Prompt research is how you find the questions buyers actually ask ChatGPT, Perplexity, Bard and Google’s AI features about your category, so you measure and optimize for the right ones. Collect real buyer wording from sales calls, support tickets, reviews and forums, use keyword data to size and sanity-check it, organize the result by funnel stage and persona, then run one baseline to see where you’re missing. Tracking that set over time is a separate job; this post is about choosing it.
How prompts differ from keywords
A keyword is a compressed guess at intent. A prompt is the intent written out, often with the buyer’s situation attached. Someone who types “time tracking software agencies” into Google might ask an assistant: “We’re a 15-person design agency that bills hourly and uses QuickBooks. What time tracking tool should we use, and what will it cost?”
That difference changes how you research:
| Keyword | Prompt | |
|---|---|---|
| Length | A few words | A sentence to a paragraph |
| Context | Implied | Stated: team size, budget, stack, constraints |
| Session | One query, then a click | Several turns, each narrowing the shortlist |
| Demand data | Search volume available | No official volume from AI platforms |
| Winning looks like | A ranked link | Being named, described accurately and cited |
Two consequences follow. First, the unit of research is a question theme with several phrasings, not one exact string, because the same need arrives worded many ways. Second, constraints matter. “Best CRM” and “best CRM for a two-person team that sells over WhatsApp” can produce very different shortlists, and your buyers ask the second kind.
Sources of real buyer questions
Start from what buyers have already said, in their own words.
Sales calls
Discovery calls are the richest source because prospects state their situation and constraints out loud. Search call recordings or notes for question patterns: “how do you compare to,” “do you integrate with,” “what happens when,” “we tried X and.” For a month, have reps ask one extra question: what did you search for or ask an AI assistant before booking this call? The answers are often close to verbatim prompts.
Support tickets and pre-sales chat
Tickets from new customers and pre-sales chat logs show the “can it do X” questions people ask before they commit. These map to validation-stage prompts about integrations, limits, migration, security and pricing edge cases.
Reviews
Your reviews and your competitors’ on G2, Capterra, Trustpilot, app marketplaces or Amazon describe the problem in buyer language and often name the alternatives considered. Negative competitor reviews are especially useful: a line like “switched because it couldn’t handle multiple currencies” becomes a prompt about exactly that.
Forums and communities
Reddit threads, industry Slack and Discord groups, Quora and niche forums are full of “what do you all use for…” questions, the same questions assistants get asked. Note the wording and the constraints people add.
Smaller sources worth a pass
Site search logs, chatbot transcripts, the free-text field on your demo form, win/loss notes and answers to “How did you hear about us?”
For each question, log the exact wording, source, persona, constraints and any competitors named. A spreadsheet is fine. With hundreds of transcripts, an LLM can do the first extraction pass, but read a sample of the raw material yourself; summaries flatten the phrasing you’re trying to keep.
Using keyword data to size and validate prompts
AI platforms don’t publish prompt volumes, so you can’t size a prompt directly. You can size the need behind it. Map each question theme to the keyword cluster that expresses the same need and use that cluster’s search volume as a rough proxy for demand. The methods in keyword research for B2B SaaS apply unchanged; you’re just using the output differently.
Three places to look:
- Search Console for conversational queries. In the Performance report, add a query filter using custom regex.
^(how|what|which|why|is|can|should|does)\bsurfaces questions, and(\S+\s+){7,}surfaces queries of eight or more words, which are often typed the way people talk to an assistant. Search Console leaves out very rare queries for privacy, so the long tail is understated. - People Also Ask and related searches. Search your category terms and record the questions Google shows.
- Question reports in your keyword tool. Most tools filter by question modifiers. Use them to find themes your qualitative sources missed.
Then validate. My rule: keep a question if it shows up in at least two sources, such as a sales call and a Reddit thread, or in one source with obvious revenue weight, like a question that comes up in most enterprise deals. A theme with no measurable search volume can still earn its place in niche B2B, where buyers are few and rarely show up in keyword tools. A theme with big volume that never appears in your calls, tickets or reviews probably belongs to someone else’s buyers.
Don’t paste keywords in as prompts. Rewrite “time tracking software for agencies” as the full question, with the context buyers actually give.
Building a prompt set by funnel stage and persona
Tag every prompt with a stage and a persona. The stage tells you what kind of answer you need to win; the persona tells you whose language to write in.
