Obility Editorial · · 4 min read
An English check can describe a market your buyers never see
Most AI visibility checks are run in English, often by someone in a head office. For a company that sells into Germany, the Nordics or the Baltics, that check may describe a shortlist its buyers never see. Two studies by the same researcher, both published on arXiv this year, changed only the language of a buying question and measured what happened. In commercial categories, the language changed which companies were named.
The local language can replace the whole shortlist
The larger study, published in June 2026 by Dmitrij Żatuchin, queried GPT-5.4, Gemini 3.1 Pro and Perplexity Sonar Pro through their APIs about 66 brands from eleven Northern, Baltic and Central European markets, in twelve languages, for 35,640 grounded answers. When a question asking for the leading companies in an industry and market moved from English into a brand’s home language, the recommendation share of local champions rose by 0.80 on a scale of 0 to 1. Global multinationals gained 0.15. How the brands were described changed far less than whether they were named at all.
The August 2026 follow-up went to the logged-out ChatGPT web interface, which is closer to what a buyer uses, across 234 runs. Asked in Norwegian from an Oslo connection for accounting software for freelancers, it named Fiken, Tripletex and Conta in all six runs and no global product. Asked in English from the same connection, it named FreshBooks, QuickBooks and Wave in all six, and Fiken in one. In the four cells where the question was asked in the country’s official language, a global supplier appeared in 1 of 24 runs. Asked in English from Istanbul and Tallinn, no local supplier appeared in any of 12 runs.
Language decides whether to localise, and location decides where
The August study separated two things teams usually file together under “international”. With the language held at Turkish and the connection moved from Istanbul to Berlin, the answers stayed in Turkish but named German suppliers, with Lexware and sevdesk each in five of six runs. Russian asked from Tallinn named Estonian suppliers and no Russian ones. The paper’s reading is that the language of the question decides whether the answer localises at all, and the buyer’s location decides which country it localises to. English did not act as a neutral default. On every connection it produced a mostly global shortlist.
The local knowledge was often there. On the Estonian connection, two of six English answers said that national VAT and e-invoicing rules favoured a locally focused system, then listed only foreign products.
The effect depends on the category
The same method found no language effect in project management software. ClickUp, Asana and Teamwork were named in every run asked in English, German, Norwegian or Turkish, and no domestic supplier appeared in any language, although Germany has several. The author expected the opposite and reports the result as a failed prediction. The explanation the data supports is regulation: freelance accounting is bound to national tax and invoicing rules, while project management is not.
That gives a working rule. If your category is shaped by national law, standards, payment habits or procurement rules, expect the shortlist to change by language. If a buyer anywhere can use the same tool, expect less change, and check rather than assume.
Build the prompt set per market, in the buyer’s language
Write the buying questions in each language your buyers use, with someone who knows how buyers in that market phrase them, rather than by translating the English set word for word. When measuring a consumer interface, run each set from a connection in that market. An API call carries no buyer location, and the August study found its API arm, asked in English, answered as if the buyer were in the United States. Record the language and location with every result, so a German result is never averaged into an English one.
Read the competitor set per market as well. A domestic rival who never appears in your English checks can be the default answer in the local language, and an English report will not show it. Neither study tested what moves a brand into a local answer. One result is a caution against assuming more English content will do it: Estonian accounting tools publish English documentation for international customers, and the English answers still named none of them. Treat any fix as a hypothesis to test in the language where you are missing.
Use the direction, and hold the exact figures loosely
Both papers come from one researcher who runs Rankfor.AI, a company that sells AI visibility measurement, and both declare it. The August study used six runs per cell, one prompt per category and one undisclosed model behind the interface, over two days. The June study used APIs with an instruction to answer in the target language and classified brands as local or global by hand. Neither identifies the mechanism, and retrieval of local language sources is one explanation the August paper cannot rule out. The direction is consistent across both. The sizes belong to those models and those dates.
In Obility, Answer Engine Insights groups prompts by audience, buying stage and market, and each observation keeps its language, geography, prompt wording and collection time, so results from different markets are compared like with like instead of blended into one number. Which engines and markets are measured is agreed with each team at the demo.