OpenAI said on June 30, 2026 that users who mostly write in a language other than English now make up more than half of active users on ChatGPT's individual plans. In other words, ChatGPT non-English users are now the majority, not the exception. That is the headline number behind this piece, and it comes with real boundaries worth stating up front: it covers the Free, Go, Plus, and Pro plans, it counts users over 18 who were active in the seven days before the month began, and it classifies each person by whichever language shows up most in their messages. It does not cover Enterprise or Codex usage, and it is not a measurement of every AI search query anywhere.
So here is the honest read. ChatGPT's individual user base is no longer English-majority. Whether AI search international usage as a whole is mostly non-English is a different question, and the evidence on that is more mixed. If you run a SaaS company and you have been treating English content as the safe default while non-English markets wait their turn, this finding is a reason to look at that plan again, not a reason to panic and translate everything overnight.
What languages are ChatGPT users actually using
OpenAI names Spanish, Portuguese, and Arabic as the leading non-English languages on ChatGPT. That is a ranking of current presence among active users, based on what OpenAI itself published, and it is specific to ChatGPT rather than a claim about language use across the whole internet.
There is a second, separate finding that is easy to mix up with the first one. Among languages with at least one million active users in June 2026, OpenAI says Uzbek, Kazakh, and Burmese had the largest percentage increases in their share of active users since July 2023. That is a growth comparison, not a size comparison. Uzbek, Kazakh, and Burmese are not being reported as the biggest ChatGPT languages, just the ones growing fastest from wherever they started. Keep those two findings separate when you think about what to prioritize: Spanish, Portuguese, and Arabic tell you where a large non-English audience already sits today, and the fast-growing group tells you where attention is shifting.
ChatGPT also supports a long list of interface languages, well beyond the ones OpenAI calls out as leading or fastest-growing. That list tells you the product can display itself in a language. It does not tell you how many people actually use it in that language, and it does not tell you the model performs equally well in every one of them. Product language support, user adoption, and search quality are three different things, and it is worth keeping them apart when you are making a roadmap decision.
Where the growth is coming from
OpenAI reports that ChatGPT adoption has grown across every continent since July 2023, with the fastest relative growth in Africa and Asia. The pattern holds when you look at it by income level too: countries with lower Human Development Index values have seen faster relative growth in weekly active users than higher-HDI countries over the same period. OpenAI also points to its low-cost Free and Go plans as part of the picture, though that alone does not prove price is the whole explanation for where growth is happening.
All of this is relative growth measured against a July 2023 baseline, not a ranking of which countries have the most total users. A market can show fast relative growth while still being small in absolute terms, so treat this as a signal to watch rather than a finished business case on its own.
It also helps to remember that ChatGPT usage is broader than search. A September 2025 OpenAI study based on 1.5 million anonymized conversations found that about 49% of messages fell into an "asking" category, with roughly 30% of usage work-related and 70% not. The same study reported that adoption growth in the lowest-income countries was more than four times the rate in the highest-income countries by May 2025. That is older context from a different study, not this year's language finding, but it points the same direction: ChatGPT's growth edge has been outside the traditional English-speaking, higher-income base for a while now.
Is AI search actually mostly English
Here the answer splits into two parts, and the split matters for what you do next.
For who is asking: no, not anymore, at least not on ChatGPT's individual plans. OpenAI's own numbers say so.
For how the answers get built: English is still doing a lot of the work behind the scenes. A February 2026 observational study by Peec AI looked at more than 10 million ChatGPT prompts and over 20 million query fan-outs, filtering for cases where the query's language matched the user's own location. It found that every non-English session it studied included at least one sub-search on the English-language web, and that close to 78% of cases pulled in additional English-language research because native-language sources alone were not judged good enough for a high-quality answer. Turkish queries switched to English most often, at 94%. Spanish had the lowest switch rate in the study at 66%, which still means two out of three Spanish searches examined pulled in English sources. Across all the research steps in the study, 43% were conducted on the English-speaking web even when the original question was asked in another language. That is a vendor's observational dataset, not an OpenAI measurement, so treat it as evidence of a pattern rather than a fixed rate that applies to every query everywhere.
Tracking AI search international usage patterns this closely is still new for most content teams, and the numbers above show why it matters. A December 2025 test by SEO researcher Glenn Gabe compared how Google Search, Google AI Overviews, Google AI Mode, Bing, ChatGPT, Perplexity, Gemini, Copilot, and Claude handled the same multilingual queries. Results varied by platform. In one French test, Google's standard results and Google AI Mode returned the French version of a page, while Google AI Overviews, ChatGPT, and Perplexity returned the US English version instead, and Claude returned the wrong language entirely. Bing and Copilot returned the correct French version. Similar mismatches turned up in Italian and Spanish tests. Google and Bing were more consistent about returning the right language version; the AI-search products varied a lot more.
Put together, an AI engine can generate its answer in your language while citing a page in a different one. The language you see on screen and the language of the source behind it are two separate things, and neither Peec AI's study nor Gabe's tests support one clean number for how English-heavy AI search is across the whole market. Nobody has published that number yet.
