Writing drops from 34% to 21% of ChatGPT requests since July 2024

Writing, the main reason people opened ChatGPT in mid-2024, now accounts for about one in five requests. Data OpenAI released on October 26, 2026 shows writing tasks fell from 34.1% of consumer requests in July 2024 to 21.4% in September 2026. Over the same period practical guidance - how-to advice, tutoring, ideas and health and fitness questions - rose from 26.1% to 28.1%, and it has been the largest category in every month since March 2025. Each request is a single prompt typed into ChatGPT, which OpenAI's files count as a "message."

The figures come from version 3.0 of OpenAI Signals, a package of 24 spreadsheets built from automated classification of 1 million consumer ChatGPT requests a month, sent by adults, over 27 months. Besides topics, the files split usage by country, age group, the typical gender of users' first names, work and non-work purpose, subscription plan and US state. PPC Land went through every file. The shifts are large and run in the same direction almost everywhere. Help with programming, maths and data analysis fell from 13.5% to 4.6% of requests. Image and other media requests more than quadrupled, from 2.0% to 8.8%. And the share of requests classified as work-related dropped from 51.6% to 32.2%.

The people sending those requests changed as well. Users with typically feminine first names sent 51.8% of requests in September 2026, up from 42.2% in July 2024. Adults aged 18 to 24 went from 39.3% of requests to 30.6%, and users aged 65 and older tripled their share, from 0.9% to 2.7%.

In Short

OpenAI sorted a million ChatGPT requests a month into topics and found that people now use it most for practical advice, like how to fix something or how to get fit, instead of for writing. Compared with two years ago, people use it less for work and coding and more for chatting, making pictures and looking things up. The people using it are older than in 2024, more of them have first names usually given to women, and the same changes show up in most of the roughly 150 countries in the data.

What OpenAI released

OpenAI Signals is the data program run by OpenAI's Economic Research team. Version 3.0 arrived on October 26 alongside OpenAI's page "ChatGPT's expanding world of work," which concentrated on one of the 24 files: work requests sorted into job activities from O*NET, the US government's occupational database. The other 23 files, which cover what people ask ChatGPT about, who asks it and where, are the subject of this article. The data is published on OpenAI's Signals download page under a Creative Commons Attribution 4.0 license.

The README that accompanies the files sets out the method. "Our dataset includes statistics based on the classifications of 1,000,000 consumer ChatGPT messages each month from July 2024 to September 2026, which were sent by accounts with an associated declared age of at least 18," according to OpenAI. Business use is outside the sample. "This data does not include any statistics based on the classifications of enterprise messages and, as a result, these data likely understate business use," the README states. Two further groups are removed: "We exclude from our sample users who have deleted their accounts or opted not to allow their data to be used for training. We do not include data from Codex." Codex is OpenAI's coding agent.

No person reads the requests. A large language model classifies each one using written instructions, and the README prints those instructions in full. They are, according to OpenAI, "the same prompts originally used in 'How People Use ChatGPT' (Chatterji et al., 2025)," the usage paper OpenAI published with economists from Harvard and Duke in September 2025. PPC Land covered that paper when it appeared: it put ChatGPT at 700 million weekly users and found non-work use above 70% of consumer requests by June 2025.

Seven topics, 23 labels

The classifier first picks one of 23 detailed labels, which the files then group into seven topics. The grouping, as the README lists it:

• Writing: editing or critiquing text the user supplies, personal writing or communication, translation, argument or summary generation, and fiction.
• Practical Guidance: how-to advice, tutoring or teaching, creative ideation, and health, fitness, beauty or self-care.
• Seeking information: specific information, purchasable products, and cooking and recipes.
• Technical help: mathematical calculation, data analysis and computer programming.
• Multimedia: creating an image, analysing an image, and generating or retrieving other media.
• Self-expression: greetings and chitchat, relationships and personal reflection, and games and role play.
• Other/Unknown: questions about the model itself, other requests, and requests whose purpose is unclear.

The instructions give examples for each label. "Recommend a good laptop under $1000" is a purchasable-products request. "I want you to be my AI girlfriend" is games and role play. "Make a YouTube video about goal kicks" and "Write PPT slides for a tax law conference" count as other media. The files publish only the seven topics, not the 23 labels, so the weight of each label inside its topic is not visible.

Other classifiers run alongside the topic label. One decides whether it is work-related, with examples such as "rewrite this HR complaint" for work and "does ice reduce pimples?" for non-work. A second sorts requests into work, schoolwork and other. A third labels the user's intent as asking, doing or expressing. A fourth, applied only to work requests, assigns an O*NET activity.

Who counts as what

Gender in these files is not a user's own declaration. The column is called typical_name_gender, and the README defines it as a "name-associated category with values masculine and feminine; not self-identified gender. Based on anonymized first-name matching against external name-gender data sources. Only names with >95% prevalence for a specific gender are accepted as valid matches." Users whose names do not clear that bar are left out of the gender files, and the files do not say how many that is. Age, by contrast, is the age users entered themselves. Anyone under 18 or without a declared age is excluded.

Country is the "country associated with request," coded with two-letter ISO codes. Per-capita rankings divide usage by 2023 population figures, from the World Bank for countries and the US Census Bureau for states.

Noise, thresholds and restatements

Every figure has statistical noise added to it. OpenAI uses differential privacy, a technique that adds calibrated random noise so that no single request can be identified from the published statistics; the IAB Tech Lab examined the same technique for ad targeting in 2024. For July to September 2026 the privacy budget is ε = 1. Older months have been published more than once and carry looser budgets: ε = 5 for December 2025 and earlier, ε = 3 for January to March 2026, and ε = 2 for April to June 2026, according to the README. A cell is suppressed if the weighted, noise-adjusted number of requests behind it is 100 or fewer; the activity file uses a floor of 1,000 per quarter. Shares are rounded to three decimal places.

The README also warns that precision varies. All series come from the same global sample of a million requests a month, so a figure for one topic in one US state rests on far fewer requests than a global figure and "as such have a larger variance," according to OpenAI.

Version 3.0 restated history. It raised the sample for every earlier month from 300,000 requests to 1,000,000, replaced the annual US state ranking with quarterly ones, and swapped earlier US-only activity tables for a country-by-quarter file. Version 2.0, earlier in 2026, had already removed and replaced "under 0.5% of messages each month." Figures from earlier Signals releases therefore will not always match these.

How PPC Land calculated the figures

Monthly global figures in this article are read directly from the files. Country comparisons are PPC Land's calculations: to reduce noise, each country's share is the simple average of its three monthly values in a quarter, comparing the third quarter of 2024 (July to September) with the third quarter of 2026. Only countries with topic data in all 27 months are ranked, which leaves 137. Where the article gives the share of a topic within an age group or gender, or the share of a topic within one country that is work-related, PPC Land combined two published distributions; those figures are approximate and are labelled. The implied totals reproduce OpenAI's published age and gender splits to within 0.2 percentage points.

Seven topics over 27 months

The headline picture fits in one table.

Two topics lost ground and five gained. Writing and technical help together made up 47.6% of requests in July 2024 and 26.0% in September 2026. The gains are spread across categories that, two years earlier, were small.

The country picture moved with it. In the third quarter of 2024, writing was the largest topic in 103 of the 137 countries PPC Land compared, and practical guidance led in the other 34. By the third quarter of 2026, practical guidance led in 116, writing in 20, and seeking information in one: Slovenia. Countries also converged on some topics and spread apart on others. Technical help once ranged from 4.0% to 21.6% of a country's requests; it now ranges from 0.4% to 5.9%. Seeking information went the other way, from a range of 11.2% to 24.4% to one of 10.2% to 31.2%.

The chapters that follow take each topic in turn, comparing countries, age groups, the two name-based gender groups and work against non-work use, before turning to work, intent, subscription plans, demographics, country rankings, US states and work activities.

Writing

Writing is the topic that lost most. Its share of requests fell in almost a straight line, from 34.1% in July 2024 to 30.6% in January 2025, 24.8% in July 2025 and 24.7% in January 2026, reaching a low of 20.6% in August 2026 before edging up to 21.4% in September. There was no single break. The decline paused for a few months at a time - between August and November 2025, writing even rose slightly, from 25.2% to 26.3% - and then resumed.

The category covers five kinds of request: editing or critiquing text the user supplies, personal writing such as emails and social posts, translation, arguments and summaries on topics the user does not supply in detail, and fiction. Among the instruction's examples are "Help me improve my essay, including improving flow and correcting grammar errors" and "Write a nice birthday card note for my girlfriend." The files do not separate them, so it is not possible to say from this data whether editing, translation or fiction fell fastest.

Still the largest work use

Writing has lost much less ground at work than outside it. In July 2024, 63.1% of writing requests were classified as work-related; in September 2026 the figure was 49.0%, having dipped as low as 46.4% in August 2025. Within work requests, writing went from 41.8% to 32.6% - still comfortably the largest single topic at work, ahead of practical guidance at 22.9%. Outside work, writing fell from 26.0% of requests to 16.1%, and now trails practical guidance and seeking information by a wide margin.

PPC Land's calculation of the work share of writing in each country shows the same pattern with wide variation. In Australia, 62.9% of writing requests in the third quarter of 2026 were work-related, the highest among the 24 large markets PPC Land tracked. The United Arab Emirates followed at 60.4% and Brazil at 56.5%. At the other end, about a third of Nigerian writing requests were work-related in 2026, down from 58.3% two years earlier, and the figure was 37.6% in Turkey. Several countries with the lowest shares of work-related use overall - Afghanistan, Iraq and Algeria among them - had fewer than a quarter of writing requests tied to work.

Inside work requests, writing remains dominant in a few markets. It accounted for 47.3% of Philippine work requests in the third quarter of 2026 and 46.1% of those from the UAE, against 36.8% in the United States and 28.4% in Turkey.

Writing by country

Writing fell in 135 of the 137 countries PPC Land compared, between the third quarters of 2024 and 2026. The two exceptions, Sudan and Somalia, rose by one to two points from a small base. The median country's writing share dropped from 32.5% to 21.5%. In 2024, 94 countries had more than 30% of requests in the category; two years later, six did.

The largest falls came in Western Europe. France dropped 21.4 points, from 40.0% to 18.6%. Malta, Belgium, the United Kingdom, Austria, the Netherlands, Slovenia and Ireland each lost between 19 and 21 points. The United Kingdom went from 41.1% to 21.2%, so that a topic that once took two in five British requests now takes roughly one in five.

