For most of its commercial life, ChatGPT has been described as an office tool: a quicker way to draft an email, tidy a report or fix a formula. OpenAI's own numbers now describe something rather different. Data the company released on Monday, October 26, shows that writing, the main reason people opened ChatGPT in mid-2024, accounted for 21.4% of consumer requests in September 2026, down from 34.1% in July 2024. Over the same 27 months the share of requests related to work fell from 51.6% to 32.2%. The largest single topic, in every month since March 2025, is the one OpenAI labels practical guidance.

PPC Land examined the release in two articles on Tuesday, October 27. The first, published at 12:22 UTC, reported that writing drops from 34% to 21% of ChatGPT requests since July 2024, working through all 24 spreadsheets in the package. The second, published at 11:32 UTC, found that Americans use ChatGPT at work most to make decisions, at about one in five requests, drawing on a companion OpenAI page titled "ChatGPT's expanding world of work".

Why should an advertising audience care about a labour economics dataset? Because ChatGPT now sells advertising, and it sells it only to users of its Free and Go plans, the two tiers whose requests have moved furthest from work. In September 2026, about three-quarters of Free requests and 57.6% of Go requests were not related to work. The release itself, PPC Land noted, "says nothing about advertising." Set beside a new link format that appeared inside ChatGPT's ads on Monday, and a viral privacy claim about a Samsung television from a researcher whose OpenAI account was suspended, the data describes a product whose users increasingly ask it to advise, decide and create, or simply to hear them out, and whose owner is still working out what that audience is worth.

Twenty-four spreadsheets and a classifier

The figures come from version 3.0 of OpenAI Signals, released on October 26 under a Creative Commons CC BY 4.0 licence. The method matters more than usual here, because no human reads the underlying conversations. An automated classifier, itself a large language model, sorts each request into one of 23 detailed labels, which the files then group into seven topics: writing, practical guidance, seeking information, technical help, multimedia, self-expression and other or unknown. Separate classifiers decide whether a message relates to work, school or neither, whether its intent is asking, doing or expressing, and which O*NET work activity it resembles.

OpenAI's README says the dataset "includes statistics based on the classifications of 1,000,000 consumer ChatGPT messages each month," sent by adults with declared ages of 18 or over. The classifier prompts are reused from How People Use ChatGPT, the 2025 paper by Aaron Chatterji and colleagues. Version 3.0 raised the monthly sample from 300,000 to 1,000,000 requests and restated earlier months, which means the new series does not line up exactly with the old one. When PPC Land covered that 2025 paper in September 2025, it reported that ChatGPT usage had reached 700 million users as personal applications dominated platform growth, with non-work requests rising from 53% to more than 70% between June 2024 and June 2025. The restated v3.0 series starts lower, with non-work requests at 48.4% in July 2024, and reaches 67.8% in September 2026.

Privacy protections shape what can be seen. Every figure carries differential-privacy noise, with a privacy budget, epsilon, of 1 for July to September 2026 and a looser budget for earlier months. Cells with 100 or fewer weighted requests are suppressed. Gender is inferred from first names using a 95% prevalence threshold, which the README is careful to describe as "not self-identified gender." Smaller subsets, it adds, "as such have a larger variance."

The exclusions are just as important. Enterprise messages are left out, as are deleted and opted-out accounts, and the README concedes that "these data likely understate business use." Codex, OpenAI's coding agent, is also outside the frame: "We do not include data from Codex." PPC Land added its own calculations, including country comparisons that average three monthly values per quarter across 137 countries, and flagged what the README does not explain, among them an overlap between the work and schoolwork fields and sharp movements in August and September 2026. The topic files do not separate the 23 labels, so label-level weights cannot be seen at all.

Writing gives way to advice

The topic shares, set side by side, tell a clear story. Between July 2024 and September 2026, writing fell from 34.1% to 21.4% of consumer requests. Practical guidance rose from 26.1% to 28.1%. Seeking information rose from 16.7% to 20.2%. Self-expression more than doubled, from 4.0% to 8.9%, and the other or unknown category rose from 3.5% to 8.1%. Two categories moved further than any of these: technical help collapsed and multimedia surged.

