• AI visibility

Guide · AI search

How should global brands approach GEO across languages?

Treat each language market as its own AI search market, with its own sources, its own competitors and sometimes its own version of Google. Research from 2025 and 2026 shows that AI assistants draw on different websites and name different brands when the language changes. The AI layer itself is also switched on unevenly from country to country. The workable model is one global playbook fed by local evidence, local coverage and local measurement.

By Underneath Team Guide Updated

The short version

  1. Google’s AI Overviews reached 229 countries in 2025, up from 7 in 2024, yet France, Turkey, Iran, China and Cuba were still left out (Aral, Li and Zuo (opens in a new tab)).
  2. In a twelve-language European study, asking in a local brand’s home language raised how often AI named it by 0.80 on a 0-to-1 scale, against 0.15 for global brands (Żatuchin (opens in a new tab)).
  3. In Tokyo, Gemini’s Japanese answers cited hotels’ own websites more often than its English answers; the authors estimate 2.27 times more such citations and link the gap to deeper Japanese-language content (Zhu and Chang (opens in a new tab)).
  4. In our four-country test, ChatGPT’s brand picks moved between countries far more for insurance, lending and tax questions (a gap of 0.180) than for global products (0.034) (our country study).
  5. In a lab test, fake brand pages fooled AI assistants about as often in English as in Chinese: 87% of the time for restaurant questions in English, against 82% in the matched Chinese category (Luo and Chen (opens in a new tab)).

Should global companies build a separate GEO strategy for each language?

Yes for evidence and measurement, no for principles: the playbook can be global, the inputs must be local. What differs by language is which sources, competitors and wording matter, not what earns trust.

The largest multilingual study so far, by Żatuchin (opens in a new tab), collected 35,640 answers from three AI assistants that search the web, about 66 European brands in twelve languages, in April and May 2026. Switching from English to a brand’s home language raised a local champion’s recommendation share by 0.80 on a 0-to-1 scale, against 0.15 for a global brand. So how much local work you need depends on your position in each market. A multinational is named at similar rates in either language. A brand that leads at home lives or dies in its home language. The author is affiliated with an AI brand-monitoring company and discloses it. Our article on English-only AI visibility checks covers what that means for audits.

The principle that travels is earned credibility. In a University of Toronto study by Chen and colleagues (opens in a new tab), a web-enabled ChatGPT model leaned on earned media: independent reviews and publishers. It drew 77.6 percent of its sources for Canadian consumer electronics questions from earned media, and 92.1 percent for the same questions about the US. The rule held in both markets; the publications behind it were local.

Why does the same content perform differently in different languages?

Because each language has its own web, and AI assistants mostly search the web in the question’s language. Your page competes against a different set of sources in every language.

Zhu and Chang (opens in a new tab) describe Tokyo’s hotel market as two largely separate webs: Booking.com and Expedia dominate English results, while Jalan, Rakuten Travel and Ikyu dominate Japanese ones. Gemini’s answers followed the language of the question. For questions about the guest experience, 62.1% of Japanese citations came from sources other than booking sites, against 50.0% in English, because the Japanese web offered more of those sources. The Toronto team found the same pattern from another angle: changing the language moved cited sources more than rewording a question within one language.

The answers themselves differ too. In the European study, the median answer ran 289 words in English but 189 in Finnish. Finnish answers resembled answers in Germanic languages more than those in Estonian, its closest linguistic relative. The author’s plausible explanation is that Finnish business coverage leans heavily on English and Swedish sources. In other words, “language” really means the information ecosystem behind it. For how each engine localizes its sources, see our guide to local-language citations.

Does automatically translating your content keep your AI search visibility?

No study we reviewed tests this directly, and the indirect evidence says translation alone is not enough. Treat any claim otherwise as unproven.

Three findings point the same way. First, the Toronto authors write about engines that switch to local sources. Their conclusion: “simply translating your own brand content is not enough; you need earned coverage in the local language.” Second, in Tokyo, content depth by language mattered. The authors report that Japanese queries produce 2.27 times more hotel-direct citations than English ones. They link this to Japanese hotel sites carrying neighborhood guides and transit details, while English versions focused on booking. A translated booking page adds no such depth.

Third, a survey of AI visibility measurement (opens in a new tab) warns that a translated question may not preserve local availability, vocabulary or regulation. The same logic applies to pages: a literal translation can miss the words local buyers use. The Tokyo team wrote its Japanese questions as natural-sounding queries rather than literal translations for exactly this reason. This is our inference from the research, not a tested result.

Is AI search even the same product in every market?

No: whether AI answers appear at all, and how often, depends heavily on the country. Your exposure is uneven before any content work begins.

