How to Do AI SEO in Multiple Languages
Write each language version for the people who speak it, give every version its own URL, and link the versions together with hreflang tags. Test real prompts in each language, because AI answers can differ between them. Machine translation alone is a risk: Google warns that mass-produced pages with little value, including translated ones, can break its spam policy 13.
- Google AI Overviews run in dozens of languages, so a non-English audience already sees AI summaries 45.
- Hreflang links language versions together, and Google ignores it unless the pages link back to each other 1.
- Google lists translating pages in bulk without adding value as scaled content abuse 3.
- We found no public data on how often assistants search in the user's language, so test your own prompts.
How should you approach AI SEO in more than one language?
Treat each language as its own market with its own questions, examples and proof. Build the pages that matter most in the language of your best customers first, and add others as demand appears. Translating every page at once is the fast route to thin pages.
Google's AI Overviews are available in many countries and dozens of languages, including English, Spanish, French, German, Hindi, Japanese and both forms of Chinese 4. Alphabet's CEO said in July 2025 that they reached more than 200 countries and territories and 40 languages 5. Your buyers may already see AI summaries in their own language.
What do you need, and how long will it take?
You need a native speaker for each language, a list of your ten most valuable pages, and a developer who can edit page headers or the sitemap. Plan one week to audit and set up the structure, then two to four weeks per language for the first batch of pages.
The speaker should be a subject-matter reader, not only a proofreader. A translator who does not know your trade will miss the words buyers actually type.
Should you write natively or translate?
Write natively for pages that win customers, and translate with human review for stable reference pages. A buyer in Madrid does not phrase a question like a buyer in Chicago. Local words, prices, laws and examples change the answer, so a literal translation often answers a question nobody asks.
Google's spam policies name the risk. Scaled content abuse covers many pages made mainly to manipulate rankings, with examples such as using generative AI to produce many pages without adding value, and automated transformations like synonymizing or translating 3. Translation is fine when it helps users. It becomes a problem when it is a switch that creates hundreds of pages nobody checked.
Google also says localized pages count as duplicates only if their main content stays untranslated 1. So a real translation does not cause a duplicate problem, but a page with only the menu translated does harm. Google warns against translating only the boilerplate while leaving the main content in one language 2.
| Page type | Best route | Why |
|---|---|---|
| Service and product pages | Native writing | Local terms, prices, proof and legal wording differ |
| Pricing and policies | Native writing, reviewed by legal | Wrong local detail damages trust |
| Definitions and how-to guides | Translation plus human edit | The facts are stable; wording needs a native pass |
| News and opinion | Native only | Context is local |
| Low-value or duplicate pages | Do not translate | Adds pages without adding value 3 |
How do you set up hreflang so the versions find each other?
Hreflang tells Google that pages are localized versions of the same content. You can add it in the page header, in an HTTP header or in the sitemap, and Google says combining methods gives no extra benefit 1. Pick one.
Four rules prevent most errors. Each version must list itself and every other version. If two pages do not point to each other, the tags are ignored. Use an ISO 639-1 language code, optionally followed by an ISO 3166-1 Alpha 2 region code, and never a region code alone. Add an x-default entry for users whose language matches no version 1.
Google also says it does not use hreflang or the HTML lang attribute to detect the language of a page 1. It reads the page content. So keep one language per page, for content and navigation, and avoid side-by-side translations 2.
Hreflang is a Google documentation topic. We did not find equivalent official guidance from OpenAI, Perplexity or Anthropic, so do not assume other assistants read the tags. Clear URLs, page language and content do the work there.
How does the language of the prompt change what AI cites?
We found no reliable public study that measures this across assistants, so the safe answer is to test. A user may ask in Spanish and the assistant may search in Spanish, in English or in both. The cited pages can differ from your English results.
Run the same ten questions in each target language, in fresh chats, and record which pages are cited and in which language. Use the query fan-out explorer to see the searches an assistant runs. If your pages are cited for the English question but not for the Spanish one, you have found a gap worth filling.
Also test mixed phrasing. Many people in multilingual markets mix languages or type local words in Latin letters. This is a modelled pattern, not something we measured, so confirm it from your own search data and customer emails before you write for it.
How do you write each version so an assistant can quote it?
