Free Growth Assessment
SEO & AEO

Prompt Research: Finding the Questions Buyers Ask AI Before They Call You

Nobody publishes how often buyers ask an AI a given question, so you can’t look the prompts up. Here’s where to find them, how to word them, and how to test the list.

By Holly Mack Last updated: November 20, 2026 13 min read

Prompt research is finding the questions your buyers ask AI assistants before they call you. You build the list from your own calls, tickets and search data, then test it across ChatGPT, Gemini and Perplexity.

A woman and a man in dark clothes reading a stack of blank printed pages together at a walnut table, one holding a pencil
The best prompts come from what buyers already said.

If you’ve typed your own category into ChatGPT and gotten back a list of names that leaves you out, the reflex is obvious. Find the right prompt. Fix the page. Run it again.

It makes sense why that feels like the job. But a step is missing, and it comes first. Which prompt?

You typed the question you would ask. Your buyer didn’t. Take a made-up 35-person accounting firm whose backup failed on Friday. It doesn’t type “best managed IT provider.” It writes something longer and messier, full of details you’d never think to include.

So the list you’re testing against is a guess, and a guess dressed up as a test plan will send you off to fix the wrong page for the wrong reason while the real question goes unheard. Guessing the prompt isn’t research. It’s a test you’ve already studied for.

This isn’t a content problem. It’s a research problem. Our AI visibility audit puts buyer questions to the assistants and logs whether you’re named, cited or missing, and those questions have to come from somewhere. This page is where, written so you can run it yourself.

Key Takeaways

The Short Version

Prompt research starts with people, not tools. The best prompts come from what buyers said on calls, in tickets and in search queries. Write each one three or four ways and test it across several runs, because assistants return a different answer nearly every time, and treat any prompt “volume” figure as a guess, since the assistants don’t publish one. The keyword half of the work is in keyword research for MSPs, and scoring what you find is in measuring MSP AI search visibility.

What Is Prompt Research, and How Is It Different From Keyword Research?

Prompt research is the work of collecting the questions real buyers put to AI assistants, rewriting them as realistic prompts, grouping them by intent, and testing whether your company shows up in the answers. It’s research on language and intent first. Measurement comes second.

Keyword research starts from a tool that reports volume. Prompt research can’t, because the assistants don’t publish it. SparkToro’s Rand Fishkin put it plainly in a November 17, 2026 analysis of a ChatGPT panel. Google and Bing report how often a keyword is searched, while ChatGPT, Claude, Perplexity, Copilot and the other assistants publish nothing like it.

Keyword Research vs Prompt Research

 Keyword researchPrompt research
You start fromA phrase with a monthly search countA buyer’s own question, often a few sentences
The data lives inKeyword tools and Search ConsoleYour calls, tickets, site search, forums and the assistants themselves
VolumeReported by Google and BingNot published by the assistants
The unit you trackOne keywordOne intent, worded several ways
What you checkA results page you can screenshotA fresh answer on every run

None of this retires keyword research. The two feed each other, and a short query in Search Console is often the seed of a longer prompt. Keyword research tells you what people typed. Prompt research tells you what they meant. Our keyword research template is the sheet for the first list. If you’d rather hand that list off, see our keyword research services.

Why You Can’t Just Look Up the Prompts

Suppose a tool did hand you every prompt your buyers type. You’d still be looking at the wrong layer. The question a buyer types is not what the engine searches. Here’s the evidence.

Google says AI Overviews and AI Mode may use a “query fan-out” technique, issuing multiple related searches across subtopics and data sources to build a response, in its guide to AI features and your website. OpenAI’s Help Center says ChatGPT search typically rewrites your query into one or more targeted queries, and may send more specific ones after it reviews the first results.

Microsoft shows the third piece. Bing Webmaster Tools added an AI Performance report in public preview on February 10, 2026, and its grounding queries show the key phrases the AI used when retrieving content that was cited. Microsoft calls the data a sample.

That gives you three layers. There’s what the buyer typed, what the engine searched and what it cited, and your pages compete on the last two while prompt research works on the first.

If you’ve ever run a service desk, you’ve seen this. The customer writes “email is broken.” The tech searches the knowledge base for “Exchange mailbox full.” Different words, same problem. You can’t fix the second without understanding the first. Sound familiar?

Where to Find the Prompts Buyers Actually Use

Sort your sources by how close they sit to a real buyer. Closer is better. The ones at the bottom are hints, not evidence.

Your own calls and notes

Start with sales calls. Discovery recordings, call notes, the first email a prospect ever sent you. Read for the sentence where they describe the problem in their own words, and for the details around it, like headcount, tools, compliance rules and what broke, because those details decide whether a prompt sounds real.

Then add one question to every discovery call. Just one. Did you ask an AI tool about this before you called, and what did you ask? The answers are short, real and free.

Support tickets, review notes and lost deals

Tickets show how customers describe a problem before anyone translates it, and lost-deal notes show the questions that ended with another provider’s name in the final decision, which is the part you’d want to understand first. Both already sit in your PSA and CRM, so pull twenty. Not two hundred.

