Multilingual SEO for AI Search: What 7 Studies Show by Market

AI search engines do not treat every language and country the same way. ChatGPT often runs part of its background research in English even when the user writes in German, Polish or Turkish, while Google's AI Overviews and AI Mode stay closer to the query's language and lean on Google's existing international signals such as separate language URLs and hreflang. So multilingual SEO for AI search is less about one global trick and more about being retrievable in the local language, being described correctly in the English sources ChatGPT also reads, and being present on the local sites each engine cites in each market.

This guide pulls together the seven studies and tests that measure this behavior (with market, sample and caveats for each), separates what Google, OpenAI and Anthropic actually document from what vendors have observed, and turns it into decisions: which URL structure to use, what to translate first, where to build local mentions, and how to track visibility per country and language. Every source was checked on October 24, 2026.

How do AI search engines handle prompts in other languages?

Each engine combines two things: the language of the prompt and the user's approximate location. How much weight each gets, and which index the engine searches, differs by engine, so the same German prompt can produce German sources on one engine and English ones on another.

What the vendors document is thin. Google says its AI features may use "query fan-out", issuing "multiple related searches across subtopics and data sources", and that a supporting link must be "indexed and eligible to be shown in Google Search with a snippet" (Google Search Central, AI features and your website). Google's AI Overviews reached more than 200 countries and territories and more than 40 languages in May 2025 (Google), and AI Mode added more than 35 languages and over 40 countries and territories in October 2025 (Google). At Google I/O in May 2026, Search head Liz Reid said AI Mode is "more multilingual by design" and that its grounding considers which content may be more useful for a user's location (Search Engine Journal).

Anthropic says consumer Claude "may use your IP address to determine coarse-grained location (city/region level)" for features like web search, and lets users turn this off (Anthropic Privacy Center). The developer APIs for both Claude and OpenAI accept an approximate user_location (city, region, country, timezone) to localize web search results (Claude web search tool docs, OpenAI web search docs). Neither company documents how the language of a rewritten search query is chosen, and no AI vendor documents an "AI hreflang".

The table below combines what each vendor documents with what independent tests have observed. Treat the observed column as directional: most of it comes from small manual tests or vendor datasets.

Two of these findings look contradictory and are not. Perplexity cites a high share of local German sources for German commercial prompts, yet returned English versions of translated pages in Gabe's test. The first measures whether an engine finds local businesses and publishers; the second measures whether it picks the right language version of one page that exists in several languages. A brand can pass one test and fail the other.

The evidence: what seven studies found

There are seven pieces of work worth reading on this topic, and every one of them has limits. Six come from vendors that sell AI visibility or translation software, and only OtterlyAI's two indexes compare markets with one method. The table lists each with its sample and its main caveat so you can weigh it yourself.

Two numbers in the field need care. The Weglot study's widely quoted "431%" and "327%" figures cannot be reproduced from the citation counts on its own page, which is why this guide quotes the raw counts instead. And Peec's 43% and 78% use different denominators (fan-out steps versus prompt runs), so they are not two measurements of the same thing.

Why does ChatGPT search in English, and what does it mean for you?

Nobody outside OpenAI knows why. OpenAI documents that ChatGPT rewrites prompts into search queries but not how it picks their language. Peec's own explanation is that English pages carry more authority signals and that searching the larger English web lowers the risk of finding nothing useful; Peec did not test either idea. Whatever the cause, the effect is measurable: Peec's examples include a German software prompt that returned no German companies and a Polish marketplace prompt where Allegro.pl, Poland's dominant marketplace, was missing while eBay appeared. If you want the mechanics of how one prompt becomes many searches, see our explainer on query fan-out and our breakdown of how ChatGPT search works.

If you are a local brand in a non-English market

Your local-language pages are only part of ChatGPT's retrieval pool. You also need to be described accurately in the English-language sources that rank for your category: English listicles and comparisons that name local providers, an English version of your key product and pricing pages, and English-language coverage in trade or international press. Peec's own recommendation is to compare your English-language presence with your competitors' and to create English versions of key pages. You do not need to translate everything into English; you need the pages that answer "best X in [country]" questions to exist in a form English-language fan-outs can find.

