Every day, people type over 2.5 billion prompts into ChatGPT alone. Google says the average AI Mode query is now three times longer than a traditional search. And according to G2, half of B2B software buyers now start their research on ChatGPT more often than on Google.
Here’s the problem: all of that demand is invisible. Google Keyword Planner can’t tell you how many people asked “what’s the best project management tool for a remote design team?” this month. No search volume tool can. The questions your customers ask AI are longer, more specific, and full of context — and almost nobody publishes data about them.
That’s what prompt research is for. It’s the AI-era version of keyword research: discovering the exact questions your potential customers ask ChatGPT, Perplexity, Gemini, Claude, and Google AI Mode — so you can show up in the answers, and track whether you do.
This guide covers 12 methods to find those prompts, ranked from the most reliable (real customer language) to the most scalable (AI-native tools and prompt databases), plus how to turn a messy list of 200 candidates into a focused tracking set.
What Is Prompt Research?
Prompt research is the process of discovering the exact questions your target audience types into AI assistants — so you can create content that gets cited in the answers and monitor whether your brand appears.
It sounds like keyword research, but the raw material is fundamentally different:
| Keyword research | Prompt research | |
|---|---|---|
| Typical query | “crm software small business” | “I run a 10-person agency and I’m drowning in client follow-ups — what should I use?” |
| Length | 3–4 words | 10–25+ words, full sentences |
| Context | Implied | Explicit (role, budget, constraints, use case) |
| Volume data | Public, reliable | No public data — only estimates |
| Where demand lives | Google’s index | Private chat logs (OpenAI, Google, Anthropic don’t share) |
| Success metric | Rankings + clicks | Mentions + citations in the answer itself |
| Results stability | Relatively stable | Probabilistic — same prompt can give different answers |
That example on the right is real, by the way — Profound uses it to illustrate how differently people talk to AI versus Google. Someone who types “best CRM” into Google will write an entire paragraph of context into ChatGPT. The brands AI recommends for each query overlap only 8–12% of the time. If you’re only doing keyword research, you’re optimizing for a shrinking slice of how buyers actually discover products.
The 4 Types of Prompts Worth Finding
Before the methods, know what you’re looking for. Nearly every commercially valuable AI prompt falls into one of four buckets:
| Type | What the user is doing | Example (for a CRM) | Funnel stage |
|---|---|---|---|
| Problem prompts | Describing a pain, no solution in mind | “How do I stop losing track of client follow-ups?” | Awareness |
| Category prompts | Asking for options in a solution category | “What are the best CRMs for small agencies?” | Consideration |
| Competitor prompts | Comparing or seeking alternatives | “HubSpot vs Pipedrive for a 5-person team?” / “HubSpot alternatives that aren’t so expensive” | Decision |
| Brand prompts | Asking about you directly | “Is Pipedrive good? What’s Pipedrive’s pricing?” | Decision / Post-sale |
Most brands obsess over competitor and brand prompts and ignore problem prompts. That’s backwards — problem prompts are where buyers form their shortlist. If you’re absent there, you often never make it to the comparison stage. A healthy tracking set covers all four, weighted toward where your gaps are.
Tier 1: Real Customer Language (Most Reliable)
The best prompts aren’t invented — they’re overheard. Start here.
Method 1: Survey your customers directly
The simplest method and the one almost nobody does: ask your customers what they ask AI.
Send a 3-question survey to your email list or best customers:
- “Have you used ChatGPT, Perplexity, Gemini, or Claude to research [your product category]?”
- “If yes — can you paste the exact question(s) you asked? Even roughly remembered helps.”
- “What did the AI recommend? Did you follow its suggestion?”
Offer a small incentive (gift card, account credit, swag). Even 20–30 responses will surface phrasing you’d never brainstorm in a doc — real buyers describe problems in ways marketing teams never guess.
Method 2: Mine sales calls and support tickets
Your sales and support conversations are a transcript of how buyers describe their problems in natural language — which is exactly how they prompt AI.