The table uses a hypothetical time tracking tool for agencies. The shares are my usual starting point, not a benchmark.
| Stage | What the buyer is doing | Example prompt | Share of set |
|---|---|---|---|
| Problem | Describing pain, no category yet | “Our agency keeps underbilling clients. How do other agencies track billable hours?” | 10-15% |
| Category | Learning the options | “What should an agency look for in time tracking software?” | 10-15% |
| Shortlist | Asking for recommendations | “Best time tracking tool for a 20-person agency that invoices through QuickBooks” | 30-35% |
| Comparison | Weighing named options | “Alternatives to [Competitor A] that handle multiple currencies” | 20-25% |
| Validation | Checking a specific vendor | “Does [your brand] integrate with Asana?” | 15-20% |
Weight toward shortlist and comparison, because that’s where an assistant decides which brands to name. Validation prompts are branded, so keep them in a separate group: they test whether answers describe you accurately, not whether you get recommended.
For personas, most B2B sets need two or three: the economic buyer, the day-to-day user and, for technical products, the evaluator who asks about security, APIs and migration. Ecommerce sets usually split by use case and shopper type instead. Write each question in the voice of the person who asks it.
A few more rules:
- 30-60 prompts is enough for most companies. Fewer misses themes; more makes the baseline too slow to run carefully.
- Write two variants for key themes: a short one and a context-rich one. Answers often change once constraints are added.
- Keep your brand out of unbranded prompts, and don’t write prompts only your product can answer.
- Record the same fields for every prompt: ID, text, stage, persona, source, keyword-cluster proxy and priority.
Running a baseline to find gaps
The baseline is a single snapshot to decide what to work on. Run every prompt on the engines your buyers use, typically ChatGPT, Perplexity, Bard, Microsoft Copilot and Google’s experimental AI-generated search results. Use a clean session, logged out where the engine allows it and with personalization such as custom instructions turned off, so your own history doesn’t shape the answer. Run each prompt more than once, because answers vary between runs.
For each run, record:
- Whether your brand is named, and where it sits in the list
- How you’re described, and whether it’s accurate
- Which competitors are named
- Which sources are cited, by URL
- Whether any of your own pages are cited
Read the results by stage and theme. “Absent from every shortlist prompt about multi-currency billing” is a finding. “Missing from prompt 23 on Perplexity” is noise.
Re-running the set on a schedule is monitoring, with its own tooling and cadence. That’s covered in how to track your brand in ChatGPT, Perplexity and Google’s AI results.
Turning gaps into content, PR and review priorities
The cited sources tell you what to fix. Group each gap by what the answers relied on:
| What the baseline shows | Likely gap | Priority action |
|---|---|---|
| Answers cite “best X” roundups you’re not in | Third-party mentions | Outreach to those publishers, partner content, analyst briefings |
| Answers cite competitors’ comparison or use-case pages | Owned content | Comparison, alternatives and use-case pages with specific facts |
| Answers cite review sites or Reddit threads | Reviews and community | A steady review request program; genuine community participation |
| You’re named but described wrongly | Entity consistency | One description across homepage, About page, profiles and listings |
| Answers are generic and cite nothing authoritative | Open field | The most specific, citable page on the topic |
Score each gap on three things: stage weight (shortlist and comparison count most), breadth (how many prompts share the pattern) and effort. Owned-content gaps usually close fastest, so ship those first. Mention and review gaps take longer, so start them in parallel rather than after.
Turning prompt gaps into a content, PR and review plan is the core of my GEO and AI search visibility work.
Refreshing the prompt set
Buyer questions drift as your market changes. Review the set every quarter:
- Pull new questions from the last quarter’s calls, tickets and reviews
- Add prompts for any new feature, market, pricing change or competitor
- Retire prompts that never come up in buyer conversations
- Check that stage and persona shares still match how you sell
- Keep a fixed core set for comparison and rotate a small exploratory group
My rule of thumb is to keep around 80% of prompts fixed so trends stay comparable, and rotate the rest to test new themes. When a rotating prompt keeps appearing in buyer conversations, promote it to the core.
Get it built
If you want to know which questions buyers ask AI about your category, and where you’re missing from the answers, start with a Growth Audit. It’s $1,500 fixed and credited if we continue. See pricing or get in touch.