Multilingual GEO in 2026 comes down to three things: correct hreflang tagging, content written natively in each target language rather than translated, and citation sources that exist inside that language’s own information ecosystem. Get those three right and AI engines are far more likely to surface the correct localized page instead of defaulting to your English homepage. Get them wrong and you lose visibility in every market outside your primary language, even when you have a fully translated site.
Do AI engines actually respect hreflang when answering in other languages?
Inconsistently, and the gap between platforms is large. Glenn Gabe of GSQi ran a December 2025 test across ChatGPT, Perplexity, Claude, Copilot, and Gemini, searching in French, Italian, and Spanish against pages that were properly translated and tagged with hreflang. Copilot and Gemini returned the correct localized URL almost every time, because both platforms run on Bing’s and Google’s existing multilingual detection systems. ChatGPT, Perplexity, and Claude did not: in most of Gabe’s tests, they answered in the local language but linked back to the US English page instead of the French, Italian, or Spanish version that actually existed.
That distinction matters for anyone spending money on translated content. Google and Bing have spent two decades building systems that read hreflang and country signals correctly, and AI Overviews and AI Mode inherit that maturity. ChatGPT, Claude, and Perplexity are working from a much younger stack, and as of Gabe’s testing, they frequently ignore hreflang entirely and pull whatever version of a page ranks highest in their training data or retrieval index, regardless of the language the user is searching in.
Running a business that serves customers in more than one language? Get the free AI visibility audit at subscribepr.com/audit and see exactly which language version of your site AI engines are citing right now, and which one they’re skipping.
Why does ChatGPT keep citing the English page when you search in Spanish?
Because ChatGPT’s retrieval layer weighs authority and freshness more heavily than language match, and your English homepage usually has more backlinks, more citations, and a longer indexing history than a newer translated subpage. When ChatGPT triggers a web search and finds two versions of a domain answering the same query, it defaults to the version its retrieval system trusts most, which is typically the original language page, not the one matching the searcher’s language setting.
This is the core failure mode in Gabe’s testing: the response text came back in the user’s chosen language, but the cited source link pointed to the English original. Perplexity showed the same pattern more often than not, and Claude frequently returned no sources at all until directly asked, then still surfaced the English URL. Only Copilot and Gemini consistently got it right, and only because they defer to Bing and Google’s language detection rather than running their own.
The practical implication: a translated page with hreflang is necessary but not sufficient for ChatGPT and Perplexity specifically. Those two platforms need the localized page to carry independent authority, its own backlinks, its own citations in local publications, and its own structured data, or they will keep defaulting to the stronger English original. AI engines cite the page they trust most, not the page technically correct for the query, the same authority logic that governs citation decisions in any single language.
Which sources do AI engines cite most, and does that change by language?
Wikipedia dominates citations across every language AI engines search in, but the specific secondary sources shift by market. A Profound study covering 680 million AI citations found Wikipedia alone accounts for 7.8% of ChatGPT’s total citations and nearly half of its top ten most-cited domains. Separately, Surfer’s AI Tracker analyzed 36 million AI Overviews and 46 million citations between March and August 2025 and found Wikipedia and YouTube were the two most-cited domains overall. A June 2025 Semrush study of AI Overviews specifically found Quora as the single most-cited source, with Reddit close behind.
The pattern holds inside non-English markets too, but the specific platform changes. Spanish-language queries pull heavily from Wikipedia en Español, Spanish-language Reddit and Quora threads, and regional news and directory sites, not the English-language equivalents of those same platforms. An AI engine answering a query in Portuguese is drawing from Portuguese Wikipedia, Brazilian forums, and local review sites, a completely separate citation graph from the English one. If your English content has strong entity signals but your Spanish or Portuguese content has none, you are invisible in that market’s AI answers regardless of how well you rank in English, because the retrieval and citation graph doesn’t cross language lines the way most site owners assume it does.
How do you build a multilingual GEO strategy that AI engines can parse?
Start with technical hreflang correctness, then build independent authority for each language version, because AI engines increasingly need both signals, not just one. The technical layer is table stakes: every translated page needs a correct hreflang tag pointing to every other language and regional variant, self-referencing tags, and a matching entry in your XML sitemap. Run a full-site crawl focused on hreflang at least quarterly, since broken or missing tags are common after redesigns and CMS migrations and they are the first thing Google and Bing’s systems fall back on when nothing else disambiguates a page.
