Prompt Research: How to Find the Prompts Your Buyers Ask AI

Prompt research is the practice of figuring out what your potential customers ask AI assistants, and then checking what those assistants answer. Think of it as keyword research for ChatGPT, Gemini, Claude, Copilot and Perplexity.

It borrows the mindset of keyword research, but the raw material is different. A keyword is a short string. A prompt is a small story: "I run a 12-person design agency, we use Google Workspace, what's the simplest project management tool that won't need training?"

Keywords vs prompts

• Length: keywords are usually a few words. Prompts are often full sentences or paragraphs.
• Context: prompts include the user's situation, budget, constraints and preferences.
• Output: a keyword returns a list of links. A prompt returns a synthesized answer that may name brands and cite sources.
• Data: keywords have search volume estimates. Prompts don't have reliable public volume data.

A growing share of people now use AI assistants somewhere in their research, especially for comparisons, recommendations and "what should I do" questions. That changes where buying decisions start.

The uncomfortable part: you can rank well on Google and still be invisible in AI answers. The models synthesize from many sources, and the brands that get named are often the ones mentioned consistently across trusted third-party sites, not just the ones with the best-optimized page.

If a buyer asks an assistant for the best tool in your category and your name isn't in the answer, you didn't lose a ranking. You lost the shortlist.Shmul

Because there is no trustworthy prompt volume data, the quality of your prompt set decides the quality of the whole exercise. Here's the process I use.

1. 1Start from buying jobs, not keywords. List the four to eight core problems your product solves and the decisions buyers make along the way.
2. 2Mine real language. Sales call notes, support tickets, demo requests, forum threads and reviews show how people describe their situation.
3. 3Translate top keywords into prompts. Your best commercial keywords are a starting point. Rewrite them the way someone would ask a smart colleague.
4. 4Add personas and constraints. The same need sounds different from a solo founder, an enterprise buyer, or someone on a tight budget.
5. 5Cover the stages. Include learning prompts, comparison prompts, recommendation prompts and "how do I" prompts.
6. 6Keep the set manageable. Fifty to two hundred prompts is plenty for most brands. You need coverage, not infinity.

The prompt types worth including

• Recommendation: what's the best X for Y.
• Comparison: X vs Y for my situation.
• Alternatives: alternatives to a competitor.
• Problem: how do I solve a specific pain.
• Validation: is X legit, is X worth it, reviews of X.

This is where most teams get misleading results. They ask a prompt once, screenshot the answer, and draw conclusions. AI answers vary between runs, models, versions, account settings, and whether web search is used.

• Run each prompt multiple times. Look at how often a brand appears, not whether it appeared once.
• Test several assistants. Visibility in one model doesn't predict visibility in another.
• Control the setup. Use clean sessions without personal memory or custom instructions where possible, so your own history doesn't skew results.
• Note whether web search was used. Answers grounded in live web results behave differently from answers from model memory.
• Record the cited sources. The list of URLs and domains an assistant leans on is often more useful than the answer text.

Once you have repeated runs, look for five things.

1. 1Mention rate. How often your brand appears across runs for each prompt.
2. 2Framing. How you are described. Accurate? Outdated? Positioned for the wrong audience?
3. 3Competitors. Who appears when you don't, and how they are described.
4. 4Sources. Which domains are cited or appear to shape the answer: review sites, comparison articles, forums, documentation.
5. 5Errors. Wrong pricing, discontinued features, confused product names. These are fixable, and they matter.

Prompt research produces a gap list. Here's how I map gaps to actions.

Gap: nobody explains your category well

Create the definitive explainer or comparison page, written answer-first, with specifics. The SEO content writing play covers the format.

Gap: competitors appear on third-party lists and you don't

This is off-site work: review platforms, comparison publishers, partner directories and digital PR. If the model's sources don't mention you, your own page alone rarely fixes it.

Gap: you appear, but with wrong or outdated information

Update your own pages, your structured data, and the most-cited third-party profiles. Consistency across sources is what eventually corrects the picture.

Gap: you are strong on Google but weak in AI

Look at entity SEO. Clear, consistent descriptions of who you are and what you do, across the web, help models connect your brand to the category.

Prompt research isn't a one-time project. Models update, sources change, and competitors move. Here's a routine that fits into a normal month.

1. 1Week one: run your core prompt set, with repeated runs, across your main assistants.
2. 2Week two: update the gap list and compare mention rates to last month.
3. 3Week three: ship one or two gap-closing actions, on-site or off-site.
4. 4Week four: refresh the prompt set with new language from sales, support and search data.