If you are an English-first brand entering a market

The English bias works in your favor on ChatGPT, which is exactly why it is a trap. Your English pages may get you mentioned in German ChatGPT answers without any German content, while Google's AI features, Copilot and Perplexity favor local sources more strongly. Measure each engine separately before concluding that a market does not need localization.

Do AI engines prefer ccTLDs, subfolders or .com domains?

Local country domains do take a large share of local citations, but a localized section of a strong .com competes too. OtterlyAI's indexes are the only large dataset on this question, and they show the split in two markets with the same method.

In Germany, 51.1% of citations for German prompts pointed to .de, .at or .ch hosts, 11.9% to a German path such as /de/ on a global domain, 1.9% to a German subdomain, and 34.1% to global domains with no German signal at all. In France the country-domain share was lower (34.4%) and the unsignaled share higher (48.6%). OtterlyAI estimates that many of those unsignaled French citations still went to French-market businesses on .com, .info or .org addresses, such as the comparison site selectra.info, the most-cited domain in its French data at 0.40% of citations. Its practical reading, which it presents as an inference rather than a test, is that engines cite the page rather than the suffix, and that international companies should consider a French or German path on an established domain before buying a new ccTLD.

Google's own guidance for international sites is the other half of the decision, because Google's AI features ground on Google's index (Google, managing multi-regional and multilingual sites):

The last row matters more for AI search than for classic SEO. Google states that Googlebot's default IPs "appear to be based in the USA" and that it crawls "without setting Accept-Language" (Google, locale-adaptive pages). A site that serves German only to German IPs may never have its German version indexed, and a page that is not indexed cannot be a supporting link in AI Overviews or AI Mode.

Does hreflang matter for AI search? What Google says

For Google's AI features, hreflang matters indirectly: it helps Google Search choose the right language version, and AI Overviews and AI Mode draw their links from Google Search. For ChatGPT, Claude and Perplexity there is no documentation either way, and the one public test found they usually ignored the translated versions.

Google describes hreflang as a way to "help Google Search point users to the most appropriate version of your page by language or region" (Google, tell Google about localized versions). Three details from the same documentation trip up multilingual sites:

• Return links are mandatory. "If two pages don't both point to each other, the tags will be ignored." Each version must also list itself.
• hreflang does not set the page language. "Google doesn't use hreflang or the HTML lang attribute to detect the language of a page." Google uses the visible content, so keep one language per page and avoid side-by-side translations.
• Codes must be valid. Language is ISO 639-1, the optional region ISO 3166-1 alpha-2. "en-UK" does nothing (the code is GB), a bare region code such as "be" is read as a language (Belarusian), and "es-419" is not supported.

Glenn Gabe's December 2025 test is the closest thing to evidence beyond Google. Using translated Google documentation and a Cloudflare post, all with hreflang, he found Google's blue links, Bing and Copilot returned the correct French, Italian and Spanish URLs; AI Mode mostly did; ChatGPT, Perplexity and Claude mostly returned the US English original. His conclusion was that hreflang appears to help Google and Bing, and through them AI Overviews, AI Mode, Gemini and Copilot, and that it is unclear whether other platforms support it at all.

One more Google feature affects multilingual visibility. Google may machine-translate the title and snippet of a result that is not in the searcher's language and open a machine-translated page when clicked, in 21 languages including Spanish, German, Hindi and Arabic (Google, translated results). Weglot's Spanish book-retailer example shows the downside: for English queries its links pointed to Google Translate's proxy, not to its own site. You can opt out with notranslate, but the better fix is a real translated page that Google can show directly.

Which content should you translate first?

Translate the pages that answer commercial questions in each market, then the evidence pages that make you credible there. Both OtterlyAI indexes measured bottom-funnel prompts and recommend translating comparisons, pricing, specifications and support pages first. Weglot's Spain and Mexico data points the same way for Google's AI Overviews, where untranslated sites were cited far more for queries in the language they already had.

Is machine translation safe?

Machine translation that a fluent reviewer checks and localizes is fine; machine translation used to mass-produce pages is not. Google's spam policies list "scraping feeds, search results, or other content to generate many pages", "including through automated transformations like synonymizing, translating, or other obfuscation techniques", as scaled content abuse (Google spam policies). The question Google asks is whether the pages add value, not which tool wrote them. In practice: translate fewer pages, review them, localize examples, prices and proof points, and keep facts such as prices and specifications identical across versions so engines do not see contradictions.

