Yes, it can. A translated page can be cited in an AI answer, and Google's documentation sets the same bar as for any other page: it must be indexed and eligible to show in Search with a snippet, with no additional technical requirements. That is the floor. What no Google document tells you is how the language of the question changes which pages get picked. The loudest claim on that comes from a translation vendor, so treat it as a hypothesis rather than a rule. Below: what's documented, what translation does and doesn't buy you, the technical hygiene, a fixed-prompt test, and a per-locale checklist.
How AI answers pick sources when the question is in another language
Start with what Google actually says. AI Overviews and AI Mode surface supporting links from Google Search, and both may use query fan-out: several related searches across subtopics and data sources, with models finding more supporting pages while drafting the answer. That matters. Google says that produces a wider and more diverse set of links than a classic search, but also notes the two features may use different models and techniques, meaning links will vary. AI Overviews appear only when additive to classic Search. They often don't trigger.
Notice what's missing. Nothing on that page mentions the query language, so everything below about language is inference, and I'll label it as such.
My reading is this. Fan-out still queries Search systems, and a search in Japanese will be matched against pages that Search understands as Japanese. Google determines a page's language from its visible content, not from the URL or lang attributes. If your Japanese page is a real Japanese page, it enters the pool. If it's English with a translated menu, it isn't a Japanese page at all. Again, that is my inference, not a statement Google makes about AI features.
The vendor claim goes further. Weglot, which sells website translation, says that asking the same question in a local language instead of English shifted a large share of citations toward local, in-language domains, most dramatically in Japanese. It also says AI systems strongly prefer citing content in the user's query language. That is plausible. But Weglot has an obvious commercial interest, and I'd want to see the underlying numbers before betting a budget on it.
Two more data points to hold loosely:
- An Ahrefs study, reported via Trovance, found 38% of pages cited in AI Overviews also ranked in the top 10 for the same query. That's from 863,000 keyword results pages and 4 million AI Overview URLs. An earlier July 2025 Ahrefs study found 76%, but it counted only the top three citations per Overview.
- Straight North notes that top organic ranking doesn't guarantee a citation, and that source patterns differ by question, geography and product experience.
Ranking in local-language results certainly helps. It does not settle anything, which is exactly why you need to test by language rather than assume.
What translation alone does and doesn't buy you
It buys existence. Without a page in the language, there is no chance of being retrieved for a query in that language. Weglot's phrasing is blunt and, as a statement about eligibility, hard to argue with: partial translation creates partial visibility, and the content you don't translate is content AI cannot cite. The classic gap is familiar. A translated homepage sitting on an English-only blog is exactly that gap.
It also buys correct duplicate treatment. Google says localised versions are only considered duplicates if the main content of the page remains untranslated. Translate the body and you have separate pages. Translate only the navigation and footer, and you have a duplicate with a new coat of paint.
What it doesn't buy:
- Local intent. People don't just use different words in German, they often ask different questions.
- Local facts. The wrong currency, a US phone number, shipping terms that don't apply in that country.
- Local corroboration. A page citing only English-language sources reads like an import.
- Special treatment. Google says specific optimisation isn't required for AI features and the same SEO foundations apply. A translation plugin is not a GEO tactic.
If you're still choosing how to translate, I've covered the mechanics in translating a site with AI and, for WordPress, translating a WordPress site in 2026. The point here is narrower. Translation is the entry fee, not the strategy.
I work with a 30-language, 91,000-page site, and at that size the temptation is to ship raw machine output everywhere and call it done. Don't do that. Decide which pages earn a native editorial pass, the ones you'd want cited, and which can sit at a lighter tier.
Locale-native answer passages and local sources
Here's where I'd spend the human hours.
An AI answer is assembled from passages. If the question is, say, how to register a small business, the passage that gets used has to answer that question in that country, in that language, in the register people actually use. A faithful translation of a UK-centric explainer won't do it. Write or rewrite the opening of each important page so a reader in that locale gets the direct answer in the first few lines, then the detail.
Local sources
Cite what a local reader would trust: a national regulator, a local standards body, a domestic trade association, or a statistic from the country's own statistics office. This does two jobs: it makes the page more useful, and it ties your content to the same local web the question is being answered from. I cannot point to a Google document stating outbound citations lift AI citation, so treat this as good editorial practice rather than a proven lever.
Per-locale entity facts
Every locale should state its own facts, in visible text, on its own pages:
- Legal entity name and registered address for that market
- Currency, tax treatment and delivery terms
- Local support hours and contact routes
- Local product names, units and date formats
- Local credentials, certifications or awards
If you mark any of this up, Google lists making sure structured data matches the visible text among the practices that remain worthwhile. So update the schema per locale too. Copying the English Organisation markup across 30 languages is how you end up with a French page claiming a London phone line.
Technical foundations: hreflang, canonicals and duplicate signals
This is the boring part, and it's where large multilingual sites quietly leak. Google says that telling it about localised versions helps it point users to the most appropriate version by language or region. Nobody has documented hreflang as an AI-citation lever, so I'd frame it as protecting the pages that feed the same index. Get the right URL into the right market and you've done your part.

hreflang rules that actually bite
- Each language version must list itself as well as every other version. Alternate URLs must be fully qualified, and they don't have to share a domain.
- If two pages don't point at each other, the tags are ignored. Bidirectional or nothing.
- If full bidirectional sets get unmanageable, you can omit some languages on some pages, but link newly added languages back to your originating or dominant language.
