Prompt Research: Find the Questions Buyers Ask AI

Your keyword tool does not know what a prompt is. It sells you hvac repair tulsa — three words, a volume number, a difficulty score. It cannot tell you that a homeowner just typed forty words into ChatGPT describing a dead compressor, a sixteen-year-old unit, a $6,000 replacement quote she does not trust, and a request for a company that will not upsell her.

Those are two different artifacts. They need two different pages. And there is no Keywords Explorer for the second one — so you build the list yourself, out of data you already own, in a spreadsheet, for free.

What is a prompt, and how is it different from a keyword?

A keyword is a one-to-four-word fragment; a prompt is a full sentence carrying constraints — and the two behave nothing alike in search. Ahrefs analyzed 146 million SERPs and found AI Overviews appear on 9.5% of single-word queries but 46.4% of queries with seven or more words (Ahrefs, September 2025 desktop data).

Real user behavior says the same thing. Pew Research Center tracked 68,879 Google searches by 900 US adults and found just 8% of one- or two-word searches produced an AI summary, rising to 53% for searches of ten words or more (Pew Research Center, March 2025 data). Length is the trigger. Length is also what a keyword tool throws away.

So the fragment and the sentence are not the same target. Here is the difference laid out:

That last row is the whole problem, and it is why most AI-search plans are built on the wrong unit. Planning your GEO strategy from a keyword export is a category error — you are optimizing for the fragment while the buyer is typing the sentence.

Is there such a thing as prompt volume, and can you buy the data?

No. Nobody publishes prompt volume, and any prompt-volume number you are shown is modeled, not measured. Google's own documentation confirms that pages appearing in AI Overviews and AI Mode are folded into ordinary reporting — they are 'included in the overall search traffic in Search Console' and reported inside the Web search type (Google Search Central, December 2025).

There is no AI search type. No AI query dimension. No export of the prompt itself. And ChatGPT, Perplexity, and Gemini publish nothing at all about what individual people type into them.

Which means every dashboard quoting you a monthly prompt volume is inferring it — from a browser panel, a clickstream sample, or its own scraper. That is not a scandal, it is just a limitation, and the vendors selling you a $500/month subscription rarely lead with it.

Verdict: buy nothing yet. The two things you actually need — the prompt list and the citation check — are both free and both manual. We cover the no-budget version in detail in how to measure AI search visibility without spending $500/mo.

Where do you find the prompts your buyers actually type?

Five sources, and four of them are already sitting in your business right now. Work them in this order — the first two do most of the work:

• Search Console long-tail queries. Your 8-plus-word queries are people typing sentences into Google. It is the closest free proxy to prompt phrasing that exists, and almost nobody scrolls that far down the report.
• Recorded sales calls. The first ninety seconds of a discovery call is a prompt read out loud, with the constraints attached. Transcribe them and steal the phrasing.
• Inbox and DM objections. Every question a prospect emailed you before booking is a prompt they were too polite to type into a search bar.
• Reddit and niche forums. Pew found that Wikipedia, YouTube, and Reddit are the most frequently cited sources in Google's AI summaries, together making up 15% of the sources listed. The threads AI is reading are also the threads where your buyers phrase their problem in their own words.
• The engines themselves. Paste a seed prompt into ChatGPT, then ask it what else someone in that situation would ask. It will hand you the fan-out — which is exactly what Google says AI Mode does internally: it issues multiple related searches across subtopics to build one answer.

Notice what is not on that list: a keyword tool. Keyword research still matters for the classic SERP, and we still do it — see keyword research for service businesses — but it is a different input for a different surface.

How do you mine 8+ word queries out of Search Console for prompt ideas?

It takes about ten minutes with one regex. Open the Performance report, set the grouping dimension to Queries, then add a query filter and choose Custom (regex) — Search Console accepts RE2 syntax (Google Search Console Help).

Use this expression to return only queries of eight words or more: ^(\S+\s){7,}\S+. Set the date range to the last 12 months, sort by impressions, and export.

If regex makes you nervous, do it in a sheet instead. Export the Queries tab, add a column that counts words — the length of the query minus its length with spaces stripped out, plus one — and filter to 8 or higher. Same result, no syntax.

Two honest caveats before you trust the output. Google omits anonymized queries from this report, and filtering by query can shift your totals, so the list is a sample, not a census. And a query with 400 impressions and zero clicks is not a failure here — it is a prompt shape, and it belongs in the set. If Search Console is new territory, start with our Search Console guide for service businesses.

How do sales calls and inbox objections become prompts?

By keeping at least two constraints and one deal-breaker in every line you extract — that is the test that separates a real prompt from a keyword you dressed up in a sentence. 'Best SEO agency' is a keyword wearing a costume. 'I run a 12-tech HVAC company in Tulsa, I got burned by a 12-month contract, and I need someone who can show me leads not reports' is a prompt.

Capture the phrasing verbatim. Do not clean up the grammar, do not remove the frustration, do not shorten it. Engines respond to phrasing, and a tidied-up prompt is a different prompt that tests a different thing.

The translation from objection to prompt to page is mechanical once you see it:

Every row on the right is a page that already exists or should. Ours are pricing and GEO. Yours will be different — but if a prompt in your set has no page to point at, you have found the actual work.

How do you build a 30-prompt tracking set and map it to pages?

Thirty prompts, split evenly into three buckets of ten, is the smallest set that survives contact with reality: small enough to re-run by hand in under an hour, wide enough to notice when something moves. Ten problem-aware prompts (the buyer knows the symptom, not the fix), ten vendor-selection prompts (they are choosing between providers), ten objection prompts (price, contract, trust).

Three composition rules, and they are not negotiable. Every prompt is at least 12 words. Every prompt carries a geography, a vertical, or a budget. Every prompt maps to exactly one page.

That last rule does more work than it looks like it does. If two of your pages both claim the same prompt, you do not have a GEO problem — you have a cannibalization problem, and no amount of schema will fix it.

The tracking sheet needs six columns and nothing else:

You do not write 30 pages. You write six to ten, each one built to answer a neighborhood of four or five related prompts in a single passage. That structural choice — one page, one whole ask, answered up front — is what passage-level optimization is for.

How often should you re-run the prompt set, and what counts as a win?

Monthly, on the same day, logged out, in a clean browser session. Weekly is theater — you will spend an hour a week generating noise. And run every prompt twice in the same session: engines are nondeterministic, and the same prompt can return different sources ten minutes apart. Count a citation only when it shows up in both runs.

A win is a bucket moving, not a lucky hit. If seven of your ten objection prompts went from Absent to Mentioned this quarter, that is real. If one prompt cited you once and never again, that is variance, and reporting it as progress is how agencies build monthly decks out of nothing.

When a prompt comes back Absent, the fix is usually boring. Google's documentation is blunt about this: to be eligible in AI Overviews or AI Mode a page 'must be indexed and eligible to be shown in Google Search with a snippet' and there are 'no additional technical requirements' (Google Search Central). Not a schema trick. A page that answers the whole ask, in text, near the top.

And the scoreboard above the prompt set is still booked calls. Share of voice across 30 prompts is a diagnostic, not a result. If a channel produces no qualified leads in 90 days, we cut it — that applies to AI search the same as anything else.

If you want the prompt set built against your own Search Console data and mapped to the pages you already have, that is part of the GEO work inside our base retainer — not a separate AI-visibility upsell, and no lock-in. Get my free audit.