Most ChatGPT use is not people typing product queries the way they would into Google. The best available large-scale study of how people use ChatGPT found that about 24% of usage is direct information seeking, 29% is practical guidance like tutoring and how-to advice, and 24% is writing. If you are trying to size AI chat as a discovery channel, that composition matters more than the overall adoption numbers, because it tells you how much of the traffic through ChatGPT even resembles search.
That study comes from researchers at OpenAI and Harvard, published as an NBER working paper in September 2025. It analyzed consumer ChatGPT usage from the product's launch through July 2025 using a privacy-preserving classifier, not a survey where people try to remember what they did. That makes it the strongest evidence we have on how people use ChatGPT, even though it is one study, on one product, over one window of time.
What people actually use ChatGPT for
Three categories account for close to 80% of everything people ask ChatGPT to do.
Practical guidance is the largest single category, at roughly 29% of overall usage. This is tutoring and teaching, general how-to advice, creative ideation, and questions about health, fitness, beauty, or self-care. Tutoring and teaching alone make up about 36% of this category, and general how-to advice makes up close to 30% of it.
Seeking information comes next, at 24% of usage as of July 2025. This includes looking up facts about people, current events, products, and recipes. It's up sharply from 14% a year earlier.
Writing is also 24% of usage. This covers more than drafting an article from a blank page. It includes editing or critiquing text someone already wrote, personal messages, translation, summaries, and fiction. A lot of that is people finishing something they already knew they wanted to say, not people discovering what to say.
Everything else, technical help, image and media generation, and personal reflection or role-play, adds up to a smaller share. Technical help (math, data analysis, and programming) sits around 5% of overall usage. Multimedia grew from 2% to more than 7% over the study period. Personal reflection and role-play together come to about 2.4%.
If you've been picturing ChatGPT as mostly a search box or mostly a coding tool, this breakdown is worth sitting with. The largest single use is people asking for guidance on a problem they're trying to work through, not people looking something up the way they'd Google it.
There's a second way the same researchers sliced the data, by what the person was trying to accomplish rather than what the conversation was about. About 49% of messages are "asking", seeking information or advice to inform a decision. About 40% are "doing", asking ChatGPT to produce something or complete a task, like drafting, planning, or coding. The remaining 11% are "expressing", reflection, conversation, or play with no clear information or task goal. These two ways of counting usage don't stack on top of each other since they classify the same messages by different criteria, but together they reinforce the same picture: ChatGPT is used far more often as an adviser than as a task engine, and used as a task engine far more often than as a companion.
How much of that usage is actually search-like
This is the question that matters for sizing AI chat usage patterns as a discovery channel, and it deserves a careful answer instead of a rounded-up one.
The 24% seeking-information figure is the closest measured proxy for search-like behavior in ChatGPT. The researchers describe it as a close substitute for web search, though they're careful to note the overlap isn't perfect. Someone asking ChatGPT what a company does, what a product costs, or what a current event means is doing something close to what they'd otherwise do in a search bar.
The 24% figure isn't ChatGPT's share of the search market, and it isn't the share of conversations that lead to a brand being discovered or a rate of website visits or purchases. The study measured what people asked ChatGPT to do. It didn't measure whether an answer mentioned a brand, cited a page, or led anywhere afterward.
It's tempting to add the 29% practical guidance share to the 24% seeking-information share and call more than half of ChatGPT usage search-like, but that's a step further than the evidence goes. Practical guidance is adjacent to search in the sense that someone working through a problem or comparing options is doing something search-adjacent, but the study doesn't say what portion of that 29% involves an actual brand or category decision. Treating it as additional search volume turns two different categories into a number nobody measured.
The honest version is two separate lines: 24% is the defensible search-like share, and 29% is a real but unmeasured adjacent opportunity. Keep them apart when you're reporting a number to your own team.
Why practical guidance matters as much as information retrieval
The practical guidance category deserves more attention than it usually gets, because it's the largest bucket and it doesn't look like a typical search query.
Someone asking ChatGPT to explain a concept, walk through how to approach a decision, or think through the tradeoffs of an option isn't typing a branded search term. But they're often at the exact stage where a brand could become relevant, before they know what to search for by name. A person trying to understand how to reduce content production costs, for instance, is in practical guidance territory long before they're comparing named vendors.
This is a different kind of AI-search opportunity than the 24% seeking-information share, and it calls for different content. A page written to answer "what is X" competes for a different slice of ChatGPT use cases than a page written to walk someone through "how should I think about X." Both matter, but conflating them into one AI-search strategy misses that the biggest category of usage is advice-shaped, not fact-shaped.
