Prompt research tells you which questions buyers ask ChatGPT, Perplexity, Gemini and Google’s AI features, and whether your brand shows up in the answers. It has two jobs: finding prompts you are missing, and measuring the prompts you already appear for. This guide covers three practical methods, plus a workflow you can run every month in OtterlyAI.
One clarification before we start. This is prompt research for marketing and GEO, not prompt engineering. Prompt engineering is about writing better instructions for a model. Prompt research is about understanding what your customers ask.
Quick answer: how do you do prompt research? Generate prompt ideas from your keywords, URLs and brand with a prompt research tool. Add real phrasing from Google Search Console, customer surveys, sales calls and communities like Reddit. Group the prompts by intent stage, keep the ones with the most buying intent, then track them over time to see where you get mentioned and cited.
What is prompt research?
Prompt research is the process of identifying the prompts your potential customers type into AI search engines, then deciding which of them are worth monitoring and optimizing for.
Think of it as keyword research for AI search. The output is not a list of short search terms. It is a list of conversational questions, grouped by intent, with an estimate of how much demand sits behind each one.
In OtterlyAI, prompts are the foundation of everything. Your brand mentions, citations, share of voice and sentiment are all measured against the prompts you track. If your prompt list does not reflect what customers really ask, your visibility data will not either.
Why you can’t just look up AI search queries
There is no Search Console for ChatGPT. OpenAI, Perplexity and Google do not publish the prompts people type into their assistants. You can’t see:
- Which prompts people use in your category
- How often a given prompt is asked
- Which brands get mentioned, recommended or clicked
That may change one day, but you should not wait for it. Every method in this guide is a way to get a reliable signal without direct access to that data.
How is prompt research different from keyword research?
Keywords are fragments typed into a search box. Prompts are full questions asked to an assistant, often with context, constraints and follow-ups.
| Keyword research | Prompt research | |
|---|---|---|
| What people type | Short phrases, e.g. “CRM small business” | Full questions, e.g. “Which CRM is easiest for a 10-person agency that uses Gmail?” |
| Intent | Guessed from a fragment | Stated directly, including budget, team size, location or use case |
| Volume data | Mature, published by Google and SEO tools | Estimated and still developing |
| What you win | A ranking position in a list of links | A mention or citation inside one AI answer |
| Stability | Rankings move slowly | Answers vary between runs, engines and countries |
| Unit of work | A keyword | A prompt, its intent stage and its fan-out sub-questions |
The constraints are what matter. A page optimized for “best CRM” can still be invisible for the agency owner above, even though that person is closer to buying.
Keyword research is not dead. Your SEO keywords show how your market talks, and they are one of the best starting points for prompt ideas. They are just no longer the full picture.
The two halves of prompt research
Every method below serves one of two jobs. Knowing which one you are doing saves a lot of frustration.
Discovery surfaces prompts you have not thought of and may not appear for yet. Measurement shows how your brand performs on prompts you already know about.
| Method | Main job | Finds prompts you don’t appear for? |
|---|---|---|
| OtterlyAI AI Prompt Research tool | Discovery | Yes |
| Customer surveys, sales calls, Reddit, ChatGPT autocomplete | Discovery | Yes |
| Google Search Console query data | Discovery and measurement | Partly: only queries where you already get impressions |
| Search Console generative AI report, Bing AI Performance | Measurement | No |
| OtterlyAI prompt monitoring | Measurement | No, it tracks the prompts you add |
You need both. Discovery fills your list. Measurement tells you which prompts to work on and whether your changes worked.
Cover all four intent stages
A good prompt list is not just “best [category]” prompts. Buyers ask AI different questions depending on how far along they are. We group prompts into four stages:
| Intent stage | What the buyer knows | Example prompt |
|---|---|---|
| Problem based | They know their problem, not the solution | How can I see if ChatGPT recommends my brand? |
| Category based | They know the solution category | What are the best AI search monitoring tools for agencies? |
| Competitor based | They use a competitor and want something better | Alternatives to [competitor] for tracking AI Overviews |
| Brand based | They know you and want details | Is OtterlyAI good for tracking Perplexity citations? |
If you want to grow, weight your list toward problem and category prompts. These are non-branded, have high buying intent, and are where new customers first meet your brand. Brand-based prompts are still worth tracking, because they show how AI describes you to people who are already interested.