Why English still shows up so much
Part of the answer is supply. W3Techs reported on September 12, 2026 that English was used by 49.5% of websites with a known content language, nearly eight times the share of the next language, Spanish, at 6.0%. Portuguese sat at 4.1%. That is a measure of how much content exists in each language, not how many people search in it or how AI engines weight it, but it explains why an engine chasing a high-quality answer keeps landing back on English sources: there is simply more English material to draw from.
There is also a quality gap that is not about volume alone. A Stanford Report piece from May 2025 pointed to uneven data availability as the core reason large language models perform less well in some languages than others. The piece contrasted Swahili, spoken by roughly 200 million people but under-represented in the digitized text these models train on, with Welsh, spoken by far fewer people but unusually well documented online. The researchers behind that work found a consistent pattern: models do better on tasks that resemble their training data and worse the further a task drifts from it. Automatic translation can help close some of that gap, but the same research warned it is not a full fix, since translation errors tend to compound on complex material. None of this measures ChatGPT usage directly, but it explains why simply having more non-English users does not automatically mean AI systems serve those users as well as English-speaking ones yet.
What this means if you have no local-language content
Here is the part that should actually change how you plan. If a market you care about speaks a language other than English, and you have nothing useful published in that language, you can end up missing from the very answers your prospects are reading, even though your product is a fit. This is where the growth among ChatGPT non-English users turns from a statistic into a content gap you can actually see.
Weglot's research, published in updated form in August 2026, gives a sense of scale for that gap, though it is vendor research and should be read as such rather than as a neutral industry benchmark. Weglot analyzed more than 1.3 million AI-search citations and ran a flagship study on Spanish-language markets in Spain and Mexico. It compared 153 high-traffic sites with no English translation against 83 sites available in both languages. For Spanish-language queries, untranslated Spanish sites picked up 17,094 citations in Google AI Overviews versus only 2,810 for the same queries answered in English, a gap Weglot described as 431%. Sites that had both a Spanish and an English version saw a much smaller gap, 22%, and translated sites picked up 24% more total citations per query than untranslated ones across both languages. In localized Mexican-market queries, 96% of Google AI Overview citations in Weglot's test came from Spanish-language sources.
Weglot's own tests also showed this is not one uniform rule across engines. ChatGPT queried the web in both English and the user's own language and showed almost no language bias in the translated-site comparisons. Google AI Mode returned roughly 30% more sources per query than Google AI Overviews and pulled 93% Spanish sources in the same tests, and only 55% of queries shared any cited source at all between AI Mode and AI Overviews. Weglot also ran location-based tests in the United States: Spanish-language searches from Buffalo and Los Angeles returned only about 32% Spanish-content citations, which suggests where the searcher is sitting affects the result as much as what language they typed in.
The pattern across all of this is not that translation guarantees a citation. It is that a market with real non-English demand and no useful local-language source is a market where you are handing AI engines nothing to cite, and every finding above is consistent with that being a real cost.
What a SaaS founder should actually do with this
This is a reason to move AI search localization earlier on your roadmap, not a reason to translate your whole site this quarter. Start with the markets where the evidence already points the same direction.
Put a market in the "do this now" bucket when most of these are true: it is already part of your customer or pipeline strategy, prospects describe their problem in that language in calls or support tickets, your product actually works for buyers there, competitors already show up in AI answers in that language, and you have real product and customer context worth publishing, not just a translated homepage. If an AI answer in that language currently says nothing about you or cites nothing of yours, that is your clearest signal.
Put a market in "test selectively" when it looks promising but you do not have enough evidence yet to justify full localization. Pick a small set of high-value pages, the ones that answer real buying questions, not your whole blog, and compare how the localized and English versions perform before committing further.
Put a market in "monitor" when it shows up in the growth data (say, one of OpenAI's fast-growing languages) but is not yet tied to your actual customer base or product availability. A usage trend on its own is not a business case.
When you do localize, start with the pages that carry the most decision weight: what the product is and the problem it solves, comparison or evaluation content where you have a real point of view, pricing and packaging for that market, and the support or security information a buyer needs before they commit. Translating a blog archive top to bottom is the wrong first move. Making the pages a buyer actually needs available in their language is the right one.
Treat translation as a first pass, not the finished job. A page that reads like a direct translation, with the wrong terminology, no local examples, and none of the context a buyer in that market would expect, is not the same as a page that was actually written for them. The research on uneven model performance backs this up: proximity to what a model has actually seen and trained on matters, and raw translation does not fully close that gap.
Once you have localized content live, measure the results of your AI search localization work the same way you would measure English content: is the brand mentioned when you ask representative questions in that language, are your own pages getting cited, which competitors or third-party sources are showing up instead, and does the answer change depending on which engine or location you test from. This is exactly the kind of tracking DeepSmith's AEO module is built for: it lets you set up prompts in a specific language and market, run them against the AI engines you care about on a schedule, and see whether your brand is mentioned, cited, or absent, broken out per prompt and per platform. That tells you whether a language gap is closing, without requiring you to guess from traffic numbers alone.
None of this means English stops mattering. It is still the language most of the web is written in, and for a lot of companies it is still the right primary language for their biggest market. The finding here is narrower and more useful than "go multilingual now": stop treating English-only coverage as a neutral default the moment you have real non-English demand, and use the evidence, not the headline number alone, to decide which language gets attention next.