Technical help, which covers programming, mathematics and data analysis, took 13.5% of requests in July 2024, one in seven. By September 2026 it took 4.6%. "The fall was steady rather than sudden," PPC Land wrote. One likely reason, PPC Land suggested, is that some of that work has moved to places the dataset does not measure. OpenAI said in February that Codex had 1.6 million weekly active users, more than triple the number at the start of 2026, as PPC Land reported when OpenAI hit 900 million weekly ChatGPT users; by July, it said Codex served more than 5 million people a week. The same pattern appears at a competitor. Anthropic's March Economic Index found coding moving from Claude.ai toward the API, and PPC Land reported in April that experienced Claude users succeed 10% more and the gap is widening, with the ten most frequent tasks falling from 24% to 19% of Claude.ai conversations between November 2025 and February 2026.

Outside the AI companies, the evidence is starker still. Stack Overflow received 1,442 new questions in July 2026, down 99% from its March 2014 peak of 207,204. A fall in technical help within consumer ChatGPT, then, is not evidence that people code less with AI. It is more consistent with coding moving to dedicated agents and developer interfaces, leaving the consumer chat window to other uses.

Multimedia is the topic that grew fastest, from 2.0% of requests to 8.8%. The timing points to a single product decision. OpenAI 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. The commercial follow-through came this autumn. MediaPost's Laurie Sullivan reported on October 6 that OpenAI added a visual ad format and made ChatGPT ads more measurable: ads that appear next to results while users generate images, clearly labelled and kept separate from the generated picture, with US testing due to start later in October with an initial group of advertisers. OpenAI's blog post, quoted by MediaPost, said the format lets advertisers "tell their story through images." The same announcement named brand-suitability pilots with DoubleVerify and Integral Ad Science, conversion pipelines from Tealium, LiveRamp and Hightouch, and incrementality tests from Haus, Measured and WorkMagic.

Images also reach into shopping. Sullivan wrote on October 19 that chatbots drive loyalty through identity and lifestyle, noting that OpenAI had launched a global virtual try-on feature in ChatGPT, built on its Images 2.5 model, which shows how clothes and accessories look on a user. Her caveat was blunt: an AI preview that does not match the item that arrives risks a situation in which "trust with the consumer can be permanently broken." A category that barely registered in mid-2024 has become, in little more than two years, both one of ChatGPT's fastest-growing uses and one of its newest pieces of ad inventory.

Talking, not asking

The intent data shows a deeper change in how people address the machine. OpenAI's classifier sorts each request into asking for information, asking the chatbot to do something, or expressing, which its instructions define as statements that are "neither asking for information, nor for the chatbot to perform a task." On PPC Land's calculation across all requests, asking fell from about 45.7% to 36.9% and doing from about 42.7% to 33.9%, while expressing rose from about 11.7% to 29.1%. The 2025 paper had put the split at 49% asking, 40% doing and 11% expressing.

The shift is sharpest outside work. Among non-work requests, expressing rose from 13.9% to 31.2%, peaking at 33.6% in July 2026; asking fell from 53.6% to 41.5%, and doing from 32.5% to 27.3%. Even at work, expressing climbed from 9.6% to 24.8%. In the United States, non-work requests moved 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% two years later.

People, in other words, increasingly tell ChatGPT things. A DuckDuckGo survey of 1,944 US adults, published in August and cited by PPC Land, found that 32% of AI users had told a chatbot something they had withheld from a close friend, parent, colleague, doctor or therapist; among self-described AI enthusiasts the figure was 56%. That is the kind of conversation OpenAI's advertising rules were written to keep at arm's length. The January rules that accompanied the ad tests bar ads near health, mental health and political topics. Whether the growth in expressing statements falls inside or outside those categories, the files do not say.

The filtering has a commercial cost that advertisers have already felt. Digiday's Krystal Scanlon, Tim Peterson and Sam Bradley reported on May 26 that as OpenAI's ChatGPT ad delivery improved, the doubts it created were not so easily fixed. OpenAI executives argued that underdelivery was the price of careful filtering: ads pass through sensitive-conversation filters, then contextual relevance checks, then the auction. One advertiser had spent $2,500 of a $250,000 commitment over four weeks. "Less than 10% of the eligible base was part of the pilot," one attendee at Digiday's Programmatic Summit said. If a growing share of conversations consists of statements rather than questions or tasks, the pool of ad-eligible moments may grow more slowly than the user base.

An older audience, and a shift in gender

The people typing have changed too. Users aged 18 to 24 sent 39.3% of requests in July 2024, rising to 49.6% in November 2024 before falling to 30.6% in September 2026. The 25 to 34 band, at 34.3% in September 2026, first overtook the youngest group in August 2025 and has led every month since December 2025. Every older band grew: 35 to 44 from 16.6% to 19.0%, 45 to 54 from 6.9% to 9.4%, 55 to 64 from 2.3% to 4.0%, and those aged 65 and over from 0.9% to 2.7%.