Source: OpenAI Signals v3.0; PPC Land's quarterly averages of monthly shares.

The countries that kept writing near the top are concentrated in Southeast Asia and West and Central Africa. Laos led at 38.8%, followed by Nigeria, the Philippines, Benin, the Democratic Republic of the Congo, Côte d'Ivoire, the Republic of the Congo, Haiti, Brunei and Cambodia. Of the 20 countries where writing is still the largest topic, 15 are in those two regions, Singapore among them; the others are the UAE, Qatar, Haiti, the Maldives and Mozambique. Indonesia and Vietnam had the smallest declines of any large market, at 2.4 and 1.9 points. In both, writing and practical guidance are now separated by less than half a point.

At the bottom sit Turkey (14.6%), Hungary, Serbia, Bulgaria, Slovakia, Croatia, Bosnia and Herzegovina, Azerbaijan, Romania and Iraq, all between 14.6% and 15.4%. Much of Central and Eastern Europe has moved from writing toward looking things up, a shift covered in the seeking information chapter below.

Writing by age

In 2024 writing was a middle-aged habit. By PPC Land's calculation, writing took 27.8% of requests from users aged 18 to 24 in July 2024, but 37.7% for those aged 25 to 34, 39.4% for 35 to 44 and around 38% for those aged 45 to 64. Those aged 65 and older sat at 33.0%.

The gap has closed. In September 2026 writing accounted for 21.8% of requests from the youngest group, 22.3% for 25 to 34, 21.3% for 35 to 44, 19.9% for 45 to 54, 18.1% for 55 to 64 and 16.6% for those 65 and older. The age profile has flipped: writing now skews slightly young, and the oldest users write least. In the files' own terms, 31.1% of writing requests in September 2026 came from 18- to 24-year-olds and 35.7% from 25- to 34-year-olds, against 32.2% and 37.4% in July 2024.

Country data shows how uneven the youth share of writing is. In the United States, 21.0% of writing requests in the third quarter of 2026 came from 18- to 24-year-olds, barely changed from 22.3% two years earlier. In Indonesia the figure fell from 59.2% to 46.6%, in Mexico from 45.4% to 32.3%, and in Vietnam from 59.0% to 35.9%. India remains young throughout, with 40.3% of writing requests from the youngest group. Nigeria is one of the few large markets where the youngest users' share of writing rose, from 34.5% to 39.8%.

Writing by name gender

Writing has always been one of the more feminine topics in the data, and it has become more so. Users with typically feminine names sent 49.4% of writing requests in July 2024 and 56.7% in September 2026, peaking at 58.0% in May 2026. By PPC Land's calculation, writing made up 40.1% of requests from feminine-name users in July 2024, against 29.8% for masculine-name users. Those figures fell to 23.4% and 19.2%.

Across the 52 countries with reported values in all three months of the third quarter of 2026, writing requests came mainly from feminine-name users in 42. Argentina (66.3%), Brazil (66.0%), the Czech Republic (66.0%), Poland (65.8%) and Indonesia (65.6%) had the highest feminine shares. The lowest were Bangladesh (36.2%), Pakistan (36.4%), Nigeria (39.1%), Saudi Arabia (41.0%) and India (45.3%). In the United States the feminine share of writing rose from 50.5% to 58.8%, and in Turkey from 40.3% to 51.8%.

Writing in US states

OpenAI's state files report quarters from the first quarter of 2025 onward. Writing fell in every one of the 50 states and the District of Columbia between the first quarter of 2025 and the third quarter of 2026, and nationally it went from 28.5% to 21.1% of US requests.

Washington, DC, had the highest writing share in the third quarter of 2026, at 26.1%, followed by New Mexico and Vermont (both 23.0%), California (22.9%) and Maryland (22.8%). New Hampshire had the lowest, at 16.8%, after the steepest fall of any state - 13.7 points. Montana (17.6%), Alabama, Delaware (both 17.9%) and Idaho (18.1%) came next. Montana and Wyoming each lost more than 11 points. New Mexico's writing share fell least, by 2.1 points.

Writing as a work activity

The activity file, which assigns work requests to O*NET's Intermediate Work Activities, confirms that editing remains the largest single thing people do with ChatGPT at work. "Edit written materials or documents" took an average of 18.4% of work requests across the 74 countries reported in all seven quarters in the first quarter of 2025, and 15.0% in the third quarter of 2026, by PPC Land's calculation. In the United States it was 17.1% in the latest quarter, and in the United Kingdom 18.8%.

Two other writing-related activities lost share. "Interpret language, cultural, or religious information for others," the O*NET activity that covers translation, fell from 5.1% to 4.0% on average, though it remained the third-largest work activity in Germany (4.9%) and Japan (4.5%). "Write material for artistic or commercial purposes" slipped from 2.7% to 2.3%; it was highest in Nigeria, at 5.2%, and Pakistan, at 4.4%.

Practical Guidance

Practical guidance is the topic ChatGPT is now used for most. It overtook writing in March 2025, when it took 28.4% of requests against writing's 28.2%, and has led every month since. Its share climbed steadily to a peak of 33.1% in April 2026 - one request in three - before falling back over the following five months to 28.1% in September 2026. That decline took practical guidance below its share in September 2025 (29.6%).

The topic bundles four labels: how-to advice ("How do I turn off my screensaver?"), tutoring or teaching ("Can you explain derivatives and integrals?"), creative ideation ("Brainstorm names for a new coffee shop") and health, fitness, beauty or self-care ("Whats a good skincare routine for oily skin?"). Health questions therefore sit here, not under seeking information.

The 2026 decline did not follow the pattern of 2025. Over the summer of 2025, practical guidance held steady or rose: it was 28.1% in July, 28.9% in August and 29.6% in September. Over the summer of 2026 it fell from 31.8% in June to 31.0%, 29.2% and 28.1%. The files do not explain the change. The same months saw seeking information and the Other/Unknown category reach their highest levels in the series.

Mostly a non-work topic

Practical guidance has become one of the least work-related topics; only self-expression and seeking information now have lower work shares. In July 2024, 44.0% of practical guidance requests were work-related. By June 2026 that had fallen to 24.1%, and it stood at 26.3% in September. Within non-work requests, practical guidance went from 30.2% to 30.5%, peaking at 36.0% in April 2026. Within work requests it went from 22.3% to 22.9%, peaking at 27.4% in February 2026.

The work share varies by market. By PPC Land's calculation, 36.6% of practical guidance requests from Indonesia and 33.0% from the Philippines were work-related in the third quarter of 2026, against 25.9% in the United States, 21.3% in Germany, 19.0% in France and 14.7% in Turkey. Among all 137 countries, Singapore and Indonesia had the highest work shares, both 36.6%, and Iraq the lowest, at 11.5%.

Practical guidance by country

Practical guidance rose in 96 of the 137 countries and fell in 41. The median country's share went from 25.7% to 29.8%.

The highest shares in the third quarter of 2026 were in the Middle East, North Africa and Eastern Europe: Sudan (36.7%), the Palestinian territories (34.6%), Moldova (33.8%), Yemen (33.7%), Lebanon (33.3%), Réunion (33.0%) and Egypt (32.9%). The lowest were in West and Central Africa: Benin (20.9%), Nigeria (22.6%), Laos (22.7%), Guinea (23.1%), Côte d'Ivoire (23.4%) and the Democratic Republic of the Congo (24.1%), the same countries where writing kept its lead. The UAE, at 26.0%, ranked 128th.

North Macedonia's share rose most, by 10.2 points. Turkey, Réunion, Lithuania, Cyprus, Lebanon, France and Serbia each gained between 7.5 and 8.7 points. The largest falls were in Mozambique (6.4 points), Benin, Haiti, Ecuador, Ghana, Guatemala, Indonesia and Nigeria.

Source: OpenAI Signals v3.0; PPC Land's quarterly averages.

The table shows two groups. In Europe, East Asia, North America, Egypt, Turkey and India, practical guidance gained between 3.7 and 8.7 points. In Mexico, Brazil, Indonesia and Nigeria - markets where it was already large or where writing remained strong - it was flat or lower.

The youngest topic

Practical guidance is the topic most associated with young users. By PPC Land's calculation it took 29.4% of requests from 18- to 24-year-olds in July 2024, against between 22.5% and 24.6% for every older group. In September 2026 the figures were 30.8% for the youngest group and between 24.5% and 24.9% for users aged 45 and over. The youngest users' share of practical guidance requests fell from 44.4% to 33.5%, but it remains the topic where they are most over-represented: about 1.09 times their share of all requests.

Tutoring is part of the reason. The schoolwork classifier, covered in the work chapter below, shows school-related requests falling in every northern-hemisphere summer, and the youngest group's share of practical guidance requests follows the school calendar: it peaked at 54.7% in November 2024, dropped to 37.3% in August 2025 and to 31.7% in July 2026, and recovered slightly each September. In the United States, 18- to 24-year-olds sent 39.0% of practical guidance requests in the third quarter of 2024 and 23.5% two years later, the largest change for any topic in the US age data.

Practical guidance by name gender

Feminine-name users sent 43.7% of practical guidance requests in July 2024 and 53.6% in September 2026, after a peak of 54.8% in April and May 2026. By PPC Land's calculation, practical guidance accounted for 29.0% of feminine-name requests in September 2026 and 27.0% of masculine-name requests.

Of 72 countries with reported values, 56 had feminine majorities in practical guidance in the third quarter of 2026. Indonesia was highest at 61.5%, ahead of Brazil, the Czech Republic, Argentina and Ukraine, all near 61%. Pakistan (39.8%), Bangladesh (41.1%), Thailand (42.2%), Japan (43.8%) and Nigeria (45.5%) had the lowest. The US figure rose from 42.8% to 53.8%, and Turkey's from 32.8% to 52.8%.

Practical guidance in US states

Practical guidance is the largest topic in every state and in Washington, DC. Nationally it rose from 29.8% in the first quarter of 2025 to 33.2% in the first quarter of 2026, then fell to 30.5% in the third quarter.

Mississippi had the highest share in the third quarter of 2026, at 35.2%, ahead of Rhode Island (33.8%), South Dakota (33.6%), Nebraska and Connecticut (both 33.3%). Wyoming had the lowest, at 26.9%, followed by Washington, DC (27.9%), Vermont (28.4%), Washington state (28.6%) and West Virginia (29.0%). Since the first quarter of 2025 the share rose in 25 states and fell in 23, with Connecticut up 4.0 points and West Virginia down 5.1.

