Guide · 10 min read

Prompt research: how to find the questions people ask AI about you

Prompt research is keyword research for AI assistants: finding the questions your customers actually ask ChatGPT, Gemini, Perplexity, Claude or Copilot before they choose a provider. Those questions are longer, more personal and more specific than search keywords. This guide covers where to find them, how to turn them into a prompt set you can track, how many to keep and how to keep the list useful over time.

Prompt research and monthly tracking are part of every RecommendedGEO plan, from $449 a month with a three-month minimum.

Prompt research: how to find the questions people ask AI about you

What prompt research is, and why keyword lists are not enough

A keyword list tells you which short phrases people type into a search box. An assistant receives something different: a sentence or a paragraph, often with a situation attached, sometimes followed by two or three follow-up questions. The answer it writes names a few brands and leaves the rest out.

Prompt research aims to reproduce those conversations closely enough to measure them. The output is a prompt set: a fixed list of questions, written the way customers write them, grouped by intent and tagged by persona, country and language. Everything you later report about AI visibility depends on how good that list is.

How prompts differ from keywords

Take a keyword like “payroll software”. A prompt on the same topic might read: “I run a 12-person design studio in Manchester, we pay freelancers too, what payroll software would you recommend that works with Xero and costs less than £100 a month?” Four differences stand out:

  • Length. Prompts are full sentences, often several. The short head term is buried inside.
  • Conversational wording. People write as they speak, with “I”, “we”, “should I” and “what would you recommend”.
  • Constraints. Budget, team size, integrations, location, deadlines and deal-breakers narrow the answer, and they change which brands get named.
  • Context and follow-ups. The first question is often followed by “and which of those is easiest to set up?”. The brand that survives the follow-up is the one that wins.

This is why you cannot simply paste your keyword list into ChatGPT. The keyword tells you the topic; prompt research tells you the situation, the constraints and the comparison the customer has in mind.

Where to find real customer questions

The best sources are the ones where customers use their own words before they buy. Collect raw questions first and clean them up later.

  • Sales calls and demos. Ask the sales team for the questions they hear in the first ten minutes, and listen to recorded calls if you have them.
  • Support tickets, chat logs and emails. Pre-sale questions and objections are gold; post-sale “how do I” questions feed problem prompts.
  • Reviews of you and your competitors. What people praise and complain about shows the criteria they compare on.
  • Reddit and niche forums. Threads asking for recommendations show real wording, constraints and the alternatives people already consider. See Reddit for AI search.
  • People Also Ask boxes in Google. They show related questions around a topic and are quick to collect by hand.
  • Search Console long queries. Filter the Performance report with a regular expression to keep queries of seven words or more, or queries that start with who, what, which, how or best. These are people already writing in prompt style.
  • Bing Webmaster Tools AI Performance. See the next section for what its grounding queries are and are not.
  • Internal site search. What visitors type into your own search box reveals missing information and the terms they use for it.

What Bing's grounding queries can tell you

Bing Webmaster Tools added an AI Performance report in public preview in February 2026. It shows how often your pages are cited in Microsoft Copilot, AI summaries in Bing and some partner integrations. It includes grounding queries, which Bing describes as the key phrases the AI used when retrieving content that was then cited.

Read that definition carefully. A grounding query is the phrase the system used to look up sources, which is not necessarily what the person typed. Bing also says the data is a sample. In June 2026 Bing announced Intents, Topics, Citation Share and Compare, also in preview, which group grounding queries by intent and theme.

For prompt research this is still useful. Grounding queries show which sub-questions an assistant breaks a request into, and which of them already lead to your pages. Treat them as hints for topics and wording, not as a list of customer prompts.

Turn raw questions into a prompt set by intent

Once you have a few hundred raw questions, remove duplicates, merge near-identical ones and rewrite each in natural first-person language. Then sort them into intents. Five groups cover most businesses:

  • Discovery: the customer does not know any brand yet. “What is the best accounting firm for startups in Bristol?”
  • Comparison: the customer weighs named options. “Brand A or Brand B for a small online shop?”
  • Alternatives: the customer wants to leave or avoid someone. “What are good alternatives to Brand A that are cheaper?”
  • Problem: the customer describes a pain, not a product. “Our invoices keep getting paid late, what should we change?”
  • Local: the answer depends on place. “Which physiotherapy clinic near Clapham treats running injuries?”

Many teams add a sixth, smaller group for brand questions (“Is Brand A reliable?”), which shows how assistants describe you once you are named. Keep it small: being described well matters less than being named in the first place.

Discovery and alternatives prompts usually carry the most commercial weight, because they are where an assistant decides which brands make the shortlist. Comparison prompts matter once you are on it. Problem prompts are where content you write can earn a citation even before the customer knows they need a supplier.

Variations by persona, country and language

The same need sounds different depending on who asks. A finance director asks about compliance and audit trails; a founder asks about price and setup time. Write the main prompts once per relevant persona instead of averaging them into one bland question.

