ChatGPT Named Different Suppliers When the Question Changed Language

An AI visibility report can miss local suppliers because it asks a different question from the customer. The product category may be right and the country may be right, yet the language can change which businesses enter the answer.

A small study of ChatGPT recommendations makes that problem concrete. From one connection in Tallinn, the researcher asked about accounting software for freelancers in Estonian, Russian and English. Each language was tested six times in logged-out web sessions on 30 August 2026.

An Estonian supplier appeared in all six Estonian-language answers and none of the six English-language answers. Russian-language answers included an Estonian supplier in four of six runs. Global suppliers appeared in all English and Russian runs, and in one Estonian run.

These counts describe the observed answers. They are not market shares or reliable estimates of how often a supplier will appear next month. Still, the contrast is useful: an English-only audit would have presented a very different supplier landscape from the Estonian questions in this sample.

Language is part of the test conditions

The study's title, The Language of the Question Selects the Market, captures the pattern in this accounting comparison. It should not be read as a universal rule that language always overrides location.

Keeping the connection fixed helped the researcher compare language without simultaneously moving the request to another country. That does not identify the internal cause of the result. The interface could select different sources, interpret the market differently or draw on different learned associations. The observed answers alone do not separate those mechanisms.

There is also a useful counterexample in the same study: its project-management comparison did not show the language effect. A category with widely used international tools can behave differently from one where local requirements matter. That is a reason to test your market, not to assume every English-language report is wrong.

Preserve the buying question when you translate it

Begin with a question a customer might actually ask about a product or service you sell. Include requirements that change the choice, such as local compliance, team size or a necessary integration, when they genuinely belong in that question.

Then ask a fluent speaker to express the same buying intent in the other relevant language. A literal keyword translation can change the meaning or sound unlike a normal customer question. If one version asks for tools suitable for Estonia and another merely asks for popular tools, the comparison mixes language with a different market instruction.

Keep a record of the exact wording. Also record the AI product, search mode, date, connection location and any location stated in the question. These details make a result inspectable. A score labeled only “AI visibility in Estonia” hides too much of the test.

Repeat each version in fresh sessions. Six repetitions were enough to describe the small study's contrast; they do not establish a generally sufficient sample size. For decisions that depend on a stable estimate, the amount of repetition should reflect the variation you observe and the difference you need to detect.

Compare suppliers and sources, not just your position

Read the full answers before reducing them to a brand count. Which local businesses appeared? Which international alternatives replaced them? Did the answer cite local material, and did that material actually concern the customer need?

A brand appearing in both languages can still be represented differently. One answer might describe a service accurately while another omits a local capability. Conversely, a missing brand can reflect a different interpretation of the question rather than a technical failure to read its website.

If your own results differ, inspect the sources and the information available about the missing suppliers. That investigation can reveal a content or discovery question worth pursuing. The experiment did not test whether translating a website causes more recommendations, so translation is not an established remedy for the observed gap.

If the supplier sets are similar, keep that result. The suspected language problem may be absent in your category or setup. The first job of this comparison is to establish whether your report represents the questions customers ask; deciding what to change on the website comes afterward.