If your website exists only in English, you may be invisible to a large share of the people asking AI tools about your products. That is the central issue in multilingual AI SEO. Answer engines respond in the user’s language and tend to favor sources that match it. This article explains what the research says, where it should be questioned, and how to build a sensible market-by-market plan.
Language acts as a visibility filter
One agency summary reports that Profound ran a study across 3.25 billion citations from seven AI models in 14 countries and concluded that query language is the dominant force shaping which sources get cited. I did not review the original study, so treat that as a claim to verify. The practical implication is easy to grasp: a French question tends to produce French sources, and an English-only page may be skipped.
Reach is also growing. Google expanded AI Mode to 40 new countries and 35 new languages in late 2025. The more languages these features cover, the more your competitors in local markets can appear while you stay absent.
What the translation study found, and what to question
The most discussed dataset comes from Weglot. Its study analyzed 1.3 million citations across Google AI Overviews and ChatGPT, focusing on Spanish-language websites in Spain and Mexico.
The headline findings:
- Untranslated Spanish sites from Spain earned 17,094 citations for Spanish queries but only 2,810 for the equivalent English queries.
- Translated sites in Spain showed a far smaller gap, with 10,046 citations versus 8,048 in English.
- Overall, translated sites achieved 327% more visibility than untranslated ones and 24% more total citations per query, and in ChatGPT the language bias almost vanished. searchenginejournalsearchenginejournal
Read this carefully. Weglot sells translation software, so it has an interest in the conclusion. The study covered a single language pair and two markets. And it is observational: sites that invest in translation may also be larger, better funded, and better optimized in other ways. The direction is plausible, but the exact multipliers should not be promised to a client.
Engines handle language versions differently
Practical testing adds a useful, more cautious picture. Glenn Gabe ran multilingual tests across Google, Bing, AI Overviews, AI Mode, and several chat tools. A summary of his results says Google and Bing generally showed the correct language version when hreflang was in place, AI Overviews sometimes served the local version and sometimes didn’t, and AI Mode was steadier. The weakest area was chat tools: ChatGPT, Perplexity, and Claude frequently replied in a different language while still linking to English sources. In one French test, the AI search platforms were all wrong other than Copilot, which leverages Bing’s systems. Contact Us About Us +2
These were small tests on specific documentation pages, and behavior changes quickly. The lesson is that you cannot assume one setup works everywhere. Verify engine by engine.
Translation quality is a trust issue
Cheap translation can cause real harm. A 2023 peer-reviewed study in the ACL’s TACL journal found that hallucinations in large multilingual translation models are remarkably fluent, meaning they read naturally even when wrong. That makes errors harder to spot without a native reviewer. The same guide advises flagging lower-resourced languages and applying a stricter quality assurance threshold to those pages.
For a business, a mistranslated price, dosage, or legal term is not a style problem. Treat translated pages as publications that need review.
A market-by-market method
Sarkar SEO suggests choosing languages by business case, not by ambition. This is our own working approach, not an industry standard.
- Rank markets by evidence. Use revenue, existing impressions in Search Console, sales inquiries, and competitor presence to choose two or three languages first.
- Start with your best twenty pages. Translate and localize the pages that already earn citations and leads, not the entire site.
- Localize intent, not just words. Adapt examples, currency, units, regulations, and the way people phrase questions. A phrase that works in Madrid may feel odd in Mexico City.
- Add local proof. Engines lean on sources in the query language, so mentions, reviews, and directory listings in that language matter too. An English-only reputation won’t carry over.
- Run a quality process. Use a glossary for brand and product terms, a native reviewer for key pages, and a named owner for each language.
Technical basics that still apply
Keep the technical layer simple and correct:
- Hreflang. One localization vendor describes it as the routing layer that prevents translated pages from competing against each other or appearing in the wrong market. Use reciprocal tags, so each language page points to the others and to itself.
- Separate, crawlable URLs per language. Avoid serving translations only through scripts or cookies that crawlers might not process.
- Consistent entities. Keep brand names, product names, and contact details aligned across languages.
- Language and region accuracy. Tag content for the dialect you mean, and don’t treat Spanish as one market.
Test in the languages your buyers use
Build a prompt set in each target language, written by someone fluent. Run it across Google’s AI features and at least two chat tools. Record which language version of your page, if any, appears, which competitors are cited, and whether the answer language matches the question. Repeat monthly. Because Gabe’s tests showed uneven behavior, your log is the best evidence you will get.
Common mistakes
- Translating the entire site at once with machine output and no review.
- Missing or one-way hreflang tags.
- Treating one language as one market.
- Assuming English citations carry over to other languages.
- Reporting multipliers from a vendor study as guaranteed outcomes.
