Multilingual GEO: How AI Citations Change With the Language You Ask In

Ask an AI engine the same question in English, Italian or Spanish and you often get different cited sources. Independent research shows models pivot to English pages when local sources are thin, and the click-through benchmarks most marketers quote were measured only in the United States. For a multilingual European business, that means every imported US number carries a language ceiling — and native-language measurement is not optional.

Why doesn’t the click-through benchmark everyone quotes apply to your market?

The single most-cited number in generative search is a US number. Pew Research (July 2025) tracked the browsing of 900 US adults in March 2025 and found that when a Google AI summary appeared, users clicked a traditional result on 8% of visits, versus 15% when no summary appeared; only 1% clicked a source link inside the summary itself. It is a precise, useful figure — for English-speaking America. It says nothing about Milan or Madrid, because Google only switched on AI Overviews in Italy and Spain on 26 March 2025, per Search Engine Land (March 2025) — the same month Pew’s data window closed. Many benchmarks now circulating literally could not have measured those markets. We treat US click-through and overlap figures as directional in our GEO vs SEO breakdown, never as your number.

Do AI engines really cite different sources depending on the query language?

The strongest independent evidence comes from academia, not vendors. A Stanford-led study (Suzgun et al., arXiv:2605.22785, February 2026) put six commercial chatbots — Gemini 3 Flash and Pro, Grok 4, Claude 4.5 Sonnet, GPT-5 and GPT-4o mini — through 2,100 same-day news questions across six regional BBC services, 12,600 model-question instances in all. The retrieval pattern was lopsided:

“Nine of ten are primarily English-language despite four of six regions covering non-English content” — Suzgun et al., Evaluating Commercial AI Chatbots as News Intermediaries, February 2026

Nine of the ten most-cited domains globally were English-language, even though four of the six regions covered non-English content. One caveat we keep front and centre: this paper measures factual accuracy on news, not brand or commercial citations. Use it as evidence of the mechanism — an English retrieval bias — not as a study of how businesses get cited.

Is this a comprehension problem or a retrieval problem?

Retrieval, and the distinction matters. The models understand the languages fine; they fail at finding good local sources to ground their answers. In the Stanford data, accuracy on Hindi-language questions fell to roughly 79%, against 88.9-91.3% for the other regions — not because the model misread Hindi, but because it reached for English pages when well-indexed local sources were missing.

“a retrieval-and-grounding failure in which models pivot to English-language sources” — Suzgun et al., February 2026

More than 70% of all errors in the study were retrieval failures. The practical lesson: if your language’s web is thinly indexed on a topic, the engine defaults to English — and to whoever is well-sourced in English. That retrieval pipeline, query fan-out and all, is the same one we unpack in how AI engines pick their citations.

What do the GEO vendors add — and how far can we trust it?

Two vendors have published the largest multilingual citation analyses so far. Treat their numbers as informed and directional, not reproducible: both are interested parties, and neither releases an open dataset.

Profound (April 2026) analysed 3.25 billion citations across seven models and 14 countries, filtering every prompt by the country’s native language, and reached a blunt conclusion:

“a country’s geography is secondary to a much more powerful force: the language of the query” — Profound, April 2026 (vendor data)

Temso AI (early 2026) went further on the source-language question, analysing 7,058,891 citations across four models, six non-English languages (Italian and Spanish among them) and 47 sectors. Its “local-language citation rate” — the share of cited pages whose language matches the prompt — splits the models wide open:

AI surface	Local-language citation rate
Google AI Overview	85.4%
Microsoft Copilot	76.7%
ChatGPT	70.2%
Grok	51.7%

Source: Temso AI, early 2026 (vendor data, no public dataset).

That is a 34-point gap between the most local model and the least. Temso reports the same split inside a single language: for Dutch prompts, Grok cited more English sources (53.5%) than Dutch (38.3%), while the identical prompts in Google AI Overview returned 81.2% Dutch. Sector matters too — local-language citation ran at 76.9% for K-12 education but just 35.5% for hotels and hospitality, because globalised verticals lean on English content. The honest gap: there is still no independent, non-vendor study measuring brand citation by language on an open dataset. The mechanism is well-evidenced; the commercial figures are vendor-supplied.

How many Europeans actually use generative AI — and does a small base mean no opportunity?

Enough to matter, and unevenly. Eurostat (December 2025) reports that 32.7% of people aged 16-74 in the EU used generative AI in the three months before the survey. The spread is enormous: Denmark (48.4%), Estonia (46.6%) and Malta (46.5%) lead, while Italy sits near the bottom at 19.9%, second-lowest after Romania (17.8%). A low base is not an absent market — it is an early one, with less entrenched competition for citations. We track those adoption curves in AI search adoption in Europe.

What does this mean for every US statistic in this blog?

Every one of them carries a language ceiling. A click-through rate, a citation-overlap percentage or a conversion figure measured on US English-language traffic tells you about a market that had AI Overviews for months longer than Europe did, in the language the models retrieve most abundantly. It is a legitimate directional signal. It is not your baseline. Where we cite a US-only figure, treat it as observed-in-the-US and assumed-elsewhere — an assumption, not a measurement, until someone measures your language. Extending any of this to a smaller market such as Slovenian, which none of these studies covered, is pure extrapolation and should be labelled as such.

What should a multilingual SME do about it?

A sequence that respects the evidence, ordered by leverage:

1. Measure in the market’s real language. Ask ChatGPT, Perplexity and Google (with AI Overviews) your customers’ questions in Italian, Spanish or Slovenian — not in English — and log who gets cited. An English baseline predicts the wrong thing.
2. Check the answer per model, not in aggregate. The same Dutch prompt swung from 38.3% to 81.2% local sources depending on the engine (Temso); assume your language behaves differently on Grok than on Google AI Overview.
3. Publish native, indexable content. Not English translated on the fly, but pages written in the market’s language that a crawler can fetch and parse — so the model has a local source to ground on instead of pivoting to English.
4. Down-weight imported benchmarks. Use US numbers to understand direction, then replace them with your own measured baseline.

The through-line is simple: because the language of the query changes what gets cited, AI visibility has to be measured in the real language of each target market, not extrapolated from English. That principle — audit first, measure in-language, label every number as measured, calculated or assumed — is exactly how our AI visibility service approaches it, on the same SEO foundations that let a crawler reach your pages in the first place.