Method 1: Generate prompt ideas with the AI Prompt Research tool
Start here. It is the fastest way to build a broad first list, and it surfaces prompts your team would not have written on its own.
OtterlyAI’s AI Prompt Research tool accepts three kinds of input. Each returns a list of prompt ideas with an Intent Volume estimate, so you can see which ideas likely carry the most demand.
- Start from your SEO keywords. Paste your best-performing keywords and OtterlyAI turns them into conversational prompts, in different languages if you need them. This is the quickest bridge between your SEO work and your GEO work.
- Start from a URL. Add a page and OtterlyAI extracts its topics and suggests matching prompts. Use your top landing pages, pages that already get referral traffic from ChatGPT, or a competitor’s key pages to see the ground they cover.
- Start from your brand. Enter your brand name, domain, industry and country. OtterlyAI suggests prompts across the intent stages, including prompts that involve your competitors.
Run all three. Keywords give you volume, URLs give you topic coverage, and the brand input gives you the competitive and positioning angle.
What to keep in mind: these are estimated prompts, not logs of real ones. Their quality depends on your inputs. That is why Methods 2 and 3 add real customer phrasing on top.
Method 2: Turn your search engine data into prompts
Your own search data is the largest free source of real, conversational questions you have. It is not ChatGPT data, but it is the closest scalable proxy available today.
Google Search Console queries
Since Google launched AI Mode, Search Console shows far more long, conversational queries. Many read exactly like AI prompts. Filter your Performance report for:
- Question words and comparisons: how, what, why, which, best, vs, alternative
- Long-tail queries with high impressions: real demand in the searcher’s own words
- Queries with impressions but zero clicks: often content gaps worth a dedicated page
Then paste the best ones into the AI Prompt Research tool as keyword inputs, or add them as prompts directly. We walk through the full process in how to extract prompts from Google Search Console, and share what we found in our analysis of real prompts in Search Console.
The limit: you only see queries where your site already earned an impression. Searches where you are invisible do not appear.
Search Console’s generative AI report and Bing AI Performance
Both Google and Microsoft now report on AI visibility, but neither gives you a list of user prompts.
- Google’s generative AI performance report shows which of your pages appeared in AI Overviews and AI Mode, with impressions by page, country and device. It has no query dimension and is still rolling out to some properties. Use it to see which pages AI already trusts, then copy their format and depth.
- Bing Webmaster Tools AI Performance shows how often Copilot and Bing’s AI answers cited you, plus “grounding queries”. A grounding query is the phrase the AI searched to find sources, not what the user typed. Treat it as a hint about retrieval phrasing, not as a prompt list.
These reports belong to the measurement half. They are useful for improving what you have, not for finding what you are missing.
Method 3: Collect real phrasing from customers and communities
Tools and search data give you scale. Customers give you the exact words. This method fills in the phrasing, constraints and pain points the other two miss.
| Source | What you get | Watch out for |
|---|---|---|
| Customer survey | Prompts in your buyers’ own words | Low response rates; only covers people who already bought |
| “Where did you hear about us?” field | Ongoing data at the moment of conversion | Shows the channel, rarely the exact prompt |
| Sales calls, demos and RFPs | Real questions, objections and comparisons | Skews to mid and late funnel |
| ICP panels (e.g. Pollfish, Wynter) | Input from non-customers in your target market | Cost at scale; answers can be generic |
| Website search logs | What visitors look for on your site | Site-specific, not global |
| ChatGPT autocomplete | Native suggestions from ChatGPT itself | No volume data; changes often |
| Copilot related queries | AI-generated follow-up topics | Not ChatGPT-specific |
| Reddit threads | Honest, natural questions in niche markets | No volume data; threads drift off topic |
| AnswerThePublic, AlsoAsked | Question clusters around a topic | Google-based, not AI-specific |
| Query fan-out tools (e.g. Qforia) | The sub-questions AI breaks a prompt into | Simulated, can over-generate |
Ask your customers directly
The simplest method is often the most valuable. Pull a segment from your CRM and send a two-minute survey with three questions:
- Where did you first hear about us? (Include “ChatGPT or another AI assistant” as an option.)
- Did you use an AI assistant while researching tools like ours?
- If yes, what did you ask it? Please share any prompts you remember.
