Prompt Language and Location Are Two Different Settings

Surely a question asked in German gets an answer about Germany. That is the objection most people raise when someone suggests that prompt language and location are two separate settings, and it sounds like common sense: the language is the market, so set the language and you have set the market. For a good part of the picture it holds. For the part that matters most to an exporter, it does not, and a small experiment published on arXiv at the end of August 2026 shows where the seam runs.

This piece takes the claim "language equals market" apart, one component at a time, and keeps only what survives.

The claim, stated fairly

The strong version goes like this. An assistant asked in Italian draws on the Italian-language web, names Italian suppliers and answers for an Italian buyer. Where the person sits when they type the question is noise. If you want to know how you look in Italy, you ask in Italian, and that is the whole method.

There is a solid reason people believe it. Language really does change the answer, and not by translating it. We have covered why in which language you should actually measure: each language is a different body of text, so it produces a different list of names. The mistake is the step that follows, from "language changes the answer" to "language is the only thing that changes it".

What the August probe changed, one setting at a time

The paper is arXiv 2608.30052, by Dmitrij Żatuchin, submitted on 30 August 2026, with data collected on 29 and 30 August. It is a preprint, not peer-reviewed, and the author discloses that his own company sells AI visibility measurement. Read it as a careful probe, not a verdict.

The design is what makes it useful. It sent 234 runs to ChatGPT, through the logged-out web interface and through the OpenAI API. It used six query languages (English, German, Norwegian, Turkish, Estonian and Russian) and four exit countries (Germany, Norway, Türkiye and Estonia). Each combination was run six times with identical wording. Two product categories were tested: accounting software for freelancers, which is bound to national VAT and invoicing rules, and project management software for a small marketing agency, which is not.

Because language and connection were varied separately, the paper can say which of the two moved what. Most informal tests cannot, because they change both at once without noticing.

Component one: language decides whether local suppliers appear

This part of the claim survives. In the regulated category, when the language of the question matched the country of the connection, a global brand took the top recommendation in only 1 of 24 runs. Local suppliers dominated. When the same connections in Estonia and Türkiye were used to ask in English, local suppliers won 0 of 6 runs in each country.

So the language of the prompt works like a switch. Ask in the local language and the assistant tries to localise. Ask in English and, at least from those two countries, it reaches for the global names.

One detail spoils the neat version. In English from Berlin and from Oslo, local suppliers did appear. The author reports the asymmetry and does not explain it. That is honest, and it is also a warning: even the component that survives does not behave the same way everywhere.

Component two: the connection decides which country

This is where the claim breaks. With the language held fixed and only the exit IP changed, the named suppliers changed country, while the answer stayed in the language of the question. A question in Turkish sent from Berlin named suppliers from the country of the connection. So did a question in Russian sent from Tallinn.

The paper sums it up in a section heading: language sets the register, location sets the market. Put more plainly: the language decides whether the assistant tries to answer locally, and the location decides which locality it picks.

For an exporter the consequence is direct. A Turkish-speaking buyer living in Germany, or a Russian-speaking one in Estonia, does not get the answer that a "Turkish" or "Russian" measurement would show. They get local suppliers for the place they are sitting in, described in their own language. A measurement that fixes the language and ignores the connection has measured someone who does not exist.

Component three: it depends on the category

The claim also assumes the effect is general. The second category says otherwise. Project management software has domestic suppliers in Germany and Estonia, and the paper names some, yet none of them was named in any language. The language effect simply did not appear.

The author's reading is that the deciding factor is whether a category is nationally regulated, not whether it has national suppliers. Accounting software has to follow each country's tax rules, so an answer that ignores the country is a wrong answer. A project tool works the same in Tallinn and in Berlin, so the assistant has no reason to localise.

For technical B2B this is worth thinking about before measuring. Products tied to national standards, certifications, invoicing formats or safety rules are the likeliest to show the effect. Products that cross borders unchanged may show little of it. That is a hypothesis from one paper, not a rule, and the only way to know for your category is to test it.

What the probe does not show

The limits are real, and the author lists most of them himself:

• One assistant only: ChatGPT. Claude, Gemini and Perplexity were not tested.
• One prompt per category and six runs per combination. That is enough to see dominance, not to measure small differences.
• The model behind the logged-out interface is not disclosed, and time of day was not controlled.
• The API arm had no geography at all: the paper treats it as having no user location, so the connection effect comes from the web interface.
• The web answers carried no citations, so the paper cannot say whether the shift came from retrieval or from the model itself.

None of this undoes the central result. It does mean the result is a strong hint about one assistant, and the other three need checking on their own terms.

Measuring prompt language and location as separate axes

If the two settings move answers in different ways, they have to be recorded separately. A useful grid has languages in one direction and connection countries in the other, and every cell states both. A row that says only "German" is missing half its label.

Through an API the location is an explicit parameter, which makes it easier to control than in a browser. OpenAI's web search tool accepts a user_location with an approximate country as a two-letter ISO code, plus optional city, region and timezone. A run that sets it and a run that does not are two different measurements, even with the same prompt. The trade-offs between the two routes are covered in API or browser scraping.

Our own numbers show the language axis on its own. Across 17,719 answers collected by PSentry between December 2025 and October 2026, the language of the prompt shifted where the cited sources came from: with Italian prompts, more than half the sources cited by ChatGPT, Gemini and Claude sat on .it domains, while with English prompts roughly three sources in four were on .com. The sample has a clear limit. The prompts are the ones our customers monitor, mostly B2B services and industry, so it is not representative of all searches, and it does not vary the connection, so it says nothing about the location axis. A separate piece will look at those domains in more detail.

The same honesty applies to the tool. PSentry scores visibility per language and market. It does not simulate where an individual buyer is connecting from. Its numbers describe the language axis well; for the location axis, the cleanest evidence is still a controlled test like the one in this paper, run on your own category.

What survives the dismantling is a narrower and more useful rule. Language decides whether the assistant looks for local suppliers. Location decides which country's suppliers it finds. Measure one without recording the other and you will not know which of the two moved your number.

What exporters ask before running this test

If I ask ChatGPT in English, does it matter which country I am connecting from?

In the August probe it did. English questions sent from Estonia and Türkiye brought up global names, while English from Berlin and Oslo still surfaced local suppliers, so the connection has to be recorded even for English runs.

Does the language effect apply to every product category?

Not in the probe. It showed clearly in accounting software, which is bound to national tax rules, and did not appear in project management software. Products tied to national standards, certifications or invoicing formats are the likeliest candidates, but that is a hypothesis to test in your own category.

Can I set the buyer's country when measuring through an API?

With OpenAI's web search tool, yes: it accepts a user_location with an approximate country as a two-letter ISO code, plus optional city, region and timezone. A run with it and a run without it are two different measurements, even with the same prompt.

Does PSentry measure where a buyer is connecting from?

No. PSentry scores visibility per language and market, but it does not simulate the location of an individual buyer. For the location axis, a controlled test on your own category remains the cleanest evidence.