Where to build local sources and mentions per market

Every market has its own set of cited publishers, comparison sites and institutions, and they are rarely the ones you would guess from the US. Brand and company pages made up about 57% of citations in all three markets OtterlyAI measured, so the remaining 43% is where local mentions come from, and its mix differs by country:

• Germany: news and media took 14.9% of citations, led by Finanztip, Chip, Handelsblatt and Test.de. Associations and NGOs took 8.9% (US 5.3%), led by Verbraucherzentrale and ADAC. Reddit took 0.60% versus 1.79% in the US; YouTube took 1.99%, slightly above the US.
• France: news and media took 10.4%, mostly trade and niche titles such as Clubic, Frandroid and MoneyVox. Government sources took 3.5%, nearly double Germany, led by Service-Public and Ameli. Consumer group Que Choisir and comparators like selectra.info were prominent. Reddit took 0.62%, YouTube 1.93%.
• Google's AI features in non-English markets: when they cite Reddit, they often cite Reddit's own machine translations. In Peec's data, translated pages were 71.6% of AI Overview Reddit citations in Sweden, 73.0% in Norway and 68.4% in Spain. An English Reddit thread about your product can surface in a Swedish AI Overview in Swedish. Our guide to Reddit and AI search visibility covers how to take part in those threads without breaking community rules.
• Markets with many languages: India is the clearest example, with AI Mode in Hindi and seven more Indian languages since October 2025. Our guide to AI search optimization in India covers that market in detail.

The practical step is the same everywhere: run your category's buying prompts in the local language, list the third-party domains each engine cites, and work through that list with PR, listings, reviews, association memberships and expert contributions. Google also lists "links from other local sites" and Business Profile signals among the ways it identifies a site's intended country, so local mentions help classic search and AI answers at the same time.

How do you track AI visibility by country and language?

Treat each country-and-language pair as a separate market with its own prompt set and baseline. A visibility score averaged across markets hides exactly the gaps this guide is about.

1. Write native prompts, not translations of your English set. People in each market phrase buying questions differently and name different competitors. Have a fluent speaker write or review them.
2. Set location where the tool lets you. Location changes answers. Claude's consumer apps use coarse IP location for web search, OpenAI's and Anthropic's APIs both accept an approximate user location, and Google's AI features are served per country. A German prompt run from a US location is a language test, not a German-market test.
3. Read each engine separately. OtterlyAI found a single-engine measurement can misread the French market by up to 16.7 points. Compare ChatGPT, Google AI Overviews, AI Mode, Gemini, Perplexity, Claude and Copilot side by side.
4. Log which domains are cited, and in which language. Note whether a citation is your local page, your English page, a local third party or a machine translation. This tells you whether to fix translation, hreflang or local mentions.
5. Re-run on the same prompts over time. AI answers vary from run to run, so compare repeated runs rather than one screenshot.
6. Check Google's own numbers per country. Search Console's Generative AI performance report has a Countries dimension ("Understand your visibility on a country basis") for AI Overviews and AI Mode impressions (Google Search Central blog). See our walkthrough of the Search Console generative AI report.

Tools that track AI answers by market

Disclosure: AI Rank Lab is our product. We list it first because this is our site, and we've kept the same criteria for every tool below. The table covers only what each tool documents publicly about country and language (checked October 24, 2026).

A market-entry checklist for AI search

Use this before you launch a new country or language, and again 60 to 90 days after launch. Each step maps to evidence or documentation above.

What this evidence cannot tell you yet

Be careful about how far you push these findings. Six of the seven studies come from companies that sell AI visibility or translation tools. The OtterlyAI indexes cover bottom-funnel commercial prompts in a 44-day window, so informational queries may behave differently. Peec's English fan-out data reflects ChatGPT in early 2026, and its translated-Reddit data shows ChatGPT's behavior changing sharply between April and June 2026, so any of these patterns can shift with a model update. Gabe's hreflang test used three pages. None of the studies isolates cause and effect: translated sites getting more citations may also be bigger, better-linked sites. Treat the numbers as a map of where to look, and use your own per-market tracking to confirm what holds for your category.

If you want to see where your brand stands in each market, AI Rank Lab can run your native-language prompts across ChatGPT, Claude, Gemini, Perplexity and Google's AI features, and set country and language for AI Overview and AI Mode checks. Start with AI visibility tracking.