The ageing shows up country by country. In Vietnam, the 18 to 24 share fell from 66.3% in the third quarter of 2024 to 36.5% in the third quarter of 2026; in Indonesia, from 62.9% to 43.8%; in the United States, from 30.9% to 21.4%. Older users also want different things. In September 2026, PPC Land calculated, practical guidance took 30.8% of requests from 18 to 24 year olds but 24.9% from those 65 and over, while seeking information rose from 17.3% among the youngest to 26.7% among 55 to 64 year olds. Multimedia was highest of all among the over-65s, at 10.4%. Technical help sat between 4.5% and 4.8% in every band under 65.

Gender has moved further still, with the caveat that it is inferred rather than declared. "Gender in these files is not a user's own declaration," PPC Land wrote. Users with typically feminine first names sent 51.8% of requests in September 2026, up from 42.2% in July 2024, and every topic gained feminine-name users. How many names were dropped for failing the 95% threshold is not published. Edison Research's AI User Metrics survey of 1,030 US adults in May 2026 offers an outside check: 65% of US adults had used an AI platform in the previous week, up from 52% in February, with 70% of men and 61% of women. By age, the figure was 78% for 18 to 34 year olds, 70% for 35 to 54 year olds and 51% for those 55 and over.

For an ad business, this is the more consequential half of the release. A product used mainly by students and young professionals for drafting and debugging is a narrow audience. One used by a population that is older, more evenly split by gender and more interested in advice and images looks much more like a mass medium.

A map from Laos to Slovenia

The country files show how uneven that medium remains. In the third quarter of 2024, work was the majority use in 44 of 137 countries. In the third quarter of 2026, it was the majority in none, and the median country's work share fell from 46.4% to 27.9%. Singapore had the highest work share, at 42.9%; Iraq the lowest, at 12.4%. Writing remained strongest in Laos, at 38.8%, followed by Nigeria at 34.8% and the Philippines at 34.6%, and weakest in Turkey, at 14.6%.

Other topics produce their own outliers. Practical guidance peaked in Sudan, at 36.7%, and was lowest in Benin, at 20.9%. Slovenia, at 31.2%, was the only country where seeking information was the top topic, while Laos, at 10.2%, sat at the bottom. India led on technical help at 5.9%. Panama led on multimedia at 12.4%. Japan led on self-expression at 17.2%, and it was the only country where asking narrowly led doing at work, by 35.0% to 34.0%. It also had the highest non-work expressing share, at 40.3%. Azerbaijan was the only country where expressing led non-work requests, at 38.8%.

Azerbaijan supplies the strangest line in the ranking file, which orders countries by requests per person using 2023 World Bank population figures. Small, wealthy countries dominate the top. Singapore ranked first in the third and fourth quarters of 2025, the Netherlands has been in the top four since the third quarter of 2025, and Estonia, Montenegro, Latvia, Andorra, Malta, Luxembourg, the UAE, the Cayman Islands and Switzerland complete the top 12. Azerbaijan ranked first in the third quarter of 2026, up from 86th in the first quarter of 2025. The files offer no explanation for the jump, and smaller country cells carry the most noise.

Inside the United States, the spread is narrower but still visible. Washington, DC led on writing at 26.1%, against 16.8% in New Hampshire. Mississippi led on practical guidance at 35.2%, Wyoming trailed at 26.9%. Delaware led on seeking information at 26.1%, New Mexico trailed at 16.8%. West Virginia led on both self-expression, at 13.1%, and multimedia, at 9.2%. Europe, meanwhile, already treats ChatGPT as infrastructure: PPC Land reported in August that ChatGPT faces EU risk rules after declaring 159.1 million users, and the European Commission designated it a Very Large Online Search Engine on August 31.

At work, a decision engine

The second file set looks only at work, and here OpenAI's framing was explicit. "Today, we are releasing OpenAI Signals data that explores 165 work activities across nearly 100 countries," its page begins. The 165 categories are Intermediate Work Activities, or IWAs, drawn from O*NET, the US occupational database, and by PPC Land's calculation they amount to just under half of its 332 IWAs. The data runs from the first quarter of 2025 to the third quarter of 2026, carries the names of Aaron Chatterji, Thomas Cunningham, David J. Deming, Zoe Hitzig and Drew Johnston, and is published under the same CC BY 4.0 licence.

