Multilingual GEO: Do AI Engines Cite Non-English Sources?
Here is the pattern we see on almost every multilingual site we audit: the English pages get cited inside AI answers, and the Spanish, German, or Indonesian versions of the same content do not — even
Multilingual GEO: Do AI Engines Cite Non-English Sources?
Bottom line: Yes - AI engines cite non-English sources, but unevenly. Multilingual GEO works because engines like Perplexity and Gemini increasingly pull answers in the user's query language, while ChatGPT still leans on English URLs. The brands that win translate entities, not just words, and keep hreflang, schema, and author signals consistent across every language version.
Here is the pattern we see on almost every multilingual site we audit: the English pages get cited inside AI answers, and the Spanish, German, or Indonesian versions of the same content do not - even when they rank fine on blue links. Teams assume it's a content-quality problem. It usually isn't. It's an entity and hreflang consistency problem that quietly disqualifies translated pages from citation.
Key Takeaways:
- —A multilingual mental-health citation audit on arXiv found ChatGPT averaged 17.5 sources per English answer but only 4.3 in Twi, returning zero sources for roughly a third of Twi responses - while Perplexity stayed within 8.6-10.2 sources across all seven languages tested.
- —Query language changes who gets cited: in a 320-response English-vs-Arabic test, local sources rose from 31.6% of citations in English to 56.7% in Arabic.
- —Engines differ sharply. Perplexity and Gemini tilt toward query-language sources; ChatGPT frequently answers in-language but still cites English URLs.
- —Google ignores your hreflang unless every language version links back to every other version, including itself - a single missing return link can orphan a translation from AI citation.
- —Entity consistency (same brand name, same schema
@id, same author) across languages matters more than translation polish for staying citable.
Do AI engines cite non-English sources?
They do - the honest answer is "yes, but it depends heavily on the engine and the query language." The clearest evidence comes from a multi-platform multilingual citation audit published on arXiv, which ran the same queries across seven languages. Non-English queries consistently returned fewer sources than English ones, but the drop wasn't uniform. ChatGPT collapsed hardest: 17.5 sources per answer in English against 4.3 in Twi, with no sources at all for about a third of Twi responses. Perplexity barely flinched, holding between 8.6 and 10.2 sources across every language in the set.
That gap is the whole game. If your category's buyers ask questions in Spanish or German, the engine they use decides whether your translated page even enters the candidate pool. A second signal reinforces this: a separate GEO research paper on arXiv found that under non-English prompts, citations tilt toward the target language - but less so for Google than for the pure chat engines.
So the real question isn't whether non-English sources get cited. It's whether yours does, in the language your market actually searches, on the engine they actually use.
Why does query language change which sources get cited?
Think about what happens before an engine writes a word. For a non-English query, many retrieval-augmented systems translate the query into English, retrieve English documents, then answer back in the user's language. That "English pivot" is why ChatGPT so often responds in fluent Spanish while linking to English URLs - informal multilingual tests by Stan Ventures documented exactly this behavior across ChatGPT, Perplexity, and Claude.
The counterweight is that some engines weight query-language sources much more aggressively. In a 320-response test comparing English and Arabic answers about Lebanon's economy, local sources made up 31.6% of citations in English but 56.7% in Arabic. Same facts, same questions - a completely different citation set depending on the language of the prompt. The author is careful to note the sample is small and doesn't prove causation, but the directional signal matches what vendor data from Profound on query-language citation deltas also reports: every non-English market underperformed English, with Portuguese and Spanish coming closest to parity.
Two mechanisms, pulling opposite ways. Your multilingual GEO strategy has to account for both - not assume one universal rule.
The language-by-engine citation behavior table
This is the part most GEO guides skip. "AI engines" is not one behavior - each engine handles non-English sources differently, and your fixes change accordingly. The table below summarizes the behavior we see in audits and that the research above supports, plus the single rule that most affects citability on each engine.
| Engine | How it handles non-English queries | Non-English source citation | Rule that keeps you citable |
|---|---|---|---|
| ChatGPT Search | Often answers in query language but pivots to English for retrieval; source count drops sharply in low-resource languages (arXiv audit) | Weakest - frequently cites English URLs even for non-English answers | Publish a genuinely localized page (not a thin translation) with query-language entities and schema; English fallback will otherwise win |
| Perplexity | Most stable across languages; holds source count and tends to cite in-language (arXiv audit) | Strongest - rewards well-structured query-language sources | Clear headings + FAQ structure in the target language; it lifts passages aggressively |
| Google Gemini / AI Overviews | Tilts to target language but less than chat engines; tied to Search eligibility (GEO paper) | Moderate - gated by indexing and hreflang correctness | Page must be indexed and snippet-eligible per Google's AI-features docs; hreflang must validate |
| Microsoft Copilot | Mirrors Bing index; in-language citation depends on Bing coverage in that locale | Moderate, locale-dependent | Verify the language version is indexed in Bing, not just Google |
The three cross-engine constants worth memorizing:
- —Hreflang consistency decides whether Google treats your French and English pages as distinct citable documents or collapses them to one and ignores the rest.
- —Entity consistency - identical brand name,
sameAslinks, schema@id, and author identity across languages - tells every engine that the translated page is "you," so authority transfers instead of resetting to zero. - —Query-language structure - question-shaped headings and answer-first passages written natively, not auto-translated - is what Perplexity and Gemini actually lift.
How do hreflang and entity consistency keep translated pages citable?
Translation quality is table stakes. The technical layer is where translated pages silently drop out of AI citation. Work these in order:
- Fix return links first. Google ignores hreflang annotations entirely unless every language version links back to every other version, including itself - this is stated plainly in Google's localized-versions documentation. One missing return link orphans the translation.