Country and language change the answer too. Assistants tend to favour local providers and local sources, so a prompt asked in Spanish from Madrid and in English from London are separate measurements. Do not machine-translate your prompt set. Ask local staff or customers how they would phrase it, because the constraints differ: VAT rules, local review sites, the competitors who actually operate there.

Keep variations under control. Two or three personas, the countries you sell in and the languages your customers use are usually enough. Each extra dimension multiplies the number of prompts to run.

How many prompts to track

Fewer prompts measured consistently beat many prompts measured once. Each prompt should be run several times per assistant, because answers vary between runs, so every prompt you add has a real cost.

  • Around 30 prompts suit a single market, a single language and one main service. This is our Starter plan.
  • Around 80 prompts cover several services, two or three personas or two markets. This is Growth.
  • Around 200 prompts make sense for several countries and languages, or a broad catalogue. This is Scale.

Whatever the size, keep a fixed core that you measure every period and a smaller test group for new prompts. That way trends stay comparable. The measurement method itself is covered in how to measure AI visibility.

Prioritise prompts by business value

Good prompt research is selective: not every prompt deserves a place in the core set. Score each candidate from 1 to 3 on three questions and keep the highest totals:

  1. Value: if an assistant recommends you here, how likely is a sale, and how large? A prompt from a buyer with budget beats a student's question.
  2. Frequency: how often do you hear this question in sales, support, reviews or forums? Real repetition beats guesswork.
  3. Reach: can you realistically influence the answer with content, listings or mentions in the next six months?

Prompt research that ends in a long list nobody scores is wasted. The scoring is what makes the set match your revenue rather than your curiosity.

A worked example with a fictional company

Example only: “Velora Heat” is an invented heat pump installer in Leeds, and every figure below is made up to show the process.

Velora's team collects 140 raw questions in three weeks: 60 from sales calls, 35 from reviews of Velora and two competitors, 25 from Reddit and local forums and 20 from Search Console queries of seven words or more. After merging duplicates, 70 remain.

They define two personas, homeowners with older houses and small landlords, and keep one market: Yorkshire, in English. They score the 70 prompts and keep 30 for the core set:

  • Discovery, 9 prompts, such as “Who installs heat pumps in old stone houses in Leeds and is actually good?”
  • Comparison, 5 prompts, such as “Velora Heat or a national installer for a three-bed semi?”
  • Alternatives, 4 prompts, such as “Alternatives to the big installers who are quicker to book?”
  • Problem, 8 prompts, such as “My heat pump bills are high in winter, is it the installation?”
  • Local, 4 prompts, such as “Heat pump installer near Headingley with good reviews?”

The first measurement shows Velora named in some problem answers but missing from almost every discovery answer, where two review sites and a national comparison page are cited repeatedly. The next steps follow from that: get onto those sources and write pages that answer the discovery questions directly.

Maintain the prompt set over time

Prompt research is not a one-off. Customers change how they ask, new competitors appear and assistants change the way they answer. A light routine keeps the set honest:

  • Monthly: add new questions from sales and support to the test group, not the core.
  • Quarterly: review the core set. Retire prompts that no longer match what you sell and promote test prompts that keep coming up.
  • After any launch, price change or new market: add the prompts that come with it.
  • Every change: log the date and reason, so a jump in a chart can be traced to the set and not mistaken for progress.

Where prompt research fits in GEO

The prompt set is the starting point of a GEO strategy. It tells you which answers to measure, which pages to write, as explained in how to write content AI cites, and which outside sources to target, such as best-of lists. If you want to see where you stand on a first set of questions, request a free GEO audit.

Frequently asked questions

Is prompt research the same as keyword research?

It has the same goal, finding out what customers look for, but the unit is different. Keywords are short phrases with search volume; prompts are full questions with constraints and context, and no tool reports reliable volumes for them yet. Keyword data is still a useful input.

Can I see what people actually type into ChatGPT about my brand?

No. AI companies do not share user prompts with site owners. The closest public signal at the time of writing is Bing's grounding queries, which are retrieval phrases rather than user prompts. That is why prompt research relies on your own customer conversations.

Should prompts include my brand name?

Most should not. Discovery, alternatives, problem and local prompts are where assistants decide whom to recommend, so they must be brand-neutral. Keep a small separate group for comparison and brand prompts where naming a brand is the point.

How often should I update my prompt set?

Add new questions monthly to a test group and review the core set every quarter. Change the core as little as possible, and record every change, so results stay comparable over time.

Do I need a paid tool for prompt research?

Not to start. The sources that matter most, sales calls, support, reviews, forums and Search Console, are free. Paid tools help later with running prompts at scale; the measuring guide explains what to compare.

The four workstreams of GEO: measuring AI answers, citable content, outside presence and technical GEO
AI assistants tracked every month: ChatGPT, Gemini, Perplexity, Claude, Copilot and Google AI Overviews
Monthly GEO process: measure, fix, publish and report

Pay with Stripe

On the live site this button takes you straight to secure card payment (Stripe). This prototype doesn't charge yet.

Scroll to Top