Keyword tools can't show what buyers ask ChatGPT. Mine sales calls, tickets, Search Console and Reddit, sort by intent, and pick the prompts worth tracking.
Muhammet İLBAŞ 10 min readPrompt research means finding the questions your buyers actually type into ChatGPT, Gemini, Claude and Perplexity when they're choosing a product. The best sources are the ones keyword tools can't see: sales-call notes, support tickets and chat logs, long question-shaped queries in Google Search Console, People Also Ask boxes, and the threads where buyers ask each other on Reddit and in communities. Collect the phrasing, rewrite it the way people talk to an assistant, sort it into intent buckets - discovery, problem-first, comparison, alternatives, validation - and track a focused set of a few dozen prompts, each with a couple of variants.
It's the step most AI visibility programs skip. Teams paste their SEO keyword list into a tracker and wonder why the results feel unrelated to the deals they win and lose. The prompts you track define what your visibility numbers mean, so they deserve the same care as the numbers. Here's how to find them.
Because people don't talk to assistants the way they type into a search box. Semrush analysed 80 million clickstream records from the second half of 2024 and found the average ChatGPT prompt was 23 words long, against 4.2 words for ChatGPT Search. A search query is a few nouns; a prompt is a situation. "crm small agency" becomes "we're a 12-person agency moving off spreadsheets, which CRM won't need an admin and works with Gmail?"
Those extra words carry the things that decide which brands come back: team size, budget, the tools someone already uses, what went wrong with the last product. Two buyers in the same category can get completely different shortlists because they described different situations. A keyword list throws all of that away.
There's also no public demand data. No AI vendor publishes prompt volumes the way Google Keyword Planner publishes search volumes, so any "prompt volume" number you see is an estimate from a panel - often clickstream data like the Semrush study above. Useful for direction, not for precision. And what people do with assistants is broad: OpenAI's own research found that practical guidance, seeking information and writing account for nearly 80% of all ChatGPT conversations. Buying questions are a slice of the information-seeking part, and they rarely look like keywords.
Finally, the engines expand what they're asked. Google says AI Overviews and AI Mode may use "query fan-out" - issuing multiple related searches across subtopics - to build one answer. A single prompt can trigger many searches behind the scenes, so the prompt, not the keyword, is the unit worth tracking.
No single source gives you a complete list. Pull from several, weighted toward the ones closest to an actual purchase.
This is the highest-signal source, and nobody else has it. Read the notes or transcripts from your last fifteen or twenty discovery calls and write down three things: how the prospect described the problem in their first few sentences, which other products they said they were looking at, and the question that nearly stopped the deal. The first becomes a problem-first prompt, the second your comparison and alternatives prompts, the third a validation prompt. Keep their words, not your positioning - prospects don't say "revenue intelligence platform", they say "something that tells me which deals are going to slip".
Pre-purchase questions in your chat widget and support inbox are prompts people asked you because they couldn't find the answer elsewhere. "Does it integrate with HubSpot?", "Can I use it for a team of three?", "Is there a free plan?" - people ask assistants exactly these questions. They make good brand-fact prompts, which check whether engines describe you correctly, and they often show the constraints buyers filter on.
Your Search Console performance report holds real queries people typed, and the long ones read a lot like prompts. Filter them with a regex: in the Performance report, add a Query filter, choose Custom (regex), and paste one of these. The filter uses RE2 syntax:
# question-shaped queries
^(what|which|how|why|is|are|can|does|do|should|best|top)\b
# long, conversational queries (7+ words)
^(\S+\s+){6,}\S+$
# comparison and alternatives intent
\b(vs|versus|alternative|alternatives|compared|instead of)\bSort by impressions and read the top few hundred. Google reports AI Overviews and AI Mode traffic inside the same "Web" search type, so these lists already reflect some of the ways people search in Google's own AI surfaces.
The People Also Ask box, autocomplete and the related searches at the foot of a results page are Google's summary of the follow-up questions people ask about a topic. Search your category's head terms in a clean browser and copy the questions. They're phrased as natural questions already, and they show what buyers worry about after the first query: price, setup time, whether the tool fits a particular team.
"What do you use for X?" threads on Reddit, Slack and Discord communities, and niche forums are the closest public record of how buyers ask peers for recommendations - which is how they ask assistants too. Read the opening posts more than the replies; the opening post is the prompt. Those threads also tend to be cited in AI answers, so you learn the question and see one of the sources at the same time.
G2, Capterra and their equivalents in your category ask reviewers what problem they were solving and why they chose the product. Those answers are buyer language at scale, and the comparison pages on those sites show which rivals buyers put side by side. If review sites feed AI citations in your category, they're worth reading twice.
Asking ChatGPT "what questions would a small agency ask when choosing a CRM?" produces a tidy list in seconds. Use it to fill gaps in the buckets, never as evidence that people ask those things. The model is guessing at demand, and a list that comes entirely from a model tends to sound like a model.