- For several same-language, different-country URLs (en-IE, en-CA, en-AU), Google suggests a generic
encatchall, andx-defaultas the fallback for unmatched languages. - Only ISO 639-1 language codes and ISO 3166-1 Alpha 2 region codes work. Mind the traps:
beis Belarusian, not Belgium. And EU, UN or UK as a region value has no effect on Google Search. - Script variants go through ISO 15924, such as
zh-Hansandzh-Hant. - Pick one method: HTML, HTTP headers or sitemap. Google says they're equivalent and using all three gives no benefit.
And a reminder that hreflang is a hint about which URL to serve. It doesn't tell Google a page's language. Google uses its own algorithms for that.
Canonicals: older advice versus current advice
Old posts trip people up here. A 2010 Google post proposed pairing rel=canonical or 301s with hreflang, but it was explicitly not for sites that fully translate each page. It targeted pages where only the template was localised. A December 2011 post then carried an update saying Google no longer recommends rel=canonical for same-language, multi-country pages.
Current rule of thumb from Google's canonical documentation: if you use hreflang, specify a canonical in the same language (or the best substitute if none exists). Make each page self-canonical, and don't send conflicting canonical signals through different methods. Redirects and rel=canonical are strong signals, sitemap inclusion is weak, and they stack. So a French page canonicalised to the English one is a self-inflicted wound.
Avoiding duplicate and conflicting signals
- Give each language its own URL. Google recommends this over cookies or browser settings, since Googlebot usually crawls from the USA without an
Accept-Languageheader. - Don't auto-redirect between language versions. Link to them instead.
- One language per page for content and navigation. No side-by-side translations.
- Skip locale-adaptive pages that swap content by perceived location. Google's international overview warns it might not crawl, index or rank all of it.
- Think hard before geotargeting a whole site to one country. It can help there, at the expense of other locales.
How to test citation by language with a fixed prompt set
You can't read this off a dashboard, so build a small, boring, repeatable experiment. I'm describing a procedure here, not reporting results.
- Pick intents, not keywords. Choose a set of questions your pages should answer. Keep the same set across every locale.
- Write each prompt natively. Have a native speaker phrase the question as a local would ask it. Word-for-word translations of English prompts test your translator, not your visibility.
- Freeze the set. Same prompts, same wording, each run. Changing prompts mid-stream makes trends meaningless.
- Run per engine and per locale. Google's AI Overviews and AI Mode can differ, and Overviews often don't trigger at all, so record when nothing appears. Repeat runs, because answers vary.
- Log the result, not a screenshot of vibes.
An illustrative log structure (the rows are placeholders, not data):
| Intent | Language | Engine | Cited URL | Language of cited page | Ours? |
|---|---|---|---|---|---|
| Intent A | German | AI Mode | (record URL) | (record) | yes/no |
| Intent A | Japanese | AI Overviews | (none shown) | n/a | n/a |
| Intent B | Spanish | AI Mode | (record URL) | (record) | yes/no |
The column worth staring at is the language of the cited page. If your German prompt keeps citing English pages (yours or someone else's), either your German page isn't competitive or it isn't indexed properly. Check indexing first. Then compare against Search Console: AI feature traffic is reported in the Performance report under the Web search type, folded into overall search traffic, so segment by country and page group to watch each locale. I've written more on tooling in measuring AI search visibility.
Honest caveat: a fixed prompt set shows you your own pattern, not a law of nature. Take small samples as directional.
The checklist
Work through this per locale, not per site. A locale that fails three of these is probably invisible, whatever the English version does.
- The page is indexed and eligible to show with a snippet (the only technical requirement Google names).
- The main content is translated, not just the template, in a single language on the page.
- Each locale has its own crawlable URL, with no cookie-based or auto-redirect language switching.
- hreflang is reciprocal, self-referencing, fully qualified, uses valid ISO codes, and is set in one method only.
x-defaultpoints to a sensible fallback or selector page.- Each page is canonical to itself, in its own language.
- The opening of each key page answers a locally phrased question directly.
- Currency, units, dates, legal entity, address and support details match the market.
- At least some cited sources are local, authoritative and in the page's language.
- Structured data matches the visible text, per locale.
- Blog and resource content is translated for priority markets, not just the core pages.
- A fixed, natively written prompt set runs per locale on a schedule, and results are logged with dates.
If keeping all of this consistent across dozens of locales is the sticking point, there's more on how I approach it on the multilingual SEO page.
FAQ
Does hreflang make AI answers cite my local page?
Not as far as any Google document says. Hreflang helps Google point users to the right language or region version in Search. The AI features' only stated requirement is an indexed, snippet-eligible page, so hreflang matters indirectly, by keeping the right URLs indexed and associated.
Is a partly translated site worse than an untranslated one?
Weglot argues partial translation means partial visibility, because untranslated content can't be cited in that language. Google's side is simpler: pages whose main content stays untranslated are the ones treated as duplicates. Either way, a half-translated section tends to be the weakest part of your international footprint.
How do I handle Latin American Spanish?
Codes like es-419 aren't supported. Use es as the generic catchall, plus country-specific values such as es-MX where you really do have separate pages for a market.
Will AI Overviews show up in every language?
I can't tell you from the documentation. Google says Overviews appear only when its systems judge them additive to classic Search and that they often don't trigger. Whether that varies by language isn't something Google's public pages address, which is one more reason to log empty results in your tests.
The one fact to carry away: Google's own documents give a translated page exactly the same entry requirement as any other, and say nothing about query language, so every stronger claim is a hypothesis you should test with your own fixed prompts. Treat the vendor figures with suspicion until you see numbers. And fix the plumbing (self-canonicals, reciprocal hreflang, real translated main content) before you spend anything on cleverness.