on it. There is a retailer whose category pages get pulled into retrieval constantly, and you do not stock with them. There is a forum where the practical questions get answered, bluntly, by people with no stake in your success.

None of this mattered to your English visibility. In German it decides it. A model answering a German question leans on German material, and that material is produced by institutions you are not part of. It is not a content problem you fix on your own site. It is an absence from someone else's territory.

The competitors change, and you will not recognise them

Run the same buying question in two languages and read who gets named. The lists overlap far less than you expect. In your export markets you will find companies you have never benchmarked against, that appear in no English answer.

They are not necessarily bigger than you. Often they are considerably smaller: one country, a narrower range, less revenue. What they have is density in that language. The trade magazine profiled them, the association lists them, the forum argues about their products. In the text the model absorbed for that market, they are close to the only available answer.

Which means your competitive analysis is watching the wrong race. Your competitor set is the home-market set, translated. The model's was assembled from what the local web says.

Two languages, two different questions

Now the mistake that quietly ruins most attempts to measure any of this. You have a prompt set that works in English, you push it through a translator, and now you have a German prompt set. It looks correct. Every word maps.

It is still the wrong instrument, because a buyer in Munich does not ask the English question in German. They ask their own question. The category often has a different everyday name, not a translation of yours at all. The qualifier that matters is a national standard rather than an international one. The comparison they want is against the incumbent in their country, not the incumbent in yours.

Feed a literal translation to a model and it will answer, cleanly and plausibly. You will get a score, and that score describes your visibility for a question nobody in that market types: a precise measurement of something that does not happen.

The fix is unglamorous. Ask someone who actually sells in that market to write the questions in their own words, without showing them the English list. Then compare what they wrote against your translation. The divergence is worth the exercise on its own: it shows how differently your category is framed somewhere you assumed you were selling the same product in another accent.

An average is a device for hiding its worst component

So stop looking at a single visibility number. A global score blends the market where you are the obvious answer with the market where you do not exist, and returns something in the middle that looks tolerable. Nobody experiences the average. A buyer in Spain experiences the Spanish answer, and it either has your name in it or it does not.

The reading that matters is market by market, side by side, with the question set localised for each. What you are looking for is a specific and rather painful shape: the markets where you have real revenue, real customers, a distributor, a sales target, and where the machines do not mention you. Those are not your weakest markets. They are frequently your best ones, which is what makes the gap dangerous. The buyers now asking a model are handed a list without you on it, and they will never call to tell you so.

That comparison is what PSentry is built to produce: a visibility score computed per language and per market rather than globally, the local competitors named in your place in each one, the trend for each market over time, and an export gap analysis that isolates the markets where you sell and the models stay silent. It runs across ChatGPT, Claude, Gemini and Perplexity, because those four do not agree with each other either.

What measurement does not do

Finding the gap does not close the gap. A dashboard showing that you are invisible in Dutch has told you something true and has changed nothing about the Dutch web. Closing it depends on third parties in that language having a reason to write about you: a journalist who covers you, a distributor who publishes a real page rather than a reprinted press release, a customer who answers a question in a forum. Some of that you can influence. None of it you can command, and anyone offering to guarantee it is describing a lever that does not exist.

Nor can anyone tell you how long it takes. It depends on how dense that market's web is, on how much of it the models absorbed and when, on whether the source that mentions you is one they retrieve at all. It is slower than a campaign and less predictable than a media buy. What you can have now is a reading of where you stand in each language you sell in, so that the market where you have quietly gone missing is a fact you know rather than a revenue line that erodes for reasons nobody at head office can explain.

Frequently Asked Questions

We publish original German content, not just translations. Why has nothing changed?

Because it is still your voice. Original or translated, content on your own domain is a company describing itself, and a model weighs it accordingly. The work is not wasted: it gives other people something to cite. On its own it is the same source with a bigger word count.

So should we not have translated the site?

Translate it. A page a buyer cannot read is a page that cannot convert. The point is not that translation is useless. It is that translation is necessary, nowhere near sufficient, and routinely treated as the whole job.

Which market should we look at first?

Not the biggest one. The one where the distance between your commercial reality and your machine visibility is widest: real orders, real customers, no mentions. That is where the risk concentrates, because you have something to lose and no early warning that you are losing it.

Can we use the English answers as a proxy for the rest?

No, and that assumption is what produces the surprise in the first place. Technical buyers do research in English in many countries, so English is worth measuring. It is one language in the set, not a stand-in for it.

If a model can translate, why does it not translate what it knows about us?

Because its knowledge is not a language-neutral fact about your company, rendered on demand into whatever language you used. It is the residue of text. Famous brands do cross language boundaries, because they are written about everywhere. A company whose entire third-party footprint is a handful of pages in one language has very little to carry across.