No public source gives a complete, independently checkable count of how many times people ask every AI engine one particular question. The evidence on this is mixed, and the fair grade is that the claim "no one can know" is only partly true. An engine could count a precisely defined question inside its own logs for a set period, but nothing public lets you verify that kind of number across engines. The AI prompt volume figures you can actually find measure something broader, such as all messages on one platform or a modeled estimate for a topic.
If you are a marketing lead trying to decide which questions to care about, that gap matters, because it is easy to read a big number and take it for the one you wanted. This article walks through what the primary sources say, what vendors mean when they sell prompt volumes, and how to prioritize when there is no exact number to lean on.
What "prompt volume" can mean
The phrase "AI prompt volume" has no single meaning, and most of the confusion comes from that. When someone gives you a number, it usually falls into one of five units, and they are not interchangeable.
- Exact-text count. How many times the identical string was typed. A reworded version does not match.
- Topic or semantic count. How many prompts a system judged to mean roughly the same thing. This depends on the grouping rules, so it is an estimate for a category and not a count of one string.
- Message count. All user messages on a platform, which includes follow-ups, task requests and statements, not only questions.
- Conversation count. Sessions, where one conversation can hold several messages and replies.
- User count. How many distinct people used a product, which says nothing about how many of them asked one thing.
Here is a small example. "What CRM should a small nonprofit use?" and "Which customer-management tool is best for a small charity?" might be grouped as one topic by one system. They are still different strings. A follow-up like "What about one under $50 a month?" only makes sense with the earlier conversation behind it, so counting it as a standalone question would misread what was asked. This example is only an illustration and is not measured data.
"Question" is also a narrow word for the big totals. OpenAI's usage research sorts messages into asking, doing and expressing, and many messages are not questions at all. So before you trust any figure, it helps to ask what the unit is, which platform it covers, which people and countries are in it, what period it spans, and how it was produced.
What the biggest platform numbers do and do not tell you
If you are looking for how many people ask ChatGPT something, the numbers most people quote come from a September 2025 working paper called How People Use ChatGPT, published through the National Bureau of Economic Research (NBER). Its authors are Aaron Chatterji, Thomas Cunningham, David J. Deming, Zoe Hitzig, Christopher Ong, Carl Yan Shan and Kevin Wadman. The paper reports that by July 2025 ChatGPT had more than 700 million weekly users sending more than 2.5 billion messages a day.
Those are platform-wide usage totals. They tell you ChatGPT is very large, and they do not tell you how many people asked any particular question. The paper's June 2025 table gives a more exact daily figure of 2,627 million messages a day across consumer plans, split into 1,911 million non-work messages (73%) and 716 million work-related messages (27%). A year earlier, in June 2024, the daily total was 451 million, with 238 million non-work (53%) and 213 million work-related (47%). Again, these are aggregate totals with a broad work or non-work label on them, and none of them is a count for one query.
The paper also looks at what people are doing. It finds that Practical Guidance, Seeking Information and Writing together made up nearly 80% of conversations in its topic analysis. On intent, it reports that the split moved from roughly even between asking and doing in mid-2024, with just under 8% expressing, to 51.6% asking, 34.6% doing and 13.8% expressing by late June 2025. Those percentages describe classified messages in the study. They do not describe the share of people asking a specific question.
The method is worth knowing about, because it shows how the numbers were built. The researchers used a privacy-preserving automated classification pipeline, and no member of the research team read message content. The conversation sample was roughly 1.1 million conversations, chosen uniformly, with one message taken from each. A separate sample covered 1.58 million messages from a subset of users. Those two samples should not be merged or described as a census of exact questions. The paper covers consumer Free, Plus and Pro plans, and it leaves out Business, Enterprise, Education and some other account types.
OpenAI also keeps a public Signals page, which was updated in October 2026. It describes a sample of messages from individual ChatGPT Free, Go, Plus and Pro accounts from July 2024 through September 2026, and it shows high-level topics, work and non-work views and message-intent categories. The page we reviewed does not offer a table of how often each exact question appears. It leaves out enterprise and Codex usage, and it leaves out countries where there is not enough data or where OpenAI does not operate. That makes it good evidence about how people behave at platform scale, and it is not a place to look up a per-question number.
What an outside panel can show, and what it cannot
A second source comes from Epoch AI, which published its ChatGPT Usage Explorer on October 1, 2026. It draws on metadata from 5,000 US YouGov panelists who agreed to share their ChatGPT histories and who had used ChatGPT at least once in 2026. The dataset holds about 8.3 million messages across about 660,000 conversations, and for some people the histories reach back to the early days of the product.
This is detailed usage data, and it still cannot tell you how often one question was asked. YouGov removed all message text before sharing it, and Epoch says it did not see transcripts or classify topics. What the explorer does show is message times, the model that replied and the tools used.
Some of what Epoch reports is useful for understanding why totals are hard to read. The median number of messages per active panelist, counting prompts and replies, rose from 14 in January 2023 to 36 in December 2025. The mean in December 2025 was 158, because a small group of heavy users contributed a large share, and the most active 10% of panelists sent 63% of prompts.