Use the same rules in every language: a direct answer in the first two sentences, clear headings, specific numbers with sources and a short FAQ. Google says AI systems understand synonyms and general meaning, so you do not need exotic phrasing or exhaustive keyword lists 6. Write the way a helpful local expert speaks.
Cite local sources. A German buyer trusts a German regulator or trade body more than a US blog. Use local currency, units, dates and addresses, and name local places. Our guide to answer-first introductions applies in any language.
Say plainly when a page was translated or machine-assisted. Google's guidance on helpful content notes that readers benefit from knowing how content was created 7.
Where should the language versions live?
A subfolder per language, like example.com/de/, is the easiest choice for most sites. Google's table lists it as easy to set up and low maintenance on the same host, with the weakness that users may not recognise the geotargeting from the URL alone. Country domains are expensive, need more infrastructure and can target only one country 2.
Our guide on choosing between a subfolder and a subdomain covers robots.txt and reporting for each option.
What mistakes should you avoid?
Avoid auto-redirecting visitors by their location, which can block crawlers and users who want another language. Avoid mixing two languages on one page, and avoid hreflang tags without return links. Avoid publishing a language you cannot support with customer service and legal accuracy.
Also avoid ignoring regional spelling. English for the UK, the US and India differs in terms and prices, and Google's hreflang format lets you separate them with region codes such as en-GB and en-US 1.
How do you check it worked?
Check in three places. In Search Console, filter by country and page to see impressions for each version. In a prompt log, record whether each language version is cited. In your enquiry form, add a language field and a source question.
Give it time. Google says changes can take hours to months and suggests waiting a few weeks 8. Review each language after four and eight weeks.
When is a second language not worth it?
Skip it when you have no customers, support or legal cover in that language, or when your buyers already search in English. A thin, unsupported language version can hurt trust more than no version at all.
Frequently asked questions
Do AI assistants read hreflang tags?
Google documents hreflang for Google Search and says it ignores the tags unless pages link back to each other 1. We found no official statement from OpenAI, Perplexity or Anthropic that their crawlers use hreflang, so rely on clear URLs and native page content for them.
Is machine translation bad for AI SEO?
Not by itself. The risk is publishing many translated pages without review or added value, which Google lists under scaled content abuse 3. Translate a small set of valuable pages, have a native expert edit them, and check facts and local rules.
Which languages do Google AI Overviews support?
Should each language have its own domain?
Usually not. A subfolder per language is easier and cheaper to run. Country domains give clear geotargeting but cost more and need more infrastructure 2. Choose them only for a clear business reason.
Sources
- Tell Google about localized versions of your page · Google Search CentralOpened October 20, 2026. Supports: Three equivalent hreflang methods; each version lists itself; return links required; ISO codes; x-default; Google does not use hreflang or lang to detect page language; translated localized pages are not duplicates. Last updated 2026-09-21.
- Managing multi-regional and multilingual sites · Google Search CentralOpened October 20, 2026. Supports: URL structure pros and cons (ccTLD, subdomain, subdirectory); one language per page; avoid side-by-side translations; avoid translating only boilerplate. Last updated 2025-12-10.
- Spam policies for Google web search · Google Search CentralOpened October 20, 2026. Supports: Scaled content abuse examples including generative AI pages without value and synonymizing or translating; doorway abuse. Last updated 2026-08-28.
- AI Overviews availability · Google Search HelpOpened October 20, 2026. Supports: AI Overviews available in many countries and dozens of languages including English, Spanish, French, German, Hindi, Japanese, Chinese (Simplified and Traditional). Page carries no date.
- Q2 earnings call: Remarks from our CEO · Google (Alphabet)Opened October 20, 2026. Supports: AI Overviews had over 2 billion monthly users across more than 200 countries and territories and 40 languages (July 2025). Company-reported.
- Optimizing your website for generative AI features on Google Search · Google Search CentralOpened October 20, 2026. Supports: AI systems understand synonyms and general meaning; no need for special phrasing or exhaustive long-tail coverage. Last updated 2026-07-10.
- Creating helpful, reliable, people-first content · Google Search CentralOpened October 20, 2026. Supports: Disclosures about how content was created, including automation or AI, are useful where readers might ask how it was made. Last updated 2026-10-05.
- SEO Starter Guide: The Basics · Google Search CentralOpened October 20, 2026. Supports: Changes can take hours to several months; wait a few weeks to assess effects. Last updated 2025-12-10.