Offer a small thank-you to lift response rates. Add question 1 as a required field on your lead forms too. One marketing team we spoke to saw ChatGPT drive only 1 to 2% of site visits, while 5 to 10% of new leads named it as their first touch.
Brainstorm with AI
AI assistants are good at simulating buyer questions across the funnel. Use our free custom GPT prompt finder, or paste this template into ChatGPT with web search turned on:
Act as an experienced brand and marketing manager. I want to monitor AI search visibility for the brand below.
Create 10 short prompts per intent stage that potential customers would ask ChatGPT when looking for a brand like this.
Brand: [brand name]
Domain: [domain]
Industry: [industry]
Country: [country]
Intent stages:
- Problem based: they describe their problem
- Category based: they search by solution category
- Competitor based: they want a better alternative to what they use
- Brand based: they are asking about this brand
Return only the prompts, one per line, grouped by stage, without quotes.Treat the output as a draft. It reflects the model’s assumptions and yours, so cross-check it against what customers and Search Console tell you.
Put the three methods into one workflow
No single method is perfect. The signal comes from layering them and repeating the process. Run discovery monthly and review your tracked prompts every week or two.
- Discover. Run the AI Prompt Research tool on your keywords, your top URLs, your brand, and two close competitors’ URLs.
- Enrich. Add conversational queries from Search Console and the phrasing you collected from customers, sales and Reddit.
- Clean and tag. Remove duplicates and near-duplicates. Tag each prompt by intent stage and topic, so you can report on groups rather than single prompts.
- Prioritize. Keep prompts with high Intent Volume, clear buying intent and a real fit with what you sell. A focused list you act on beats a huge list you only watch.
- Track. Add the shortlist to a brand report in OtterlyAI. You will see brand mentions, citations, competitors and sentiment across ChatGPT, Google AI Overviews, Google AI Mode, Perplexity, Gemini and Copilot.
- Act on the gaps. For prompts where competitors appear and you don’t, check which pages the AI cites. Those sources are your content brief: publish or update pages, and pursue coverage on the third-party sites AI trusts.
- Repeat. Re-run discovery, retire prompts that no longer matter, and add new ones.
Tips for tracking prompts in OtterlyAI
- Add important prompts early. Data collection starts the day you add a prompt. There is no history from before that date.
- Get the wording right first. Prompt text can’t be edited. Changing the wording means deleting the prompt and adding a new one, and the old history does not carry over. Tags and brand report assignments can be changed anytime.
- Only tracked prompts produce data. OtterlyAI shows mentions and citations for the prompts you monitor, so gaps in your list become gaps in your reporting.
FAQ
Can I see the prompts real users type into ChatGPT or Perplexity?
No. AI search engines do not publish query data, and no tool has access to a log of real user prompts. The methods in this guide are the best substitutes available.
How many prompts should I track?
Start with a focused set that covers all four intent stages for your most important product or market. A practical starting point is 25 to 50 prompts, weighted toward problem and category prompts. Expand by product line, persona or country once you know which prompts drive results.
How does OtterlyAI score the prompts it suggests?
Each suggested prompt comes with an Intent Volume estimate. Learn how Intent Volume works.
Why do AI answers change between checks?
AI answers vary by engine, country, timing and phrasing, and engines often split one prompt into several sub-queries. That is why you track trends across many prompts over time, rather than reading one answer as a fixed ranking.
Is keyword research still relevant?
Yes. Keywords show how your market phrases things, and they are one of the best inputs for generating prompts. Prompt research builds on keyword research rather than replacing it.
How often should I redo prompt research?
Run discovery about once a month and review your tracked prompts every one to two weeks. Revisit your list whenever you launch a product, enter a new market or see a new competitor appear in AI answers.
Your first prompt research session
Prompt research will not give you a perfect log of what customers typed. It gives you something more useful: a list of well-chosen prompts, grounded in real demand, that you can measure and improve. Plan around trends and direction, not exact numbers.
Here is a 45-minute exercise to get started:
- Run your brand, your top three keywords and one competitor URL through the AI Prompt Research tool
- Pull ten conversational queries from Search Console
- Tag everything by intent stage and pick the 25 prompts with the clearest buying intent
- Add them to a brand report in OtterlyAI and check where competitors appear and you don’t
Within a few days, you will know which prompts you already win, which ones competitors own, and which pages to build next.
Ready to find the prompts that matter? Start your free OtterlyAI trial and run your first prompt research in minutes.