- Sales call recordings (Gong, Chorus, Zoom recordings): search for “how do I”, “we’re struggling with”, “what happens if”, “is it better than”
- Support tickets and chat logs (Intercom, Zendesk): recurring questions are recurring prompts
- Onboarding calls: the problems new customers describe at signup are the awareness-stage prompts that brought them in
As Convert’s guide puts it: look for prompts starting with “How do I…”, “What happens if…”, and “Is this better than…” — those map directly to top, middle, and bottom-funnel AI queries.
Method 3: Your own site search and chatbot logs
If your site has a search bar or a support chatbot, you already collect prompt-shaped data every day. Export the last 90 days of internal search queries. Full-sentence queries (“how do I integrate X with Slack”) are prompts people will also ask AI. Short ones are keywords you can expand into prompts (see Method 5).
Tier 2: Search Data You Already Own
Google data doesn’t tell you what people ask AI — but it’s the best behavioral proxy you have, and it’s free.
Method 4: Filter Google Search Console for question queries
GSC is a goldmine of real questions people already ask about your niche. The trick is filtering for them with regex.
In GSC → Performance → Search results → + New → Query → Custom (regex), paste this to isolate question queries:
^(who|what|what's|where|when|why|how|can|could|is|are|does|do|should|which|best)(\s|$)
And this one to surface the long, conversational, AI-style queries (10+ words) that increasingly show up as Google folds AI Mode data into Search Console:
^(\b\w+\b\s+){9,}\b\w+\b$
Sort by impressions, export, and you’ve got a list of real demand phrased in natural language. Pay special attention to queries where you get impressions but few clicks — those are often queries where an AI Overview or featured answer is absorbing the click, and exactly where AI visibility matters most.
Method 5: Convert your existing keyword list into prompts
Your SEO keyword list isn’t obsolete — it’s raw material. Take your top commercial keywords and transform them the way a real person would ask an AI:
| Keyword | Becomes this prompt |
|---|---|
| “crm software pricing” | “How much does CRM software cost for a 10-person team? Any hidden fees?” |
| “email marketing tools” | “What are the best email marketing tools for a small ecommerce store in 2026?” |
| “mailchimp alternatives” | “I’m outgrowing Mailchimp — what should I switch to that has better automation?” |
| “how to track employee time” | “What’s the easiest way for a remote team of 8 to track time without being annoying about it?” |
The transformation rules: add a persona (who is asking), add a constraint (budget, team size, platform), and phrase it as a full question. Keyword tools help scale this — in Ahrefs or Semrush, filter any keyword report by the “Questions” modifier to get the question-shaped starting points.
Method 6: People Also Ask, AnswerThePublic, and AlsoAsked
Google’s People Also Ask boxes are machine-generated clusters of related questions — search your category terms and expand the PAA boxes (each click loads more). AnswerThePublic and AlsoAsked automate this, giving you question maps around any seed term. These are Google-sourced, not AI-sourced — but they’re phrased as natural questions, which makes them strong prompt candidates and great raw material for Methods 10 and 11.
Tier 3: Community Mining
AI engines lean heavily on community content when forming recommendations — Reddit is the single most-cited source across AI platforms in several citation studies. The same threads AI learns from tell you what people ask.
Method 7: Reddit
Go to the subreddits where your buyers live and search for:
- “what do you use for”
- “recommend” / “recommendations”
- “alternative to”
- “worth it?”
- “how do you handle / deal with”
Sort by relevance or by new. Buying-advice threads (“What’s the best X for Y?”) contain the exact phrasing people use — and since Reddit threads heavily influence AI answers, prompts that appear as thread titles are very likely prompts that matter. Quora, niche forums, Facebook groups, Slack/Discord communities, and even YouTube comment sections work the same way.
Method 8: Review sites (G2, Capterra, Trustpilot)
Reviews are unfiltered buyer language at scale. Two mining tactics:
- Your competitors’ 3-star reviews: phrases like “I wish it could…”, “great but doesn’t do…”, “switched because…” reveal the exact needs behind alternative-seeking prompts (“X alternatives that can do Y”)
- “Pros/Cons” sections: the use cases reviewers praise are category prompts (“best tool for [use case]”)
This works beyond SaaS — Amazon reviews, Trustpilot, and app store reviews all contain the same raw material.