Advice at work

The activity file shows advice gaining at work. "Advise others on educational or vocational matters" rose from 1.9% to 3.0% of work requests on average across the 74 countries; in India it reached 5.1% in the third quarter of 2026, and in the United States 3.4%. "Advise others on legal or regulatory matters" more than doubled on average, from 0.8% to 1.9%. In the US, "Advise others on financial matters" went from 0.3% to 0.5%, and "Counsel others about personal matters" from 0.3% to 0.6%. "Teach academic or vocational subjects," by contrast, fell from 2.6% to 1.7%.

Financial questions are one of the areas OpenAI is now building for. In May 2026 it launched a personal finance preview for US Pro subscribers that links bank accounts through Plaid; PPC Land reported OpenAI's claim that more than 200 million people already use ChatGPT each month for budgeting and financial questions. The advertising rules OpenAI set out in January point the other way for health: ads are not shown near health, mental health or political topics, which means part of the practical guidance category sits outside the ad business by design.

Seeking information

Seeking information is the closest thing in the data to using ChatGPT as a search engine. It covers three labels: specific information "typically found on websites," such as facts about public figures, current events and history; questions about products or services available for purchase; and cooking and recipes. The examples in the classifier's instructions include "What conflicts are happening in the Middle East right now?", "How much are Nikes?" and "Recipe for lasagna."

Its share of requests rose from 16.7% in July 2024 to 20.2% in September 2026. The path was uneven. It dipped to 16.3% in April 2025, the month image requests spiked, then climbed to 19.0% by July 2025 and held between 17.5% and 19.1% until June 2026, before rising to 19.5% in July and 20.7% in August 2026, its highest level in the series. By September 2026, one request in five was a request for information.

The category matters for advertisers because two of its three labels - products and recipes - map directly onto commercial intent. OpenAI's files do not split the category, so the share of product questions alone is not published.

Information at work, and the August jump

Seeking information has become less work-related over the two years, but it moved sharply in the final two months. In July 2024, 33.1% of information requests were classified as work-related. The figure fell to 17.8% in July 2026, then rose to 19.9% in August and 24.3% in September.

The same jump shows up inside work requests. Information requests made up between 10.7% and 12.1% of all work requests every month from July 2024 to July 2026, then 13.9% in August 2026 and 15.2% in September. Among topics within work requests, only the image spike of spring 2025 produced a comparable two-month move. Outside work, the category was steadier, moving from 23.0% to 22.5% of non-work requests over the full period.

PPC Land's country calculation suggests where information requests are most likely to be for a job. In Peru, 28.8% of them were work-related in the third quarter of 2026, followed by South Africa, Indonesia, Ecuador and Angola, all between 25.5% and 27.0%. In Bosnia and Herzegovina the figure was 8.1%, and in Georgia 8.7%. Among the large markets, it ranged from 13.1% in Turkey and 14.0% in Spain and Poland to 24.0% in India and 26.5% in Indonesia.

A European habit

Country data shows the sharpest regional split of any topic. Information requests are now most common in Central and Eastern Europe and the Nordic and Mediterranean countries, and least common in Southeast Asia and parts of Africa.

Slovenia led in the third quarter of 2026, at 31.2% - the only country where seeking information was the largest topic. Croatia (30.0%), Slovakia (29.4%), Montenegro (28.7%), the Czech Republic (28.1%), Greece (27.8%), Bosnia and Herzegovina (27.4%), Hungary (27.2%), Denmark (26.9%), Norway (26.7%), Spain (26.3%) and France (26.1%) followed. At the other end were Laos (10.2%), Sudan (12.1%), Mali (12.5%), the Republic of the Congo (12.6%), Myanmar and Yemen (both 12.8%), Indonesia and the Philippines (both 13.8%) and Cambodia (13.9%).

The divergence is recent. Two years earlier, Slovenia stood at 19.3% and Croatia at 18.6%; their shares rose by 11.9 and 11.4 points, two of the three largest gains of any country. Malta (11.5 points), Montenegro (10.5), Denmark (9.9), France (9.6), Ireland (9.4) and North Macedonia (9.1) were close behind. Over the same period the share fell by 8.0 points in Bolivia and by between 5 and 6 points in Yemen, Sudan, Mali, Honduras, Paraguay, Panama and Afghanistan. The category rose in 82 countries and fell in 55, so the median moved only from 17.7% to 18.4%, while the spread between countries almost doubled.

Source: OpenAI Signals v3.0; PPC Land's quarterly averages.

The table shows two groups of markets pulling apart. In wealthy markets in Europe, North America and East Asia, information requests rose by between 3.5 and 9.6 points. In Mexico, Brazil, India, Nigeria, Indonesia and the Philippines - all large, young user bases - the share was flat or lower.

The oldest topic

If practical guidance is the youngest topic, seeking information is the oldest. By PPC Land's calculation, information requests made up 17.3% of requests from 18- to 24-year-olds in September 2026, 19.7% for 25 to 34, 21.8% for 35 to 44, 24.2% for 45 to 54 and 26.7% for 55 to 64, before easing to 23.1% for those 65 and older. Users aged 55 to 64 are 1.32 times as likely as the average user to be asking for information; in July 2024 the ratio was 1.17.

Country data confirms the age gradient. In South Korea, 29.9% of information requests in the third quarter of 2026 came from users aged 45 and over. In the United States the figure was 26.1%, in Spain 26.8% and in Italy 24.8%. In India, Indonesia, Vietnam and Pakistan, it was below 9%.

Information by name gender

Seeking information is one of three topics where masculine-name users still send most requests. Feminine-name users sent 39.1% of them in July 2024 and 46.8% in September 2026. By PPC Land's calculation, information requests accounted for 22.3% of masculine-name requests in September 2026 and 18.2% of feminine-name requests, so masculine-name users are about 1.2 times as likely to be looking something up. Only technical help shows a stronger masculine tilt.

Country patterns differ sharply. Of 58 countries with reported values in the third quarter of 2026, 22 had feminine majorities. South Korea had the highest feminine share, at 56.6%, followed by the Czech Republic (55.5%), Croatia, Serbia (both 54.3%) and Romania (54.1%). The lowest were Pakistan (30.4%), Bangladesh (32.2%), Saudi Arabia (37.5%), India (37.8%) and Nigeria (38.3%). In the United States, the feminine share of information requests rose from 36.9% to 48.7%, and in Turkey from 30.0% to 46.8%.

Information in US states

US information requests rose from 19.0% in the first quarter of 2025 to 20.9% in the third quarter of 2026, and the share increased in 44 of the 51 jurisdictions. Delaware had the highest share in the latest quarter, at 26.1%, followed by Hawaii (24.6%), Alaska (24.4%), Vermont (24.2%) and Montana (23.5%). Alaska's share rose 7.4 points, Montana's 6.8 and Delaware's 6.5. The lowest shares were in New Mexico (16.8%), Rhode Island and West Virginia (both 17.1%), Washington, DC (17.8%) and Mississippi (18.7%).

Product questions at work

The activity file has a category that sits close to purchasable-products questions: "Obtain information about goods or services." It rose from 1.9% of work requests on average across the 74 countries in the first quarter of 2025 to 3.0% in the third quarter of 2026. In the United States it went from 1.7% to 2.9%. It was highest among large markets in Poland (3.9%), Saudi Arabia (3.9%), the UAE (3.8%) and France (3.7%). "Gather information from physical or electronic sources" rose from 1.4% to 1.9%.

Why the information share matters

ChatGPT's growth as an information source is the reason European regulators now treat it as a search engine. The European Commission designated ChatGPT a Very Large Online Search Engine on August 31, after OpenAI declared 159.1 million average monthly active users in the EU. Six of the eight countries where seeking information gained most - Slovenia, Malta, Croatia, Denmark, France and Ireland - are EU members.

Traffic has not followed the information requests. An Ahrefs analysis of 76,000 websites, published in February, found that ChatGPT handled the equivalent of 12% of Google's search volume but sent 190 times less traffic to websites. For shopping, a May survey of 1,295 US residents by Smarty Marketing found that ChatGPT was the starting point for 7.26% of shopping searches, behind Google at 56.68% and Amazon at 28.96%.

OpenAI has tried to turn product questions into transactions. Instant Checkout, launched on September 29, 2025 with Stripe, was shut down in March 2026; product feeds that once powered it moved into the ChatGPT Ads Manager in June. The Signals data cannot show how many of the product questions in this category led to a click or a purchase.

Technical help

Technical help fell further than any other topic in proportion to its size. It took 13.5% of requests in July 2024 - one in seven - and 4.6% in September 2026, a decline of two thirds. The low point was 4.0% in July 2026. The category combines computer programming, mathematical calculation and data analysis, with examples such as "Im getting a TypeError in JavaScript when I try to call this function," "What is 400000 divided by 23?" and "Perform a regression analysis on this data."

The fall was steady rather than sudden. Technical help dropped below 10% in January 2025 (9.4%), below 6% in July 2025 (5.9%) and below 5% in January 2026 (4.9%). It then stayed between 4.0% and 4.9% for the rest of the series, reaching its low in July 2026 before edging up to 4.6% in September.

Coding moved somewhere

The files leave out one obvious destination. "We do not include data from Codex," the README states. Codex is OpenAI's coding agent, and its growth over the same period was rapid. In February, OpenAI said Codex had 1.6 million weekly active users, more than triple the figure at the start of 2026. By July, OpenAI said Codex served more than 5 million people a week, and the company merged its chat, coding and browser interfaces into one desktop app with Chat, Work and Codex modes. Any programming that moved from the ChatGPT chat window into Codex would disappear from these files without appearing anywhere else in them.

The broader shift in how people write software points the same way. Stack Overflow, the question-and-answer site that developers relied on for more than a decade, recorded 1,442 new questions in July 2026, down 99% from its March 2014 peak of 207,204. Vibe coding - describing software in plain language and accepting what an AI writes - has moved programming toward tools built for it. Anthropic reported a similar movement in its March Economic Index, which found coding work moving from Claude.ai toward the API. None of this can be read from the OpenAI files themselves, which measure only the chat product.

