ChatGPT prompts for customer research

ChatGPT prompts for customer research

Prompts that synthesize real customer evidence you provide — interviews, tickets, reviews, threads — instead of asking a model to imagine your customer.

Start here: do not ask ChatGPT to imagine your customer

The single most common way customer research with ChatGPT goes wrong is asking the model to invent a persona, guess what your buyers think, or summarize a market it has never observed. A plain ChatGPT answer comes from training data with no sources. When you ask it to describe your ideal customer from nothing, it produces confident fiction — a tidy, plausible persona that nobody in your market actually matches.

Synthetic personas built from thin air are a known trap. They feel like research because they are written in the voice of research, but there is no evidence underneath them. Decisions made on that basis are decisions made on a guess wearing a lab coat.

The prompts on this page are built on the opposite principle. Every one of them either reasons over real customer material you paste in — interview transcripts, support tickets, reviews, survey verbatims, real community threads — or helps you design research to go collect that material. The model becomes a synthesis engine for evidence you already have, not a generator of evidence you wish you had.

What counts as real input

Before any prompt below is useful, you need raw voice-of-customer material. The trustworthy sources are the ones where customers spoke in their own words for their own reasons: recorded and transcribed interviews, discovery-call notes, support and ticket threads, app-store and product reviews, open-ended survey responses, and unprompted discussion in the communities where your buyers actually hang out.

That last category is the hardest to gather by hand and often the most honest, because people post there to help each other rather than to answer your questions. A grounded Reddit research tool like rawneed gives you sourced raw material — real threads classified by pain, willingness to pay, sentiment, and tools mentioned, each linked back to its original post — which you can paste straight into the synthesis prompts here. The principle holds either way: feed the model real language, and make it cite which input each conclusion came from.

The prompt library

PromptWhat it doesHow to ground it
Below are [N] customer interview transcripts. Identify the recurring themes across them. For each theme, give it a short name, describe it in one sentence, and quote two verbatim lines from different interviews that support it. Do not include any theme that appears in only one transcript. Transcripts: [paste]Clusters interviews into evidence-backed themes with quotes, suppressing one-off noise.Paste full transcripts. The two-quote and cross-interview rules force grounding.
Here are [N] support tickets. Extract the distinct pain points customers describe. Group near-duplicates. For each pain point, count how many tickets mention it and quote one representative line. Rank by frequency. Tickets: [paste]Turns raw tickets into a ranked, counted pain-point list.Paste real ticket bodies. Ask for counts so you can sanity-check the tally against the source.
From the reviews below, extract the jobs the customer was trying to get done — the underlying goal, not the feature they mention. Phrase each as: when [situation], I want to [motivation], so I can [outcome]. Quote the line each job is drawn from. Reviews: [paste]Derives jobs-to-be-done framed as real situations rather than feature requests.Paste reviews or call notes. The quote requirement ties each job to an actual sentence.
Using only the material below, draft a 12-question customer interview guide. Order questions from broad context to specific behavior. Avoid leading questions and questions answerable yes or no. Note which existing finding each question is trying to confirm or challenge. Material: [paste]Designs a non-leading interview guide grounded in what you already know.Paste prior findings or a rough hypothesis so questions probe real gaps, not generic ones.
Write a discovery-call guide for selling [product] to [role]. For each section, include one open question that surfaces current pain and one that surfaces what they have already tried. Base the pain areas on the verbatims below, not on assumptions. Verbatims: [paste]Produces a discovery-call structure anchored to observed pain.Paste real customer verbatims describing the problem area.
Build an evidence-based persona from the material below. Include only attributes you can support with a quote or a clear pattern across multiple sources. For each attribute, cite the source. Mark anything you cannot support as Unknown — do not fill gaps with assumptions. Material: [paste]Constructs a persona strictly from evidence, leaving genuine gaps visible.Paste a mix of interviews, tickets, and reviews. The Unknown rule blocks fabrication.
From the text below, pull the exact words and phrases customers use to describe their problem, their goal, and their current workaround. Keep their wording — do not paraphrase into marketing language. Group the phrases under those three headings. Text: [paste]Harvests voice-of-customer language for copy in the customer's own words.Paste raw threads, reviews, or transcripts. Forbidding paraphrase preserves real phrasing.
Here is a list of feature requests pulled from customer messages. Cluster them by the underlying need, not the surface wording. Name each cluster, list which requests fall under it, and quote one request per cluster. Flag any request that does not fit a cluster. Requests: [paste]Clusters feature requests by intent so the roadmap reflects needs, not phrasing.Paste verbatim requests. The single-quote-per-cluster rule keeps clusters traceable.
Below are survey open-ended responses. Code each response into one or more categories that emerge from the data itself — do not impose a preset list. Report each category, its response count, and two example verbatims. Responses: [paste]Performs emergent open coding of free-text survey answers.Paste raw open-ended answers only — not the multiple-choice fields.
From the threads below, identify where customers express willingness to pay, hesitation about price, or mention what they currently pay for an alternative. Quote each instance and label it pays / would-pay / price-objection. Do not infer willingness to pay where it is not stated. Threads: [paste]Surfaces real pricing and willingness-to-pay signals with their exact quotes.Paste real community threads or call transcripts. The no-inference rule prevents invented demand.
Compare the customer language below against my current landing-page copy (pasted after). List places where my copy uses internal or vendor language that customers never use, and suggest the customer phrasing to replace it. Customer language: [paste]. My copy: [paste]Audits your copy against actual customer vocabulary and flags mismatches.Paste harvested customer phrases plus your live copy. Both inputs must be real.
Read the material below and list the three strongest counter-signals — evidence that contradicts the assumption that [your assumption]. Quote each. If there is no contradicting evidence, say so plainly rather than inventing it. Material: [paste]Stress-tests a belief by hunting for disconfirming evidence in your data.Paste the same corpus you used to form the assumption, so the test is fair.

