AI search volume is not directly measurable the way Google search volume is. ChatGPT, Perplexity, Gemini and Claude publish no per-prompt counts, prompts are long and almost never repeated word for word, and no assistant runs an advertising auction that would require public numbers. Every “AI search volume” figure on the market is therefore an estimate, built from user panels, clickstream samples or modelled from Google demand. What you can measure reliably is demand for the underlying question and how often your brand appears in the answers. SQSEO works from that angle: it finds the real questions people ask Google, ChatGPT and Perplexity, then measures your share of the answers, instead of pretending a prompt has a precise monthly count.
Why Google volumes exist and AI volumes do not
Google publishes search volume because it sells ads against keywords. Keyword Planner exists to help advertisers forecast and bid, and even there the numbers are rounded into ranges for most accounts. The keyword is the unit of the auction, so the keyword gets a count.
Chat assistants have no equivalent unit. A prompt such as “I run a small Shopify store selling ceramic mugs, which email tool is cheapest once I pass two thousand subscribers?” is effectively unique. Turning millions of prompts like it into a stable number requires grouping them by meaning, deciding which ones count as the same question, and choosing a sample. Each of those is a modelling decision. Two vendors making different decisions will publish different numbers for the same topic, and neither is wrong in the way a miscounted Google volume would be.
What the published usage data actually tells us
There is solid aggregate data, just not at the prompt level. The most detailed public study is the NBER working paper How People Use ChatGPT, written with OpenAI researchers. It reports that ChatGPT reached about 10% of the world’s adult population by July 2025, that non-work messages grew from 53% to more than 70% of usage, and that practical guidance, seeking information and writing together account for nearly 80% of conversations.
In the United States, Pew Research Center found that 34% of adults had used ChatGPT, roughly double the share in 2023.
These figures tell you the channel is large and that information seeking is a major use. They do not tell you how many people asked about your category last month. That gap is where estimates live.
How AI search volume estimates are built
| Method | How it works | Main strength | Main weakness |
|---|---|---|---|
| User panels | Opted-in users share their prompts through a browser extension or app | Real prompts, real wording | Small, skewed samples; heavy users and some countries overrepresented |
| Clickstream data | Traffic data sold by data brokers, inferred from referrals and visits | Large samples | Sees clicks out of assistants, not the questions inside them |
| Modelled from Google | Google volume for a topic multiplied by an assumed AI share | Stable and cheap | An assumption dressed as a measurement |
| Public conversation datasets | Research datasets of logged chats, such as WildChat | Transparent, inspectable | Not representative of commercial assistants today |
| Vendor query logs | An engine reports its own aggregates | Accurate for that engine | Almost never published at topic level |
Research datasets are useful for understanding how people phrase requests. The WildChat study collected about a million real user conversations with chatbots and shows how varied, multi-turn and personal prompts are. It also shows why prompt volume is hard: very few conversations repeat exactly.
When a tool shows “12,000 monthly prompts” for a topic, ask which row of the table produced it. Most commercial numbers combine a panel with a model, and the size of the panel decides how much the number moves from month to month.
Why do two AI volume tools disagree so much?
If you check the same topic in two tracking tools, the numbers rarely agree, and sometimes they differ by an order of magnitude. That is not a sign that one tool is broken. It is the predictable result of four choices each vendor makes differently.
The panel. A tool built on a browser extension sees the prompts of people who installed that extension. Those users skew toward particular countries, professions and levels of AI use. Scaling a few thousand of them up to a national figure multiplies every bias in the sample.
The grouping. Deciding that “best CRM for a five person agency” and “which CRM should a tiny agency use” are the same question is reasonable. Deciding that “CRM for agencies” also belongs in that group is a judgment. Wider groups produce bigger numbers.
The engines covered. Some tools estimate ChatGPT only. Others add Perplexity, Gemini, Copilot and Google’s AI Mode, each with its own assumed share of the market.
The refresh. A monthly figure from a small panel moves sharply from one month to the next simply because different people happened to be active. Smoothing hides this, but it also hides real change.
None of these choices is dishonest. They simply mean that a prompt volume is a property of the method as much as of the market, which is why comparing numbers across tools tells you about the tools rather than the topic.
What you can measure reliably instead
Three measurements are observable today, and together they answer the question volume was supposed to answer: is this topic worth the effort?
Demand for the question. Google long-tail data is still the best public proxy for what people want to know, because the same people ask both channels, often in similar words. Community threads on Reddit and forums add the phrasing real buyers use. Our guide to prompt-based keyword research shows how to turn those into a prompt set.
Your presence in answers. Run a fixed set of prompts across engines repeatedly and count how often you are named, cited or recommended. Answers vary between runs, so single checks mislead; repeated runs give a share you can track. The verifiability study of generative search engines found that cited sources often did not fully support the statements attached to them, which is one more reason to measure what engines actually say rather than assume.