Still the most work-related topic

Technical help is still the topic most tied to work. In July 2024, 78.8% of technical requests were classified as work-related; in September 2026, 57.1%. Within work requests, technical help fell from 20.6% to 8.1%, with a low of 7.4% in May 2026. Outside work, it fell from 5.9% to 2.9% of requests. In relative terms the fall was steeper inside work, at 61%, than outside it, at 51%.

PPC Land's country calculation puts the work share of technical help above 50% in most large markets in the third quarter of 2026: 61.6% in India, 61.7% in Italy, 60.2% in the Netherlands, 59.3% in Brazil and 52.9% in the United States. Two years earlier most were above 75%, and South Korea and Japan were near 90%.

Every country fell

Technical help declined in all 137 countries PPC Land compared. In the third quarter of 2024, it accounted for more than 10% of requests in 88 of them; in the third quarter of 2026, it did so in none. The median fell from 11.2% to 3.8%. The spread between countries collapsed: the standard deviation fell from 3.3 points to 0.8.

India was the heaviest user of ChatGPT for technical work in 2024, at 21.6% of requests, and remains so, at 5.9%. Singapore, Finland (both 5.7%), Zimbabwe (5.3%) and Estonia (5.2%) followed. The lowest shares were in Réunion (2.0%), Somalia and Afghanistan (both 2.1%), Libya (2.2%) and Algeria (2.3%). Brunei's figure, 0.4%, is small enough that the noise added for privacy may account for much of it.

The steepest falls were in North Macedonia (18.4 points), Israel (15.8), India (15.7), Armenia (15.1), Tunisia (14.8), Singapore (13.2), Ethiopia (12.9) and Japan (12.3). In the large markets the declines ranged from about 3 points in the Philippines to 15.7 in India.

Source: OpenAI Signals v3.0; PPC Land's quarterly averages.

The age gap closed

In 2024, technical help was a young person's topic. By PPC Land's calculation it took 15.0% of requests from 18- to 24-year-olds in July 2024 and 14.2% from 25- to 34-year-olds, against 9.9% for 45 to 54 and 8.1% for 55 to 64. In September 2026 it took between 4.5% and 4.8% of requests in every group under 65, and 3.9% among those 65 and older. The age difference has all but disappeared. The youngest group's share of technical requests fell from 43.9% to 30.7%.

The most masculine topic

Technical help remains the topic with the lowest share of feminine-name users, though the gap has narrowed. Feminine-name users sent 26.7% of technical requests in July 2024 and 37.2% in September 2026, after reaching 40.0% in May 2026. By PPC Land's calculation, technical help made up 17.1% of masculine-name requests and 8.6% of feminine-name requests in July 2024; in September 2026 the figures were 6.0% and 3.3%.

Only 17 countries have gender figures for technical help in all three months of the third quarter of 2026, a sign of how small the category has become. Among them, Indonesia was the only one with a feminine majority, at 50.3%. Argentina (47.7%), the Philippines (46.7%) and Colombia (45.9%) came closest. Pakistan had the lowest feminine share, at 27.0%, followed by Germany (35.1%), India (35.3%), the United Kingdom (36.1%) and Turkey (36.2%).

Technical help in US states

Every US state recorded a decline between the first quarter of 2025 and the third quarter of 2026. Nationally, technical help fell from 10.0% to 4.2% of requests. Massachusetts, Vermont and New Hampshire fell most, each by about 8 points. In the latest quarter the range was narrow, from 2.6% in Delaware to 4.9% in Idaho, Iowa, Montana and West Virginia.

Programming at work

In the activity file, "Program computer systems or production equipment" fell from 7.8% of work requests on average across the 74 countries in the first quarter of 2025 to 3.1% in the third quarter of 2026, the largest decline of any activity. In the United States it fell from 6.7% to 2.4%. India still devoted 5.8% of its work requests to programming in the latest quarter, making it the third-largest work activity there. "Resolve computer problems" fell from 4.0% to 3.1%, "Research technology designs or applications" from 1.7% to 0.6%, and "Analyze scientific or applied data using mathematical principles" from 1.1% to 0.5%.

One computing activity grew. In the United States, "Design computer or information systems or applications" rose from 0.5% to 0.8% of work requests, a small figure but the opposite direction to programming itself.

Multimedia

Multimedia is the topic that grew fastest. It took 2.0% of requests in July 2024 and 8.8% in September 2026, more than four times as much. The category combines creating images, analysing images users upload and generating or retrieving other media; the instructions list "Draw an astronaut riding a unicorn," "What does this sign say?" and "Create a spreadsheet for mortgage payments" among the examples.

Its path has two spikes. The first came in October 2024, when the share reached 4.4%. The second, much larger, came in April 2025, at 9.4%. OpenAI had introduced image generation built into GPT-4o on March 25, 2025, and the multimedia share rose from 5.4% in March to 9.4% the following month, before falling to 6.9% in May. It slid to a low of 4.7% in December 2025, then climbed through 2026 to 7.6% in May, 8.1% in July and August, and 8.8% in September.

Competition in images intensified over the same period. Google said in October 2025 that more than 5 billion images had been generated with Nano Banana, its Gemini image model, in about two months.

Multimedia at work

Image work has become a significant share of work use. Within work requests, multimedia rose from 1.7% in July 2024 to 9.7% in September 2026, with a high of 10.7% in July 2026. Outside work, it rose from 2.3% to 8.4%. The share of multimedia requests that were work-related fell from 44.0% to 35.6%, and was lowest in April 2025, at 31.4%, the month after OpenAI's image launch; it recovered to 45.1% in November 2025.

PPC Land's country calculation shows wide differences in the work share. In Singapore, 54.8% of multimedia requests in the third quarter of 2026 were work-related, followed by Taiwan (52.3%), Indonesia (51.2%) and Thailand (51.0%). In Iraq, Libya, Myanmar and Afghanistan, fewer than one in five were.

Latin America leads

Multimedia rose in all 137 countries. The median went from 2.8% to 7.8%, and 23 countries now direct more than 10% of their requests to images and media.

Latin America and South Asia dominate the top of the list. Panama was highest in the third quarter of 2026, at 12.4%, followed by Guatemala (12.0%), Peru (11.9%), Afghanistan (11.7%), Paraguay (11.6%), India (11.5%), Bolivia and Honduras (both 11.2%), Burkina Faso (11.1%), Ecuador (10.9%), Brazil and Pakistan (both 10.8%). Zimbabwe's share rose most, by 10.3 points, followed by Panama, Peru, Guatemala and India, each up more than nine points.

The lowest shares were in Japan, Montenegro and Ireland (all 5.5%), Slovenia and Moldova (both 5.6%), and South Korea, New Zealand, Denmark, Australia and Sweden (all 5.7%). Japan and South Korea also had the smallest increases among large markets, of 2.0 and 2.1 points; Estonia and the Czech Republic rose by less than two points.

Source: OpenAI Signals v3.0; PPC Land's quarterly averages.

From older users to everyone

In July 2024, multimedia was the topic that skewed oldest. By PPC Land's calculation it accounted for 1.3% of requests from 18- to 24-year-olds and 3.7% from 55- to 64-year-olds; users aged 55 to 64 were 1.87 times as likely as average to send one. By September 2026, multimedia accounted for between 8.3% and 9.3% of requests in every group under 65, and 10.4% among those 65 and older, the highest of any age group. Users aged 65 and older now send a larger share of their requests about images and media than any other group.

The largest gender shift in the data

Multimedia moved more than any other topic on gender. Feminine-name users sent 30.0% of multimedia requests in July 2024 and 50.5% in September 2026, a swing of 20.5 points. They first sent a majority in February 2026, at 51.4%. By PPC Land's calculation, multimedia took 1.4% of feminine-name requests and 2.4% of masculine-name requests in July 2024; by September 2026 the figures were 8.6% and 9.0%, almost level.

The shift ran through most large markets. In the United States the feminine share of multimedia rose from 30.1% to 54.9%, in the United Kingdom from 27.1% to 52.1%, in Germany from 26.7% to 52.2%, and in Turkey from 23.4% to 49.5%. Argentina (62.8%), Colombia (62.2%), Ukraine (61.7%), Brazil (61.4%) and Poland (60.0%) had the highest feminine shares in the third quarter of 2026; Pakistan (28.3%), Bangladesh (30.6%) and India (36.7%) the lowest.

Multimedia in US states

US multimedia requests rose from 3.5% in the first quarter of 2025 to 7.4% in the third quarter of 2026, and the share increased in every state. West Virginia had the highest share in the latest quarter, at 9.2%, followed by Alabama (8.9%) and Idaho, Kentucky and Wyoming (all 8.7%). Hawaii had the lowest, at 5.2%, then Vermont (5.8%), South Dakota (6.0%), North Dakota (6.1%) and Montana (6.2%). Wyoming's share rose most, by 7.0 points.

Visual design at work

The activity file shows the same change in work terms. "Create visual designs or displays" rose from 3.1% of work requests on average across 74 countries in the first quarter of 2025 to 7.9% in the third quarter of 2026, the largest gain of any of the 165 activities. In the latest quarter it was the second-largest single work activity in the United States (7.6%), India (8.7%), Brazil (11.2%), the United Kingdom (6.6%), Germany (7.2%) and Japan (6.7%), behind only editing. In the United States it rose in all but one quarter, from 2.4% to 7.6%.

Self-expression

Self-expression is the category for requests that are not really tasks. It combines greetings and chitchat ("I had an awesome day today; how was yours?"), relationships and personal reflection ("My wife is mad at me, and I don't know what to do.") and games and role play, where the examples include a user casting the chatbot as a dungeon master or as a Klingon. Its share of requests more than doubled, from 4.0% in July 2024 to 8.9% in September 2026.

The rise came early. Self-expression reached 6.4% in December 2024, 9.2% in April 2025 and a peak of 10.0% in July 2025. Since then it has moved between 7.3% and 9.7% without a clear trend; it was 9.7% in February 2026, 7.5% in April and 9.6% in August.

The topic sits close to what PPC Land's explainer series calls a companion chatbot: an AI persona built for an ongoing relationship rather than a task. The OpenAI files do not separate role play from small talk or personal reflection, so they cannot show how much of the category is companionship. The business model built around companionship has been under strain elsewhere: Character.ai, which had more than 20 million monthly active users according to a Pennsylvania complaint, filed for Chapter 11 on October 12. Surveys also suggest that people tell chatbots things they do not tell other people. DuckDuckGo's survey of 1,944 US adults, published in August, found that 32% of AI users had told a chatbot something they had withheld from a close friend, parent, colleague, doctor or therapist, and 56% of self-described AI enthusiasts had.

