New research shows ChatGPT strongly favors English pages, unlike Google AI and Copilot. Here’s what that means for your site.
If you feel English pages outrank (or “outcite”) your local-language ones in AI answers, you’re not imagining it.
We analyzed data from Copilot, Google AI, and ChatGPT across dozens of multilingual sites to measure bias and determine whether adding an /en/ folder to your site actually helps.
Anyone who runs a site outside the English-speaking world has seen it. You ask an AI assistant something in Dutch, French, or Spanish. The sources it relies on mostly come back in English.
Multilingual sites with an English version often see those pages quoted in AI answers, even if real-world visitors largely ignore them. Some SEOs have already pointed to one important reason: A non-English prompt gets split into English “fan-out” queries before anything gets retrieved.
In Common Crawl’s The AI Visibility Audit, Stephen Burns goes one step further, stating, “A high-quality English version of your cornerstone pages is one of the few content-side moves that widens your footprint.”
Because of its relevance to many of our agency clients, we wanted to test this. Two questions shaped our research:
- For multilingual sites that already have an English version, how much does AI over-rely on it?
- For sites without an English folder, are they losing AI visibility because of it?
Where ChatGPT’s English bias comes from
The clearest evidence of that bias appears in OpenAI’s bot behavior, as reflected in our server logs. OpenAI runs three crawlers, and treating them as one hides what’s really happening.
- GPTBot collects training data.
- OAI-SearchBot is used to “surface websites in search results in ChatGPT’s search features,” according to OpenAI’s guidelines.
- ChatGPT-User performs the live retrieval used to answer a real user prompt.
Both OAI-SearchBot and ChatGPT-User determine which URLs end up cited.
Split the logs by bot, and the English preference is anything but flat. It rises sharply the closer you get to the moment of citation.
Look at it from left to right. GPTBot, the training crawler, hardly touches English on multilingual sites. It grabs everything in every language, and it actually sits below organic demand for English.
From a training perspective, this makes complete sense. It was already trained primarily on English data, so it needs to “learn” other languages. OAI-SearchBot lands in the middle.
Then comes ChatGPT-User, the bot whose visit shows up right before a citation. It retrieves the English version of a page more than any of the others. Its median English share is about three times that of Googlebot on the same sites.
So, “ChatGPT prefers English” isn’t quite the right way to describe it. The retrieval bot prefers English. The training bot doesn’t. That gap is the whole point. It tells you the bias enters at answer time, not somewhere deep in the training data.
While most of our clients are in Europe (Belgium, the Netherlands, France, etc.), we see the same pattern on Canadian and Indian websites.
The over-indexing ranges from 1.9 to 6.5 times. The median is around 2.6. ChatGPT fetches English pages 65% to 79% of the time.
That’s not noise. The assistant is choosing which version of your brand it reads before it generates an answer.
Dig deeper: The web is multilingual – so why does search still speak just a few languages?
See where your brand appears in AI search, where competitors are winning, and what it takes to become the answer AI recommends.
See your AI visibility
Not every AI model favors English equally
Not every AI model answers prompts the same way. We ran the same English over-indexing test on Copilot (using Bing Webmaster Tools’ AI Performance tab) and Google AI Mode and AI Overviews (using the new report in Google Search Console). The bias mostly disappears.
We turned the bias into a single index: English’s share of AI citations divided by its share of traditional search impressions. A value of 1.0 is neutral, meaning AI cites English at the same rate the site earns English search impressions. Anything above 1.0 means English is getting more than its share.
Copilot comes in at 1.07 across 272 Bing properties, making it essentially neutral. Google AI comes in at 0.79 in our GSC panel, meaning it uses slightly less English than search demand would suggest while closely tracking the local language. ChatGPT, by comparison, comes in at roughly 2.6.
The reason is likely architectural. Copilot and Google AI use search indexes that have been multilingual for years. They inherit a long history of serving the right language to the right user, with hreflang tags and the answer layer largely preserving that behavior.
ChatGPT retrieves content live within a model whose center of gravity is English, without that multilingual foundation. The bias belongs to the LLM retrieval path, not search indexing in general.
For anyone doing this work, that’s good news in a strange way. The problem has a name and a shape. It’s a ChatGPT problem (and likely applies to LLMs that retrieve content the same way OpenAI does).
Dig deeper: What multilingual regions reveal about the future of AI search
Why ‘English-first’ might be the new ‘mobile-first’
Years ago, Google introduced mobile-first indexing, where only your mobile version was used for indexing.
For LLMs like ChatGPT, the same now applies to English-language folders. Publish an English version, and you should assume some LLMs will treat it as your primaryon is to your actual audience
This English-first approach comes with a risk for websites that added an English version as an afterthought. One bank in our data left an outdated English PDF of its rates online long after the Dutch and French pages were updated.
You can guess which file the assistant pulled and quoted. A stale English asset doesn’t just sit there. It becomes the answer served to customers who never asked for English in the first place.
Get the newsletter search marketers rely on.
See terms.
Should you add an English folder?
If LLMs reach for English, the obvious move is to give them English to reach for. We tested whether that pays off by comparing AI visibility with traditional search visibility.
If we assume a 1:1 correlation between traditional visibility (Google and Bing impressions) and AI visibility, then compare sites with an /en/ folder against those without one, an interesting pattern emerges.
In Copilot, sites with an English folder received 892 citations per 10,000 impressions. Sites without one received 585. That’s an aggregate difference of 52%.
