AI visibility reports can overlook local businesses simply because the test prompt doesn't match how actual customers ask questions. Even when the product category and physical location are identical, the language of the prompt can shift which companies appear in the response.

A small study of ChatGPT illustrates this clearly. Using a single connection in Tallinn on 30 August 2026, a researcher asked for freelancer accounting software recommendations in Estonian, Russian, and English. Each language was tested six times in logged-out web sessions.

The resulting supplier lists varied by language:

  • Estonian prompts: An Estonian supplier appeared in all six responses. Global suppliers appeared in one.
  • Russian prompts: An Estonian supplier appeared in four of the six responses. Global suppliers appeared in all six.
  • English prompts: No Estonian suppliers appeared in any of the six responses. Global suppliers appeared in all six.

These figures only reflect this specific set of responses; they are not market share estimates or predictions of future model behavior. However, the takeaway for testing is clear: in this sample, an English-only audit would have shown a completely different competitive landscape than an Estonian-language audit.

Language is a test variable

The paper, titled The Language of the Question Selects the Market, summarizes what happened in the accounting software test, though it shouldn't be taken as a blanket rule that language always overrides location.

By keeping the Tallinn connection constant, the researcher isolated prompt language from IP location. Still, looking at the final answers doesn't tell us why the model behaved this way. ChatGPT might be pulling from different retrieval sources, inferring different market needs, or relying on different training associations depending on the language. The outputs alone can't separate those internal factors.

The study also includes a helpful counterexample: when the researcher ran a parallel test for project-management software, the language effect disappeared. Categories dominated by international tools can behave differently than categories tied to local tax and compliance rules. Instead of assuming every English-language report is flawed, you have to test how your specific category behaves.

Keep buying intent consistent across translations

To test your own market, start with a realistic customer question for your product or service. Include practical constraints—like local compliance, team size, or specific integrations—only if a real buyer would naturally include them.

Next, have a fluent speaker translate that underlying buying intent into your other target languages. Avoid literal keyword translations, which can warp the meaning or sound unnatural. Also make sure the geographic scope stays consistent: if one prompt asks for tools suited for Estonia and another just asks for popular tools, you are changing the market instructions, not just the language.

Document the exact prompts alongside the test conditions: the AI model, search mode, date, IP location, and any geographic terms in the prompt. A generic metric like "AI visibility in Estonia" hides too many variables to be useful without this context.

Run each language version multiple times in fresh sessions. While six runs were enough to show a stark contrast in the Tallinn test, six isn't a magic number for every audit. If you need a stable baseline for decision-making, scale your sample size based on how much the answers vary and how small a difference you are trying to measure.

Look at the full response, not just brand mentions

Before boiling the results down to a simple count of brand mentions, read the actual text of the responses. Note which local businesses show up, which global competitors replace them, and whether the cited sources actually address the user's need.

Even when a brand shows up across multiple languages, its framing can shift. One language might get an accurate summary of your service, while another drops a key local feature. Similarly, if your brand is missing in one language, it might mean the model interpreted the user's intent differently—not necessarily that a crawler failed to read your site.

If you find a gap between languages, look closely at the sources the AI cited and what information exists about the missing suppliers in that language. That can point you toward real content or discovery gaps. Just keep in mind that this study didn't test remedies; it does not prove that simply translating your website will get you recommended more often.

Finally, if your tests show the exact same suppliers across languages, accept that finding. The language effect might not apply to your software category or setup. The primary goal here is to make sure your tracking reflects how real customers ask questions—deciding what to change on your website comes later.