Your own site and Search Console

Search Console’s Queries list shows what people typed before they landed on you. Filter it for words like how, what, best, cost and vs, and look at the long ones.

Short queries are seeds. Long queries are close to a prompt already, which is why this is where the keyword work and the prompt work meet.

One catch. Google counts AI Overviews and AI Mode traffic in the same Performance report, inside the “Web” search type, so there’s no separate list of AI prompts hiding in there. Site search and chat transcripts on your own pages are worth a look too.

Bing’s grounding queries

If your site is in Bing Webmaster Tools, open the AI Performance report. Grounding queries are the engine’s phrases, not your buyer’s, so they run shorter and more topical than a real prompt, which makes them a useful clue and a poor stand-in. Still, they show the phrases the engine used when it retrieved content it later cited.

Read them as a clue about the topics the engines connect to you. Then write the human version.

Forums, reviews and community threads

Buyers vent in public before they ask an assistant. Threads on Reddit and in MSP communities, review sites and the comments under an industry post all describe the problem in raw words, which is exactly the unpolished phrasing a real prompt tends to start from. Treat them as language samples. One thread isn’t a trend.

Weaker Sources Are Ideas, Not Evidence

SourceWhat it gives youWatch for
Asking an assistant what buyers ask about your categoryA quick list of plausible questionsIt reflects what the model has seen, not your buyers, and the answer changes between runs
A tool that shows a volume number beside each promptA way to rank a long listThe assistants publish no volume, so ask how the figure is produced
A published study of how people use ChatGPTThe shape of everyday behaviorSparkToro’s November 2026 panel is U.S. consumers, and says nothing about how a 60-person firm shops for managed IT
Your own promptsA fast first draftYou know the answer, so you write the question that returns it

That study deserves a closer look, because people lean on it. It covered 118,163 prompts from 12,321 U.S. ChatGPT panel users between January 1 and October 29, 2026, and only 6.9% of those prompts showed someone researching a brand to buy from.

The panel skews consumer, younger, female and lower income, and it leaves out API use. The author also notes his own ChatGPT history looks nothing like an average user’s. Marketers are outliers. So is anyone who builds a prompt list from their own habits.

How to Turn Raw Questions Into a Prompt List

Overhead view of two hands sorting blank index cards into four groups on a dark slate desk beside a pen
Group by what the buyer wants, not by how they worded it.

You now have a pile of fragments. Sentences from calls, long Search Console queries, a few forum threads. Four moves turn the pile into a list you can test.

Keep the constraints

Rewrite each fragment as the sentence a person would actually say out loud to a colleague, and keep every detail that made it specific, including headcount, location, tools, compliance, budget and what broke. Details are what separate a real prompt from a keyword wearing a question mark. Details win.

Write three or four phrasings per intent

In a February 11, 2026 follow-up, SparkToro described asking hundreds of people to write prompts for the same underlying intent. The average semantic similarity across those prompts was 0.081, so the wording varied dramatically, even though everyone wanted the same kind of answer. The tools often returned overlapping groups of brands anyway.

Read that twice. People word the same need very differently, and the assistants often still understand it. One phrasing per intent tells you next to nothing about how often you show up for real buyers.

Group by intent, not wording

Sorting by wording gives you a list that looks long and says little, while sorting by what the buyer is trying to do gives you a short list that covers real ground, and for an MSP buyer six intents cover a lot of it. Here they are.

An Intent Table for an MSP Buyer

IntentWhat the buyer wantsExample prompt (made up)
Find a providerA shortlistWe’re a 35-person dental group in Phoenix. Which managed IT providers are worth talking to, and what should we ask them?
Compare pathsA choice between two optionsShould a 60-person law firm keep an in-house IT person or move to a managed service provider?
Price itA range before a callWhat does managed IT cost per user per month for a small medical practice, and what gets left out of the quote?
Replace a providerA way outOur MSP is slow to respond. How do we know whether to switch, and what does switching involve?
Check a riskProof of compliance or securityWhat should a managed IT provider handle for a small clinic’s HIPAA requirements?
Vet a proposalA second opinionWe got a proposal from a managed IT provider. What should we check in the contract before we sign?

Every prompt in that table is invented to show the shape. None is a customer quote, and none comes from C4 data. Use your own calls instead.

Keep the list a size you’ll rerun

My working view, and not a benchmark, is five to seven intents with three or four phrasings each, which comes to roughly 20 to 30 prompts, and that’s enough to start because a list you’ll rerun every quarter beats a list of 200 you ran once. Start there.

A prompt list is a map of intent. It isn’t a keyword list in a trench coat.

How to Test a Prompt List Before You Trust It

Now run it. And run each prompt more than once. One answer is a sample of one.

