What is the study, and who ran it?
The paper is “How People Use ChatGPT”, National Bureau of Economic Research Working Paper 34255, dated September 2025. The authors are Aaron Chatterji, Thomas Cunningham, David J. Deming, Zoe Hitzig, Christopher Ong, Carl Yan Shan and Kevin Wadman. It is the first economics paper written on internal ChatGPT message data, and several of its authors were working at OpenAI, which supplied that data. The study was approved by the Harvard institutional review board under IRB25-0983.
The sample is approximately 1.1 million conversations, drawn from 15 May 2024 to 26 June 2025 and reweighted to reflect total message volumes on each day. The classification was done by automated classifiers running inside OpenAI's infrastructure. The paper states plainly that “no user messages were observed by humans during any part of the work on this paper”.
Sampling excludes several groups, and the exclusions matter. Users who opted out of sharing their messages for model training are excluded, as are users who self-report as under 18, logged-out users, deleted conversations and accounts that were deactivated or banned. The privacy-protective design and the representativeness of the sample are in direct tension: the people most careful about what they share are the people least likely to be counted.
The scale being described is worth stating because it sets what a percentage point is worth. By July 2025 ChatGPT had been used weekly by more than 700 million users, collectively sending more than 2.5 billion messages a day — about 29,000 messages a second. The paper puts adoption at around 10% of the world's adult population.
This is behavioural data rather than survey data, which is its main advantage over everything else published on the question. Nobody is being asked to recall what they typed. The corresponding disadvantage is that no outside researcher can replicate it, because no outside researcher has the messages.
What are people actually asking about?
The classifier assigns each message to one of 24 categories, which the authors aggregate into seven topics. Three of the seven dominate. The paper describes their combined share as “about 77%” in the body, “nearly 78%” in the conclusion and “nearly 80%” in the abstract — a looseness worth noticing in a paper this widely quoted.
The movement inside that block is the more interesting part. Practical Guidance — how-to advice, tutoring, creative ideation, health and fitness — stayed flat at roughly 29% of usage across the year. Writing fell from 36% of all usage in July 2024 to 24% a year later. Seeking Information rose from 14% to 24% over exactly the same period, very nearly swapping places with Writing.
% of all messages, approximately one year apart
Source: Chatterji, Cunningham, Deming, Hitzig, Ong, Shan and Wadman, 2025, How People Use ChatGPT, NBER Working Paper 34255, Figure 7.
Values for July 2024 and roughly a year later, as described in the paper's text. Bars are scaled to the largest value shown.
Below the seven topics, the paper names granular shares for some sub-categories and not others. Tutoring or teaching is 10.2% of all user messages and 36% of Practical Guidance. General how-to advice is another 8.5%. Computer programming is 4.2%, mathematical calculation 3% and data analysis 0.4%. Relationships and personal reflection is 1.9%, and games and role play 0.4%.
Those last two are worth dwelling on, because the popular account of chatbot usage is that it is mostly companionship and therapy. On this data it is not: 2.4% of all messages fall into those two categories combined. The authors explicitly contrast this with an earlier estimate that therapy and companionship was the single most prevalent use case, and attribute the discrepancy to that study having been built from Reddit, Quora and online articles rather than from messages.
The second classifier asks a different question: what kind of output does the user want? Asking is seeking information or advice to be better informed or make a better decision. Doing is asking for a task to be performed. Expressing is neither. About 49% of messages are Asking, 40% Doing and 11% Expressing. The paper's own concluding section misprints Expressing as 1%, which is a typographical error against the 11% given in the body and in the figure.
Asking is growing faster than Doing, and Asking messages are rated higher quality both by an automated satisfaction classifier and by direct user feedback. For anyone whose interest is whether these systems function as an information intermediary, that is the single most relevant trend line in the paper.
% of all ChatGPT messages, by intent classifier
Source: Chatterji et al., 2025, NBER Working Paper 34255, Section 5.3 and Figure 12.
Shares do not sum to 100% because of rounding. The paper's conclusion misstates Expressing as 1%.
How much of it is commercial — products, brands, recommendations?
This is the question a founder actually wants answered, and it is the one place the paper declines to give a number.
