Most AI SEO advice quietly assumes an English-speaking audience. If you serve customers in several countries, that assumption can cost you visibility. AI answer engines treat languages differently, and your English page may not be what a French, Hindi or Japanese searcher is shown.
This article looks at what the evidence says about multilingual AI search, and what a worldwide site should do about it.
AI Search Is Already Global
Google’s AI Mode is no longer English-only. In October 2025, Google announced it was launching in more than 35 new languages and over 40 new countries and territories, reaching over 200 countries and territories in total. Google said its model was built to grasp the subtleties of local languages.
That is Google’s own claim about its own product. The next section tests it against independent data.
AI Engines Prefer Local-Language Sources, Unevenly
A study from Temso, titled “Lost in Translation,” examined over 7 million AI-generated citations across four models, with prompts in seven languages. It measured how often each engine cited English sources in response to non-English prompts:
- Google AI Overviews: 8.7%
- Microsoft Copilot: 15.4%
- ChatGPT: 23.7%
- Grok: 39.6%
The practical reading is that Google’s AI features lean strongly toward local-language pages, while some other engines fall back to English far more often. The study’s own note is worth remembering: the prompt set is commercially oriented, so results may differ for other kinds of questions. It also comes from a single vendor, so treat the numbers as indicative.
If your audience uses Google, a quality page in their language matters a great deal. If they use other assistants, an English page may still get cited. Either way, relying on English alone is a gamble.
Language Versions Do Not Always Match
Search consultant Glenn Gabe tested how different tools handle multilingual queries using Google’s own documentation. As summarised by Stan Ventures, Google and Bing generally showed the correct language version where hreflang was in place, while AI Overviews and AI Mode often got it right, but not always. In one test, AI Mode and ChatGPT surfaced both Spanish and English URLs, while Perplexity and Claude returned the English page. stanventuresstanventures
This is a small test, not a large study, so do not generalise too far. It does show that serving the right language version cannot be taken for granted in AI answers.
What a Worldwide Site Should Do
- Write natively, not mechanically. Machine translation makes pages exist, but it does not make them useful. Local searchers use different words, ask different questions and care about different details. Have a native speaker adapt examples, units, prices, regulations and tone.
One translation vendor reports that translated sites earn on average 24% more total citations per query, including a 16% increase in citations for the original content. That figure comes from a company that sells translation, so I would not plan around it. The logic is still sound: AI can only surface content that exists in the user’s language.
- Implement hreflang correctly. Google’s documentation uses hreflang tags to signal which version belongs to which language and country. Give every version a self-referencing tag, keep the annotations consistent across versions, and test with real searches. A vendor guide warns that misconfigured hreflang can cause the wrong regional content to appear in AI Overviews. I could not verify that against Google’s text, so treat it as a plausible risk, not a proven one.
- Build local-language mentions. Earlier AI SEO research suggests off-site references influence citations. For each market, that means reviews, forum answers, directories and press in the local language. A glowing English review does little for a Spanish query.
- Prioritise markets by platform. Do not translate everything at once. Work out which AI tools your customers in each country actually use, and start where the audience and your commercial value overlap.
- Test every language yourself. Ask your target questions in each language across several AI engines, ideally from a device set to that region. Record which URL is cited, in which language, and whether the answer describes your business correctly. Repeat monthly.
Mistakes to Avoid
- Auto-redirecting by IP address. It can hide language versions from crawlers and from users who prefer another language.
- Translating only the main text. Menus, titles, structured data, image text and metadata should match the language of the page.
- Duplicating English pages with a thin translation. Quality matters more than coverage.
- Assuming one rollout fits all. One source notes that AI systems trained mainly on English data tend to have weaker knowledge and citation patterns for non-English content. Expect more variation in smaller languages. overthetopseo
How Sarkar SEO Approaches Multilingual AI Search
At Sarkar SEO, we start international projects by testing, not translating. Sarkar SEO runs the same set of buyer questions in each target language and records which pages the engines cite, so the work begins from evidence about each market.
That often shows that two countries speaking the same language need different pages, and that one well-built local page beats five rushed translations. Sarkar SEO then sequences the rollout by commercial value, so the first markets pay for the next. If you want help doing this for your own site, Sarkar SEO can map it across your markets.