SparkToro and Gumshoe.ai had 600 volunteers run 12 prompts through ChatGPT, Claude and Google’s AI a combined 2,961 times, as reported in their January 27, 2026 study. For ChatGPT and Google there was less than a 1 in 100 chance that any two responses to the same prompt gave the same list of brands. The order and the length of the lists varied too.

Two limits. SparkToro’s co-researcher works at Gumshoe.ai, an AI tracking startup, a link the author discloses. And the prompts asked for brands and products across several sectors, so the numbers aren’t a measure of IT buying.

Still, the author, who began the project as a skeptic of AI tracking and expected the results to be too random to measure anything, came away with one usable idea. How often a brand appears across many runs of the same prompt is a reasonable measure. Its position in any one list isn’t.

So the test is small. And repeatable.

  • Run each prompt several times on each assistant you care about. SparkToro ran each prompt 60 to 100 times, which is more than a small team will do, so say how many runs you used whenever you report a result.
  • Write down whether you were named, who was named instead, and which pages were cited.
  • Report a share of runs, such as named in 4 of 10. Never a rank.
  • Drop any prompt that returns nothing relevant, and any a real buyer would never say.

The scoring, the monthly routine and the reporting live in our measurement guide for MSPs. The aim here is smaller. Find out whether the list is good enough to score.

Once you know the questions, the page work comes next. Our AEO checklist for MSP and IT company pages covers it, and LLM optimization covers the wider job of getting named. For a finished list to start from, use our AI search prompt library.

Prompt Research Mistakes That Cost a Quarter

  • Writing it alone. If every prompt carries your category name and your preferred wording, you wrote a test you can pass. A prompt list built in the marketing office tests the marketing office.
  • One phrasing per intent. Hundreds of people worded one need very differently in SparkToro’s study.
  • Trusting a single run.
  • Paying for a volume number. Nobody can verify it, and Fishkin calls the demand guesses brands are forced into “impossible-to-verify.”
  • Treating ChatGPT as Google. In that same panel, 6.9% of prompts researched a brand to buy from. SparkToro’s author reads the data as a reason to watch Google’s AI features and Bing alongside ChatGPT, which is his view, not a rule.
  • Researching once. Buyers, products and platforms change, and a list from last spring describes last spring.

Bigger isn’t better. A bigger list of guesses is still a list of guesses.

Who Owns Prompt Research, and How Often to Redo It

Ownership is simple. Give it to two people, someone in sales or customer success who hears buyers talk and someone in marketing who can run the tests and fix the pages, because one person alone ends up with half a list.

Redo it every quarter, and again after a change in what you sell or a major platform update. Not yearly. That’s my working rule, not a published standard.

And yes, someone will suggest asking the chatbot to write the whole list. It can help with the phrasings. It can’t tell you what your buyers said.

If you can’t pull the call notes, can’t run the tests, or don’t have time to repeat them, that’s where a partner earns the fee. Here’s how we run the audit, which starts from the same buyer questions.

The better question isn’t which prompt to fix. It’s which questions you haven’t heard yet. Start with your next five discovery calls.

Ask each buyer what they asked an AI before they booked. Write down the exact words. That list is worth more than any volume number. If you’d like a second set of ears on it, book a growth strategy session.

What Owners Ask Before They Build a Prompt List

What is prompt research, in plain terms?

It’s finding out what your buyers actually ask AI assistants, then turning those questions into a list you can test. Think keyword research, except the unit is a full question and nobody publishes the volume.

The output is a set of prompts grouped by intent, with a record of whether each assistant names you.

Can I see how many people ask ChatGPT a certain question?

Not from the assistants themselves. SparkToro’s November 2026 analysis says ChatGPT, Claude, Perplexity and Copilot publish nothing like the search volume Google and Bing report.

Any figure you see beside a prompt comes from a model or a sample. Treat it as an estimate, and ask how it was built before you plan around it.

Is this just keyword research with a new name?

Related, but the unit changes. Keyword research counts phrases people searched. Prompt research collects full questions, with their details, and groups them by what the buyer wants.

They feed each other. A long query in Search Console is often a prompt in disguise.

How many prompts do I need to start?

Twenty to thirty, spread over five to seven intents, is a workable start. That’s a working view, not a benchmark.

Size matters less than whether you’ll rerun it. A short list you test every quarter teaches you more than a long one you ran once.

Do I need a paid tool to do this?

Not to begin. Call notes, Search Console, a spreadsheet and a few assistant accounts will get a first list built and tested.

A tool starts to earn its cost when you want the same prompts run many times on a schedule, and before you buy one, check how it produces any volume figure it shows.

How often should the list be redone?

Every quarter, plus after any big change in what you sell or in how a platform answers. That’s my working rule, not a published standard.

Ask the next five buyers what they asked an AI first. If the answers don’t match your list, it’s out of date.

Prompt Research AI Search Buyer Questions SEO & AEO
Share
READY TO ACT?

Find the questions your buyers are really asking

Talk to us about turning your sales calls and search data into a prompt list you can test, so you know which questions to build your pages for.

Book a Growth Strategy Call