Tier 4: AI-Native Methods (Most Scalable)
Now the fun part: using AI itself to find what people ask AI.
Method 9: Perplexity’s related questions
Perplexity appends a set of follow-up questions to almost every answer — and those are generated from real query patterns.
- Go to perplexity.ai
- Ask one of your seed questions (e.g., from Method 4)
- Scroll to the related/follow-up questions at the bottom of the answer
- Click the most relevant ones and repeat — two or three levels deep
In ten minutes you’ll have 15–20 prompts phrased the way an AI search engine expects users to ask. ChatGPT suggests follow-ups inside conversations too — they’re less visible, but worth capturing.
Method 10: Ask the AI to generate your prompt list (meta-prompt)
LLMs are surprisingly good at simulating what your customers would ask — as long as you give them structure. Here’s a copy-paste template (adapted from the one Otterly published in their prompt research guide):
Act as an experienced brand strategist. I want to monitor how visible my
brand is in AI search. Generate 25 prompts that potential customers would
realistically type into ChatGPT before buying a product like mine.
Brand: [your brand]
What it does: [2-sentence description]
Industry: [industry]
Target customer: [who they are, company size, role]
Country: [market]
Generate prompts across these 4 intent stages:
- Problem-based (5): they describe the pain, no product category mentioned
- Category-based (8): they ask for the best/top tools in the category
- Competitor-based (6): they compare or seek alternatives to [competitor 1], [competitor 2]
- Brand-based (6): they ask about my brand, pricing, reviews, alternatives
Write each prompt the way a real person types it: conversational, specific,
with context like team size, budget, or use case. No keyword-speak.
Treat the output as candidates, not ground truth — these are simulated prompts, so validate the best ones by actually running them (Method 12) and cross-checking against real language from Tiers 1–3.
Method 11: Query fan-out tools (see what AI searches behind the scenes)
When someone asks ChatGPT or Perplexity a question, the engine doesn’t run one search — it silently breaks the prompt into several sub-queries (“query fan-out”), searches each one, and synthesizes the answer. Peec AI’s analysis of query fan-outs found that ChatGPT runs ~2.1 hidden searches per prompt, Grok runs ~6.8, and Perplexity ~1.4 — and that ChatGPT most often injects the word “best” into its hidden searches, even when the user never typed it.
Fan-out data tells you two things: the sub-topics an AI considers necessary to answer your prompt, and the keyword-shaped version of the question (useful for content optimization). Ways to see it:
- Qforia (by iPullRank) — free tool that simulates fan-out for Google AI Overviews and AI Mode
- LLMrefs Query Fan-Out Generator — free, simulates multiple AI engines
- SE Ranking’s ChatGPT Fan-Out Query Extractor — free, extracts the actual queries ChatGPT generates
- Manual method: open Chrome DevTools → Network tab while ChatGPT answers, and watch the search queries fire in the requests (technical, but shows the real thing)
This matters beyond research: since ChatGPT combines results from its fan-out searches (using Reciprocal Rank Fusion), content that ranks for several of the sub-queries is far more likely to be cited than content that ranks for just one.
Method 12: Prompt databases and AI visibility tools
Finally, the category that didn’t exist two years ago: tools with their own prompt datasets.