The headline is that Americans use ChatGPT at work above all to decide things. In the first quarter of 2026, making decisions and solving problems accounted for 21.6% of US work requests, down slightly from 22.8% a year earlier. Thinking creatively took 15.6%, up from 11.8%. Documenting and recording information took 11.8%, up from 10.6%. Getting information fell from 11.6% to 9.9%. Providing consultation and advice rose from 5.5% to 9.0%. The sharpest fall came in interacting with computers, which dropped from 11.7% to 6.7%.

Germany, the only other country PPC Land set out in detail, moved in the same directions with different weights. Decisions and problem-solving fell from 22.0% to 19.4% of German work requests, interacting with computers from 12.7% to 8.4%, while consultation and advice rose from 7.1% to 11.7% and creative thinking from 11.3% to 13.1%. The largest gap between the two countries was in interpreting the meaning of information for others, at 8.6% in Germany against 4.0% in the US in the first quarter of 2026. OpenAI offered a cautious reading: "These differences might reflect variation in the kind of work done across countries or varying access to AI tools." It did not explain the German gap, or the fall in computer tasks.

The comparison with the wider economy is the most striking chart in the release. In the first quarter of 2025, a broader planning and decisions grouping accounted for 35.8% of US work requests to ChatGPT against 9.3% of economy-wide activity, nearly four times as much. The largest single block on the economy side is Other IWA, at 43.1%, covering activities that did not appear often enough in ChatGPT data to be reported. OpenAI wrote that the two distributions differ "largely because about 43% of the employment-weighted task mix" falls into that residual, and that "These omitted categories did not meet our threshold for consistent representation in the original U.S. ChatGPT sample."

Three caveats limit what the work data can prove. It covers individual ChatGPT plans only: "The data covers work-related messages on individual ChatGPT plans and does not include enterprise messages or Codex use." A drop in computer tasks may therefore partly reflect work moving to Codex, which the files cannot show. The data reports shares, not volumes, so a falling share can coexist with rising absolute use. And small shares carry more noise-related uncertainty.

A specialisation story on thin support

OpenAI drew a broader conclusion from the data, and PPC Land pushed back on it. "Connecting this data to IWAs revealed that work-related use became less concentrated in the most common activities," OpenAI wrote, adding that countries "tended to specialize" in their work over time. Across the 74 countries observed in every quarter, the top ten activities did fall from 56.0% of work requests in the first quarter of 2025 to 51.1% in the third quarter of 2026. But the between-country measure on which the specialisation claim rests, an average total variation distance, rose only from 22.2% to 23.0% over seven quarters, and stood at about the same level in the third quarter of 2025 as in the third quarter of 2026. The within-country chart, meanwhile, shows work spreading across more activities rather than narrowing.

On privacy, OpenAI's page is reassuring in tone and sparse in detail. "To protect users' privacy, researchers do not read individual conversations," it says. "Automated systems classify messages after identifying information is removed." Results are aggregated, small groups withheld and statistical noise added. The page does not name the classifier model, explain how work-relatedness is determined, give sample sizes or state the size of the noise. A February 2026 Signals document said differential privacy was used and that users under 18 or who had opted out of training were excluded; the October page does not confirm whether those exclusions apply to version 3.0.

The release joins a crowded shelf of company-produced labour studies, none directly comparable. Anthropic's Economic Index, launched in February 2025, found artificial intelligence used in 36% of occupations for at least a quarter of their tasks. Microsoft's study of about 200,000 Copilot conversations, in its final version of July 22, 2025, found that interpreters and translators had the highest overlap with frequent AI use, at 98% of their work activities, as PPC Land reported when a Microsoft study showed AI affects 75% of major occupations in marketing communications. Google's ATLAS v1.0 report, built on 14.65 million interactions and published on July 23, 2026, found that AI touches 68% of jobs but only 21% of their tasks, measured as a median share of tasks performed with Gemini. Each company measures its own product, with its own classifier and its own definitions.

The labour-market evidence that does not come from AI companies is less comfortable. A paper dated July 26, 2026 cited Stanford Digital Economy Lab payroll data covering more than 25 million US workers, and PPC Land reported in September that young workers lost 16% of AI-exposed jobs: employment for workers aged 22 to 25 in the most exposed occupations fell 16% between October 2022 and September 2025. OpenAI's release does not engage with employment effects at all.

