Prompt research is finding the questions your buyers ask ChatGPT and other AI assistants about your category, in their own words, so you can see whether the answers name you. It looks like keyword research, and most tools present it that way, with a volume next to each prompt. But there is no keyword planner for ChatGPT. OpenAI does not publish how often anything is asked, and every “prompt volume” you will see is a vendor’s estimate. So good prompt research starts somewhere else: with the questions buyers already ask you, in places you can see. This post shows where to find them, how to judge the volume figures, and how to turn the questions into a set you can track.
What prompt research is, and how it differs from keyword research
The goal is the same as keyword research: know what your market asks, then be the answer. The material is different. People type more into an assistant. Google said early testers of AI Mode asked “queries that are two to three times the length of traditional searches”[13], and a year later that “the average AI Mode search is triple the length of a traditional Search query”[14]. Longer questions carry context a keyword never did: who is asking, their budget, what they already use.
| Keyword research | Prompt research | |
|---|---|---|
| Typical input | crm for agencies | Which CRM should a 10-person marketing agency use to track client deals? |
| Volume data | Published by Google and Bing through their ads tools | Not published by OpenAI. Vendor estimates only |
| Unit you target | One phrase and its close variants | One need, asked many ways |
| What you check | Your position on a results page | Whether the answer names you, and how often across runs |
| Stability | Rankings move slowly | Answers change between runs of the same prompt |
Is there prompt volume data for ChatGPT?
Not from OpenAI. We checked what OpenAI offers the people with the strongest claim to it: advertisers. Its help center says “advertisers do not have access to your chats, chat history, memories, or personal details”, and “only receive aggregated, non-identifying information about how their ads perform, such as total views or clicks”[1]. Advertisers can give “context hints” about relevant topics, but OpenAI says these “are not exact-match keywords”, and reporting covers “impressions, clicks, spend, click-through rate (CTR), average CPC, average CPM, and conversions”[2]. The ads reporting API splits results by country and product, with no query or keyword dimension[3]. If OpenAI does not give query data to paying advertisers, it is safe to assume nobody outside OpenAI has it first-hand.
What OpenAI has published is the shape of usage, not the questions. A working paper by OpenAI and academic economists found that by July 2025 “18 billion messages were being sent each week by 700 million users”, and that Seeking Information grew “from 14% to 24% of all usage” in a year. The authors describe that category, which includes searching for “products”, as “a very close substitute for web search”[4]. Useful context; no help in choosing your prompts.
Where the tools’ prompt volumes come from
Several tools show a volume next to prompts or topics anyway. Here is how each says it builds the number, in its own words:
| Data behind it | What you get | |
|---|---|---|
| Semrush | "Third-party data on real AI interactions" and machine learning models; prompts from "AI search clickstream data" | Volume for topics, not single prompts |
| Profound | Double opt-in consumer panels, with "probabilistic modeling to correct for demographic and geographic biases" | Prompt volumes, updated weekly |
| Ahrefs | Google search volume of the parent keyword, multiplied by each AI platform's estimated usage | "AI adjusted volume"; prompts from People Also Ask |
| Peec AI | Search trends, AI conversation data and industry signals | A relative demand score from 1 to 5 |
| Evertune | "Real-world usage data from 25 million internet users" | Prompt volume estimates by brand and category |
Sources: Semrush[5], Profound[7], Ahrefs[8], Peec AI[10] and Evertune[11].
The more candid vendors say what this means. Semrush explains that “individual prompts are often too specific and unique to measure directly”, which is why it reports topics[5], and admits “we don’t have years of historical search volume, CPC, or trend data for AI prompts”[6]. Ahrefs calls weighting by Google volume “a modeling choice, not a measured relationship”[9]. One agency puts it bluntly: “There is no search volume for the prompts. We do not know what the users enter specifically.”[12]
Seven places to find the questions buyers really ask
If volume cannot be measured, evidence of the question can. These sources show real questions in real words. Work through them in this order; the first ones are closest to buyers who pay you.
1. Sales calls and demos
The best prompts are the questions prospects ask before they buy. Read call notes or transcripts for three things: how they describe the problem, what else they looked at, and what nearly stopped them. “We need something our clients can log into” becomes a prompt; so does “is X cheaper than Y for a small team?” Ask your sales team which question they hear every week. That is your first prompt.
2. Support tickets and chat logs
Pre-sale chat and onboarding tickets show the jobs people hire your product for, and the words they use. Questions that start “can it…” or “does it work with…” are the must-haves buyers add to a prompt when they ask an assistant to shortlist.
3. Long queries in Google Search Console
Search Console shows the queries your pages already appear for. Traffic from AI Overviews and AI Mode is included there, in the Performance report under the Web search type[15]. The separate Generative AI report shows impressions in AI Overviews and AI Mode by page, country, date and device, but it has no queries view[16]. So the long-query route is the way in. In the Performance report, add a query filter, choose custom regex, and paste one of these. Search Console uses RE2 syntax, matches partially and ignores case by default[17].
^(who|what|which|why|how|when|where|is|are|can|does|do|should|best|top|vs|alternative)\b
^(\S+\s+){7,}\S+$Sort by impressions, export, and read them. Long, specific queries are the closest thing in your own data to how people talk to an assistant.
