top of page

Multilingual Generative Engine Optimisation: Why Brands Found in English Can Go Missing in German AI Answers

Oct 28
6 min read

This analysis was produced by AI, TELL ME!, a full-cycle AEO & GEO agency with our own AI search monitoring platform.

Summary

A brand that shows up in every English answer about its category can be missing from the German answer to the same question, and multilingual generative engine optimisation starts from that gap. In an August 2026 test of accounting software, ChatGPT named global tools such as FreshBooks and Wave in the English runs and local suppliers in the German, Norwegian, Turkish and Estonian runs, and the six German runs from Berlin held no global tool. Project management software showed no such split in Berlin, so the risk depends on the category. The sections below cover which categories carry the risk, how far the research goes, why translation alone is not enough and how to measure German visibility.

Cover titled English vs German AI answers: for accounting software for freelancers asked from a Berlin connection, ChatGPT named FreshBooks and Wave in 6 of 6 English runs and Lexware, sevdesk and Papierkram in 6 of 6 German runs, and the same split appeared in Oslo, Istanbul and Tallinn. The gap is the case for multilingual generative engine optimisation.

Most checks of how LLMs (large language models) describe a brand run in English, because the team running them reads English. A brand named in an English ChatGPT answer gets filed as visible, and the German-speaking buyer who types the same question in German never enters the report. Studies from 2026 show that the two buyers can receive different shortlists from the same model. AI, TELL ME!'s Austrian ski brands study ran its prompts in German, so its shares describe the German shortlist only.

Query language changed which suppliers ChatGPT named

Query language changed which suppliers ChatGPT named, and the location of the connection changed which country counted as local. Dmitrij Żatuchin of the Estonian Entrepreneurship University of Applied Sciences reports this in an August 2026 preprint with 234 controlled runs against ChatGPT across four exit countries and six query languages. Asked for accounting software for freelancers from a Berlin connection, ChatGPT named FreshBooks and Wave in all six English runs. The six German runs named Lexware, sevdesk and Papierkram every time and named no global tool. The same split appeared in the other three countries, and FreshBooks was named in all six English runs in every one of them.

Connection (local language)

English answer

Local-language answer

Berlin (German)

FreshBooks and Wave in 6 of 6 runs; Lexware and sevdesk in 3 of 6

Lexware, sevdesk and Papierkram in 6 of 6 runs; no global tool

Oslo (Norwegian)

FreshBooks, QuickBooks and Wave in 6 of 6 runs; Fiken in 1 of 6

Fiken, Tripletex and Conta in 6 of 6 runs; no global tool

Istanbul (Turkish)

FreshBooks, QuickBooks and Xero in 6 of 6 runs; no local supplier

Paraşüt and Logo in 6 of 6 runs, Mikro in 3 of 6; no global tool

Tallinn (Estonian)

FreshBooks, QuickBooks and Wave in 6 of 6 runs; no local supplier

Merit Aktiva in 6 of 6 runs, SimplBooks and Directo in 5 of 6; Xero in 1 of 6

Across the four cells where the query language matched the country, a global supplier appeared in 1 of 24 runs. A Turkish question from a Berlin connection returned German suppliers, written in Turkish, which separates the effect of language from the effect of location. Russian asked from Tallinn landed in between, with an Estonian supplier in 4 of 6 runs and a global one in all six.

English-only monitoring shows the shortlist of one market

English-only monitoring shows the shortlist of one market. A June 2026 study by the same author put 66 brands from eleven European markets to three search-grounded models in twelve languages, 35,640 responses in all. Moving the question from English to a brand's home language raised the recommendation share of national champions by 0.80 on a scale from 0 to 1 and the share of global brands by 0.15. Sentiment moved far less.

The author rates German among the best-resourced web languages, so thin training data is an unlikely reason for a German gap. The brands with most to lose are those that sell into Germany without being German, a group neither study tested directly.

Which categories lose visibility in German?

In the Berlin runs, the category tied to national rules changed with the language and the category free of national obligations did not. Accounting software for freelancers depends on national VAT and e-invoicing rules, and its German answer replaced the global tools with German ones. Project management software carries no such obligation, and its answer stayed the same in both languages.

