Ask ChatGPT a question in French about a page that exists in both French and English, and there's a good chance you'll get the English version anyway. SEO consultant Glenn Gabe tested this in December 2025 across Google, Bing, and five AI platforms, and found Claude returned the U.S. English URL in every one of his three language tests, even when a local-language page existed and ranked normally in Google. Hreflang, the tag built to solve exactly this problem, barely registers with most AI answer engines. The mechanics behind that gap matter for any brand expanding into new markets, before French or German content quietly stops getting cited.
What hreflang actually does, and why Google built it
Hreflang is an HTML attribute that tells Google which language or regional version of a page to serve a given searcher, nothing more. It doesn't translate content, and it doesn't rank a page on its own. A site adds an alternate-language link to each variant, and Google uses that cluster to swap in the best-matched sibling once ranking has already happened. Google's Gary Illyes made that distinction directly in a 2019 Reddit AMA, as Search Engine Journal reported at the time: hreflang brings no ranking benefit per se, it only improves which already-ranked page gets shown to a given user. That model assumes one engine ranking pages and filtering the result by locale. It's the model Google was built around when hreflang shipped in 2011. Most AI answer engines don't work that way.
What GPTBot, ClaudeBot, and PerplexityBot's documentation says about language
OpenAI's and Anthropic's published crawler documentation says nothing about hreflang, language variants, or which localized URL to prefer. GPTBot and ClaudeBot are documented purely in terms of robots.txt directives, crawl pacing, and published IP ranges, with no mention of language selection anywhere in either page. OpenAI's GPTBot documentation covers user-agent strings, shared crawls with its search bot, and the roughly 24 hours it takes for a robots.txt change to take effect, but language never comes up. Anthropic's support article on ClaudeBot is just as narrow: it follows robots.txt, respects an optional crawl-delay setting, and paces requests per domain. PerplexityBot's own public guidance is no different.
Robots.txt is still the only lever these three document publicly, and what a bot actually obeys there varies more than most sites assume once you check it directly. Confirming that traffic claiming one of these user agents is the real crawler, not an impersonator spoofing the name, is a separate and increasingly necessary check before drawing conclusions from server logs.
The test data: AI engines defaulting to English even when a local page exists
Gabe's December 2025 testing found AI platforms routinely returned the U.S. English version of a page even when a correctly tagged local-language version existed, outranked nothing, and was the version Google itself served. The pattern held across French, Italian, and Spanish queries, and it was most consistent on Claude.
In one test, a French-language query about a Google help document returned the English URL from ChatGPT, Perplexity, and Claude, even though Google's own 10 blue links and AI Mode served the French page without hesitation, as Gabe documented. An Italian-language query produced the same split: ChatGPT and Claude answered with the English page, while Perplexity returned the correct Italian URL, something Gabe called rare in his own testing. A third test, in Spanish, had ChatGPT cite both language versions at once, while Perplexity and Claude again defaulted to English.
Hreflang answers a question AI engines never ask: which single page should rank. They don't rank one page, they synthesize an answer from several, and a signal built for ranking has nothing left to attach to.
Why low-resource languages lose more ground than French or Spanish
A 2026 ICML paper on multilingual retrieval found that when a chatbot has to choose which language version of a document to cite, it favors English even when a same-content document in another language is sitting in the same context window. The size of that bias depends heavily on which language is competing.
Dayeon Ki and six co-authors tested six open-weight models, including LLaMA, Qwen, Gemma, and Aya, across eight languages, and found citation accuracy for a Swahili-language document dropped by an average of 23.9% compared to the identical English version, rising to 37.6% for one model. Bengali dropped by an average of 18.0%. Spanish and French, both well-represented in training data, dropped by only 8.1% and 8.8%. The researchers describe this as a language preference bias, and it got worse, not better, when the non-English document was placed in the middle of a longer context rather than at the start or end. In one experiment using 231 real-world queries, models sometimes cited an irrelevant English document ahead of a genuinely relevant one in another language, meaning language outweighed relevance outright.