4. Bing’s grounding queries
This is the one first-party record of what an AI engine searched for when it used your pages. Bing Webmaster Tools’ AI Performance report covers Microsoft Copilot, AI summaries in Bing and some partner integrations. Its grounding queries show “the key phrases the AI used when retrieving content that was referenced in AI-generated answers”; Bing notes the data “represents a sample of overall citation activity”[18]. Since June 2026 the queries are grouped into intents such as “Informational, Commercial, Navigational” and into topics, with a citation share for each[19]. These are the engine’s searches, not the user’s prompts, but they show what the prompts were about. The Bing Webmaster Tools guide covers setup.
5. People Also Ask
The questions Google shows under “People also ask” are real searches, already in question form. Ahrefs builds its own AI prompt set from them: its documentation says “prompts are sourced from People Also Ask questions”[8]. Search your main category terms and note every question a buyer, rather than a student, would ask. The People Also Ask guide explains how the box works.
6. Reddit and forums
Threads where people ask for recommendations are prompts written by buyers, with the context included. They also feed the answers: Semrush found that 12.6% of ChatGPT search answers in its sample contained Reddit links, and that “over half of all cited Reddit content comes from Q&A threads”[20]. Search Reddit for your category plus “recommend”, “alternative” or “vs”, and copy the question in the first post, not the title.
7. The assistants themselves
Ask ChatGPT what someone would ask before choosing a product like yours, and look at the follow-up questions it suggests after an answer. This is useful for wording and for gaps, but it is the model’s guess, not data. Use it to fill in a list built from the six sources above, never as the list.
Turn the questions into a prompt set
The rewrite is where most of the value is. An illustrative example for a CRM aimed at agencies:
| Tracked prompt | Intent | |
|---|---|---|
| Sales call: "we lose track of which client asked for what" | How can a small marketing agency keep track of client requests and deals? Which tools help? | Problem |
| Search Console: "best crm for marketing agency with client portal" | What is the best CRM for a marketing agency that needs a client portal? | Best-of |
| Reddit: "moving off HubSpot, too expensive for 8 people" | What are cheaper alternatives to HubSpot for an 8-person agency? | Alternatives |
| Support ticket: "does it sync with Gmail?" | Which agency CRMs work well with Gmail and Google Calendar? | Best-of with a must-have |
| People Also Ask: "how much does a crm cost per month" | How much should a 10-person agency expect to pay for a CRM each month? | Cost |
Two rules keep the set honest. Most prompts should not contain your brand name: a question that names you measures whether the engine knows you, not whether it would recommend you to a stranger. And the set should change rarely. Our guide to choosing prompts for LLM tracking covers the mix and how many answers a reliable number needs.
Prioritising without volume
Without a trustworthy volume column, rank prompts by evidence and value instead:
- How many sources it appeared in. A question heard on sales calls, seen in Search Console and asked on Reddit is common, whatever a tool says.
- How close it is to a purchase. “Best X for Y”, “alternatives to Z” and “X vs Z” sit near the decision. Definitions sit far from it.
- Whether the answer recommends anyone. Some questions get an explanation and no names. Those matter less for visibility, however often they are asked.
- A rough size check. Google search volume for the underlying topic, or one tool’s prompt volume, is fine as a tiebreaker between two topics. It is not a reason to drop a question your buyers ask every week.
Test the wording before you freeze it
Wording changes answers. “Best CRM for agencies” and a full question with a team size and a must-have can produce different lists, and the same prompt can produce different lists on two runs. So before you lock a prompt in, run it, and a close variant, a few times each. If the two variants name very different brands, buyers asking either way will see different answers: keep both. If they name the same ones, keep the one closer to your buyers’ words. Our post on why ChatGPT gives different answers explains why one run proves little.
From research to tracking
A prompt set is only useful if you keep asking it. CoreCited runs your set every week across ChatGPT, Claude, Perplexity, Gemini, Google AI Overviews and AI Mode and records who each answer names. The Starter plan tracks 40 prompts and the Growth plan 120; see AI visibility tracking and the pricing page for every limit. To turn the answers into a share of voice with honest error bars, see the AI share of voice guide.
Questions people ask
What is prompt research?
Prompt research is finding the questions people ask AI assistants such as ChatGPT, Gemini and Perplexity about your category, in their own words, so you can check whether the answers name you and track them over time. It is to AI answers what keyword research is to search results, with longer, more specific questions and no official volume data.
Is there search volume data for ChatGPT prompts?
Not from OpenAI. Its ads reporting gives advertisers impressions, clicks, spend and conversions, with no query or keyword breakdown, and it says advertisers do not see conversations. Tools that show prompt volume estimate it from consumer panels, clickstream data or Google search volume multiplied by a ratio. Treat their figures as direction, not counts.
How do I do ChatGPT keyword research?
Start from the questions buyers already ask you: sales calls, support tickets, long queries in Search Console, People Also Ask, Reddit threads and Bing's grounding queries. Rewrite each as a buyer would ask an assistant, with who they are and what they need. Then group them by intent and test them in ChatGPT to see which brands each answer names.
What is prompt volume?
A vendor estimate of how often people ask AI assistants about a topic or prompt. Semrush estimates it at topic level from clickstream data and machine learning, Profound from consumer panels, Ahrefs from Google search volume times each platform's estimated usage, and Peec gives a relative score from 1 to 5. None is a count from OpenAI.
How many prompts should I track?
Enough to cover each stage a buyer goes through, usually a few dozen to start, mostly without your brand name. More prompts make the overall rate more reliable than more runs of a few prompts. Keep the set fixed once you start, so month-on-month changes reflect the answers rather than your questions.