TL;DR
- Prompt research is keyword research for AI search. It means finding the real questions your customers ask ChatGPT, Gemini, Perplexity and Google’s AI answers, in their own words.
- Prompts are longer and more specific than keywords. According to Semrush’s US clickstream data, prompts that trigger a web search in ChatGPT almost doubled in length in a year, from 4.7 to 8.7 words.
- No tool can show you real prompt volume yet. The best sources are your own: sales calls, support emails, reviews, forums and the long queries in Search Console.
- The output is a prompt library: a simple table of 20 to 50 questions, grouped by customer type and buying stage, that you test every month.
- Below: where to find real prompts, the columns I use, example prompts and the mistakes I see most often.
By Vyara Stefcheva · Independent expert in AI search and GEO. I run prompt research for brands in several European markets and languages with my own testing tool. This guide is based on that client work. About me · Updated October 2026
Every week someone asks me the same thing in a different form: “What keywords should I target for ChatGPT?”
The honest answer is that keywords are the wrong unit. People don’t type three words into ChatGPT. They describe their situation, add their budget and their city, and ask for a recommendation. If you want to know whether AI recommends your business, you first need to know what people actually ask. That work is called prompt research.
What is prompt research?
Prompt research is the process of finding, collecting and grouping the questions your potential customers ask AI assistants, so you can test whether and how those assistants mention your business. It is the AI search equivalent of keyword research, but its unit is a full question in natural language, not a short search term.
The result of prompt research is a prompt library: a list of realistic questions you test regularly, from neutral accounts, in the languages and markets you sell in.
From keyword to prompt · pick an example
↓ how the same person asks ChatGPT
Prompt research vs keyword research
| Keyword research | Prompt research | |
|---|---|---|
| Unit | A short search term | A full question, often several sentences |
| Data source | Search volume from tools | Your customers, your sales team, forums, long queries |
| Volume data | Estimated, but widely available | Not reliably available yet |
| Intent | Usually one per keyword | Often three or four in one prompt |
| Result you track | Ranking position | Whether you are mentioned, where and how accurately |
Keyword research still matters. Google is not going anywhere, and good SEO helps AI visibility. But keyword data alone will not tell you how people talk to an assistant.
Why it matters now
ChatGPT passed 700 million weekly users in 2025, and OpenAI’s own research found that seeking information grew to 24.4% of conversations by mid-2025, up from 14% a year earlier. More of those conversations end with a recommendation: which tool, which agency, which clinic.
The questions are also getting longer. In its analysis of more than a billion lines of US clickstream data, Semrush found that prompts that trigger a web search in ChatGPT grew from 4.7 to 8.7 words between early 2025 and early 2026. A keyword like “accountant Berlin” becomes “I’m a freelance designer who just moved to Berlin, which accountant can handle German taxes in English and doesn’t charge a fortune?”
That second question contains a location, a profession, a language, a budget concern and a request for a recommendation. You won’t find it in any keyword tool.
The shift in two numbers
Sources: Semrush, US clickstream data, Jan–Feb 2025 vs Jan–Feb 2026; OpenAI’s study of ChatGPT usage, 2025.
Where real prompts come from
As I explained in my article on how AI visibility tools estimate demand, most prompt numbers you see are calculated, not measured. So I start with sources that come from real people.
- Sales and discovery calls. Write down the first question new clients ask, word for word. It is usually very close to what they asked an AI assistant before they called you.
- Support emails and chat logs. They show the problems, comparisons and doubts people have before and after buying.
- Reviews, yours and your competitors’. Reviews tell you which details matter enough for people to write about them: speed, price, language, location, a specific service.
- Forums, Reddit and Facebook groups. Look for “can anyone recommend” posts in your category. The phrasing is often exactly how people talk to ChatGPT.
- Long queries in Search Console. Filter queries of ten or more words with the regex
^(\S+\s+){9,}\S+$. Many of these come from conversational search in Google. - Bing Webmaster Tools. The AI Performance report shows the grounding queries behind citations in Copilot. They are not the user’s original prompt, but they show what the assistant searched for.
- The assistant itself, with care. Asking ChatGPT “what would a customer ask you about X” gives you ideas, not data. Use it to fill gaps, never as your only source.
^(\S+\s+){9,}\S+$Search Console → Performance → Query → Custom (regex). Shows queries of ten words or more.
How to build a prompt library
A prompt library doesn’t need special software. A spreadsheet is enough. These are the columns I use with clients:
| Column | What goes in it |
|---|---|
| Prompt | The question, in the customer’s words and language |
| Customer type | Who is asking: a founder, a procurement manager, a parent |
| Buying stage | Problem, research, comparison or decision |
| Market | Country, city and language |
| Platform | ChatGPT, Gemini, Perplexity, Google AI Overviews |
| Mentioned? | Yes or no, and in which position |
| Who appears instead | The competitors named in the answer |
| Sources cited | The websites the assistant links to |
| Accuracy | Whether what it says about you is correct |
| Date | When you ran the test |
Free template
The same columns as a spreadsheet, with five example rows. Opens in Excel, Numbers or Google Sheets (File, then Import).
Start with 20 to 30 prompts and grow from there. For the testing routine itself, see How to Measure AI Visibility.
Prompt types to include
- Direct: “What is the best CRM for a small real estate agency?”
- Comparison: “Is HubSpot or Pipedrive better for a team of five salespeople?”
- Problem-based: “Our leads go cold after the first call. What software helps with follow-ups?”
- Local: “Which digital marketing agency in Lisbon works with B2B tech companies?”
- Constraint-heavy: “I need an accountant who speaks English, works with freelancers and costs under €100 a month.”
- Follow-up: “Why did you recommend them? Are there any complaints about them?”
The follow-up type is the one most people forget. Customers rarely stop at the first answer, and the second answer is often where your reputation is decided.
Try it: build a realistic prompt for your business
Tip: open the links in a private window, so the answer isn’t personalised to your own account.
Common mistakes
- Turning keywords into questions. “Best accountant Berlin?” is still a keyword with a question mark.
- Testing from your own account. Assistants personalise answers. Use private windows or neutral accounts, from the right location.
- Running the test once. Answers change. Trends across many prompts and months matter, single answers don’t.
- Trusting estimated volume. Treat prompt volume numbers from tools as direction, not as fact.
- Researching only in English. If your customers ask in German, Romanian or Bulgarian, your prompt library has to be in those languages too.
FAQ
Is there such a thing as prompt search volume?
Not in a reliable form. AI platforms don’t publish what people ask them. Tools estimate volume from clickstream panels or model it, so the numbers are useful for direction, not as exact demand.
How many prompts do I need?
Twenty to thirty is enough to start for one market and one product category. Larger businesses with several categories and markets usually need between 60 and 200.
Can a tool do prompt research for me?
Tools can track prompts and suggest new ones, but they don’t know your customers. The best prompts come from your sales conversations, reviews and support, and then a tool can help you run them at scale.
How often should I update my prompt library?
Review it every quarter and test it every month. Add new questions when you launch a product, enter a market or hear a new objection from customers.
Your prompt research checklist · 0/6 done
My take
Prompt research is not a new skill. It’s the old skill of listening to customers, applied to a new channel. The businesses that already know exactly how their buyers talk have a head start, because they don’t need a tool to guess.
If you only do one thing after reading this, write down the last ten questions your new customers asked you, word for word. Then ask ChatGPT the same questions and see who it recommends.
Want to know where you stand? My AI visibility audit starts with exactly this kind of prompt research, in the languages and markets you sell in.
About The Author