- Profound’s Prompt Volumes draws on a dataset of 1.3B+ real user conversations to estimate what people ask, broken down by platform and intent
- Ahrefs Brand Radar maintains 372M+ monthly prompts derived from People Also Ask data and its 110B-keyword index, expanded via semantic fan-out
- Semrush’s Prompt Research surfaces related prompts and intent variants for any seed topic
- RankBits takes a different angle: paste your domain and it auto-generates the realistic AI-search prompts for your specific brand, category, and competitors — then immediately shows you the answers across 13 engines, so research and measurement happen in one step. Run a free scan →
All 12 Methods at a Glance
| # | Method | Data type | Cost | Best for |
|---|---|---|---|---|
| 1 | Customer survey | Real prompts | Incentive cost | Exact phrasing from actual buyers |
| 2 | Sales calls & support tickets | Real language | Free (existing) | Problem prompts, buyer vocabulary |
| 3 | Site search / chatbot logs | Real queries | Free (existing) | Product-specific questions |
| 4 | GSC regex filters | Real Google queries | Free | Question demand you already touch |
| 5 | Keyword → prompt conversion | Proxy | Free / tool cost | Scaling an existing SEO list |
| 6 | PAA / AnswerThePublic / AlsoAsked | Proxy (Google) | Free–cheap | Question clusters around a topic |
| 7 | Reddit & communities | Real language | Free | Category + competitor prompts |
| 8 | Review sites | Real language | Free | Competitor/alternative prompts |
| 9 | Perplexity related questions | AI-suggested | Free | Fast expansion of a seed prompt |
| 10 | Meta-prompt generation | Simulated | Free | Structured first draft in minutes |
| 11 | Query fan-out tools | AI behavior | Free | Sub-queries and content angles |
| 12 | Prompt databases / tools | Estimated real data | Varies | Volume estimates, validation, tracking |
The Uncomfortable Truth About Prompt Volume
You will notice one thing missing from this guide: a “prompt volume” number you can trust. That’s deliberate, and you should be skeptical of anyone who sells you one.
No AI platform publishes query data. The tools that show “prompt volume” estimate it from opt-in user panels and then extrapolate — sometimes aggressively. One marketer’s comparison of a bottom-funnel SaaS term found estimates ranging from 40 monthly searches (Ahrefs) to 250,800 (panel-based) for the same query. Ahrefs’ own analysis suggests ChatGPT’s search-like interactions are roughly 12% of Google’s volume — meaningful, but not the firehose some dashboards imply.
The practical takeaway:
- Use volume estimates directionally, to compare topics — never as precise counts
- Intent beats volume. A prompt asked by 200 people/month that names your exact use case is worth more than a generic one asked 20,000 times
- You can’t go wrong tracking the prompts your customers actually told you they ask (Tier 1), regardless of what any database says
From 200 Candidates to a Tracking Set You Can Manage
Research produces volume; tracking requires focus. AI answers are probabilistic, so SE Ranking recommends starting with 20–40 prompts run across 2–3 engines for at least 30 days. Here’s how to filter:
1. Deduplicate by “answer space.” Two prompts belong together if the AI would construct essentially the same answer for both. “What are the best CRMs for freelancers?” and “Top CRM tools for solo consultants?” are one answer space — keep one. “Best CRM for freelancers” and “CRM for agencies with 20+ clients” are different answer spaces — keep both.
2. Run each candidate through this 5-question filter:
- Would this prompt reliably make the AI name brands/tools? (Purely informational prompts often produce zero brand mentions — they’re content ideas, not visibility prompts.)
- Is it meaningfully different from prompts already in the set?
- Does it map to a real business outcome (a product, use case, or funnel stage)?
- Would appearing in this answer plausibly influence a buyer?
- Could you realistically win it? (Check who currently gets cited — if it’s all Wikipedia, government sites, and G2, deprioritize. If it’s listicles, Reddit threads, and mid-authority blogs, you have a path.)
3. Balance the set. Cover all four prompt types from earlier, weighted toward your gap. Most brands land around 40% category, 25% competitor, 25% problem, 10% brand.
4. Set a review cadence. Semrush advises a quarterly review: add prompts when you launch products, enter categories, or hear new customer language; remove prompts that produced zero mention/citation signal for 60+ days or that duplicate a better-performing prompt’s intent.
6 Prompt Research Mistakes to Avoid
- Tracking only branded prompts. “Is [my brand] good?” tells you about reputation, not discovery. The prompts that grow your business are the ones where the user hasn’t heard of you yet.
- Tracking 10 phrasings of the same question. It inflates your prompt count while measuring one answer space 10 times. Deduplicate ruthlessly.
- Chasing volume over intent. There is no reliable prompt volume data (see above). High-intent beats high-volume every time.
- Brainstorming in a doc and stopping there. Simulated prompts drift from real ones fast. Always anchor at least part of your set in Tier 1–3 sources.
- Checking one engine and calling it done. Only ~11% of domains get cited by both ChatGPT and Perplexity — the engines disagree wildly. A prompt where you win on ChatGPT may be a total loss on Gemini.