The authority layer is what ChatGPT, Perplexity, and Claude actually respond to, and it’s the piece most companies skip. Each language version of your site needs its own backlinks from publications in that language, its own local business citations, and its own structured data, including Organization schema with sameAs links to the local versions of Wikidata, LinkedIn, and industry directories where they exist. We cover the sameAs mechanics in entity SEO for AI search, and the same logic applies per language: an entity confirmed in French sources is a different signal than the same entity confirmed only in English ones. Schema markup should also be duplicated and localized per language page rather than left pointing back to a single English version, a detail we walk through in schema markup for AI search.
Content structure matters as much for translated pages as it does for English ones. Each localized page needs a direct answer in its first sentences, clear headers phrased as the questions native speakers actually type, and FAQ sections addressing local variations of the same query, the same format guidance in how to optimize content for AI search applies fully once translated, and skipping it on localized pages is one of the most common reasons a translated site performs worse in AI answers than the original.
Should you use human translation or native content teams for AI visibility?
Native content teams outperform translation, because AI engines assess content quality and intent match independently in each language, and direct translation rarely captures either. A page translated word for word from English into Spanish carries English sentence structure, English examples, and English search intent baked in, none of which match how a Spanish speaker actually phrases a question or what data points they expect to see. AI engines retrieving Spanish-language content are comparing it against other natively written Spanish content, not against your English original, so a translated page is competing at a structural disadvantage before authority is even considered.
The fix is a native-first workflow: brief a native speaker or in-market team on the topic and the target query, let them write the page as if no English version existed, then check it against your English page for factual and brand consistency rather than forcing it through a translation pipeline. This costs more per page than machine or vendor translation, but it’s the difference between a page that merely exists in Spanish and a page that reads like it was written for a Spanish-speaking audience, which is what AI engines are increasingly able to tell apart.
What does per-language citation building look like in practice?
It looks like running the same citation strategy you’d run in English, but inside that language’s own platforms, not the English versions of them. For a business targeting Spanish-speaking markets, that means securing mentions in Spanish-language trade publications, contributing to Spanish Wikipedia where your entity qualifies, engaging authentically in Spanish-language Reddit and Quora communities relevant to your category, and claiming local business directories in the target country rather than assuming your US listings carry over.
This work compounds with the authority-building your English site already does, but it does not substitute for it. A company with strong PR placements and Reddit presence in English and nothing in Spanish will still get cited by name when an English speaker asks ChatGPT for a recommendation, and passed over entirely when the same question comes in Spanish, because the two citation graphs are functionally separate. Treat each target language as its own AEO campaign with its own sources, its own entity signals, and its own measurement, not as a translation layer sitting on top of one English strategy.
Frequently asked questions
Does hreflang alone guarantee AI engines cite the right language page? No. Hreflang helps Google, Bing, AI Overviews, and AI Mode, all of which read the tag directly or inherit it through Google and Bing’s systems. ChatGPT, Perplexity, and Claude showed much weaker hreflang handling in December 2025 testing by Glenn Gabe of GSQi, frequently citing the English original even when the response was in the target language. Hreflang is necessary but needs independent page authority behind it to work across all platforms.
Is machine translation good enough for AI search visibility? Machine translation gets content into another language but rarely matches native search intent, phrasing, or examples, and AI engines assess quality per language rather than crediting a good English page for a weak translated one. Native content written for that market performs measurably better in retrieval and citation testing than direct translation.
Which AI platforms handle multilingual queries best right now? Copilot and Gemini perform the most consistently because they run on Bing’s and Google’s mature multilingual detection systems. Google’s AI Mode also performed well in independent testing. ChatGPT, Claude, and Perplexity were the least reliable at surfacing the correct language version of a page as of testing in late 2025.
Do I need separate schema markup for each language version of my site? Yes. Each localized page should carry its own structured data, including Organization schema with sameAs links to the local-language versions of Wikidata, directories, and profiles where they exist, rather than pointing back to a single English schema block. This gives AI engines independent verification for each language, not just the original.
How do I find out which language version of my site AI engines are actually citing? Search your target queries directly in ChatGPT, Perplexity, Claude, Copilot, and Gemini with your preferred language setting changed to match each market, and check whether the cited link matches the correct localized page. Repeat this quarterly, since AI platforms update their retrieval systems frequently and results shift.
Most businesses treat translation as the finish line for international AI visibility. It’s the starting point. Want a clear picture of where your business shows up in AI answers across every market you serve? Start the free audit at subscribepr.com/audit and get the gaps mapped by language, not just by domain.
Tagged