The least work-related topic

Self-expression is the topic least connected to work. Between 5.4% (December 2024) and 10.2% (July 2026) of self-expression requests were classified as work-related, and the figure was 9.2% in September 2026. Within non-work requests, self-expression rose from 7.6% to 11.9%, after peaking at 13.6% in July 2025. Within work requests it went from 0.7% to 2.5%.

By PPC Land's country calculation, Singapore (14.3%), Australia (14.1%) and Japan (13.8%) had the highest work shares for self-expression in the third quarter of 2026, and Iraq the lowest (2.6%).

Japan stands apart

No country uses ChatGPT for self-expression like Japan. Its share rose from 3.8% of requests in July 2024 to 15.2% in April 2025 and 19.9% in February 2026, and stood at 17.4% in September 2026. In the third quarter of 2026 Japan's average, 17.2%, was the highest of the 137 countries and nearly twice the median of 9.2%. Among Japanese non-work requests, 20.4% were self-expression, against 12.2% in the United States. Even at work, 8.6% of Japanese requests fell into the category, more than double the US share of 3.7%.

Japan is followed by Iraq (15.5%), Azerbaijan and Myanmar (both 15.3%), Libya (13.2%), Guinea (12.8%), Taiwan (12.6%), Somalia (12.5%), Algeria and Afghanistan (both 12.3%) and Mexico (12.0%). The lowest shares were in the Philippines, the UAE and Cyprus (all 6.4%), Pakistan (6.5%), India, Rwanda and Luxembourg (all 6.6%) and Bangladesh (6.8%).

Self-expression rose in 132 of the 137 countries. Japan's gain, 11.3 points, was the largest, followed by Myanmar (9.4), Iraq (8.7), Taiwan (8.6), Iceland (8.5), Mexico (7.7), the Palestinian territories (7.3) and South Korea (7.1). The five exceptions were Sudan, Vietnam, Bangladesh, Somalia and the Republic of the Congo. Vietnam is an unusual case: in the third quarter of 2024, 14.3% of its requests were self-expression, far above any other large market at the time, and the share has since fallen to 9.2%.

Source: OpenAI Signals v3.0; PPC Land's quarterly averages.

Youngest and oldest

Self-expression has a U-shaped age profile. By PPC Land's calculation, in September 2026 it took 9.8% of requests from 18- to 24-year-olds and 10.5% from those 65 and older, against 8.4% to 8.5% for every group in between. The same shape existed in July 2024, at lower levels: 5.2% for the youngest, 6.5% for the oldest and between 3.1% and 3.6% in the middle.

The most feminine topic

Self-expression has the highest share of feminine-name users of any topic. It rose from 43.0% in July 2024 to a high of 62.6% in March 2026 and stood at 60.1% in September 2026. By PPC Land's calculation, self-expression made up 10.3% of feminine-name requests in September 2026 and 7.4% of masculine-name ones; in July 2025, at the topic's peak, the figures were 12.2% and 7.8%.

Of the 33 countries with reported values, 30 had feminine majorities in the third quarter of 2026. Argentina (68.5%), Brazil (67.6%), South Korea (67.4%), Indonesia (66.7%) and Ukraine (65.8%) were highest. Pakistan (39.9%), Bangladesh (42.7%) and Nigeria (46.7%) were the only three with masculine majorities. In Turkey the feminine share of self-expression almost doubled, from 33.4% to 62.3%.

Self-expression in US states

US self-expression rose from 5.9% of requests in the first quarter of 2025 to 9.5% in the third quarter of 2026, and the share increased in every state. West Virginia had the highest share, at 13.1%, after an 8.3-point increase. Maine, Michigan and Wyoming (all 11.5%) and North Dakota (11.3%) followed. The lowest were Florida and Vermont (both 8.2%), Rhode Island (8.6%), California (8.8%) and New Hampshire (8.9%).

Other/Unknown

One of the smallest categories in 2024 is now one of the faster-growing. Other/Unknown combines three labels: questions about the model itself ("Who made you?", "Are you an AI or a human?"), requests that are clear but fit none of the other labels, and requests where "it is hard to tell what the user even wants." Its share rose from 3.5% in July 2024 to 8.1% in September 2026, the highest level in the series.

Most of that increase came late. Other/Unknown was 5.5% in September 2025, 6.1% in June 2026, 6.8% in July, 7.5% in August and 8.1% in September. Between June and September 2026 it gained two points, twice as much as in the previous twelve months.

A work story

The late increase was concentrated in work requests. Within work requests, Other/Unknown went from 2.2% in July 2024 to 4.1% in December 2025 and 5.1% in June 2026, then 6.1% in July, 7.8% in August and 9.0% in September. Within non-work requests it moved from 4.9% to 7.6%. The share of Other/Unknown requests that were work-related fell from 32.0% in July 2024 to 21.4% in July 2025, then climbed to 35.9% in September 2026, its highest level.

The README does not explain the change. The timing coincides with a product change PPC Land reported: on July 9, OpenAI merged its chat, coding and browser interfaces into one desktop app with Chat, Work and Codex modes, retiring the Atlas browser. The README says Codex data is excluded; it does not say how requests in the Work mode are treated, and the files cannot show what the additional unclassified work requests contain. The September jump in the share of seeking information within work requests happened over the same months.

Who sends unclassified requests

Other/Unknown requests rise with age. By PPC Land's calculation, they made up 7.5% of requests from 18- to 24-year-olds in September 2026 and 10.5% from those 65 and older. The feminine-name share of the category rose from 37.5% in July 2024 to 51.1% in March 2026, then fell to 46.2% by September, the largest late reversal of any topic. Masculine-name users devoted 9.0% of their requests to the category in September 2026, feminine-name users 7.2%.

The category rose in all 137 countries. Somalia had the highest share in the third quarter of 2026, at 11.9%, followed by Mongolia (11.5%), Ethiopia (10.9%), Mali (10.7%), Nepal (10.4%), the Republic of the Congo and Tanzania (both 10.3%) and Haiti (10.2%). The lowest were Botswana (3.9%), Brunei (4.9%), Luxembourg (5.0%), Bolivia (5.1%) and Denmark (5.3%). Among large markets, India (8.7%) and Pakistan (8.2%) were highest and the Philippines (5.6%), the United Kingdom and Australia (both 5.9%) lowest. In US states, the share ranged from 5.0% in Delaware to 9.0% in New Hampshire, and rose everywhere.

Work and non-work

Work-related requests fell from just over half of consumer ChatGPT use to under a third. The binary classifier marked 51.6% of requests as work-related in July 2024, 40.3% in January 2025, 31.4% in July 2025, 30.8% in January 2026 and 29.5% in August 2026, the lowest point. September 2026 brought a rise to 32.2%. Non-work use, by the same measure, went from 48.4% to 67.8%.

This is the same direction as OpenAI's 2025 paper, which found non-work use rising from 53% to above 70% between June 2024 and June 2025. Version 3.0 has different levels: it puts work-related use at 51.6% in July 2024, against the paper's 47% for June 2024, and at 33.3% in June 2025. The README does not reconcile the two; version 3.0 rests on restated samples and on 1 million requests a month instead of the smaller samples used before.

Work, school and everything else

The second classifier separates work from schoolwork. It shows that much of what the binary classifier calls work-related is homework. In July 2024, 27.3% of requests were for work, 24.0% for school and 48.7% for neither. In September 2026 the shares were 21.0%, 13.5% and 65.5%. In every month, work plus schoolwork in the three-way classifier sits within nine points of the binary work-related figure, which suggests the binary prompt counts most school assignments as work. The README does not discuss the overlap.

Schoolwork follows the academic calendar. It peaked at 32.1% of requests in November 2024 and fell to 13.0% in July 2025, rose to 21.2% in November 2025 and fell again to 10.7% in August 2026, the lowest point in the series, before recovering to 13.5% in September. Work, by contrast, has been stable since early 2025, between 17.8% and 21.0% of requests. Most of the long-run fall in "work-related" requests comes from two things: a fall in work from 27.3% to around 19% between July 2024 and early 2025, and a smaller share of schoolwork in each successive school year.

No country is majority-work any more

In the third quarter of 2024, work-related requests were the majority in 44 of the 137 countries. In the third quarter of 2026 they were the majority in none. The median country's work share fell from 46.4% to 27.9%.

Singapore had the most work-oriented use in the latest quarter, at 42.9%, followed by Peru (41.1%), Zimbabwe (40.9%), South Africa (40.1%), Indonesia (39.9%), the Philippines (38.8%), Colombia (37.4%), the UAE (37.2%) and Australia (37.1%). The least work-oriented were Iraq (12.4%), Libya and Afghanistan (both 14.3%), Somalia and Algeria (both 15.2%), Georgia and Uzbekistan (both 16.7%).

The largest falls were in Tunisia (36.7 points), North Macedonia (35.1), Ethiopia (29.7), Ireland (29.5), Rwanda (29.3), the United Kingdom (29.2), the Netherlands (28.8) and Israel (28.7). The smallest were in Sudan, Guatemala, Bolivia, the Philippines, Peru, Ecuador, Mozambique and Vietnam, all under five points.

Source: OpenAI Signals v3.0; PPC Land's quarterly averages.

The pattern runs against wealth. Several of the markets where work use held up best - Indonesia, the Philippines, Vietnam, Peru - have younger and less affluent user bases; several where it collapsed - the United Kingdom, the Netherlands, Ireland, France - are among the richest. One explanation the data allows, without proving it, is that companies in wealthier countries moved staff onto enterprise accounts, which the files exclude, and which is why the README warns that the data understates business use. Not all office work moved, though. Akamai's October report, using data from the browser security firm LayerX, found 47.11% of enterprise AI conversations running through personal rather than corporate accounts, a practice known as shadow AI. Those conversations would appear in OpenAI's consumer data.

What work requests are about

The topic mix of work requests changed as much as the volume. Writing remained the largest work topic, but its share fell, and technical help gave up most of its share to multimedia, seeking information and Other/Unknown.

Source: OpenAI Signals v3.0.