None of this means every how-to article you publish will get pulled into an AI answer. It means a strategy that only targets explicit product or fact queries is aiming at less than a quarter of how people actually use ChatGPT.
What people use ChatGPT for at work
Work and non-work usage look different enough that they're worth separating out. About 30% of consumer ChatGPT usage is work-related, and about 70% is not, and the non-work share has grown faster over the study period.
Inside work-related usage, the mix shifts. Writing jumps to about 40% of work messages, the single largest work category, which lines up with how often people use ChatGPT to draft, edit, or polish something for their job. Practical guidance holds at about 24% of work messages. Technical help, which includes programming, sits just over 10% of work messages, down from 18% a year earlier as more coding work has likely moved to dedicated tools and editors outside the consumer ChatGPT product.
The takeaway for a marketing team is to keep these two views separate. All-consumer usage tells you the overall shape of how people use ChatGPT, including everyday questions about products, health, recipes, and learning that have nothing to do with anyone's job. Work-related usage tells you something narrower, about how professionals draft and research. If you're planning content aimed only at professional or B2B prompts, you're deliberately ignoring the 70% of usage that isn't work at all.
What this means for sizing the channel
Here's how to put the composition to work instead of just reading it.
Start with the 24% seeking-information share as your baseline for search-like demand, and don't multiply it by total ChatGPT users to produce an audience number. The study gives you a composition, not a reach estimate, and there's no published conversion rate from a ChatGPT conversation to a citation, a click, or a sale.
Separate your prompts into cohorts instead of tracking one blended AI-visibility number. Informational questions, product or category research, how-to and practical-guidance questions, comparison questions, and work-related research each behave differently, and lumping them into a single score hides which ones are actually worth writing for.
Write for decision support, not just rankings. Since nearly half of all usage (49%) falls into what the researchers call "asking", seeking information or advice to inform a decision, content that explains a problem clearly, defines the relevant terms, compares options honestly, and names the tradeoffs is doing the kind of work this usage pattern rewards. That's an editorial implication from the data, not a proven ranking formula, so treat it as a reasonable bet rather than a guarantee.
Keep production separate from discovery. Writing is 24% of all usage and 40% of work usage, but that's people using ChatGPT to produce or polish content, not people discovering a brand through it. A tool like DeepSmith that tracks how AI engines answer questions in your category and produces the content to close the gaps is built around that discovery side specifically, not around the fact that people also use ChatGPT as a drafting tool.
Don't assume the same composition holds everywhere. This data describes consumer ChatGPT. It doesn't tell you how people use Gemini, Claude, Perplexity, or Google's AI features, and it may not hold in 2026 the way it held in mid-2025. Use it as the best current benchmark, not a permanent constant.
Where the numbers stop
It's worth being specific about what this study does and doesn't establish, because AI-search discussion tends to round uncertain numbers into confident ones.
The study is strongly supported for describing what people ask ChatGPT to do. It's genuinely useful for that: the classification comes from actual usage, not self-reported recall, and the sample is large. But it says nothing about how often a conversation mentions a brand, how often ChatGPT cites a website, how many people click through to a cited source, or how often an information-seeking conversation ends in a purchase. Those numbers don't exist yet in this research, and anyone telling you a specific conversion or citation rate from this study is going further than the data supports.
A few caveats worth keeping in mind if you're citing this yourself. It's a working paper, the strongest evidence available right now but not a settled industry census. The figures cover consumer ChatGPT specifically, not API usage, code editors, or agent-based tools that might show a very different mix. And the percentages come from two different taxonomies (what the conversation was about, and what the person was trying to accomplish) that shouldn't be added together across categories.
It's also worth remembering these AI chat usage patterns describe one product at one point in time. Gemini, Claude, and Perplexity may show a different mix of ChatGPT use cases entirely, since each product attracts a slightly different audience and is built around different strengths. Treat this study as the clearest picture available of how people use ChatGPT specifically, and be cautious about assuming the same composition holds across every AI chat product your customers might reach for.
The realistic read is this: ChatGPT is not primarily a search engine, and it's not primarily a coding tool either. It's a mix weighted toward advice and writing, with a real but bounded search-like slice sitting at about a quarter of usage. Size your AI-search plans against that slice, treat the practical-guidance category as a genuine but unmeasured adjacent opportunity, and don't let round numbers about "AI search taking over" stand in for what the actual usage data shows.
If you want to see where your own brand actually sits inside that mix, rather than guessing, DeepSmith tracks how ChatGPT and other AI engines answer the questions your buyers ask and produces the content to close the gaps it finds. You can start a free trial and check your own AI search presence directly instead of extrapolating from someone else's usage study.