- Confirm each version is indexed and snippet-eligible. AI Overviews pull only from pages eligible to appear in Search with a snippet. An un-indexed translation cannot be cited, period. Check this in Search Console per language directory.
- Keep the entity identical across languages. Same organization name, same
sameAsprofiles, same author@id. If your German page lists a different author entity than your English page, engines can't merge the authority. (Our walkthrough on building a knowledge-graph presence from scratch covers the schema specifics.) - Translate the metadata, not just the body. Untranslated titles and meta descriptions are one of the most common reasons a translated page reads as duplicate or low-quality to crawlers.
- Make the language switcher crawlable. A JavaScript-only switcher that bots can't follow means engines never discover the alternate versions - the same failure mode we flag for any JavaScript rendering issue with AI crawlers.
- Rewrite structure natively per language. Auto-translated H2s rarely match how a native speaker phrases the question. Question-shaped headings in the target language are what map to that market's People Also Ask and get lifted into answers.
If your international setup is more involved - subdomains, ccTLDs, or domain strategy - our deeper guide on international SEO, hreflang, and the mistakes that cost rankings goes past the AI-citation layer into the full architecture.
What should you fix first on a multilingual site?
Don't boil the ocean. Prioritize by what disqualifies pages fastest versus what merely limits upside. This is the order we run it in an audit:
| Priority | Fix | Why it's ranked here |
|---|---|---|
| 1 | Hreflang return links + indexation per language | A broken return link or missing index entry removes the page from the candidate pool entirely |
| 2 | Entity + schema consistency across languages | Without it, authority resets to zero on every translation |
| 3 | Native-language structure (headings, FAQ, answer-first) | Determines whether Perplexity/Gemini can lift your passage |
| 4 | Localized examples, prices, and sources | AI answers for non-English queries favor genuinely localized detail |
| 5 | Per-language citation monitoring | You can't improve what you don't measure across engines and languages |
For step 5, you need to track citations per language and per engine rather than eyeballing one market. That's exactly what we built the AI Citation Tracker for - it watches whether your pages get cited across ChatGPT, Perplexity, and Gemini so you can see which language versions are actually entering answers. If you just want a fast read on whether a single URL is AI-Overview eligible, the free AI Overview Checker is the quicker first look.
This whole discipline is an extension of standard GEO, not a separate field - if you're still building the foundation, start with our complete 2026 guide to Generative Engine Optimization, then layer the multilingual rules on top. For engine-specific depth, our breakdown of how to become a source Perplexity trusts pairs well with this, since Perplexity is the most reliable non-English citer.
How We Assessed This
The recommendations here combine published research with what we see in client audits. For the research layer, we leaned on primary and semi-primary sources: the arXiv multilingual citation audit for per-engine source counts across seven languages, a GEO research paper on arXiv for query-language citation tilt, and an independent 320-response English-vs-Arabic citation test for how local-source share shifts with query language. Technical rules on hreflang and snippet eligibility are drawn directly from Google's official Search documentation, not secondhand summaries.
On the audit side, SEO Magics works with growth-stage SaaS, DTC, and agency clients on 12-month optimization cycles, and multilingual sites surface the same failure modes repeatedly - broken hreflang return links, un-indexed translations, and inconsistent entity schema that resets authority on every language version. We cross-check indexation in Search Console, validate hreflang with crawl tooling, and monitor citations per engine and per language rather than assuming one market's performance predicts another's. Where we couldn't verify a figure against a namable source, we've written the claim qualitatively instead of inventing a number.
Frequently Asked Questions
Do AI engines prefer English sources over other languages?
For retrieval, several engines lean English - ChatGPT especially, which often pivots a non-English query to English before retrieving. But the arXiv audit shows Perplexity stays stable across languages, so "AI prefers English" is only half true. It's engine-dependent, and Perplexity and Gemini reward well-structured query-language sources.
Will translating my content get it cited in AI answers?
Translation is necessary but not sufficient. A translated page that's un-indexed, missing hreflang return links, or carrying a different entity identity than your English page usually won't be cited. Fix the technical and entity layer first, then the translation quality matters.
Is hreflang still relevant for AI search?
Yes - arguably more than for blue links. Google ignores hreflang unless every version links back to every other version including itself, and AI Overviews only cite pages eligible to appear in Search with a snippet. Broken hreflang can collapse your translations into one page and drop the rest from citation.
Which AI engine cites non-English sources most reliably?
Based on the multilingual audit data, Perplexity is the most stable across languages, holding its source count where ChatGPT's collapses. Gemini tilts toward query-language sources but is gated by Search indexing. If a market is strategically important, prioritize Perplexity and Gemini optimization.
Should I write native content per language or translate my English pages?
For AI citation, genuinely localized content wins - native question-shaped headings, local examples, and local sources. Auto-translation preserves meaning but rarely matches how native speakers phrase questions, which is what Perplexity and Gemini lift into answers. Translate the body, but rewrite the structure.
How do I track whether my translated pages are getting cited?
Monitor per language and per engine, not just one market. A citation tracking tool shows whether each language version enters answers across ChatGPT, Perplexity, and Gemini, so you can see which translations are working and which are invisible.
Get your multilingual site cited in every market
Most international sites we audit are leaving AI citations on the table in their non-English markets - not because the content is weak, but because hreflang, indexation, and entity consistency quietly disqualify the translations. SEO Magics is an AI-native agency built to get brands cited inside ChatGPT, Perplexity, and Google AI Overviews, and our AI SEO service is where multilingual GEO lives. Want more depth on adjacent topics first? The SEO Magics journal covers the full GEO stack. When you're ready to find the gaps in your own languages, book a strategy call and we'll map which markets and engines you're missing.