- Set and forget. Prompts drift, competitors publish, models update. One Reddit practitioner found a thread that drove their Perplexity citations dropped off in about six weeks. Prompt research is a quarterly habit, not a one-time project.
From Research to Measurement
Prompt research is only half the job — the other half is knowing whether AI actually recommends you for those prompts. That’s the part that’s hard to do by hand: 25 prompts × 13 engines × repeat runs (because answers vary) quickly becomes hundreds of manual checks a month.
RankBits closes the loop: enter your domain and it generates a realistic prompt set for your brand automatically — then you can edit it with the prompts you found using the methods in this guide. It runs them across ChatGPT, Perplexity, Gemini, Claude, Google AI Overviews, AI Mode, Copilot, Grok, DeepSeek, and the source engines, and tracks your AI visibility score, mentions, citations, and competitor share of voice over time.
If you haven’t checked your baseline yet, start with our 5-minute manual audit — then automate it.
FAQ
How many prompts should I track? Start with 20–40. Fewer than 15 and single-answer randomness dominates your data; more than 50 and you’re paying (in money or attention) for noise. Grow the set as you expand into new categories or find gaps.
Is there real “prompt volume” data like keyword volume? Not reliably. AI platforms don’t publish query logs; third-party numbers come from extrapolated panels and can differ by 25x or more for the same query. Treat any prompt volume figure as directional, and prioritize intent and business relevance instead.
Do I need different prompts for ChatGPT vs Perplexity vs Gemini? The same prompt set works across engines — that’s the point of tracking (you’re measuring where each engine sends people for the same demand). What differs is how engines answer, which is why one prompt can have completely different winners per engine.
Can I just reuse my SEO keyword list? Partially. Convert keywords into full questions with persona and context (Method 5) — a bare keyword typed into ChatGPT doesn’t reflect how people actually use it. Only about 8–12% of ChatGPT’s answers overlap with Google’s results, so keyword rankings alone won’t tell you if you’re visible in AI.
How often should I redo prompt research? Review quarterly. Add prompts when you launch products or hear new customer language; prune ones with 60+ days of no signal. And re-validate a handful of core prompts manually every month — engine behavior changes faster than keyword rankings ever did.
What’s the fastest way to start right now? Ten minutes: run the GSC question regex (Method 4), paste your domain into RankBits to see its auto-generated prompt set, and ask Perplexity one of your category questions to harvest its related questions (Method 9). Merge, dedupe by answer space, and you have a v1 tracking list today.
Sources
- Reuters / DemandSage — ChatGPT Statistics 2026 (900M weekly users, 2.5B daily prompts)
- Search Engine Land — “Queries in AI Mode are three times longer than traditional searches” (Alphabet Q4 2025 earnings call)
- Google — How AI Mode is changing the way people search (May 2026)
- G2 — G2 and Reddit partner (B2B buyer AI research survey, Oct 2025)
- Otterly — The Ultimate Guide to Prompt Research: 17 Ways to Find Prompts for ChatGPT & AI Search
- SE Ranking — How to Choose Prompts to Track for AI Visibility (2026)
- Semrush — How to decide which AI search prompts to track
- Profound — How to Design Prompts for AI Visibility Tracking
- Ahrefs — How to Choose the Best Prompts to Monitor Your AI Search Visibility
- Ahrefs — Brand Radar Methodology
- Profound — Comparing Profound vs. Ahrefs for AEO & AI Visibility
- Peec AI — Patterns we see in ChatGPT query fanouts
- Goodie — Query Fan-Out: A Misunderstood Concept in AEO & SEO
- SE Ranking — ChatGPT Fan-Out Query Extractor (free tool)
- LLMrefs — Free Query Fan-Out Generator
- Convert — How to Find Out What Your Buyers Are Asking the LLMs
- Omniscient Digital — How I Created the Perfect Prompt Set for AI Visibility Research
- Steve Toth — Unreliable AI Prompt Volume Metrics (LinkedIn)
- iBeam Consulting — Find Question Keywords in Google Search Console with Regex