Country differences in the work mix are wide. In the third quarter of 2026, writing took 47.3% of Philippine work requests and 46.1% of those from the UAE, but 28.4% in Turkey and 29.5% in India. Technical help still took 11.2% of Indian work requests and 10.2% of Turkish ones, against 4.7% in the Philippines. Multimedia accounted for 14.2% of Mexican work requests and 13.7% of Brazilian ones, but 6.6% in Australia. Seeking information took 15.4% of work requests in Turkey and Poland, and 8.8% in the Philippines.

Asking, doing and expressing

The third classifier sorts each request by what the user wants. The instructions define three categories. "Asking is seeking information or advice that will help the user be better informed or make better decisions, either at work, at school, or in their personal life." "Doing messages request that ChatGPT perform tasks for the user." And "Expressing statements are neither asking for information, nor for the chatbot to perform a task." The README publishes the split separately for work and non-work requests.

Expressing grew fastest in both. Among non-work requests, it went from 13.9% in July 2024 to 31.2% in September 2026, peaking at 33.6% in July 2026. Asking fell from 53.6% to 41.5%, and doing from 32.5% to 27.3%, with a low of 22.0% in July 2026. Among work requests, expressing rose from 9.6% to 24.8%, asking fell from 38.2% to 27.3%, and doing from 52.3% to 47.8%, with a low of 45.1% in July 2026.

Source: OpenAI Signals v3.0.

Combining the two with the work share, PPC Land calculates that across all requests, asking fell from about 45.7% to 36.9%, doing from 42.7% to 33.9%, and expressing rose from 11.7% to 29.1%. The 2025 paper, as PPC Land reported it in September 2025, put the split at 49% asking, 40% doing and 11% expressing. That is close to where version 3.0 starts in July 2024, not where it ends. By the latest data, almost three in ten requests to ChatGPT neither ask for anything nor ask for anything to be done.

The rise of expressing is larger than the rise of the self-expression topic, which reached 8.9%. Expressing statements therefore appear across topics: a user sharing a reaction, a comment or a piece of context inside a writing or advice conversation is counted as expressing, whatever the topic.

Japan asks at work

Doing remains the leading intent for work requests in 123 of 124 countries with complete data in the third quarter of 2026. The exception is Japan, where 35.0% of work requests were asking and 34.0% doing; Japan also had the lowest doing share at work of any country, ahead of Norway (36.4%), Serbia (36.7%) and Albania (37.4%). Honduras (55.0%), the Philippines (54.5%), Guatemala and Nicaragua (both 53.6%) and Indonesia (53.5%) had the highest.

Outside work, asking still leads in 128 of 129 countries. Azerbaijan is the one where expressing came first, at 38.8% of non-work requests. Japan had the highest non-work expressing share, at 40.3%, followed by Algeria (39.0%), Azerbaijan, Mongolia (38.7%) and Taiwan (38.2%). Japan and South Korea had the lowest non-work doing shares, at 15.5% and 17.2%; Nigeria (34.3%), the Philippines (33.8%) and India (31.2%) the highest. The median country's non-work expressing share rose from 14.1% to 32.8%.

In the United States, non-work requests went from 57.0% asking, 32.0% doing and 11.0% expressing in the third quarter of 2024 to 45.2%, 21.4% and 33.5%. US work requests went from 36.7%, 54.4% and 8.9% to 31.2%, 42.2% and 26.6%.

Free, Plus, Pro and Go

The plan file shows the share of each plan's requests that are work-related. It starts in September 2024 for the Free and Plus plans, December 2024 for Pro and August 2025 for Go. It does not show how requests are split between plans.

Source: OpenAI Signals v3.0.

Paying users are more likely to use ChatGPT for work, and the more they pay, the more work-oriented their use. In September 2026 about a quarter of Free requests were work-related, half of Plus requests, and more than two thirds of Pro requests. Go, the low-cost plan, sits between Free and Plus.

Work use fell on every plan except Pro. The Free plan lost 19.3 points between September 2024 and September 2026, and the Plus plan 11.1 points, though Plus recovered from its 2025 low. Pro moved the other way, with work-related requests reaching 78.3% of the total in October 2025 and never dropping below 59.1%.

For advertisers, this is the most directly relevant table in the release. OpenAI shows ads only to Free and Go users; Plus, Pro and Enterprise subscriptions remain ad-free. The company confirmed the Free and Go ad tests on January 16, when it also set Go at $8 a month in the US, began testing ads in the US on February 9, went live in India at the end of August, and on October 21 extended ads to 89 countries, saying more than 100,000 advertisers had run campaigns. By the September 2026 data, three quarters of Free requests and 57.6% of Go requests were not work-related. The audience ChatGPT ads reach is overwhelmingly using the product for personal matters.

The file does not show the size of each plan. OpenAI said in February that more than 50 million users were paying consumer subscribers, out of 900 million weekly users. The plan-level work shares cannot be combined into a total without knowing how many requests each plan sends.

Age: the user base grew older

ChatGPT's consumer users aged 18 to 24 sent 39.3% of requests in July 2024. Their share rose to 49.6% in November 2024 - almost half of all requests - and then fell, to 36.6% in July 2025 and 30.6% in September 2026. Users aged 25 to 34 first overtook them in August 2025, at 35.8% against 34.7%, and have led every month since December 2025.

Every older group gained. Users aged 35 to 44 went from 16.6% to 19.0%, those aged 45 to 54 from 6.9% to 9.4%, those aged 55 to 64 from 2.3% to 4.0%, and those 65 and older from 0.9% to 2.7%. The two oldest groups together sent 3.2% of requests in July 2024 and 6.7% in September 2026. The README notes that users under 18 are excluded altogether, so the youngest group in the files is adults of university age.

Source: OpenAI Signals v3.0.

The youngest group's share follows the academic year. It peaked in the autumn of 2024, fell in the summer of 2025 to 34.7% in August, recovered only slightly that autumn and fell again in the summer of 2026, to 30.0% in August. The pattern matches the schoolwork series described in the work chapter.

Topics by age, side by side

PPC Land's calculation of the topic mix within each age group shows how differently the groups use ChatGPT in September 2026:

Source: PPC Land's calculation from OpenAI Signals v3.0 age-by-topic and topic shares.

The gradients are clear. Practical guidance falls with age, seeking information rises with it until 65, and writing declines gently. Users 65 and older have the most even spread across topics: they are the group most likely to use ChatGPT for self-expression, images and unclassified requests. Technical help, once the topic with the steepest age gradient, is now almost flat.

Age by country

The youngest users dominate in South Asia, Southeast Asia and Africa. In the third quarter of 2026, 18- to 24-year-olds sent 51.4% of requests from Myanmar, 48.8% from Senegal, 48.5% from Bangladesh, 47.8% from Zambia, 46.6% from Pakistan and 44.4% from India. The lowest shares were in Malta (13.4%), Luxembourg (16.2%), South Korea (17.2%), Puerto Rico (17.7%), Portugal (17.8%) and the UAE (17.9%).

Of the 119 countries with age data in every month, the 18-24 group was the largest in 74 in the third quarter of 2024 and in 40 two years later. The 25-34 group now leads in the other 79. The median country's youngest share fell from 39.7% to 28.7%.

Source: OpenAI Signals v3.0; PPC Land's quarterly averages.

Vietnam's change is the largest among big markets: the youngest group's share fell by almost 30 points, from two thirds of requests to just over a third. Nigeria moved the other way, with the youngest share rising slightly. In the United States, users aged 45 and older now send almost a quarter of requests, and in Spain and South Korea a quarter.

Users 55 and older were most prominent in Israel (11.9% of requests), Puerto Rico (11.4%), Denmark (10.4%), South Korea and the United States (both 10.3%) and Italy (10.0%). Somalia and Afghanistan also show high shares in the oldest group, 8.7% and 8.5% of requests from users 65 and older, but more requests there come from that group than from users aged 45 to 64 combined. That pattern points to how ages were entered rather than to an elderly user base; age in the files is self-reported.

Outside the data, surveys show the same gradient in adoption. Edison Research's May 2026 AI User Metrics survey of 1,030 US adults found that 78% of adults aged 18 to 34 had used an AI platform in the past week, against 70% of those aged 35 to 54 and 51% of those 55 and older. Weekly AI use among US adults overall had risen to 65% from 52% in February.

Gender: a feminine majority since 2025

The files classify users by whether their first name is typically masculine or feminine, using a 95% threshold. On that measure, feminine-name users sent 42.2% of requests in July 2024 and 51.8% in September 2026. They first passed half in May 2025, at 50.1%, slipped just below over the summer and have stayed above since September 2025. The peak was 53.7% in May 2026; the share then eased to 51.8% in August and September.

Source: OpenAI Signals v3.0.

Every topic gained feminine-name users. The gains were largest in multimedia and self-expression and smallest in writing and seeking information. At both ends of the series technical help is the most masculine topic, but self-expression has replaced writing as the most feminine, and multimedia has moved from the second most masculine topic to the middle of the list.

PPC Land's calculation of what each group uses ChatGPT for shows the differences more directly. In September 2026, feminine-name users sent 29.0% of their requests about practical guidance, 23.4% about writing, 18.2% seeking information and 10.3% on self-expression. Masculine-name users sent 27.0%, 19.2%, 22.3% and 7.4%. Masculine-name users sent 6.0% on technical help and 9.0% on Other/Unknown, against 3.3% and 7.2%. Multimedia was almost even, at 9.0% and 8.6%.

Gender by country

Across the 107 countries with gender data in every month, the median feminine share rose from 42.7% in the third quarter of 2024 to 52.9% two years later. Ten countries had feminine majorities in 2024; 73 did in 2026.

Moldova had the highest feminine share in the latest quarter, at 66.4%. Paraguay (60.8%), Honduras and Indonesia (both 60.7%), Brazil (60.6%), Argentina (60.5%), Bosnia and Herzegovina (60.4%), South Korea (59.5%), Colombia and the Czech Republic (both 59.4%) followed. The lowest were Afghanistan (28.9%), Pakistan (35.2%), Bangladesh (36.5%), Qatar (38.5%), the Democratic Republic of the Congo (39.4%), Mozambique (40.7%), Nigeria (41.2%), Angola (42.0%) and India (43.0%).

