Does Your AI Visibility Change in Another Language?

Does your AI visibility change when someone asks in another language?

As of September 2026, yes, and by more than most teams would guess. Recent analysis shows that the same question asked in a different language returns a substantially different set of cited sources, and that a brand's apparent standing in AI answers shifts with the language of the query rather than staying fixed.

If you sell into more than one country and you only check your AI visibility in English, you are measuring one version of a picture that has several. That is the practical finding, and it has been measured at reasonable scale this year.

We should be careful about the evidence base, because it is new and partly vendor produced. This piece says what has actually been measured, who measured it, and what we think a B2B team should do about it now.

What was actually measured?

Two separate pieces of work, with different methods. The first is an analysis published by Profound in March 2026 covering 3.25 billion AI citations spanning 7 models, 14 countries and 5 domain classifications. The models included ChatGPT, Claude, Google AI Mode, Google AI Overviews, Google Gemini, Microsoft Copilot and Perplexity.

The second is an academic preprint posted to arXiv on June 22, 2026 by Dmitrij Zatuchin of Estonian Entrepreneurship University of Applied Sciences and Rankfor.AI, titled "The Language Blind Spot". It queried three assistants about 66 brands across 12 European languages, producing 35,640 responses.

Treat these differently. Profound is a vendor publishing analysis of its own data, which is worth reading with that in mind. The preprint has a stated method and sample but has not been through peer review. Neither is a settled result, and both point the same way.

How much does the source mix actually change?

Enough to change your conclusions. Profound's analysis found that Google AI Overviews showed a 22.5% social citation rate in Mexico, while in Arabic speaking markets the social platform mix inverted entirely, with Instagram reaching 29% of citations against 4.6% in English baselines.

ChatGPT moved in the opposite direction. Profound reports its social citation rate dropping from around 10% to between 3% and 5% outside English, and attributes this to its heavy reliance on Reddit, which accounted for between 51% and 76% of ChatGPT's social citations in every country studied.

That last figure is the one we find most useful. It says that an assistant's language behaviour is largely inherited from the language distribution of the sources it leans on, not from a deliberate language policy.

Does the brand's standing change, or just the sources?

Both, according to the preprint. Its central claim is that AI constructed reputation is language bound, with sentiment varying by language. It reports Germanic and English responses as the most critical, and Uralic and Baltic language responses as more positive.

Its most actionable finding concerns who benefits from switching. Moving from English to a brand's home language raised recommendation rates substantially for local champions, by 0.80, and barely at all for global multinationals, by 0.15.

Read that as a competitive asymmetry rather than a curiosity. If you are the strong local player in a market, you are more visible when buyers ask in their own language. If you are the global brand, your advantage is largest in English and does not travel automatically.

Which matters more, the language or the model?

The model, on the preprint's own evidence. It reports that response stability varied far more by model choice than by language choice, which is a genuinely important caveat and one the paper makes about its own findings.

So before you build a language strategy, check whether your variation is a language effect or a model effect. Those need different responses, and mistaking one for the other wastes a quarter.

The practical test is simple. Run the same question in the same language across three assistants. If the answers differ more than the same assistant differs across languages, your problem is model coverage, not localisation.

What does this mean for a B2B company selling into Europe?

That your monitoring is probably incomplete. The preprint's own conclusion is that English only AI reputation monitoring leaves a measurable language blind spot. If your sales team reports that prospects in one market arrive with strange ideas about your product, this is now a plausible explanation rather than an anecdote.

Start by checking rather than building. Take your ten most important questions, ask them in the two or three languages your main markets use, and read what comes back. That is an afternoon of work and it tells you whether you have a problem worth spending on.

Doing this properly is an extension of the visibility measurement we described in how visible are brands in AI answers, with language added as a second dimension.

Does publishing translated content fix it?

Only if the translated content is genuinely indexable and genuinely good, and this is where most sites fail before AI enters the picture. Google's guidance on multilingual sites, last updated on December 10, 2025, is explicit that it determines a page's language from the visible content, stating "We don't use any code-level language information such as lang attributes, or the URL."

Google also says it does not vary crawler source location to detect page variations, so locale variants have to be declared with hreflang annotations, country specific domains or explicit links. A site that serves language by browser detection alone is largely invisible in every language but one.

It further advises against automatically redirecting users based on perceived language preference, because that prevents users and search engines from reaching all versions. That single pattern breaks more multilingual sites than any content problem, and the mechanics are covered in our hreflang guide.

Which URL structure should you use?

Google names three workable options and one to avoid. Country specific domains give the clearest geotargeting but cost the most to run. Subdomains are easy to set up and can sit on different servers. Subdirectories are the lowest maintenance but tie you to one server location. URL parameters are not recommended, because the variants are difficult to segment.

For most B2B software companies we would take subdirectories, because the language versions then inherit the authority of one domain rather than splitting it across several. That matters more when citation is the goal than when local ranking is.

Decide this before you translate anything. Changing URL structure after you have three language versions is a migration, not an edit.

What should you not conclude from this research?

That translating your site will make AI assistants recommend you. Nothing in either piece of work shows that, and the preprint specifically found that switching query language raised recommendation rates without a corresponding shift in sentiment, meaning visibility moved and opinion did not.

Also resist the temptation to read the sentiment differences by language family as a claim about those markets. A model answering more critically in Germanic languages is a fact about the model's training and sources, not about German buyers.

And treat every specific percentage here as provisional. One is vendor analysis, one is an unreviewed preprint, both are months old in a field that changes quarterly, and neither has been replicated.

What would we actually do this quarter?

Three things, in order of cost. Check your top questions in your main markets' languages and write down what you find. Fix any automatic language redirect on your site, because it is cheap and it blocks everything downstream. Then decide whether translation is worth funding based on what the check showed, not on the research.

If you do translate, translate the pages that answer questions rather than the pages that describe your company. Answer shaped content is what gets retrieved, in every language, and marketing pages are not.

Keep the monitoring going afterwards. A language blind spot is not a one time fix, because the source mix behind each assistant keeps moving, which is the underlying reason the numbers differ at all. That ongoing measurement is covered in measuring AI search traffic.

What is the honest summary?

Language changes which sources AI assistants cite, and it changes how visible a brand appears, with local players gaining more from their home language than multinationals do. That much has now been measured twice, by different parties, with different methods.

How much it is worth to you is not settled, and anyone selling you a multilingual AI visibility programme on the back of these two studies is moving faster than the evidence. Check your own markets first. The check is cheap and the programme is not.

We build multilingual sites and the measurement around them for B2B companies, so we tend to see this as a structural problem before it is a content one. If you want a read on whether your site is even reachable in your other markets, we are happy to look. You can find us at phoenix.studio.