The citation gap shows up in production AI engines too, not just lab tests
A separate audit published in August 2026 tested ChatGPT, Perplexity, and Google AI Overview directly, not research models, by asking health questions in seven languages and counting what each platform actually cited. The gap it found was just as real as the language-preference research suggests, and in places it was worse.
Across 1,140 responses and nearly 16,000 citations, English queries returned a mean of 12.3 sources per response, compared with 9.2 for the other six languages combined, according to the study. The shortfall wasn't evenly spread: ChatGPT gave zero citations at all for a third of its Twi-language responses, and Google AI Overview did the same for 30% of its Hindi responses. When the researchers counted sources hosted on a country-specific domain as a rough proxy for genuine local content, high-resource languages pulled from local domains 38.0% of the time, against 3.3% for low-resource ones. Their conclusion wasn't that AI engines discriminate by language on purpose, it's that they cite whatever authoritative content actually exists in that language, and for most languages, not enough of it does. If you want to see which crawler is actually reaching your translated pages in the first place, checking your own server logs is the more reliable starting point than guessing from the outside.
What to use instead of hreflang for AI visibility
Hreflang isn't wrong to keep, it's just not what AI engines read. The more effective lever is giving each language version of your brand its own verifiable entity signal and making sure there's enough real, non-translated content in that language for an engine to find.
That starts with structured data that ties every language variant of a brand back to one entity: an Organization schema block on each locale's homepage, linked through sameAs to the equivalent profile in other languages and to verifiable external profiles like Wikidata. We've covered the mechanics of this kind of structured data in our JSON-LD guide for GEO in more depth. The second lever is content supply, not translation quality: the Sources of Truth researchers found that citation gaps track where authoritative native-language content is thin, not where it's merely untranslated, so a page translated well but hosted nowhere else in that language still loses out to a thinner page backed by a dozen other local sources.
Even within Google's own surfaces, the behavior split in Gabe's test: AI Mode served the correct French URL, while AI Overviews returned no result at all when the query ran from France, the kind of surface-by-surface inconsistency that makes guessing at hreflang's reach a poor bet for any brand relying on it alone.
Frequently Asked Questions
Does hreflang work for AI search engines like ChatGPT or Perplexity?
Not reliably. Hreflang is read by Google and, as a weaker signal, by Bing, but OpenAI's and Anthropic's own crawler documentation makes no mention of it, and testing by SEO consultant Glenn Gabe found ChatGPT and Claude defaulting to English-language URLs even when a correctly tagged local-language page existed. Perplexity was inconsistent. Treat hreflang as a Google-and-Bing signal, not a universal AI visibility one.
Why does Claude or ChatGPT answer in English even when I ask in French?
Because these engines don't filter one ranked result by locale the way Google does. They retrieve and synthesize from whichever sources rank highest in their own process, and English-language sources dominate that process for most languages. Research on multilingual retrieval has found this bias gets worse for lower-resource languages, and stronger the further into a document the cited content sits.
What should I do instead of relying on hreflang for AI visibility?
Give each language version of your brand its own verifiable entity signal, using Organization schema linked through sameAs across language variants and to profiles like Wikidata, and invest in genuinely native-language content rather than translated pages alone. Research on AI citation patterns across languages found that citation gaps track the supply of authoritative native-language content, not translation quality.
Do GPTBot, ClaudeBot, and PerplexityBot crawl language variants of a page differently?
There's no public evidence that they do. None of the three bots' published documentation mentions language-specific crawling behavior, hreflang, or a stated preference between locale variants of the same page. All three are documented purely in terms of robots.txt rules, crawl pacing, and IP ranges.
Does hreflang still matter for SEO even if AI engines ignore it?
Yes. Google's Gary Illyes has confirmed hreflang isn't a ranking signal on its own, but it still determines which language or regional URL Google serves a given searcher, and Bing treats it as a secondary signal too. It remains worth maintaining for classic search traffic, even where it does nothing for AI-generated answers.
How many languages did the research behind this article actually test?
Two separate studies, covering different ground. An ICML 2026 paper on multilingual retrieval tested eight languages across six open-weight AI models, while a 2026 citation audit tested ChatGPT, Perplexity, and Google AI Overview directly across seven languages using real health-related queries, collecting more than 15,000 citations.