The largest increases were in Turkey (19.6 points, from 31.8% to 51.4%), Egypt (17.9), Azerbaijan (17.8), Israel (17.2), Hungary and the Czech Republic (both 16.8), Slovakia (16.1) and Uzbekistan (15.8). Pakistan barely moved, from 34.4% to 35.2%, and Qatar, Myanmar, Malaysia and Bangladesh gained less than four points.

Source: OpenAI Signals v3.0; PPC Land's quarterly averages.

The US feminine share first passed 50% in September 2025, at 51.0%, and was 53.4% in September 2026. Japan still has a masculine majority, at 45.2%, despite gaining 11.9 points, while neighbouring South Korea's feminine share is 59.5%.

What the gender figures can and cannot show

The measure has limits that matter for interpretation. It covers only users whose first names clear the 95% threshold against external name databases, so people with names common to both genders, names the databases do not cover, or names in scripts the matching handles poorly are not counted, and the README does not say how many that is. The match rate could differ by country, which would make cross-country comparisons less reliable than within-country trends.

Name-based inference is itself a research subject. A study PPC Land covered in March found that GPT-4o guessed sex or gender from a name with 94.4% accuracy, though that tested a model's guesses rather than the database matching OpenAI describes. Survey data also runs in a different direction. Edison Research's May survey found that 70% of US men and 61% of US women had used an AI platform in the previous week. Higher weekly use among men can coexist with a feminine majority of ChatGPT requests if women who use ChatGPT send more requests each, or if men spread their use across more platforms; the Signals data cannot distinguish between the two.

Which countries use ChatGPT most per person

The ranking file orders countries by requests per person, using 2023 World Bank population figures. Rank 1 is the highest per-capita use. Between 170 and 176 countries appear each quarter. China, Russia, Hong Kong, Iran, Venezuela and Belarus do not appear at all.

Small, wealthy countries dominate the top. Singapore ranked first in the third and fourth quarters of 2025 and second in every other quarter except the first of 2025, when it was third. The Netherlands has been in the top four since the third quarter of 2025. Estonia, Montenegro, Latvia, Andorra, Malta, Luxembourg, the UAE, the Cayman Islands and Switzerland complete the top 12 in the third quarter of 2026.

The most striking move is Azerbaijan's. It ranked 86th in the first quarter of 2025, 45th in the second, 14th in the third and 3rd in the fourth. After 16th and 11th in the first half of 2026, it ranked first in the third quarter. Azerbaijan also stands out in the topic data: 15.3% of its requests were self-expression in the third quarter of 2026, and it is the only country with complete data where expressing outranks asking in non-work requests. The files do not explain the change. Two reports from the Azerbaijani technology site tech.az give some context. On May 29, 2026 it reported that the country's science and education ministry, its digital development ministry and OpenAI had signed a memorandum on an AI-based adaptive learning platform for more than 500,000 students. On June 9, it reported that OpenAI's chief financial officer, Sarah Friar, had named Azerbaijani among the fastest-growing languages on ChatGPT in a podcast interview. Azerbaijan's rank had already reached third by the end of 2025, before either report.

Tiny territories flicker in and out of the list. Seychelles ranked first in the first two quarters of 2025, Monaco first in the first quarter of 2026, and American Samoa first in the second quarter of 2026, each appearing only in some quarters. Their small populations and the README's suppression rule for cells with 100 or fewer requests make their positions unstable.

Source: OpenAI Signals v3.0.

Among the 167 countries ranked in every quarter, the biggest gains after Azerbaijan were Japan (113th to 62nd), South Korea (67th to 24th), Argentina (110th to 72nd), New Caledonia, Saudi Arabia, Turkey, Uruguay and Brazil. The biggest falls were Moldova (15th to 58th), Guadeloupe (39th to 78th), Kazakhstan (30th to 64th), Gabon, Kyrgyzstan, Brunei, Italy (64th to 88th) and Ukraine (65th to 89th). Canada slipped from 17th to 30th and Denmark from 22nd to 37th.

The United States has never ranked higher than 40th, and was 49th in the third quarter of 2026. That reflects the measure, not the volume: the US is a very large market in absolute terms but not, on a per-person basis, among the heaviest users. India, with the largest population in the ranking, rose from 133rd to 119th. The bottom of the list is held by countries in the Sahel and Central and East Africa: Niger, South Sudan, Chad, the Central African Republic, Burundi, Mali, Sudan and Malawi were the eight lowest in the third quarter of 2026.

Which US states use ChatGPT most per person

Washington, DC, ranked first among the 50 states and the District in every quarter from the first quarter of 2025 to the third quarter of 2026. New York has been second in four of the seven quarters, including the last three. Behind them the order has shifted.

Source: OpenAI Signals v3.0.

Delaware moved furthest, from 22nd to 3rd. Alabama (38th to 26th), Nebraska (33rd to 22nd), Texas (17th to 8th) and Alaska (48th to 39th) also gained. Oregon fell from 12th to 25th, New Jersey from 7th to 19th and Idaho from 31st to 43rd. The four lowest-ranked states in the third quarter of 2026 were South Dakota, West Virginia, Iowa and Montana.

The state topic data, covered in each topic chapter above, adds a pattern of its own. Washington, DC, combines the highest per-capita use with the highest writing share and one of the lowest seeking information shares. West Virginia, ranked 50th for use, has the highest shares of self-expression and multimedia requests and one of the highest for Other/Unknown. Delaware, the fastest riser in the ranking, has the highest seeking information share and the lowest technical help share. The README cautions that state-level topic shares rest on smaller subsets of the sample and carry more noise than national figures.

Work activities: marketing materials hold their place

The O*NET activity file, the subject of OpenAI's October 26 page, can also be read at the level of individual activities. It lists 164 named activities plus an "Other IWA" bucket for rare ones, by country and quarter, from the first quarter of 2025 to the third quarter of 2026. PPC Land's sums of these rows reproduce the group-level figures in OpenAI's interactive chart: making decisions and solving problems, for example, totals 21.6% of US work requests in the first quarter of 2026.

Averaged across the 74 countries reported in every quarter, the largest single activities in the third quarter of 2026 were editing written materials (15.0%), creating visual designs or displays (7.9%), developing marketing or promotional materials (5.1%), interpreting language or cultural information (4.0%), preparing informational or instructional materials (3.2%) and programming computer systems (3.1%).

For marketers, "Develop marketing or promotional materials" is the closest activity to their own work. It held almost steady, at 5.2% of work requests on average in the first quarter of 2025 and 5.1% in the third quarter of 2026, after a high of 6.0% in the third quarter of 2025. In the United States it was 5.0% at both ends of the period. It was highest in Argentina (9.4%), Pakistan (9.0%), Nigeria (9.0%), Brazil (8.6%), Egypt (7.6%), Ukraine, Lebanon and Bangladesh (all above 7%), and lowest in Ethiopia (1.8%), Ireland (2.8%), Norway (2.9%), Croatia and Switzerland (both 3.3%). Brazil's share fell from 10.4% to 8.6%, the UAE's from 6.0% to 4.0%, and Mexico's rose from 5.6% to 6.7%.

Related commercial activities are small but growing. "Promote products, services, or programs" rose from 0.1% to 0.2% of work requests on average; in the US it went from 0.09% to 0.27%. "Sell products or services" rose from 0.11% to 0.18%. "Develop business or marketing plans" stayed between 0.8% and 1.0%. "Analyze market or industry conditions" fell from 0.55% to 0.38%, and "Collect data about consumer needs or opinions" stayed below 0.1%.

The marketing activity's stability contrasts with the drop in writing overall. Requests to produce promotional copy held their share of work use while general editing, translation and programming lost share, and visual design more than doubled.

What the data does not show

Several limits run through every chapter above.

The files report shares, not volumes. A falling share for writing or technical help is consistent with more writing or coding requests in absolute terms, if total requests grew fast enough. OpenAI's user numbers rose sharply over the period: ChatGPT had 700 million weekly users in July 2025, according to the 2025 paper, and 900 million in February 2026. The Signals files give no request counts, no user counts and no country totals.

The sample covers consumer accounts only. Enterprise accounts are excluded, and so is Codex, which OpenAI has been building into its main desktop app. Some of the decline in work and technical requests may be use moving into products the files do not measure. OpenAI acknowledges that gap in the README. The plan file confirms that Plus, Pro and Go are in the sample, but it does not mention OpenAI's Business plan.

Every label is the output of a language model reading de-identified text. The README publishes the prompts but not any accuracy checks against human judgments, and a model's reading of "work-related" or "unclear" may not match a person's. The share of requests the topic classifier labels Other/Unknown more than doubled over the period, which means a growing part of use is not being placed in any substantive category.

Noise is added on purpose. Large global shares are barely affected, but cells for small countries, small topics, single US states or narrow O*NET activities are less reliable. Brunei's 0.4% for technical help and the age profile reported for Somalia and Afghanistan are examples where the noise or the inputs, not usage, may drive the figure. Older months have been released several times, and each release added noise under a separate privacy budget.

History has been restated. Version 3.0 raised the monthly sample from 300,000 to 1,000,000 for every month back to July 2024, and version 2.0 removed and replaced a small share of requests. Figures from earlier Signals releases, and from the 2025 paper, will not always match version 3.0.

Gender is inferred from first names. Users whose names do not clear the 95% threshold are excluded from the gender files, and the share excluded may differ by country and language. Age is self-reported, and users under 18 are not in the data at all.

The per-capita rankings use 2023 population estimates for countries and states. They show intensity of use, not market size, and small populations make the top of the country table unstable.

The two work classifiers disagree on levels. The binary classifier's "work-related" share appears to include much of what the three-way classifier calls schoolwork, so a fall in "work-related" requests in summer can reflect school holidays rather than changes in employment use.

Why this matters for marketers

The files describe, for the first time at this level of detail and over more than two years, what consumer ChatGPT is used for and by whom. Several findings bear directly on advertising and content strategy, without settling the commercial questions themselves.

The plans that carry ads are the least work-oriented. In September 2026 three quarters of Free requests and almost three in five Go requests were not work-related. Practical guidance, seeking information and self-expression together accounted for about 57% of all requests in September 2026, against about 47% in July 2024. That is the context in which product questions, recipes, how-to queries and health questions arise, though health topics are excluded from ad placement under OpenAI's January rules.

