TL;DR
- OpenAI published Signals v3.0 on 26 October 2026, built from 1 million classified consumer ChatGPT messages a month, July 2024 to September 2026, excluding enterprise and Codex.
- Writing fell from 34.1% to 21.4% of messages and seeking information rose from 16.7% to 20.2%; the two crossed only in August 2026, at 20.7% versus 20.6%.
- Work-related messages dropped from 51.6% to 32.2%; on the Free plan the work share fell from 44.0% in September 2024 to 24.7% in September 2026.
- In Thailand, seeking information fell from 18.6% to 14.5% while writing stayed the top topic at 27.0% in September 2026.
- Signals does not measure web searches or citations; the link to AI visibility is analysis, not a finding in the data.
OpenAI's Signals v3.0 data, published on 26 October 2026, shows that "seeking information" grew from 16.7% of consumer ChatGPT messages in July 2024 to 20.2% in September 2026, while writing requests fell from 34.1% to 21.4% over the same period. Work-related messages dropped from 51.6% to 32.2%. For brands that want to be named in ChatGPT answers, the shift matters because more of the traffic now looks like questions with factual answers, including questions about products.
What OpenAI released on 26 October 2026
OpenAI's Economic Research team updated its Signals individual data page on 26 October 2026 and labelled the release "OpenAI Signals v3.0" in its suggested citation. The page covers how people use ChatGPT on individual plans: the topics of their messages, whether a message is work-related, which plan sent it, the self-reported age of the user, and usage by country. The data is free to download as CSV files under a Creative Commons CC BY 4.0 licence, which allows reuse with attribution.
On the same day OpenAI published a companion post, "ChatGPT's expanding world of work", which adds a new layer to the data. It maps work-related messages to 165 "Intermediate Work Activities" from the US O*NET framework across nearly 100 countries, quarterly from the first quarter of 2025 to the third quarter of 2026. OpenAI lists "developing marketing content" among the example activities people bring to ChatGPT.
How the Signals dataset is built
The OpenAI Signals README, version 3.0, sets out the method. The statistics are based on classifying 1,000,000 consumer ChatGPT messages each month from July 2024 to September 2026, sent by accounts with a declared age of at least 18. The page says the sample covers messages from ChatGPT Free, Go, Plus and Pro accounts, which OpenAI describes as accounts generally managed by individuals rather than institutions.
Several groups are left out. Enterprise messages are excluded, and the README says that as a result "these data likely understate business use". Codex is excluded. Users who deleted their accounts or opted out of having their data used for training are excluded too.
The classification is automated. LLM-based classifiers sort each message into a topic and other labels, and the README says researchers did not view raw message text; the classifiers ran on de-identified data that had already been scrubbed of personally identifying information. OpenAI then adds differential-privacy noise to the published figures. The package contains 24 CSV files, with shares rounded to the third decimal place.
One caveat from the README is worth keeping in mind when reading any single country. Every series is drawn from the same global sample of 1 million messages a month, so a series built on a small slice of it, such as one country or one US state, has a larger variance than the global figures.
The global topic shift, July 2024 to September 2026
The table below uses the monthly global series behind OpenAI's chart "Share of messages by conversation topic over time in individual ChatGPT data". Each figure is the share of all sampled messages in that month.
| Topic or measure | July 2024 | September 2026 |
|---|---|---|
| Writing | 34.1% | 21.4% |
| Practical guidance | 26.1% | 28.1% |
| Seeking information | 16.7% | 20.2% |
| Technical help | 13.5% | 4.6% |
| Self-expression | 4.0% | 8.9% |
| Multimedia | 2.0% | 8.8% |
| Work-related share of all messages | 51.6% | 32.2% |
Practical guidance, which covers how-to advice, tutoring, creative ideas and health, fitness and self-care, has been the largest topic in every month since March 2025 in OpenAI's series. It peaked at 33.1% in April 2026 and stood at 28.1% in September 2026.
The August 2026 crossover, and why it is a thin one
In August 2026, seeking information reached 20.7% of messages and writing 20.6%. It is the only month in the 27-month series where information-seeking sits above writing. In September 2026 writing was back in front, 21.4% to 20.2%.
A 0.1-point gap in a dataset with deliberately added privacy noise is not a clean milestone, and it would be a stretch to say information-seeking has overtaken writing. The more durable reading is the trend. Writing has fallen by almost 13 points in two years and information-seeking has gained 3.5 points, so the two categories that were 17 points apart in July 2024 are now roughly level.
Work use is falling as a share, especially on the free plan
Work-related messages made up 51.6% of the sample in July 2024 and 32.2% in September 2026. OpenAI's plan-level chart, which starts in September 2024 rather than July, shows where the drop sits.
- Free: 44.0% of messages were work-related in September 2024, against 24.7% in September 2026.
- Plus: 61.9% in September 2024, against 50.8% in September 2026.
- Pro: 68.1% in September 2026. The Pro series begins in January 2025, at 63.1%.
- Go: 42.4% in September 2026. The Go series begins in September 2025, at 53.8%.
Because enterprise accounts are excluded, a falling work share inside individual plans does not mean less work happens in ChatGPT overall. It means the personal, non-work side of consumer use has grown faster.
Inside work messages, information-seeking is rising too
OpenAI also publishes the topic mix of work-related messages alone. Here seeking information was 10.7% in July 2024 and 11.4% in July 2026, then rose to 13.9% in August and 15.2% in September 2026. Writing remains the biggest work topic but shrank from 41.8% to 32.6%. Technical help inside work messages fell from 20.6% to 8.1%.
OpenAI did not say why technical help has fallen so far. One factor the data itself points to is scope: Codex is excluded, so coding work that has moved to Codex would not appear here. OpenAI did not state that link, and it should be read as a possibility rather than a finding.
