The Language You Use With AI Decides Which Brands Exist

Asked in Norwegian, ChatGPT named three Norwegian tools and no American ones. Asked in English from the same desk, it did the reverse. The language of a question now decides which companies exist in the answer.

Someone in Oslo asked ChatGPT, in Norwegian, to recommend accounting software for a freelancer. It named Fiken, Tripletex and Conta, three Norwegian products, and no American one. Minutes later, from the same connection, the same person asked the same question in English. The list now read FreshBooks, QuickBooks and Wave, and the Norwegian tools had gone.

Nothing about the buyer had changed. Same city, same category, same machine. The language of the question had selected the market, which is the finding of a controlled study posted to arXiv by a researcher in Tallinn who ran 234 such tests across four countries and six languages.

I run ESBO Ltd, a link building and brand mentions agency with a publisher network in 53 languages, so the result surprised me less than it should have. Clients find out the same way every time: a company that shows up reliably in English answers asks why German buyers never see it, and the honest reply is that German buyers are not asking in English. We started placing Spanish-language mentions of our own agency this year for the same reason.

The language of the question picks the market

The counts are lopsided. Where the question was asked in the official language of the country, a global brand appeared in one of 24 runs. Asked in English from Istanbul and from Tallinn, a local supplier appeared in none of 12.

The model was not short of knowledge. In two of the six English answers on the Estonian connection, it noted that the user was in Estonia and that local tax rules argued for a local system, then named five foreign products anyway. It knew. It did not say.

Earlier work by the same researcher put a size on the effect. Across 12,933 answers, the language of the question explained 26.5% of the variation in responses, and the brand identity explained 1.5%. The study has limits worth stating: one model, six runs per test, and an author who sells AI visibility measurement for a living and says so in the paper. Read it as a loud signal rather than settled law.

English behaves like a market of its own

These systems assemble answers from what they have read, which is sorted by language. An English question pulls from the English web, where the global brands have spent two decades accumulating coverage. A Norwegian question pulls from Norwegian pages, where the local tools are the ones being reviewed and compared.

English does not even act as a neutral default. Regarding the Turkish connection, it returned 0 Turkish suppliers, while every other language tested returned local results. A separate paper comparing Chinese and Western models gave the pattern a name, the existence gap: a brand documented in one language exists for the models that read that language and barely registers for the rest.

The visibility report is measuring the wrong country

For an American company that sells only to English speakers, none of this matters. For everyone else it changes what the monthly visibility report is counting. The study’s author put it plainly: a company selling into Germany or Norway that checks its AI visibility in English is checking a market it does not compete in, and its real rivals are invisible to its own instrument.

The reverse holds for local champions. They own the local-language answer and disappear the moment a buyer types in English, which is how procurement teams and international buyers tend to type. One answer per language, and your brand is in some of them. Most companies have never checked which.

Get named in the buyer’s language, then measure in it

The repair is the work I keep describing in these pages, done in a second language. Translating your website helps less than you hope, because what you publish about yourself is the weakest signal AI has, and the Estonian suppliers in the study already publish English documentation without that moving the English answer. What moves it is other people naming you in the buyer’s language: the local press, the local roundups, the local review platforms and the local podcasts whose transcripts the models read.

That sounds expensive and often is not. Our pricing data across 16,625 publishers puts the median sponsored article at $655 in Spanish across 1,642 sites, $480 in Polish and about $490 across Central and Eastern Europe, close to the $452 an American-audience English site charges and well under the $1,040 for a British one. The Nordic languages and German are the exception, with Swedish, Norwegian, Finnish and German sites all at $1,230 or more, and in our experience far fewer brands compete for any of these placements than for the English ones.

Then measure where you sell. Take the five questions a buyer would ask about your category, have them translated (machine translation will do for a first pass, and a native speaker should do the second) and run each one several times in each language. Several times matters: the study found the top recommendation changed across identical runs on four of six prompts, so one screenshot is one draw from a distribution. Log which brands appear in each language and which sources get cited, and you have a map of where your name needs to be printed next.

The person in Oslo got a competent answer in both languages. The difference was which companies were allowed to exist, decided by a choice nobody made deliberately: the language the question happened to be typed in. Your customers make that choice every day. If you only ever check in English, you have no idea what they see.

Asked in Norwegian, ChatGPT named three Norwegian tools and no American ones. Asked in English from the same desk, it did the reverse. The language of a question now decides which companies exist in the answer.

Someone in Oslo asked ChatGPT, in Norwegian, to recommend accounting software for a freelancer. It named Fiken, Tripletex and Conta, three Norwegian products, and no American one. Minutes later, from the same connection, the same person asked the same question in English. The list now read FreshBooks, QuickBooks and Wave, and the Norwegian tools had gone.

Nothing about the buyer had changed. Same city, same category, same machine. The language of the question had selected the market, which is the finding of a controlled study posted to arXiv by a researcher in Tallinn who ran 234 such tests across four countries and six languages.