If Your Site Is Not in English, ChatGPT May Research Your Topic Without You

Type a question in Polish and the assistant answers in Polish. What you never see is the set of background searches it ran to get there, and a good share of those may have been issued in English. Your Polish page can be the best answer on the web and still lose the slot, because the search that found the competition was never run in your language.

This matters if you publish outside English. It also has a ceiling, because the sharpest evidence comes from companies that sell products in this exact space. Read it as a strong hint, test it cheaply, and resist the instinct to translate everything on a Friday afternoon.

Two measurements that are easy to mistake for each other

There are two separate things being counted in this debate, and most write-ups blur them.

• The research step. Which language the assistant uses for the hidden searches it issues while composing an answer. Our explainer on query fan-out covers how that step works.
• The citation step. Which language the sources are in by the time they appear under the answer the reader actually sees.

A model could do its reading in English and still surface local sources, or the reverse. Keeping the two apart is the difference between a sound plan and a translation invoice you did not need.

The research step tilts towards English

Peec AI, which sells visibility tracking, worked through more than 10 million prompts and 20 million fan-outs, keeping only sessions where the person's location matched the language of the query, Polish questions from Poland and so on. Its finding: 43% of the fan-out searches behind non-English prompts were issued on the English web, and English research appeared in close to 78% of non-English sessions (Peec AI).

Tomek Rudzki, who wrote it up, describes the pattern as an assistant that opens in your language and then does much of the real digging in English. The spread by language is wide, from Turkish at the top to Spanish at the bottom, and nothing Peec measured came in under three fifths (Peec AI). A Spanish site is therefore not exempt because its number was the kindest one in the set.

Peec attributes the tilt partly to authority signals, on the grounds that English pages carry more links and more citations. The post also states that roughly half of all web content is English without naming where that comes from, so treat the explanation as the author's reading rather than a measured result.

A larger 2026 dataset points the other way

Here is where it gets interesting. An analysis published by Temso covered about 350,000 AI responses and 7,058,891 individual citations, collected across 12 countries and 47 industries in early 2026, with six non-English prompt languages: Italian, German, Swedish, Spanish, French and Dutch (Temso).

Its headline is close to the opposite of the one above. Measured on citations rather than hidden searches, Google AI Overviews matched the prompt's language 85.4% of the time, Microsoft Copilot 76.7%, ChatGPT 70.2% and Grok 51.7%. ChatGPT reached for English sources in 23.7% of citations, and Grok in 39.6% (Temso).

The variation underneath those averages is the useful part. Dutch prompts on Grok were the weakest combination in the set, with Dutch sources at 38.3% against English at 53.5%, while French on Copilot reached 87.2%. By sector, school education sat near the top of local-language citation at 76.9% and hotels and hospitality at the bottom around 35.5% (Temso). Your exposure depends on which assistant your buyers use and which industry you sell into, not on a single global average.

A third dataset, from a completely different field

The pattern turns up where nobody was selling a translation product. A team of ten researchers audited citation behaviour across ChatGPT, Perplexity and Google AI Overview, collecting 15,942 citations from 1,140 responses across 1,713 distinct domains, using twenty questions in English and a subset translated into six further languages of varying resource levels (Nguyen and colleagues, arXiv, August 2026).

Two findings carry over. Non-English queries surfaced fewer sources than the same questions asked in English. And they were routed to resources appropriate to the asker's language at a significantly lower rate (arXiv). The subject matter is health information, so nothing in it should be read as guidance about any topic other than citation behaviour. As a methodological check it is valuable: an independent group, a stated sample, a preprint you can read in full, and a direction that matches the vendor claims rather than contradicting them.

Peec and Temso cannot both be the whole story

Our read is that both are right about different steps, and that the research step is the one worth planning around. Hidden searches in English decide which candidates enter the shortlist. Citations in your language decide what the reader sees afterwards. A high local-language citation rate is cold comfort if the English-language competitor made the shortlist and shaped the summary your prospect read, with your local page cited politely underneath.

That reading leads to an unfashionable priority order. Earning a mention on the English-language pages that already win your topic beats translating your own site, because the mention affects the shortlist and the translation mostly affects the attribution. Comparison pages, industry publications and the forums your category argues in are the targets. Our guide to how products get discovered in AI search sets out that work.

You can check this yourself, and the test is specific. If, when you run your own buyer questions, the sources shown are overwhelmingly in your language and your competitors are local companies rather than English-language incumbents, then the research step is not costing you anything and the whole argument collapses into a translation question you can answer on quality grounds alone. Sectors differ enormously here, as the industry spread in the Temso data shows, so run the test before accepting either our ordering or anyone else's.

What translated sites earn, and what that does not establish

Weglot, a translation vendor, looked at 1.3 million AI citations across several AI systems. Its chief marketing officer, Eugène Ernoult, reports that translated sites averaged roughly 24% more citations per query, and that sites with no translation lost a few percent of citations when the query itself ran in English (Eugène Ernoult, on the DeepL blog).

Read the direction of that claim carefully. It says translated sites earn more. It does not say translating caused the gain, and a company that funds translation usually funds several other things that attract citations at the same time. The post does not describe its method beyond the citation count, and as far as we could find, neither it nor the Peec analysis has been reproduced by an outside party.

A test you can finish before lunch

1. Pick five questions a buyer would actually type, in your language. Not your brand name. Discovery and comparison questions.
2. Ask each one in the assistants your market uses, in a clean session, and record which assistant you used.
3. Open every source listed and tally two columns: language of the source, and whether it belongs to you, a local competitor or an English-language incumbent.
4. Write the English-language winners into a target list. Those are the pages a mention would be worth earning on.
5. Repeat next month. One run is a snapshot, and answers move between runs.

If the test shows English incumbents shaping your answers, build English versions of the handful of pages that define what you sell, written for an English reader rather than passed through a machine. Our notes on keeping AI-assisted writing in your brand's voice apply directly. If you add languages, tag and structure them properly, because the usual failures are mechanical: see fixing a broken hreflang setup.

Where all of this breaks

None of the evidence shows a penalty for publishing outside English. It shows English sources entering the same contest, which is a different problem with a different fix. Nor does any of it tell you how much of the effect reaches your market: a niche with a hundred strong English pages behaves nothing like one with three.

Do not translate a whole site on the strength of two vendor studies and a preprint. Bulk translated pages that no native speaker has reviewed read badly to humans, and search guidance has long warned about pages produced at scale that add little for the reader. Keep your local language as well. Your existing customers never stopped searching in it.

FAQs about non-English sites and AI search

Is there a penalty for publishing in a language other than English?

No evidence points to one. What the studies show is that English-language sources often compete for the same answer slot, either through the assistant's hidden research step or through the citations it displays.

Which study should I believe about English-language research?

They measure different things. Peec AI counted the background searches an assistant issues; Temso counted the citations that reach the reader. Both can hold at once, and the two together are more useful than either alone.

Is translating my whole site the right first move?

Rarely. Start with English versions of the few pages that define what you sell, and with mentions on the English-language pages that already win your topic. Expand only once your own query test shows it is paying.

How reliable are these figures?

Two of the three come from companies selling related products, and neither has been independently reproduced. The academic audit is independent but covers one subject area. Treat all three as directions worth testing, not settled measurements.