Notice the shared pattern: paste real material, demand quotes or citations, and explicitly forbid the model from filling gaps. Strip those constraints and you are back to confident fiction.

A workflow that keeps the output trustworthy

  1. 1

    Gather real material first

    Collect transcripts, tickets, reviews, survey verbatims, or sourced community threads before you open a prompt. If you have no real input, your task is collection, not prompting.

  2. 2

    Paste, do not summarize

    Feed the model the raw text. If you summarize first, you have already injected your own bias and removed the exact language that makes the output useful.

  3. 3

    Demand citations in the prompt

    Every prompt above asks for quotes or source labels. Keep that. It lets you trace each conclusion back to a real sentence and catch anything invented.

  4. 4

    Spot-check against the source

    Pick two or three claims and find the underlying quote yourself. If a quote does not exist in your pasted material, discard that part of the output and tighten the prompt.

  5. 5

    Separate findings from guesses

    Keep what the evidence supports apart from what you are inferring. Label inferences as inferences so they do not harden into facts in the next deck.

Honest caveats

Where these prompts help and where they do not.

  • ChatGPT cannot tell you what customers think on its own. Without pasted material it answers from training data with no sources and will fabricate plausible detail. Every prompt here depends on real input.
  • An AI persona is not a substitute for a real customer. A persona built from evidence is a summary of people you have observed; it is not a person you can ask new questions, and treating it as a stand-in for live conversations is how teams drift from reality.
  • The model can still misattribute or paraphrase a quote even when told not to. Spot-check quotes against your source text before quoting them publicly.
  • Synthesis reflects the sample you feed it. If your tickets skew toward angry users or your reviews skew toward fans, the themes will skew the same way. Note what your sample over- and under-represents.
  • Frequency counts the model produces are estimates, not exact tallies. Use them to rank, not to report precise numbers.
  • These prompts help you understand customers you can already hear. They do not replace going to find customers you cannot yet hear — that still takes real outreach or sourced research.

The input is the hard part

The prompts are easy. Getting enough honest, sourced customer language to feed them is the real work. If you want to see how grounded customer research material is gathered and verified before it ever reaches a synthesis prompt, read how we approach it.

Read our methodology

Frequently asked questions

What are the best ChatGPT prompts for customer research?

The best ones make ChatGPT reason over real customer material you paste in — interview transcripts, support tickets, reviews, survey verbatims, or real community threads — and require it to quote the source for each conclusion. The prompt library above gives around a dozen, covering theme synthesis, pain-point extraction, jobs-to-be-done, interview guides, evidence-based personas, voice-of-customer language, and feature-request clustering. Prompts that ask ChatGPT to invent a persona from memory are not reliable.

Can ChatGPT do customer research on its own?

No. A plain ChatGPT answer comes from training data with no sources, so asking it to describe your customers from nothing produces confident fiction. It is genuinely useful as a synthesis engine for real evidence you provide, and for designing research such as interview and discovery-call guides, but it cannot observe your market for you.

How do I use ChatGPT to create a customer persona without making it up?

Paste real customer material and instruct the model to include only attributes it can support with a quote or a clear cross-source pattern, to cite the source for each attribute, and to mark anything unsupported as Unknown rather than filling the gap. The evidence-based persona prompt above does exactly this. The Unknown rule is what prevents the model from inventing traits.

Why are AI-generated customer personas a bad idea?

Personas generated from thin air are confident fiction. They read like research but have no evidence underneath, so they describe a customer who may not exist in your market. They are also not a substitute for talking to real customers — a persona cannot answer a new question. Build personas from real data, and keep them clearly separate from live conversations.

What customer data should I feed ChatGPT for research?

Use sources where customers spoke in their own words: recorded interview transcripts, discovery-call notes, support and ticket threads, product and app-store reviews, open-ended survey responses, and unprompted discussion in communities your buyers use. Paste the raw text rather than a summary, since the exact wording is what makes voice-of-customer synthesis useful.

Keep reading

Use case

Find the pain points your customers never put in a ticket

Surface the frustrations customers vent everywhere except your inbox.

Read →
Use case

Write content about what your audience actually asks

Write about the questions your audience is actually asking.

Read →
Guide

ChatGPT Prompts for Market Research

A copy-pasteable prompt library organized by research stage — planning, instrument design, synthesis, competitor analysis, positioning — built around one rule: ground every prompt in material you supply.

Read →
Guide

ChatGPT for Market Research

What ChatGPT genuinely does well for market research, what it gets dangerously wrong, and the grounded workflow that keeps it useful.

Read →
Guide

Survey alternatives: better ways to learn why

Surveys only tell you what you thought to ask, and people misreport what they will do. A field guide to the realistic alternatives and when each one fits.

Read →
Guide

How to fact-check AI research

An analyst built a slide on a stat the model gave her. The citation looked real. She opened it — the source said the opposite. Here is the workflow that catches that before it ships.

Read →
Guide

Reddit research tool: the honest guide to every type

Reddit is the most candid place on the internet, and the hardest to read at scale. This guide maps every type of Reddit research tool — from free keyword alerts to structured-report engines — so you can pick the one that fits the question you are actually asking.

Read →
Guide

Customer research without recruiting a single participant

Recruiting is why most customer research never happens. Here’s what you can learn without scheduling a single participant.

Read →

Validate what people actually say, not what you wish they would.