• AI visibility

Guide · AI search

Do AI engines cite local-language sources in other languages?

Mostly yes: when asked in another language, most AI engines switch to websites in that language, but how far they switch depends on the engine. In one 2025 study, ChatGPT and Perplexity sourced almost entirely from the local-language web, while Claude kept reusing English sites. For brands selling abroad, that means earning coverage in each market’s own media, not only translating the website.

By Underneath Team Guide Updated

The short version

  1. In a 2025 University of Toronto study of 100 buyer questions in five languages, ChatGPT and Perplexity drew almost entirely on local-language sites, while Claude reused English authority sites.
  2. In Tokyo, Gemini cited Japanese websites 68.4% of the time for Japanese questions and 6.4% for English questions about the same hotels.
  3. Hotel websites made up 11.0% of Gemini’s Japanese citations against 8.2% of English ones, which the authors link to richer Japanese content.
  4. Across twelve European languages, a 2026 study found that asking in a brand’s home language raised its recommendation share by 0.80 on a 0-to-1 scale for local champions, against 0.15 for global brands.
  5. Even within English, country matters: in our test, ChatGPT answers outside the US that cited a local website named local brands 45.0% of the time, against 3.1% when they cited none.

Do AI engines switch to local-language websites?

Most do, but to very different degrees. The engine you ask matters as much as the language.

Chen and colleagues (opens in a new tab) at the University of Toronto translated 100 English buyer questions, ten in each of ten consumer categories, into Chinese, Japanese, German, French and Spanish. They put them to Google, Gemini, Claude, ChatGPT and Perplexity and checked the language of every cited website. Their summary: “GPT and Perplexity heavily localize, sourcing almost entirely from the target language’s ecosystem. Claude, by contrast, reuses English-language authority domains across languages.”

Gemini sat in between. For German and Spanish questions, close to 50% of its cited websites were still in English. Across engines, Japanese and French questions produced the strongest switch to local sites.

How different are the sources in each language?

Very different for most engines: the same question in two languages often shares almost no cited websites. Claude is the exception.

In the Toronto study, Google’s overlap between the English and translated versions peaked at about 0.11, meaning around a tenth of websites were shared. ChatGPT’s overlap was near zero everywhere, as it moved to a different set of sites in each language. Gemini reached about 0.32 at best, for English and German, and Perplexity about 0.22, for German laptop questions. Claude showed much higher overlap in every category.

Brand lists moved less than sources. The authors found brand overlap “much higher than domain overlap”, especially in categories led by global brands such as cameras and laptops. Different sources can still lead to similar recommendations when a few brands dominate worldwide.

What happens in a real local market?

A study in Tokyo shows two almost separate web worlds, one English and one Japanese. Gemini drew on whichever matched the question’s language.

Zhu and Chang (opens in a new tab) asked Gemini 2.5 Flash 156 hotel questions about Tokyo in English and Japanese in March 2026, collecting 1,357 citations. Japanese questions cited Japanese websites 68.4% of the time; English questions only 6.4%. The kinds of sources differed too. Travel blogs were 22.9% of English non-booking-site citations but 1.2% of Japanese ones, while Japanese travel agencies took 12.9% and appeared not at all in English.

Hotels’ own websites did better in Japanese: 11.0% of Japanese citations against 8.2% of English ones. The authors note that Japanese hotel sites tend to carry deeper content, such as neighborhood guides and transit information, while English versions focus on booking. The study covers one city, one engine and one month, and it is correlational.

Does the number of sources change by language too?

Sometimes, and not in the same direction for every engine. Language can change how many citations an answer carries.

In an analysis of a public dataset by Zhang, He and Yao (opens in a new tab), ChatGPT averaged 7.77 citations for Chinese questions and 7.03 for English. Google’s AI search showed the opposite: 11.57 citations in English against 7.53 in Chinese. The authors conclude that language effects have to be measured engine by engine. Chinese-built AI models are a separate question again; see how brands fare in Chinese AI models.

Does the same apply between English-speaking countries?

Yes: even in English, the user’s country shifts the sources cited, and local sources go with local brands. Location matters, not just language.

In our country study, we asked ChatGPT and Gemini the same 40 English buyer questions from the US, UK, Canada and Australia on 28 September 2026. Two ChatGPT answers from the same country shared 0.534 of their cited websites; answers from different countries shared 0.324. Outside the US, ChatGPT answers citing at least one local website had a local-brand share of 45.0%, against 3.1% for answers citing none. This is an association: we did not change the sources to test cause.

What should you do about it?

Build credibility in each market’s own media and test in each language, rather than relying on translated pages. Concretely:

  1. List the sites each engine cites for your category in each target language and country.
  2. Earn coverage from respected local publishers and review sites, not just English trade press.
  3. Give local pages real substance, such as local guides, prices and service details, not just translations. Our guide to GEO across languages for global brands covers market-by-market planning.
  4. Keep English authority strong too, since some engines, such as Claude in one study, reuse English sources.
  5. Track AI answers in every language and country you sell in; an English-only view can mislead.

If you want help planning multilingual AI visibility, see our generative engine optimization service.

What does the research not tell us yet?

The research shows that engines localize differently, but not why, or how stable that is. The gaps:

  • Each study covers one period; engine behavior changes, and the 2025 Toronto results may already be out of date.
  • The language studies use buyer-ranking questions, hotel questions or brand questions, not every kind of query.
  • No study we reviewed tests whether earning local coverage causes more AI citations; the evidence is observational.
  • Coverage of languages is uneven: five major languages, Japanese hotels and twelve European languages.
  • One key multilingual study has an author affiliated with a monitoring company.

Frequently asked questions

Does ChatGPT use local sources when you ask in another language?

In one 2025 study, yes: it sourced almost entirely from the target language’s websites. Its cited sites barely overlapped between English and other languages.

Is translating my website enough for AI search abroad?

Probably not. The Toronto authors conclude that earned coverage in the local language is needed for engines that localize, such as ChatGPT and Perplexity.

Which AI engine relies most on English sources?

Claude, in the 2025 Toronto study. It reused English-language authority sites across all five languages tested.

Does asking from another country change AI answers even in English?

Yes. In our test, ChatGPT answers from the same country shared 0.534 of their cited websites, against 0.324 across countries.

Sources

Free strategy call

Some questions are easier to answer about your own business.

Bring the one that matters most. On a free 30-minute call we’ll take a first look at it and send you a short written read afterward.

Ask about my businessHow we work