Seeking Information is built from exactly three classifier categories: Specific Info, Purchasable Products and Cooking and Recipes. The appendix gives the classifier's own definition of the middle one — “inquiries about products or services available for purchase” — and the example prompts it was validated against, which include “What's the best streaming service?”, “Recommend a good laptop under $1000” and “How much are Nikes?”. That is precisely the query shape that AI-visibility services are sold against.
The paper's text publishes shares for tutoring, how-to advice, computer programming, mathematical calculation, data analysis, relationships and games. It publishes none for Purchasable Products. That number exists only as a bar in Figure 9, and we are not going to read a value off a chart and present it as a finding. If you have seen a precise percentage for product queries attributed to this study, that is where it came from.
What can be said from the printed figures is less dramatic and more useful. Seeking Information as a whole is 24% of messages and has been the fastest-growing topic in the sample. Its largest constituent is Specific Info, defined in the appendix as “providing specific information typically found on websites, including information about well-known individuals, current events, historical events, and other facts and knowledge”. A company being looked up — who runs it, whether it is real, what it was in the news for — lands there, not in the shopping category.
That reframes the exposure. A business is not competing for a slice of a small shopping bucket. It is potentially present across Specific Info, across the how-to and advice half of Practical Guidance, and inside the 49% of messages classified as Asking. None of those are measured by brand, so none of them can be converted into a share for any individual company — but the surface is considerably wider than the product category alone suggests.
Does Claude's data show the same thing?
Anthropic publishes a comparable series, the Anthropic Economic Index, and its numbers are close enough in shape to be useful and different enough to be instructive.
The March 2026 report, published on 24 March 2026 and covering 5 to 12 February 2026, analysed about one million conversations across Claude.ai and Anthropic's first-party API. Personal use rose from 35% to 42% of Claude.ai conversations, with the report naming sports, product comparisons and home maintenance as examples of what that growth consists of. Coursework fell from 19% to 12%. The ten most common tasks accounted for 19% of Claude.ai traffic, down from 24% in November 2025 — usage getting broader rather than narrower.
The June 2026 report, published on 26 June 2026 and covering 10 April to 10 June 2026, classified conversations by the artefact they produced. Explanations were the most common output at 17% of conversations, followed by documents and reports at 15% and guidance at 11%. Personal use ran at roughly 35% of conversations on weekdays and roughly 50% at weekends.
The divergence is in coding. The ChatGPT paper puts computer programming at 4.2% of messages; it notes that an Anthropic study reported 33% of work-related Claude conversations mapped to programming, and attributes the gap to a genuinely different user base rather than to a measurement error. Anthropic says the same thing about itself, stating that its users skew heavily towards computer, mathematical and management occupations, and that its linked survey is not representative of the general population.
Two datasets, two products, two sets of caveats — and one consistent finding: explanation, guidance and information-seeking are what most of the volume is. Neither company publishes a clean figure for how often a specific brand or company is the subject of a query, and neither is likely to.
If only part of it is commercial, is assistant visibility worth anything?
Three things follow, and the second one is the one most likely to be left out of a pitch.
First, the addressable surface is real. Roughly half of all messages are people trying to become better informed before deciding something, and the topic that captures most of that has been the fastest-growing category in the dataset. If a company is described wrongly inside that surface, the error is being delivered at conversational scale. That is a reputational exposure whether or not anyone is selling you a service to fix it.
Second, none of that traffic comes back to you. The Pew Research Center tracked the browsing of 900 US adults across 68,879 Google searches in March 2025 and found that users clicked a traditional search result on 8% of visits when an AI summary was present, against 15% when it was not — and clicked a link inside the summary on 1% of visits. The Reuters Institute's 2026 Digital News Report found that across 27 markets just 4% of respondents say they always or often click through from an AI chatbot, against 19% from search and 17% from social media. Being present in an AI answer is a reputation channel, not an acquisition channel, and anyone modelling it as referral traffic is modelling the wrong thing. We have written separately about what Search Console actually reports for AI answers, which is impressions without clicks.
Third, nothing in either usage dataset tells you how a company gets into those answers. Message-volume research describes demand. It says nothing about retrieval, and the evidence on retrieval is thinner and more model-dependent than the category admits — see which sources AI assistants actually cite and how assistants decide which companies to recommend. Those are separate questions with separate and weaker evidence bases.
Media Morpheus sells media and podcast placement, partly on the argument that coverage improves how a company is described by AI systems. The usage data above supports the premise that the surface is large and information-shaped. It does not, on its own, support any claim about what moves a company into it.