Category (Berlin connection)

English answer

German answer

Language effect

Accounting software for freelancers

FreshBooks and Wave in 6 of 6 runs; Lexware and sevdesk in 3 of 6

Lexware, sevdesk and Papierkram in 6 of 6 runs; no global tool

Global suppliers replaced by local ones

Project management software

ClickUp, Asana and Teamwork in 6 of 6 runs; no domestic supplier

ClickUp, Asana and Teamwork in 6 of 6 runs; no domestic supplier

None observed

Żatuchin reads the pattern as a regulation effect, because Germany has domestic project management suppliers and none was named in either language. Separating regulation from domestic supply would need a third category, which the test did not include. A German run for each category shows which case applies to a brand.

How far the evidence goes

The evidence shows that a language gap exists in at least one category and does not show how large it is across a market. The August test used one model, one prompt per category and six identical runs per cell. It did not control the time of day, and the provider does not disclose which model sits behind the logged-out interface.

Both of Żatuchin's studies are arXiv preprints, and the August paper does not report any peer review. The author discloses that his company, which sells AI visibility measurement, funded the data collection, so the findings need replication by someone else.

An ICML 2026 study by researchers at the University of Maryland and Johns Hopkins University offers independent support: it tested six open-weight models on eight languages besides English and found that models often prefer to cite the document written in the query's language, and that this preference sometimes outweighs relevance. The study used machine-translated documents and did not test German, so it offers a candidate mechanism for the August result and no proof of it.

Does translating the website close the gap?

A translated website falls short on its own, because engines differ in how much they draw on local-language sources. In a September 2025 paper, researchers at the University of Toronto translated 100 English queries into five languages including German and put them to four AI search engines. ChatGPT's cited domains barely overlapped between languages, which means it swapped one set of sources for another. Claude kept reusing English-language domains, and Gemini's citations sat closer to an even split between English and the language of the prompt.

The Toronto authors advise against relying on translation alone: coverage in authoritative German publications matters for the engines that localise, and top-tier English coverage matters for the engines that reuse English sources. Engines change their sourcing often, so each language needs a fresh test.

Multilingual generative engine optimisation in four steps

Multilingual generative engine optimisation repeats single-language work with language added as a variable at every step.

  1. Write a German prompt set in the category words German buyers use, with the prompt types from the guide to checking whether ChatGPT recommends a brand, for example "What is the best accounting software for freelancers in Germany?" next to "Welche Buchhaltungssoftware ist die beste für Freiberufler in Deutschland?".

  2. Run every prompt at least six times in each language on several engines, because the top recommendation changed across six identical runs on four of six prompts in Żatuchin's test, and the paper lists 20 to 30 runs per cell among the checks still to be done.

  3. Compare the German list with the English list and with the German competitor set, category by category.

  4. Close the gaps in the German record with German pages that state the offer plainly and with coverage in German publications, directories and review platforms, then re-run the same prompts.

AI, TELL ME! is a full-cycle AEO and GEO agency that sets up monitoring per market, with prompts in English, German, Arabic and other languages on request. Its Third-Party GEO service, which includes PR for GEO, covers the editorial coverage, directories and review platforms behind step 4.

FAQ

Why does a brand appear in English AI answers and not in German ones?

In the August 2026 test, ChatGPT named global accounting tools in English and local suppliers in German, Norwegian, Turkish and Estonian. The result shows the gap and leaves its cause untested, and German-language coverage is the part a brand can influence.

Does language or location decide which suppliers a model names?

Both act, on different things: the language of the question decides whether local suppliers appear, and the location of the connection decides which country's suppliers count as local.

Which categories are affected by the language gap?

Of the two categories tested in Berlin, accounting software changed completely and project management software did not. Accounting depends on national VAT and e-invoicing rules, which is the author's reading and not a proven cause.

Does translating a website fix German AI visibility?

A translated website does not fix German AI visibility on its own. In the Toronto study, ChatGPT swapped its source sites by language while Claude kept citing English-language domains, so German and English coverage each matter on different engines.

How should a brand monitor German AI visibility?

A brand needs a separate German prompt set in the category words German buyers use, run at least six times on each engine and compared with the English results.

AEO Monitoring from AI, TELL ME! is the free first step, and its prompts can run in German.

bottom of page