Nor does ChatGPT's own position look as dominant as the volume of its data might imply. Marketers, a group ChatGPT once had almost to itself, have spread out. PPC Land reported on October 19 that ChatGPT dropped to 74.5% of marketers as Claude climbed to 72.3%, in a Basis survey of 184 agency, brand and publisher professionals; a year earlier the figures were 88.6% and 17.9%. Comscore's US panel data, reported by PPC Land in September, showed that ChatGPT lost 20 points of US AI prompt share to Gemini and Claude, falling from 70% in January 2026 to 50% in June. And an Akamai report using LayerX data found that 47% of enterprise AI chats run on personal accounts, at 47.11%. Work done on personal logins, a practice PPC Land files under shadow AI, sits in the individual-plan data OpenAI measured rather than in the enterprise data it excluded.

The silence on advertising

Then there is the subject the release omits. ChatGPT shows ads only to Free and Go users; Plus, Pro and Enterprise remain ad-free. OpenAI confirmed the Free and Go ad tests on January 16, as PPC Land reported when ChatGPT opened ads to 700 million users, set Go at $8 a month in the US, and began US ad tests on February 9. Ads went live in India at the end of August, and on October 21 PPC Land reported that advertisers in 26 more countries gained ChatGPT ads, bringing the total to 89, and OpenAI says more than 100,000 advertisers have run campaigns. In September 2026, about three-quarters of Free requests and 57.6% of Go requests were non-work. The tiers that carry ads, in short, are the tiers least likely to be used for work, the use that built ChatGPT's reputation.

That audience is large but has not yet proved to be a reliable source of clicks. PPC Land reported in February that ChatGPT sends 190x less traffic than Google despite 12% search volume, based on an Ahrefs analysis of 76,000 websites. A Smarty Marketing survey of 1,295 US residents in May found that ChatGPT was the starting point for 7.26% of shopping searches, against 56.68% for Google and 28.96% for Amazon. OpenAI's own commerce experiments have been revised along the way: Instant Checkout was shut down in March 2026, and product feeds moved into the ChatGPT Ads Manager in June.

The company's projections, by contrast, are large. Adweek's Trishla Ostwal reported on July 13 that OpenAI's ad business is on pace to miss its own forecast by 90%, an analyst says. OpenAI had projected $2.5 billion in ad revenue for 2026 and $100 billion by 2030; eMarketer expected US standalone chatbots, including ChatGPT, Microsoft's Copilot app, Google AI Mode and Amazon's shopping assistant, to earn under $1 billion in ad revenue in 2026 and about $5.41 billion by 2030. Digiday's Scanlon had reported on July 30 that OpenAI's ChatGPT reached the coupon stage of building an ad business, with a $50 credit for new advertisers who spent $50 within 14 days.

There are signs of improvement on OpenAI's own metrics. Adweek's Kendra Barnett reported on June 23 that OpenAI sees fewer users dismissing ChatGPT ads as it scales: the rate at which users close the ad units they see had dropped by 50% since launch. "We've seen that relevancy improve," Denise Dresser, OpenAI's chief revenue officer, said on stage at Cannes Lions. The company is also trying formats built for conversation rather than for search. Digiday's Scanlon reported on September 14 that OpenAI's next ChatGPT ad format is click to chat, not to site: a click opens a branded chat with the advertiser's own AI agent inside ChatGPT, tested by Wayfair in what OpenAI calls its Sponsored Agent pilot. Sarah Friar, OpenAI's chief financial officer, had called the existing ads a "basic starting point," adding: "This is all before we've really launched a format that feels truly endemic to AI."

Seen through the Signals data, the logic of that format is plain. If fewer people come to ChatGPT to produce a document and more come to work through a decision, an ad that joins the conversation fits the behaviour better than a link that ends it. Whether advertisers will pay for that, and whether users will accept it in the middle of the more personal exchanges the data now records, remains open.

A dotted line to Amazon

The format in use today is still mostly the plain text unit, and on Monday it gained a small but telling feature. Search Engine Roundtable's Barry Schwartz reported at 11:31 UTC on October 26 that ChatGPT ads now carry a Learn more arrow link, and some are served through Amazon DSP. Brodie Clark, who publishes SERP Alert notes, spotted the format and posted screenshots on X. "ChatGPT is now experimenting with a new ad format that features a learn more link with an arrow within text-based ads," he wrote, adding that "the 'learn more' text also features a dotted underline." The link does not appear on every ad, and Schwartz published a screenshot of his own.

