Prompt Volume: How to Find What People Actually Ask AI About Your Category

Every AI visibility plan runs into the same wall early: you cannot optimise for questions you haven’t identified, and no engine publishes what people ask it. Search volume tools measure keyword demand in Google. Nobody sells you a reliable count of ChatGPT prompts. So teams either guess, or they buy a number that was itself a guess.

There is a better answer, and it is free. It’s also honest about what it can and cannot tell you.

Why AI prompt volume cannot be measured directly

Conversational engines do not release query data. Their inputs are longer, messier and more personal than search queries, frequently multi-turn, and often shaped by context the user pasted in. Even if a vendor sampled prompts, that sample would not generalise across engines, regions or user segments.

Anyone quoting you a precise monthly prompt count for a specific phrase is modelling, not measuring. Treat those figures as directional at best and never put them in a board pack as fact.

What you can measure instead

You can measure the shape of curiosity in your category — the questions people are actively formulating — using public sources that reflect real human input. These are proxies. Used carefully, they’re good enough to build a content plan on, because you need relative priority, not absolute counts.

1. Harvest Google Autocomplete across question modifiers

Autocomplete surfaces real, frequency-weighted query prefixes. Most people harvest it lazily by typing one seed. The method that works is systematic expansion.

1. List your seed terms: your category, your product type, the problem you solve, the job title of your buyer, and your main alternatives.
2. For each seed, prepend question modifiers one at a time: how to, how do, what is, why is, when should, who needs, which, best, cheapest, alternative to, is it worth, do I need.
3. Then append modifiers after the seed: vs, cost, pricing, problems, requirements, for small business, for enterprise, checklist, template, example.
4. Run the alphabet sweep: seed plus each letter a–z, which forces autocomplete to reveal completions it would otherwise rank below the fold.
5. Record every suggestion verbatim, tagged by the seed and modifier that produced it.

Done properly across ten seeds this yields several hundred real phrasings in an afternoon. The value is in the phrasing, not the count — you’re learning the vocabulary your market uses, which is exactly what a language model is matching against.

2. Mine People Also Ask recursively

Search your top harvested queries and record the People Also Ask questions. Expanding one PAA item generates more, so the tree grows quickly. Two or three levels deep is usually enough before it drifts off-topic.

PAA is particularly valuable because it’s already in question form and already reflects what search systems judge to be the adjacent intent. That adjacency map is closer to how a conversational engine follows up than a keyword list is.

3. Read the questions communities actually ask

Go where practitioners in your category talk: subreddits, Stack Exchange sites, industry forums, LinkedIn comment threads, association discussion boards, review-site Q&A sections, and the questions asked at conferences and webinars.

Community questions carry something autocomplete strips out: context, constraints and frustration. “How do I do X when the vendor won’t export Y and my team has no budget” is the real prompt. Those long, situational questions map far more closely to how people talk to AI than three-word keywords do.

4. Harvest your own first-party question sources

• Sales call recordings and discovery notes — the questions asked before a deal.
• Support tickets and chat logs — the questions asked after.
• Your site search box — often the most neglected question archive in the business.
• Inbound email and RFP documents — the formal version of the same questions.
• Webinar and event Q&A transcripts.

This source is the one your competitors cannot copy, and it’s weighted toward buyers rather than browsers. If you only do one of the four, do this one.

5. Ask the engines what they think people ask

Prompt several AI engines directly: “What questions do buyers typically ask before choosing a [category] provider?” This isn’t evidence of demand — it is the model’s prior. But it is genuinely useful, because that prior is close to what the model will treat as the canonical question set when it answers someone.

Turning raw questions into a prioritised list

You now have a few hundred questions and no volume data. Score them instead, on three axes you can actually judge.

• Commercial proximity — how close is this question to a purchase decision? Score 1–3. “What is X” scores 1; “what does X cost to implement” scores 3.
• Answer scarcity — run the question in two or three AI engines. If the answer is thin, hedged, or cites nobody credible, the ground is open. Score 1–3.
• Right to answer — do you have genuine, specific, first-hand knowledge here that others do not? Score 1–3.

Multiply the three. Anything scoring 18 or above is a priority page. This ranking beats a fabricated volume estimate because every input is something you verified yourself.

The caveats you must state out loud

Be explicit with stakeholders about the limits of this method, or someone will eventually present it as harder data than it’s.

• Autocomplete and PAA reflect search behaviour, not AI prompt behaviour. They correlate; they aren’t the same population.
• Both are personalised and localised. Results differ by location, device, account state and time. Use a clean session and record your conditions.
• Autocomplete is filtered. Sensitive, regulated and disparaging phrasings are suppressed, so entire real question categories will be invisible.
• None of this yields counts. It yields a ranked list of plausible questions, which is a different and more honest thing.
• AI prompts are longer and more conversational than the fragments these tools return. Treat harvested phrases as the seed of a question, then write the fuller version a person would actually type.

Making it repeatable

Question demand shifts as your category changes. Re-run the harvest quarterly, keep the master list in one place, and mark which questions you have covered, which you own in AI answers, and which remain open. That list becomes your content roadmap and your visibility scorecard at the same time.

Once you’ve the prompt set, you need to know who currently gets cited for it. Our ARIA citation tracker runs a fixed prompt set across engines on a schedule so the answer scarcity score stops being a manual check, and the AI citability scorer tells you whether the page you wrote in response is actually extractable. Both sit inside a broader AI visibility programme.

Frequently asked questions

Are there tools that report real AI prompt volume?

Several vendors publish prompt-volume estimates, but these are modelled from search data, panels or samples rather than measured from engine logs. They can be useful directionally. Ask any vendor exactly what their number is derived from before you rely on it.

How many questions should we start with?

A working prompt set of 30 to 60 questions is enough to run a meaningful visibility programme. Harvest widely, then cut to the ones that score highest on commercial proximity and right to answer.

Should we target long conversational prompts or short keywords?

Write for the long form. Conversational prompts are how people address AI engines, and a page that answers the full situational question will also cover the short keyword inside it. The reverse is not true.

How often should we refresh the question list?

Quarterly for most categories, and immediately after any material change — new regulation, a new entrant, a pricing shift or a technology change. Those events create new questions faster than any harvest cadence will catch on its own.