What "seeking information" actually contains
The category name hides a detail that matters for marketers. According to the README, the coarse "Seeking information" topic is built from three fine-grained labels: specific information, purchasable products, and cooking and recipes. Product questions therefore sit inside the category that is growing.
The published files do not split the three apart, so the data cannot tell you how much of the 20.2% is product research and how much is recipes or general facts. What it does show is that the bucket containing product questions grew while writing shrank.
The September 2026 split between work and non-work adds context: 75.7% of seeking-information messages were non-work and 24.3% were work-related. Technical help leaned the other way, at 57.1% work-related.
What the Thailand figures show
Thailand has its own series in the Signals data, and it does not follow the global pattern on information-seeking. The figures below come from the country view of the same topic chart. They rest on a much smaller sample than the global numbers, so month-to-month moves of a point or two should not be over-read.
| Thailand measure | July 2024 | September 2026 |
|---|---|---|
| Writing | 33.4% | 27.0% |
| Practical guidance | 25.4% | 26.3% |
| Seeking information | 18.6% | 14.5% |
| Multimedia | 2.5% | 10.0% |
| Self-expression | 3.5% | 10.2% |
| Work-related share of all messages | 44.0% | 35.5% |
Three differences stand out. Information-seeking in Thailand fell from 18.6% to 14.5%, while the global share rose. Writing is still the largest single topic in Thailand in September 2026, at 27.0%, just ahead of practical guidance. And the work-related share in Thailand, 35.5% in September 2026, is above the global 32.2%.
On the asking, doing and expressing split for September 2026, 39.6% of Thailand's non-work messages were "asking" against 41.5% globally, and 28.4% of Thailand's work messages were "asking" against 27.3% globally. OpenAI defines asking as seeking information or clarification, doing as asking ChatGPT to produce an output, and expressing as sharing views or feelings without seeking information or action.
The per-capita ranking places Thailand 106th in the third quarter of 2026, by messages per person, where rank 1 means the highest usage. Thailand's quarterly rank has moved between 95th and 112th since the first quarter of 2025. Singapore ranked 2nd in the same quarter.
What the data does not tell you
None of the 24 files described in the README measures whether ChatGPT searched the web, showed sources or cited a website in its answer. Signals tells you what people ask about, not where the answers come from. Any claim that this release proves ChatGPT is sending more traffic to websites goes beyond the data.
The data also leaves out enterprise accounts, Codex and anyone under 18, and the topic labels are inferred by classifiers rather than reported by users. OpenAI did not publish Thailand-specific commentary, and it did not explain the drop in Thai information-seeking.
Analysis: what more information-seeking means for being cited
This section is Relevant Audience's analysis, not something OpenAI said. A request to rewrite an email needs no outside source. A question about a fact, a product or a price often does, and those are the answers where ChatGPT may search the web and show citations. A larger share of information-seeking messages therefore means more of the conversations where a brand can appear as a named source, even though Signals itself does not count citations.
That makes the practical work of generative engine optimisation more relevant, not less. Pages that answer one clear question near the top, carry specific facts with dates and figures, and keep product details such as prices, specifications and policies consistent across the site are easier for an AI system to quote accurately. A team planning ChatGPT SEO work can use the topic mix as a rough guide: practical guidance and information-seeking together made up 48.3% of messages in September 2026, and both are question-shaped.
For Thailand the picture is less tidy. Thai users' information-seeking share has fallen while multimedia and self-expression grew, so the case for AI visibility work in the Thai market rests more on the global trend and on how Thai buyers research specific categories than on this Thai series alone. A broader AI SEO plan that also covers Google's AI features is the safer bet while the Thai numbers are this noisy.
What this means for Thai marketers
- Treat the global shift as a signal of direction, not a forecast for Thailand. The Thai series moved the other way on information-seeking.
- Check how ChatGPT currently answers the product and policy questions buyers ask about your category, in Thai and in English, and note which sources it names.
- Make sure the facts an assistant would need, such as prices, delivery terms and specifications, are stated plainly on crawlable pages rather than only in images or PDFs.
- If you want to test Thai usage yourself, the CSV files are free to download and the country, topic and work splits are all there.
FAQ
What is OpenAI Signals v3.0?
OpenAI Signals v3.0 is a dataset OpenAI published on 26 October 2026 that describes how people use ChatGPT on individual plans, based on 1 million classified consumer messages a month from July 2024 to September 2026. It covers topics, work use, plans, age, name-based gender and countries, and it is licensed under CC BY 4.0.
Has information-seeking overtaken writing in ChatGPT?
Only in one month: in August 2026 information-seeking was 20.7% of messages and writing 20.6%, but in September 2026 writing was back ahead at 21.4% against 20.2%. The gap is small enough that the privacy noise OpenAI adds could account for it, so the safer statement is that the two are now roughly level after writing fell from 34.1%.
Does the data show ChatGPT citing more websites?
No, Signals does not measure web searches or citations at all. It classifies what users ask about. Any link to citations or referral traffic is analysis, not something the data shows.
Is there Thailand data in OpenAI Signals?
Yes, Thailand has its own country series. In September 2026 information-seeking was 14.5% of Thai messages, down from 18.6% in July 2024, and writing was the largest topic at 27.0%. Country series rest on smaller samples than the global figures, so treat small moves with caution.
Do marketers need to do anything because of this release?
Nothing changes in any ad account or tool, because this is a research dataset, not a product change. It is a useful prompt to review how your pages answer the factual and product questions that make up a growing share of ChatGPT use worldwide.
The Signals data is one of the few first-party views of what people actually ask ChatGPT, and the direction is clear: fewer writing jobs, more questions. If you want a second view on how your brand shows up when those questions are about your category, Relevant Audience can review your pages against the answers ChatGPT gives today.