The more interesting detail came from Juozas Kaziukėnas, who reported on LinkedIn that some of these ads are "served by Amazon DSP," which he could tell "because the click target URL goes through Amazon's click-tracking server." The address of his LinkedIn post refers to Delta, although Schwartz's article does not name the advertiser. "There is barely any detail on how this integration works - and I've not seen any other advertisers use it yet," he wrote. His reading, which Schwartz presented as unconfirmed, is that advertisers buy through the DSP, which handles targeting, while creative control stays the same as for regular ChatGPT ads; Amazon, in that view, is not running its own formats inside ChatGPT. The arrangement itself was announced in September, when PPC Land reported that Amazon Ads gained a ChatGPT ad surface as Delta Vacations tested first. What changed on Monday is that the plumbing became visible in a live ad.

A dotted underline is a small thing. It also arrives days before OpenAI begins changing ads in larger ways: PPC Land reported on October 21 that OpenAI will rewrite ChatGPT ads by default as advertisers get days to opt out, with the switch due on October 30 in OpenAI's email to advertisers and October 29 on its help page.

A television, a decompiler and a suspended account

If the Signals files describe what most people do with ChatGPT, the week's other story shows what happens at the edge. AdExchanger's daily roundup on Tuesday, October 27, examined a viral claim that a Samsung smart TV listens to its owners. An X user named log4j, who says they are a cybersecurity researcher, claims to have reverse-engineered the software of a $3,000 Samsung TV using ChatGPT and other decompiling tools. The poster labelled the result "PRIVACY HORROR," describing an always-on microphone that listens for barking dogs or crying babies and reports back to Samsung to serve ads. AdExchanger noted that the claims come from the poster and have not been independently verified.

OpenAI then entered the story. log4j shared an email from OpenAI saying the account had been suspended for "Cyber Abuse." AdExchanger did not state what activity triggered the suspension, and did not confirm any link between the suspension and the television analysis. The suspension, it observed, fuelled suspicion that large technology companies were coordinating against the poster.

The more prosaic explanation came from TechRadar, which AdExchanger cited: the listening features are publicly promoted in Samsung's marketing as family and pet care tools, notifying other devices when a loved one seems to be in distress. Samsung reportedly discontinued the baby-crying feature last month because it could not tell whether a baby was crying on the television or in the room. TechRadar's remaining point, which AdExchanger carried, was that Samsung could have been clearer about sending that data to its servers and about what it is used for.

The episode is a useful counterweight to the dataset. Technical help may have fallen to 4.6% of consumer ChatGPT requests, but the requests that remain include some of the most sensitive, and enforcement against them happens account by account, in emails like the one log4j posted, rather than anywhere aggregate files can show. Security was also on OpenAI's agenda elsewhere in the week: Search Engine Roundtable's daily recap for October 26 pointed to an OpenAI post on securing its frontier model training environment, shared on X by Greg Brockman.

What the files show, and what they leave out

Put together, the week's ChatGPT stories describe a product in transition. The office assistant of 2024 has become, on OpenAI's own measurement, something closer to a general adviser: consulted on decisions, used to make images, told things people would not tell their friends, and increasingly used by older people and, on OpenAI's name-based inference, by women. At work, its users ask it to decide and to think creatively more than to type. The coding that once filled one request in seven has moved to Codex and to rivals' developer tools.

That is a more valuable audience for advertising, and a more delicate one. OpenAI sells ads only on the plans whose users are least focused on work, which is consistent with the data, and it is building formats such as image-side units and agent chats that suit the behaviour the data records. Yet the same data shows the fastest growth in exactly the kinds of exchange where ads are hardest to place well: personal statements, private worries, requests for advice. The release that documents all of this "says nothing about advertising." Given what it shows, that silence may be the most informative line in it.

By the numbers

51.8%
Share of September 2026 ChatGPT requests sent by users with typically feminine first names, up from 42.2%.
Writing drops from 34% to 21% of ChatGPT requests since July 2024

1 million
Consumer ChatGPT messages classified each month for Signals, up from 300,000 before version 3.0.
Writing drops from 34% to 21% of ChatGPT requests since July 2024

165
Work activities OpenAI mapped across nearly 100 countries, just under half of O*NET's 332.
Americans use ChatGPT at work most to make decisions, 1 in 5 requests

$8
Monthly US price of ChatGPT Go, one of the two plans that carry ads.
Writing drops from 34% to 21% of ChatGPT requests since July 2024

March 2025
Month practical guidance became ChatGPT's largest topic, a lead it has held every month since.
Writing drops from 34% to 21% of ChatGPT requests since July 2024