A raw list of a few hundred questions isn't a tracking plan. Deduplicate it - many questions are the same need in different words - and sort what's left by what the buyer is trying to do. These are the buckets we use:
| Bucket | What the buyer is doing | Typical shape |
|---|---|---|
| Discovery | Building a first shortlist | "best [category] for [buyer]", "what do [buyers] use for [job]?" |
| Problem-first | Describing a pain, not a category | "how do I stop [problem]?", "is there a tool that…" |
| Comparison | Choosing between named options | "[A] vs [B] for [use case]" |
| Alternatives | Leaving or avoiding an incumbent | "[competitor] alternatives", "cheaper than [competitor]" |
| Validation | Close to deciding, checking risk | "is [category] worth it for…", "what to check before buying…" |
| Brand facts | Checking you specifically | "does [you] integrate with [tool]?", "[you] pricing" |
The first five buckets measure visibility: whether you're named when the buyer hasn't mentioned you. Leave your brand out of those prompts - naming yourself primes the answer and tells you nothing. The brand-facts bucket measures accuracy: whether what the engines say about you is right. Track it separately, and if it's wrong, follow the correction playbook.
Weight the list toward the bottom of the funnel. Discovery prompts are the most competitive and the most variable; comparison, alternatives and validation prompts are where buyers who are about to pay ask their final questions, and where a missing mention costs the most.
Write them the way the buyer would say them, with the context they'd give. A few rules hold up well:
Fill in your category and buyer. You get draft prompts by intent bucket - edit them against the real phrasing in your sales notes before you track anything.
Fewer than you think. For one product in one market, 20 to 40 well-chosen prompts covering all the buckets is a solid core - enough to see patterns, few enough that every prompt is one you'd act on. That's our recommendation, not a law. Add more when you have more product lines, more markets or distinct buyer personas, each of which is in effect a separate list.
The reason not to go bigger straight away is cost and attention. Every prompt is multiplied by the engines you track, the samples per run and the cadence. AI answers also change from run to run, so the same prompt sampled repeatedly tells you more than hundreds of prompts sampled once. A smaller list tracked properly beats a large list nobody reads.
Treat a buyer need as the unit and the phrasings as samples of it. If you track "best CRM for small agencies" and "we're a small agency, which CRM should we use?" as two separate lines, a difference between them looks like news. Grouped under one topic, the same results become a more honest number: how often you're named when someone asks this, however they word it. Report at the topic level and look at the individual phrasings when something moves.
Keep a frozen core. When you reword prompts every month, your trend line measures your edits as much as the engines. Keep most of the list fixed for at least a quarter, and add new phrasings alongside the old ones rather than replacing them. That way your visibility rate and share of voice stay comparable over time.
Quarterly is a sensible rhythm, plus whenever something big happens - a launch, a new competitor, a pricing change. Each review, retire prompts that no longer match how deals start, add new phrasing from recent sales calls and tickets, and check that the competitor names in your comparison prompts are still the ones buyers mention. The list should read like a summary of your last quarter's sales conversations. If it doesn't, it's drifted.
Zene tracks the questions you choose across ChatGPT, Gemini, Claude, Perplexity and Grok. On paid plans it also suggests questions from real search demand - People Also Ask, related searches and autocomplete for your category - and can ask a question as a specific persona ("I run a small agency…"), so the buyer's context is part of the prompt. Your question quota is one pool shared by all your brands. The sources in this article remain the core of it, though. No tool knows what your prospects said on last Tuesday's demo call.
Got a first list? See where you stand on it. Run a free audit on ChatGPT and Gemini in about 90 seconds, or try a single question with the AI visibility checker. When you're ready to track the full set, compare plans.
Prompt research is finding the questions real buyers ask AI assistants like ChatGPT, Gemini, Claude and Perplexity when choosing a product, so you can track whether those answers name you. It replaces the keyword list for AI visibility work: prompts are longer, conversational and full of context such as team size, budget and existing tools.
No AI vendor publishes prompt data, so triangulate. The best sources are your own sales-call notes and demo transcripts, support tickets and pre-sales chats, long question-shaped queries in Google Search Console (filter with a regex), People Also Ask and autocomplete, Reddit and community threads where buyers ask for recommendations, and review sites in your category.
For one product in one market, 20 to 40 well-chosen prompts spread across intent buckets is a solid core. Add more for extra product lines, markets or personas. Because AI answers vary from run to run, sampling a focused set repeatedly tells you more than tracking hundreds of prompts once.
Not the visibility prompts. Discovery, comparison, alternatives and validation prompts should leave your brand out, because naming yourself primes the answer and hides whether the engine recommends you unprompted. Track brand-named prompts separately, to check that what the engines say about you is accurate.
Not from the AI vendors. Any prompt-volume figure comes from estimates based on panels or clickstream data, which are useful for direction but not precise. A Semrush clickstream study of the second half of 2024 found the average ChatGPT prompt was 23 words long, which is why keyword volumes map poorly onto prompts.
GA4's new AI Assistant channel misses some assistants and app clicks land in Direct. The custom channel group, the regex, and what GA4 still can't see.
Muhammet İLBAŞ 9 minHow SaaS products get named when buyers ask ChatGPT, Claude or Perplexity for the best software: category clarity, comparison pages, G2, crawlable pricing.
Muhammet İLBAŞ 10 min"Best meditation apps", "app to plan a trip" - questions that used to be store searches are now asked in ChatGPT, and the answer is a shortlist of three to five names. Here's how assistants actually pick those names - listing signals, reputation bars, citable sources - and the six-step playbook for getting your app into the shortlist, with a free checker that grades where you stand today.
Muhammet İLBAŞ 12 minPut this guide into practice - get your free visibility score in minutes.
