Short answer. On multilingual sites, ChatGPT fetches the English version about 2.6 times more often than the site's language mix would predict, reaching for it 65–79% of the time. Copilot is close to neutral at 1.07, and Google AI slightly under-weights English at 0.79. The bias is engine-specific, which means "AI visibility" is not one problem across markets — it is a different problem on each engine, and your German page's main competitor is frequently your own English one.
The first serious measurement of language preference across AI answer surfaces was published on 5 August 2026, and its headline is that the engines behave nothing like each other. That single finding invalidates most multilingual AI visibility plans, which assume one bias and one fix.
This piece sets out what was measured, where the bias comes from, what it costs, and what actually works — which differs by engine.
How large is the bias, and on which engines?
OMcollective's study, reported by Search Engine Land, measured how often AI surfaces fetched the English version of a multilingual site relative to English's share of that site.
| ChatGPT | Copilot | Google AI | |
|---|---|---|---|
| English over-index factor | 2.6× median (range 1.9–6.5) | 1.07 | 0.79 |
| How often the English version is fetched | 65–79% | — | — |
| Direction | Strongly favours English | Effectively neutral | Slightly favours local |
The study also found that sites with an English folder drew 892 Copilot citations per 10,000 impressions against 585 without — a 52% difference.
The sample sizes are small and that matters. The panels were 272 Bing properties for Copilot, 26 multilingual properties for ChatGPT, and a five-site subpanel on a 30-day trailing window. Twenty-six properties is not a census. The reason it is worth acting on anyway is that it measures something nobody else has measured at all, and the effect size on ChatGPT sits far outside the noise floor. Treat the direction as reliable and the exact multiplier as provisional.
Where does the bias come from?
None of this is a policy decision by the engines. It falls out of how a retrieval pipeline is built, and it compounds at each stage rather than appearing at one.
- The training corpus is skewed. The web is disproportionately English, so a model's internal representation of almost every topic is anchored to English-language sources.
- Query rewriting drifts to English. A retrieval system that rewrites a user's question into search queries will often produce English queries even for a non-English question, because the model's strongest associations are English.
- Retrieval then finds English documents. Given English queries, the index returns English pages.
- Ranking signals favour the older, larger corpus. English pages on a given topic tend to carry more inbound references, more derivative coverage and longer histories.
The practical consequence is uncomfortable: a French page can be the better answer to a French question and lose to an English page on the same site.
How does this show up on a P&L?
- Local-language pages get built and never cited. The most expensive failure mode. Money is spent on translation, the pages are correct, and the engine keeps quoting the English original.
- The brand is described in the wrong register. When the English page is the cited one, buyers in other markets read positioning, pricing framing and compliance language written for a different market entirely.
- Market-level reporting becomes meaningless. If visibility is measured on English prompts only, non-English markets appear to perform exactly as well as the home market — because they were never measured.
What actually works, by engine?
There is no single fix, because there is no single bias.
| ChatGPT | Copilot | Google AI | |
|---|---|---|---|
| Measured English bias | 2.6× (strong) | 1.07 (neutral) | 0.79 (slightly reversed) |
| Publish an English version? | Yes — it is what gets fetched | Yes — 52% more citations per impression | Lower priority |
| Local-language investment | Still required, for accuracy of description | Strong return | Strongest return |
| What to measure | Whether the English or the local page is cited | Citations per impression, by folder | Local-language prompt set |
Across all three, four things hold regardless of engine.
- Run the prompt set in the language of the market. Not translated from an English list — written by someone who asks questions that way. A translated prompt set measures your translation, not your market.
- Publish an English version and a local version, and make the relationship explicit. Correct
hreflang, distinct canonical URLs, no automatic redirect that hides one from a crawler. If the engine is going to prefer English, the English page should at least be accurate for that market. - Have a native reviewer in the market sign off the claims. Machine translation produces text that is fluent and subtly wrong on exactly the things a buyer checks: regulatory language, service names, units, entity names. Those are the passages an engine lifts.
- Measure engines and markets as a matrix, never averaged. One number that averages ChatGPT-in-English with Google-in-Japanese describes nothing that exists.
Are your markets even running these engines?
A multilingual plan built only on ChatGPT, Gemini and Copilot silently assumes every market uses them. Several large ones do not: Naver is reported between roughly 42% and 63% of Korean search depending on the measurement source, and Baidu held around 45% of all-device Chinese search in April 2026. Where the dominant surface is a domestic portal rather than a Western answer engine, the English-bias question is downstream of a larger one — whether you are measuring the engine your buyers actually use. That is a separate scoping exercise, covered in AEO in the markets Google does not own.
The compounding problem is worth naming, though. In a Japanese or Korean market a brand can face both at once: the local dominant engine may not be the one being measured, and on the engines that are being measured, its local-language page is competing against its own English page.
Does translation fix the bigger half?
No — it addresses the minority of the problem in every market.
Roughly 85% of AI references point to third-party platforms rather than brand-owned sites. If most citations point away from your domain, then in every market the decisive question is what the local third-party sources say about you: the local trade press, the local directories, the local forums and question sites.
Those are not translation projects. They are presence projects, and they are the part that cannot be centralised.
What this analysis does not settle
- The multipliers are provisional. Small panels, one publication date. The direction on ChatGPT is well outside noise; the exact 2.6 is not a constant to plan a budget around.
- Engine behaviour changes without notice. A language strategy tuned to one quarter's engine behaviour will need re-testing against the next.
- Nothing here promises a citation in any language. No engine offers submission, placement or a guarantee. The realistic goal is being the accurate, retrievable, locally-reviewed source when the dice land.
How Lifewood approaches this
Lifewood reports multilingual AI visibility as a matrix of engine by market, never as a single score, because a blended figure hides the only cell you could act on. Each cell is measured on a question set written natively in that market's language rather than translated, and the report records whether the English page or the local page was the one cited — which is the specific reading this study makes possible and most programmes never take.
Where the English page is being fetched, the response is not to abandon the local page. It is to make the English page accurate for that market, keep hreflang and canonicals clean so both remain reachable, and have a native reviewer in-market sign off the claims on both, because the passages engines lift are precisely the ones machine translation gets subtly wrong.
That is a staffing model rather than a tooling one. 50+ languages, 40+ delivery centres across 30+ countries and 56,788 registered contributors are what make in-market authorship and review the default rather than the exception. See multilingual AI visibility services and GEO services.
Sources and further reading
- OMcollective, English-language bias across AI platforms, 5 August 2026, reported by Search Engine Land — every over-index figure above, with its stated sample limits.
- StatCounter, Internet Trend and regional trackers, search engine share by country 2026, compiled by Geotargetly.
- SerpSculpt, Search engine statistics by country, 2026.
- Omnibound, Answer Engine Optimization statistics 2026 — third-party citation share.