That gap between the median and the mean is a good thing to keep in mind with any volume number. A total can go up because a few people are very busy, and not only because more people asked. Epoch is also clear about the limits of its panel. Joining was opt-in, the data is retrospective, unbalanced and unweighted, people had to have used ChatGPT in 2026 to be included, and the panelists were older on average than ChatGPT's overall user base. So these are findings about that panel and they should not be read as counts for the whole population.
What vendors mean by prompt volumes
When a tracking vendor sells prompt volumes, they are usually selling an estimate built from samples and models. That can still be useful, as long as you read it as what it is. Two vendors describe their methods in public documentation, and they make good examples.
Profound
Profound's Prompt Volumes documentation, which we reviewed in November 2026, says its data comes from licensed double-opt-in consumer panel providers. It describes millions of active AI assistant users and tens of millions of real prompts each month. Profound says it cleans the data and applies probabilistic modeling to correct for demographic and geographic bias, and it says synthetic or simulated prompts are not used.
So the underlying data is real user interaction from panels, and it is also cleaned and modeled. That is different from saying every user of an engine is counted. The public documentation does not spell out the exact denominator behind a displayed number, a full sampling frame, a confidence interval, or whether a displayed figure is a raw count, a normalized estimate or an extrapolated total for the population. Because of that, a Profound topic volume should not be described as an audited count of an exact question.
The documentation says you enter a short, broad keyword to see how often a topic appears in AI conversations, with phrase-match and exact-match options, intent breakdowns and regional views. Those controls help you define what you are searching for, and they still do not turn a topic measure into a count of one identical question. Profound's pages are also not fully consistent with each other. One names ChatGPT, Gemini and Perplexity as sources, while a Help Center FAQ names ChatGPT, Gemini, Claude and Perplexity as available in the US. The start dates for non-US history differ between pages as well, so confirm those details on the current product page before relying on them. The FAQ says data updates weekly with a lag of under two weeks, and that coverage is strongest in the US, UK and select European markets.
Semrush
Semrush's AI Visibility data documentation, also reviewed in November 2026, says its AI prompt database holds more than 317 million prompts and responses across ChatGPT, Gemini, Google AI Overviews and Google AI Mode. It says the database is updated daily and covers 117 regional databases. That 317 million figure describes the size of the database. It is not the number of people who asked a question, and it is not the volume of any one question.
Semrush says responses are captured from real requests and not through LLM APIs, and that it sources real prompts from AI-search clickstream data and Google's keyword dataset for AI Overviews. Duplicates are removed and phrasing is simplified while intent is kept. For its Prompt Research report, it groups related prompts that move in the same semantic direction and estimates topic-level volume by combining third-party data on real AI interactions with machine-learning models. That is a modeled measure of topic demand. The documentation page we reviewed does not give a publication date, and the method details are Semrush's own description of its work, so they are not independent validation.
Why these numbers are hard to compare
If you set these sources next to each other, a few reasons keep coming up for why one figure cannot stand in for another.
- The unit changes the answer. A message total, a conversation total, a user total, an exact-string count and a topic estimate answer different questions.
- Grouping is a choice. One vendor may merge several phrasings into one topic, and another may keep them apart.
- Context matters. A short follow-up may mean nothing without the conversation before it.
- Heavy users move totals. Epoch's panel shows how far the mean can sit from the median.
- Populations differ. Plan types, countries, platforms and history windows vary from source to source.
- Sampling and modeling are not full observation. A panel adjusted by a model can be a fair estimate, and it is still a different thing from an engine counting a defined question in all its logs.
- A precise figure may not answer your question. "More than 2.5 billion messages a day" is exact enough to quote, and it still tells you nothing about how many people asked one question.
It would go too far to say exact counting is technically impossible. The evidence supports a narrower conclusion. The public sources reviewed do not publish a complete global exact-question count, and the published platform totals and commercial estimates cannot stand in for one.
The table below puts the main figures next to what each one does and does not show.
| Figure | Source and date | What it measures | What it does not establish |
|---|---|---|---|
| More than 700 million weekly users and more than 2.5 billion messages per day by July 2025 | NBER, How People Use ChatGPT, September 2025 | Aggregate ChatGPT consumer usage | Users or repetitions of any specific question |
| 2,627 million daily messages in June 2025, with 1,911 million non-work and 716 million work-related | NBER, September 2025 | Daily platform-wide message total and work classification | Exact-query demand |
| 51.6% asking, 34.6% doing, 13.8% expressing by late June 2025 | NBER, September 2025 | Share of classified message intents | Share asking a particular question |
| 5,000 US YouGov panelists, about 8.3 million messages across about 660,000 conversations | Epoch AI, October 1, 2026 | Opt-in panel metadata | A representative US population or topic counts |
| Most active 10% of panelists sent 63% of prompts | Epoch AI, October 1, 2026 | Activity concentration in an unweighted panel | The spread across all ChatGPT users |
| More than 317 million prompts and responses, 117 regional databases | Semrush documentation, no publication date, reviewed November 2026 | Stated database scale and regional coverage | Users, monthly prompts, or volume for one exact question |
| Tens of millions of real prompts per month from opt-in panels, adjusted by probabilistic modeling | Profound Help Center, no publication date, reviewed November 2026 | Vendor-described panel input and processing | A population denominator or a count for each question |
What is still unknown
A few things are honestly open, and it is better to say so than to guess.