We ran the same analysis on Google to validate the pattern. In the GSC Gen AI panel, sites with an English folder saw a 28% aggregate lift in AI impressions. After removing outliers, the median per-site lift dropped to 9%. Again, not every platform behaves the same way.
OpenAI’s bot data confirms that the English bias is strongest in ChatGPT. Sites with an English folder saw a 122% uplift.
What this means for multilingual websites
On balance, the data supports Burns’ claim. For most of the multilingual sites we analyzed, a well-maintained English version of cornerstone pages expanded their AI footprint. The effect was most visible in ChatGPT, evident in Copilot, and directionally similar across Google’s AI surfaces.
Two things keep us from recommending that every site owner add an /en/ folder.
- An English folder is a commitment, not a one-time task. The day it goes live, expect LLMs to treat it as canonical. An outdated English page is worse than no English page at all. Just look at the bank and its zombie PDF. Only add an English version if you can keep it as current as your primary-language content.
- The advantage appears smaller in more mature models, and OpenAI is increasingly training on other languages. That suggests the benefit of English content may diminish over time.
Only add an English version to your website if you can give it the attention it requires.
Dig deeper: Why log file analysis matters for AI crawlers and search visibility
If AI can’t find you, customers won’t either.
Track your visibility across AI search, uncover missed opportunities, and grow your presence where customers are asking questions.
See your AI visibility
Methodology and caveats
Sources
We used three separate datasets.
- Copilot: Bing Webmaster Tools AI citation data compared with search impression data across 272 properties, including 31 with full page-level
/en/splits for the representation index. - Google AI: Search Console AI impression and search impression data across a panel of multilingual Belgian and European sites.
- ChatGPT: Raw OpenAI server log hits for GPTBot, OAI-SearchBot, and ChatGPT-User across 26 multilingual properties, using Googlebot Smartphone as the organic demand baseline. A five-site subpanel covers a clean trailing 30-day window.
Definitions
“English” refers to locale paths matching /en, /en-xx, /xx-en, or an English-only root. The representation index equals English’s share of AI citations divided by its share of search impressions, where 1.0 is neutral.
Per-bot English shares use per-site medians on properties with at least 150 to 200 hits for each bot, preventing a few large domains from skewing the average. Over-indexing ratios compare each bot with the same site’s Googlebot English share.
Caveats
AI citation reporting is still young and noisy. Bing and GSC define AI impressions and citations differently, so the cross-platform numbers indicate direction rather than like-for-like magnitude. Log-based bot identification relies on user agent strings and reverse DNS, and it can miss spoofed or proxied traffic.
Some analyses, including the per-bot splits and parts of the GSC panel, rely on small samples ranging from single digits to the low double digits of sites, so treat those results as indicative. Brand and domain names were anonymized by vertical.
Research and analysis by OMcollective. (Disclosure: I serve as OMcollective’s strategy director.)
Topics on this page
Artificial intelligenceChatGPTEnglishMicrosoft CopilotAI OverviewsBelgiumCanadaCommon CrawlDutch languageEuropeFranceFrenchGoogleGoogle AIGoogle SearchGoogle Search ConsoleGooglebotInformation retrievalLarge language modelMicrosoftMicrosoft BingMultilingualismNetherlandsOpenAIPDFSearch engine indexingSearch engine optimizationSpanishURL
+25 more
Contributing authors are invited to create content for Search Engine Land and are chosen for their expertise and contribution to the search community. Our contributors work under the oversight of the editorial staff and contributions are checked for quality and relevance to our readers. Search Engine Land is owned by Semrush. Contributor was not asked to make any direct or indirect mentions of Semrush. The opinions they express are their own.
About the Author
Pieter Serraris is Strategy Director and SEO expert at OMcollective, a Belgian based online marketing agency. He is a regular speaker at international SEO events (BrightonSEO, SEO Vibes Summit, OMconference and Search ‘n Stuff). His main focus is technical SEO and log file analysis and clarifying SEO and AI search visibility data to his clients.
${fullHTML}
`;
const collapseEl = document.getElementById(collapseId);
const truncatedWrap = bioEl.querySelector(‘.authorTruncatedBio’);
const lessLink = bioEl.querySelector(‘.authorReadLessLink’);
// If Bootstrap’s Collapse is available, wire it up. Otherwise, a tiny fallback.
const hasBootstrap = typeof bootstrap !== ‘undefined’ && bootstrap.Collapse;
if (hasBootstrap) {
collapseEl.addEventListener(‘show.bs.collapse’, () => {
truncatedWrap.classList.add(‘d-none’);
});
collapseEl.addEventListener(‘hide.bs.collapse’, () => {
truncatedWrap.classList.remove(‘d-none’);
});
collapseEl.addEventListener(‘shown.bs.collapse’, () => {
if (lessLink) {
lessLink.onclick = (e) => {
e.preventDefault();
bootstrap.Collapse.getOrCreateInstance(collapseEl).hide();
};
}
});
} else {
// Minimal JS fallback if Bootstrap isn’t present
const moreLink = bioEl.querySelector(‘.authorReadMoreLink’);
moreLink.addEventListener(‘click’, (e) => {
e.preventDefault();
collapseEl.style.display = ‘block’;
truncatedWrap.style.display = ‘none’;
});
if (lessLink) {
lessLink.addEventListener(‘click’, (e) => {
e.preventDefault();
collapseEl.style.display = ‘none’;
truncatedWrap.style.display = ”;
});
}
}
}
// Initialize all author bio blocks on the page (supports multiple posts/cards)
document.querySelectorAll(‘[id^=”authorBio”]’).forEach(initAuthorBio);
})();