The outcome. Referral sessions from assistants, growth in branded search, and sign-ups or calls where the customer says they found you through an AI answer. These lag, but they are the numbers a business decision rests on. Our post on knowing whether you rank in ChatGPT covers the setup.
How to prioritise topics without AI volumes
A practical scoring approach uses what you can observe and ignores what you cannot.
- Start from questions, not keywords. Collect the questions buyers ask about your category, from search suggestions, People Also Ask boxes, sales calls and support tickets.
- Score buyer value. A question asked right before a purchase decision is worth more than an educational one, regardless of size. The signals are covered in what defines a commercial intent prompt.
- Check current answers. Run each question in two or three assistants. Who is named? Are you absent, mentioned, or recommended?
- Estimate your chance. Questions where answers name small or mid-sized competitors are winnable; mapping SaaS comparisons in AI shows how to read who is named. Questions dominated by a few giants take longer.
- Use Google demand as a tiebreaker. Between two equally valuable questions, prefer the one with visible search demand, since it signals that people care.
This turns an unanswerable question, “how many prompts?”, into answerable ones: how valuable, how winnable, and how visible are we now. It also keeps the team focused on work it controls. Nobody can make a topic more popular, but anyone can make a page answer a question more completely, publish the facts buyers filter on, and earn coverage in the sources assistants cite.
Why small numbers are not a reason to skip a question
Many of the highest-value AI questions would show zero volume in any tool. Prompts are specific because people describe their situation: their budget, their platform, their team size. That specificity is exactly what makes them valuable, because a person describing their situation is usually close to a decision.
The GEO paper from Princeton and collaborators found that adding concrete statistics, quotations and citations to content improved its visibility in generative answers by up to 40% in their experiments. The lesson for prioritisation is that winnability depends heavily on how well your content answers the specific question, which is within your control, while volume is not. We make the same argument for traditional search in is long-tail keyword research dead.
Does Google demand predict AI demand?
Partly, and better than most people expect. The same person who types a short query into Google often asks a longer version of the same question in ChatGPT. Topics with no Google demand at all rarely turn out to be large in AI assistants, and topics with strong Google demand are almost always asked in assistants too.
Where the two diverge is in the shape of the question. AI prompts carry more context: the budget, the industry, the constraints, the comparison the person is weighing. Google captures the head of a topic well and the conversational tail poorly. So use Google data to confirm that a topic matters, and use real prompt phrasing, from panels, communities and your own customer conversations, to decide what the content must actually answer.
There is one systematic exception. Tasks people delegate to assistants rather than research, such as drafting an email or rewriting a paragraph, show up heavily in AI usage and barely at all in search. The NBER study’s finding that writing is one of the three biggest uses reflects this. Those tasks are rarely where a brand gets recommended, so the gap matters less for visibility than it seems.
How should you report AI visibility without volume?
Teams still need a number for the monthly report, and an invented one does more harm than none. A cleaner structure uses three figures that are honest about what they measure.
Coverage: the share of your tracked prompt set where your brand appears at least once across repeated runs. This answers “are we in the conversation at all?”
Share of answers: across all runs of all tracked prompts, the percentage that name you, compared with each main competitor. This answers “how often, relative to them?” Our guide to increasing your ChatGPT share of voice explains the method.
Outcome signals: referral sessions from assistants, branded search trends and self-reported attribution on sign-up forms.
Report the prompt set itself alongside the numbers, so anyone reading can see which questions were asked. A visibility figure without its prompt list is as meaningless as a ranking report without its keywords.
When should you buy an AI volume tool?
Buy one when you need to rank many topics against each other and the vendor explains its method well enough that you trust the relative order. A transparent panel with stated size and countries is useful for comparison inside the tool.
Before paying, ask four questions: how many people are in the panel, which countries they come from, how prompts are grouped into topics, and how much the figure for a stable topic moved over the last six months. A vendor confident in its method answers all four in a sentence each.
Do not buy one to forecast traffic, to compare against Google volumes, or to report precise numbers to a board. The figures are not stable enough for that, and two tools will disagree. If the budget is limited, spend it on measuring your presence in answers first, because that number moves when your work works.
Key takeaways
AI search volume is not directly measurable: no assistant publishes per-prompt counts, prompts rarely repeat, and every figure on the market is an estimate from panels, clickstream or modelling. Aggregate studies confirm the channel is large and information-heavy, with ChatGPT reaching about a tenth of the world’s adults by mid-2025. Measure what you can observe instead: demand for the question, your share of answers across repeated runs, and business outcomes. Prioritise by buyer value and winnability, use Google demand as a tiebreaker, and treat any prompt volume as a relative signal inside one tool.