- No exact count for any named question across all AI platforms turned up in the public sources reviewed.
- The public materials do not show whether any engine privately keeps a queryable exact-question count, or whether it would ever publish one.
- Profound's public documentation does not fully describe its sampling frame, the denominator behind a displayed volume, an uncertainty interval, or whether a given number is normalized or extrapolated.
Given that, it is safest not to treat weekly users or daily messages as the volume of a specific question, and not to call a topic estimate an exact prompt volume.
How to prioritize without an exact number
A calmer option is to let the absence of a dependable count push you toward relevance and evidence. What follows is a decision framework and not a validated formula, and nothing here says one signal predicts traffic.
Start with the outcome. Write down what you want the content to help you achieve before you choose any prompts or topics. That one sentence settles a lot of later arguments. If you need a starting point for the prompt set itself, the GEO templates include a copyable brief, page audit and prompt set.
Define who you are talking to and whether you fit. Name the audience, the need, and whether you have a credible product or service for that need. A question that matches your buyer and your offer can be worth more than a loosely related topic with a bigger volume estimate attached.
Separate the intent. Label each question as informational, commercial, transactional or navigational, and do not assume they carry the same value to you. If you want a way to group and label what you track, our guide on how to organize tracked prompts by topic, funnel stage and persona walks through one.
Read the answers, not just a score. Look at what the model recommends, what criteria it states, which brands it names and which sources it cites. A mention and a citation are different things to observe, and reading prompt-level win and loss data against competitors shows how to compare them per question. The guide to what prompt tracking is explains how a list of questions turns into mention and citation rates and where those rates can mislead you.
Compare like with like. When you test responses, keep the exact wording, platform, date, market and conversation context the same each time. What you get is an observation under those conditions. It is not an estimate of how many people asked the prompt. Our guide to tracking the same prompts across ChatGPT, Gemini and Perplexity shows how to keep those conditions steady.
Keep the small denominators visible. Moz's AI visibility workflow, which has no date on the page we reviewed, recommends showing the prompt count next to any percentage, because one response can move a percentage a lot when the set is small. A rate with no count beside it can look more certain than it is.
Keep the action list short. The same Moz workflow suggests picking a small set of actions based on business impact, how often something shows up, how confident you are in the evidence, and effort. Treat those as criteria for choosing and not as a statistical formula. Tracking data also tends to sit unused when nobody owns it, which is covered in why prompt tracking data goes unused.
A simple rule you can use is to start where buyer relevance, a believable fit with your product, solid evidence and real business impact all line up. If you do use a vendor's volume estimate, treat it as one directional input, and only when the vendor states the platform, geography, period and unit, and whether the figure is sampled, modeled or topic-level.
Once you have a short list, the next step is to work out how many prompts and engines you need to cover it, and our piece on estimating prompt and engine needs before you buy a plan covers that. You can also pick which questions to check first by looking at the four types of prompts worth tracking.
Turning a prompt list into something you can measure
Since you cannot get a demand number for a question, the thing you can control is your own set of questions and how consistently you check them. DeepSmith checks the questions you define on a schedule and reports your mention rate and citation rate for each one, with the answer history kept, so you can see how you show up for the questions that matter to your buyers. That gives you an observed result for your own list, and it does not claim to tell you how many people ask each question. If you are weighing tools for this job, the best prompt tracking tools roundup compares how they handle prompt sets.
If you would like to try that on your own prompts, you can start a free trial of DeepSmith and set up a starter list from your product and audience.
What would change this verdict
The grade here is mixed, and it could move. A platform that published a question-level metric with a clear definition, scope, period and method would make the picture stronger. So would a transparent, independently auditable cross-platform dataset. That is a standard for better evidence and is not a claim that either is planned. Our overview of the best AI visibility tools shows how different tools measure what they can. Until then, the sensible reading is that platform totals tell you AI use is huge, vendor estimates give you a direction for topics, and neither tells you how many people asked your question this week.
Related guides
- How to Run a 14-Day Proof of Concept With an AI Visibility Tool: Run a structured two-week trial with a fixed prompt set, success criteria and a day-by-day plan so you can compare tools on equal terms.
- A Governance Model for Multi-Brand AEO and GEO Programs: Set up a governance model for several brands: decide who owns prompts, content standards and reporting, what stays central versus local, and how approvals and exceptions get handled.
- Hidden Costs in AI Visibility Software: Overages, Add-On Engines and Seats: Spot the extra charges that inflate AI visibility software bills, including prompt overages, add-on engines, extra seats, history limits and export fees, and know which contract questions surface.
- How to Monitor LLM Brand Mentions Without Buying Software: Monitor brand mentions across LLMs using a spreadsheet and a fixed routine.
- Per-Engine Trackers vs Cross-Engine Trackers: Which Approach Fits?: Decide between single engine and cross engine trackers by weighing depth of data, coverage, cost, reporting effort and your team's priorities, and pick the approach that matches how.