Aral, Li and Zuo (opens in a new tab) ran the same searches on Google in 2024 and 2025 across 243 countries. AI Overviews, the AI summary at the top of Google’s results, went from 7 countries to 229. In most countries that had them in 2025, they appeared on 55% to 70% of searches, but in Iceland only 7.4%. France, Turkey, Iran, China and Cuba were excluded altogether, even as 222 new countries were added. Among early markets, Japan saw a 78% rise in AI answers in a year. The country explained more of whether a search showed AI than its topic or wording.

The timing gap has real costs. Khosravi and Yoganarasimhan (opens in a new tab) used Google’s staggered rollout, US default in May 2024 and the EU only in early 2025, as a natural test. Monthly search traffic to English Wikipedia articles fell by approximately 5.45% relative to the same articles in German. The same page can feel AI’s effect a year earlier in one language than in another.

Which markets and categories need the most local attention?

Markets where you are a local champion, and categories where products, prices or rules are national. Global consumer products move least.

In our country study, we asked ChatGPT and Gemini the same 40 English buyer questions five times each from the US, UK, Canada and Australia in September 2026. ChatGPT’s brand picks moved between countries far more for insurance, lending and tax questions (a gap of 0.180 in brand overlap) than for global products such as headphones and CRM software (0.034). For small business accounting software, it led with QuickBooks in the US and Canada and with Xero in the UK and Australia.

The assistant also adapted the answer itself. ChatGPT mentioned the country, its currency or its regulators in 68.8% of UK answers, though the question named no country. If your category depends on local rules or prices, the AI answer in each market is effectively a different answer. China adds a further split, which our guide on brand visibility in Chinese versus Western AI models covers.

Do the risks to your brand cross languages too?

Yes: in one lab test, fake pages misled AI assistants about as often in English as in Chinese. Brand protection has to run in every language.

Luo and Chen (opens in a new tab) swapped real products for invented ones in the top three web pages an AI assistant read, then counted how often twelve AI models recommended the fake. They ran it mainly in Chinese, then repeated it in English with 360 trials. In English, the fake was recommended 87% of the time for San Francisco restaurants and 43% for smartphones; the matched Chinese figures were 82% and 23%. Categories where the AI knew less about real brands were the most exposed.

This was a controlled setup, not the live web. Still, a market where your brand has thin local coverage is the kind of market where false pages have the least to compete with.

What should you do about it?

Run one global GEO program with a local evidence base and local measurement in every priority market. In practice:

  1. Classify each market. Decide whether you are a local champion or a global brand there; the European study shows that this sets how much local work you need.
  2. Check the AI layer first. Confirm whether and how often AI answers appear in each country before setting targets.
  3. Map the local sources. List the publishers, review sites and directories that AI answers cite in each language, and earn coverage there, not only in English trade press.
  4. Write natively, not literally. Give local pages real local substance: prices, rules, places, the words buyers use. Have a reviewer in each market check them.
  5. Measure each market separately. Run your buyer questions in each language and location, more than once, and report them separately rather than as one global number.
  6. Watch for impersonation in every language. Fake pages worked in both languages tested, so monitor what AI says about you in each market.

If you want help planning a multi-market program, see our generative engine optimization service.

What does the research not tell us yet?

Most of the strategy questions are untested: no study measures whether translated or localized content changes AI visibility. The main gaps:

  • No controlled test compares machine-translated pages with natively written ones for AI citations.
  • Brand-level evidence covers twelve European languages, five major world languages, Japanese hotels and four English-speaking countries. Most other markets have only data on whether Google shows AI answers.
  • The largest multilingual study was written by an author affiliated with a brand-monitoring company and covers one window in April and May 2026.
  • The link between local sources and local brands is an association; no study has changed a market’s sources to test cause.
  • AI products and rollouts change quickly; country-level availability in 2025 may already differ today.

Frequently asked questions

Should we translate our website for AI search in other markets?

Translate, but do not stop there. No study shows translation alone keeps AI visibility, and the research points to local earned coverage and locally useful content as what AI assistants cite.

Do AI assistants recommend different brands in different languages?

Yes. In 35,640 answers about 66 European brands, asking in a local champion’s home language raised how often it was named by 0.80 on a 0-to-1 scale, against 0.15 for global brands.

Do Google AI Overviews appear in every country?

No. A study of 243 countries found them in 229 in 2025, with France, Turkey, Iran, China and Cuba excluded. They appeared on 55% to 70% of searches in most countries, but 7.4% in Iceland.

Is our English content enough for AI visibility in Europe?

For a global brand it may carry some weight; for a local leader it is not enough. The European study found local champions’ visibility depends mainly on answers in their home language.

Sources

Free strategy call

Some questions are easier to answer about your own business.

Bring the one that matters most. On a free 30-minute call we’ll take a first look at it and send you a short written read afterward.

Ask about my businessHow we work