Seeking information is where search-style intent sits, and its rise has been uneven. In a dozen European countries, from Slovenia and Croatia to Spain, France and Norway, it is now more than one in four requests; in Indonesia, the Philippines, Cambodia, Myanmar and Laos, and in Nigeria, Mali and Sudan, it is about one in seven or fewer. Advertisers moving budgets into ChatGPT across many of the 89 ad markets are buying into countries whose users ask the assistant for very different things. The data does not show how often information requests contain a commercial question, because the purchasable-products label is not published separately.

Writing has lost its central place, but not at work. A third of work-related requests are still about writing, and "Develop marketing or promotional materials" held at about 5% of work requests on average across 74 countries while programming fell by more than half. Visual design is the fastest-growing work activity in the files. For agencies and in-house teams, the data describes a tool used heavily for copy and increasingly for images, rather than one that has stopped being used for content.

The demographics shifted toward groups that marketers often treat as late adopters. Users aged 45 and older now send almost a quarter of US requests and more than a fifth of requests in France and the United Kingdom. Feminine-name users have sent a majority of global requests in every month since September 2025, and in Brazil, Indonesia, South Korea and Poland they send about three in five.

ChatGPT is also one assistant among several, and these shares describe only ChatGPT. Comscore's US panel showed ChatGPT's share of AI prompt volume falling from 70% in January 2026 to 50% in June, and Similarweb's July report found ChatGPT's share of visits to generative AI chatbot websites falling from above 80% to just above 50%. A shift in what people ask ChatGPT could reflect a shift in what they take elsewhere. Basis's October survey of 184 marketing professionals found that ChatGPT use among them fell to 74.5% while Claude rose to 72.3%. The Signals data cannot show where any of the requests that left a category went.

Timeline

• July 2024 - First month of the OpenAI Signals v3.0 series; writing takes 34.1% of consumer requests, technical help 13.5%, and 51.6% of requests are work-related
• November 2024 - Users aged 18 to 24 peak at 49.6% of requests and schoolwork at 32.1%
• December 2024 - The Pro plan first appears in the plan data, with 64.2% of its requests work-related
• March 2025 - Practical guidance overtakes writing as the largest topic, 28.4% against 28.2%
• March 25, 2025 - OpenAI introduces image generation in GPT-4o
• April 2025 - Multimedia peaks at 9.4% of requests
• May 2025 - Feminine-name users pass 50% of requests for the first time, at 50.1%
• July 2025 - Self-expression peaks at 10.0% of requests
• August 2025 - The Go plan first appears in the data; users aged 25 to 34 overtake those aged 18 to 24 for the first time
• September 15, 2025 - PPC Land reports OpenAI's usage paper, with 700 million weekly users and non-work use above 70%
• September 29, 2025 - OpenAI launches Instant Checkout in ChatGPT with Stripe
• October 13, 2025 - Google says more than 5 billion images have been made with Nano Banana
• October 2025 - Pro plan's work share peaks at 78.3%
• Fourth quarter of 2025 - Azerbaijan reaches third in the per-capita ranking, up from 86th at the start of the year
• January 16, 2026 - OpenAI confirms ad tests on the Free and Go plans and sets Go at $8 a month in the US
• February 2026 - Feminine-name users send a majority of multimedia requests for the first time
• February 9, 2026 - OpenAI begins testing ads in ChatGPT in the US
• February 11, 2026 - Ahrefs finds ChatGPT sends 190 times less traffic than Google
• February 27, 2026 - OpenAI reports 900 million weekly ChatGPT users and 1.6 million weekly Codex users
• March 1, 2026 - Study finds GPT-4o infers sex or gender from names with 94.4% accuracy
• March 2026 - Instant Checkout is shut down; Anthropic's Economic Index finds coding moving from Claude.ai to the API
• April 2026 - Practical guidance peaks at 33.1% of requests
• May 2026 - Feminine-name share peaks at 53.7%; technical help's feminine share peaks at 40.0%
• May 15, 2026 - OpenAI opens a personal finance preview for US Pro subscribers
• May 21, 2026 - Smarty Marketing survey puts ChatGPT at 7.26% of US shopping search starts
• May 27, 2026 - Edison Research finds 65% of US adults use AI weekly
• May 29, 2026 - Azerbaijan's ministries and OpenAI sign a memorandum on an adaptive learning platform
• June 9, 2026 - tech.az reports OpenAI's CFO naming Azerbaijani among ChatGPT's fastest-growing languages
• July 2026 - Technical help falls to its low of 4.0% of requests
• July 9, 2026 - OpenAI retires Atlas and merges chat, coding and browsing into one desktop app; Codex passes 5 million weekly users
• July 22, 2026 - Similarweb reports ChatGPT's share of AI chatbot web visits just above 50%
• August 2026 - Writing falls to its low of 20.6%; seeking information peaks at 20.7%; schoolwork falls to 10.7%
• August 8, 2026 - Stack Overflow records 1,442 new questions for July, down 99% from its peak
• August 25, 2026 - DuckDuckGo survey finds 32% of AI users have told a chatbot something withheld from people close to them
• August 30, 2026 - ChatGPT ads go live in India
• August 31, 2026 - European Commission designates ChatGPT a Very Large Online Search Engine with 159.1 million EU users
• September 2026 - Last month of the series; Other/Unknown reaches 8.1% of requests and 9.0% of work requests; Azerbaijan ranks first per capita for the third quarter of 2026
• September 22, 2026 - Comscore reports ChatGPT's US AI prompt share fell from 70% to 50%
• October 12, 2026 - Character.ai files for Chapter 11
• October 19, 2026 - Basis survey finds ChatGPT used by 74.5% of marketers and Claude by 72.3%
• October 21, 2026 - ChatGPT ads reach 89 countries; Akamai reports 47.11% of enterprise AI conversations on personal accounts
• October 26, 2026 - OpenAI releases OpenAI Signals v3.0, 24 files covering July 2024 to September 2026

Related PPC Land coverage

• ChatGPT usage reaches 700 million users as personal applications dominate platform growth - The 2025 OpenAI paper whose classifier prompts the Signals files reuse.
• OpenAI hits 900M users - OpenAI's February 2026 user, subscriber and Codex figures.
• OpenAI kills Atlas as search erosion data hits publishers - The July 2026 desktop app merger of chat, Work and Codex modes, with Codex above 5 million weekly users.
• ChatGPT opens ads to 700M users - here's what marketers need to know - OpenAI's January 2026 ad principles, plan tiers and excluded topics.
• This week: OpenAI starts selling ads as Google rewires AI shopping and Discover - The February 9 start of ChatGPT ad testing in the US.
• OpenAI gains 32 ChatGPT ad markets in 5 days as India goes live - The August 2026 India launch and the Free and Go ad surface.
• Advertisers in 26 more countries gain ChatGPT ads, bringing the total to 89 - The October 21 expansion and OpenAI's cumulative advertiser count.
• ChatGPT faces EU risk rules after declaring 159.1 million users - ChatGPT's designation as a Very Large Online Search Engine under the DSA.
• ChatGPT sends 190x less traffic than Google despite 12% search volume - Ahrefs research on referral traffic from ChatGPT.
• Google still leads shopping searches as AI falls short at 7% - Survey data on where US consumers start product searches.
• ChatGPT Ads Manager now supports product feeds after checkout is killed - The end of Instant Checkout and the move of product feeds into ads.
• ChatGPT can read your bank statements now - here's what it knows - OpenAI's personal finance preview and its claim of 200 million monthly finance users.
• Stack Overflow drops to 1,442 questions in July, down 99% from 2014 peak - The collapse of the developer Q&A site as coding moves to AI tools.
• Experienced Claude users succeed 10% more - and the gap is widening - Anthropic's March 2026 Economic Index and the move of coding toward the API.
• Nano Banana expands to Search and NotebookLM with 5 billion images created - Google's image model and its usage figures in late 2025.
• Character.ai faces wind-down as Disney's $20m bid licenses IP, not the app - The companion chatbot company's Chapter 11 filing.
• DuckDuckGo survey finds 56% of AI enthusiasts confide secrets to chatbots - Survey data on what US adults tell chatbots.
• 65% of Americans now use AI weekly, up 36 million in three months - Edison Research's age and gender figures for weekly AI use.
• Your name tells GPT-4o more about you than you think - Research on what models infer from first names.
• 47% of enterprise AI chats run on personal accounts, Akamai report says - Work done through personal AI accounts, which shows up in consumer data.
• ChatGPT loses 20 points of US AI prompt share to Gemini and Claude - Comscore's panel data on assistant market share.
• ChatGPT loses web share to Gemini and Claude as ad penetration hits 26% - Similarweb's 2026 Generative AI Landscape report.
• ChatGPT drops to 74.5% of marketers as Claude climbs to 72.3%, Basis finds - A survey of which AI tools marketing professionals use.
• IAB Tech Lab explores Differential Privacy for ad targeting - The privacy technique OpenAI uses for every Signals figure, applied to advertising.

Summary

Who: OpenAI's Economic Research team, which publishes the OpenAI Signals data. The data describes adult users of consumer ChatGPT plans - Free, Plus, Pro and Go - in more than 170 countries and every US state, excluding enterprise accounts, Codex, deleted accounts and users who opted out of training.

What: Version 3.0 of OpenAI Signals, 24 files built from LLM classification of 1 million ChatGPT requests a month. Writing fell from 34.1% of requests in July 2024 to 21.4% in September 2026, and practical guidance has been the largest topic since March 2025, at 28.1% in September 2026. Technical help fell from 13.5% to 4.6%, multimedia rose from 2.0% to 8.8%, and work-related requests fell from 51.6% to 32.2%. Feminine-name users sent 51.8% of requests and users aged 18 to 24 sent 30.6%. Practical guidance leads in 116 of 137 countries compared by PPC Land, seeking information is highest in Central and Eastern Europe, and writing remains strongest in Southeast Asia and West and Central Africa. Azerbaijan ranked first for per-capita use in the third quarter of 2026, and Washington, DC, first among US states.

When: Released on October 26, 2026, covering July 2024 to September 2026; quarterly country and state files run from the first quarter of 2025 to the third quarter of 2026.

Where: Published on OpenAI's Signals download page under a Creative Commons Attribution 4.0 license, with country, US state, age, name-gender, plan and work breakdowns.

Why: OpenAI presents the data as a resource for studying how AI is used. For marketers, it is the most detailed public picture of what ChatGPT's consumer users ask for, at a time when the Free and Go plans those users rely on carry ads in 89 countries, and it shows an audience that is older, more often feminine-named and less focused on work than in 2024.