Methodology and limitations
Primary figures are from “How People Use ChatGPT” by Aaron Chatterji, Thomas Cunningham, David J. Deming, Zoe Hitzig, Christopher Ong, Carl Yan Shan and Kevin Wadman, NBER Working Paper 34255, September 2025, based on approximately 1.1 million sampled conversations from 15 May 2024 to 26 June 2025, reweighted to daily message volumes. Five limits apply. First, NBER working papers state on their cover that they have not been peer-reviewed, and this one has not. Second, the data is OpenAI's own and the classifiers run inside OpenAI's infrastructure, so no independent party can replicate the sample; several authors were affiliated with OpenAI at the time of writing. Third, the sample excludes users who opted out of sharing messages for training, self-reported under-18s, logged-out users, deleted conversations and deactivated or banned accounts, which biases against the most privacy-conscious users. Fourth, the classifications are produced by a language model rather than by human coders, and the paper's own validation appendix reports confusion between Seeking Information and Practical Guidance, which are conceptually adjacent. Fifth, it measures one product in one period; the Anthropic figures cited alongside it come from a user base Anthropic itself describes as unrepresentative, skewed towards computer, mathematical and management occupations. No percentage is quoted here for the Purchasable Products category because the paper does not publish one in its text.
Conclusion
The honest summary is that people mostly ask AI assistants to explain things, advise them and look things up, and that this has been growing while the writing-assistant use case has been shrinking. Practical Guidance at 29%, Seeking Information at 24% and an intent classifier putting 49% of messages in the Asking bucket describe an information surface rather than a shopping surface. The commercial slice inside that is real, unmeasured in public, and smaller than the marketing implies — and the one number everyone quotes for it does not appear in the paper. For a company, the useful conclusion is that AI assistants are now a place where you are described to people who will not click through to check, which makes accuracy in that description a reputational question rather than a traffic one.
Frequently Asked Questions
What is the study, and who ran it?
“How People Use ChatGPT”, NBER Working Paper 34255, September 2025, by Chatterji, Cunningham, Deming, Hitzig, Ong, Shan and Wadman. It classifies about 1.1 million sampled conversations from 15 May 2024 to 26 June 2025. It is a working paper and has not been peer-reviewed.
What are people actually asking about?
Practical Guidance, Seeking Information and Writing together account for about 77% of messages. Practical Guidance held steady near 29%; Seeking Information grew from 14% to 24% over the year to June 2025; Writing fell from 36% to 24%.
How much of it is commercial — products, brands, recommendations?
The paper never prints a number. Purchasable Products is one of three sub-categories inside Seeking Information, and its share appears only inside a bar chart. Any precise product-query percentage in circulation has been read off that chart rather than quoted from the text.
Does Claude's data show the same thing?
Broadly, with a heavy coding skew. Anthropic reported personal use rising from 35% to 42% of Claude.ai conversations by February 2026, and explanations as the single most common output at 17% of conversations in its June 2026 report. Anthropic states its user base is not representative.
If only part of it is commercial, is assistant visibility worth anything?
It is worth something as a reputation surface and almost nothing as a traffic channel. Pew found users clicked a link inside an AI summary on 1% of visits, and the Reuters Institute found 4% say they often click through from a chatbot against 19% from search.
Is this study peer-reviewed?
No. NBER working papers are circulated for discussion and state on their cover that they have not been peer-reviewed. Several authors were affiliated with OpenAI, which supplied the data, and no outside party can replicate the sample.
Sources
- academic Chatterji, A., Cunningham, T., Deming, D. J., Hitzig, Z., Ong, C., Shan, C. Y., and Wadman, K., 2025. How People Use ChatGPT. NBER Working Paper 34255, September 2025. Working paper, not peer-reviewed.
- vendor research Anthropic, 2026. Anthropic Economic Index report, published 24 March 2026, covering 5–12 February 2026.
- vendor research Anthropic, 2026. Anthropic Economic Index report: Cadences, published 26 June 2026, covering 10 April to 10 June 2026.
- survey research Pew Research Center, 22 July 2025. Google users are less likely to click on links when an AI summary appears in the results. Browsing data from 900 US adults, March 2025.
- survey research Reuters Institute for the Study of Journalism, June 2026. Emerging uses of AI chatbots for news and what it means for journalism, Digital News Report 2026.