How to Do Market Research With ChatGPT in 30 Minutes

Most guides on how to do market research with ChatGPT give you a list of 30 prompts and stop there. The prompts are usually fine. The trouble starts later. You paste a market-size figure into a client deck, someone asks where it came from, and the link is missing or says something else.

This guide is the version we teach: a 30-minute sprint that ends with a one-page brief, and you have opened the source behind every number in it yourself. It works for a consultant scoping a new engagement, a marketer testing a positioning idea, or an owner checking whether a segment deserves a closer look. It won't replace customer interviews, but it will tell you which interviews to run.

What ChatGPT is good at in market research, and what it isn't

ChatGPT is fast at turning a fuzzy question into a structured one. With web search on, it finds and summarizes public sources. It can compare competitors using material you give it, and it can pull patterns out of reviews, transcripts and open-ended survey answers.

It is weak at exact numbers it hasn't actually retrieved, at niche markets with little public data, and at anything that depends on what your customers really think.

The citation problem has been measured. A peer-reviewed study in The American Economist found that GPT-4 produced false citations more than 20% of the time on economics topics, and the rate went up as questions got narrower (Buchanan et al., 2024). Market research questions are usually narrow. So the workflow below treats the model as a fast research assistant, and every claim it makes gets checked.

The 30-minute market research sprint

Minutes	Step	What you do	Output
0–5	Frame	Turn the topic into a decision, no search yet	1 research question, 4–6 sub-questions
5–15	Desk research	Web search or deep research with strict source rules	Table of claims with links and dates
15–22	Competitors	Feed in pricing pages and reviews you collected	Comparison table, buyer quotes, gaps
22–27	Claim check	New chat reviews the table, you open the top 3 links	Confirmed vs. unverified rows
27–30	Brief	Synthesis from confirmed material only	One page: answer, unknowns, hypotheses

Two practical notes. A deep research task can take anywhere from a few minutes to half an hour, so launch it at minute 5 and do the competitor step while it runs. Also check your plan. As of October 2026, OpenAI's pricing page lists deep research as limited on Free and Go, expanded on Plus ($20/month) and maximum on Pro, and a counter in the product shows how many tasks you have left. Regular search in ChatGPT is enough for a first pass.

Step 1: Turn the topic into a decision (minutes 0–5)

"Research the pet care market" gets you a generic essay. A decision gets you a research plan. Start with search off:

You are helping me scope a market research question. Do not research anything yet.

Decision I need to make: [e.g., whether to run a 6-week test campaign for segment X next quarter]
What I already know: [3-5 bullets]
Who reads the result: [client partner / CMO / myself]
Time budget: 30 minutes

Ask me up to 5 clarifying questions first. Then rewrite my decision as:
1. One research question
2. 4-6 sub-questions whose answers would change the decision
3. For each sub-question, the type of source that could answer it (official statistics, company filings, official pricing pages, customer reviews, trade associations, analyst reports)

Answer the clarifying questions honestly, and "I don't know" counts as an answer. The sub-questions it gives back are your scope. Anything outside them waits for another day.

Step 2: Desk research with sources (minutes 5–15)

Turn on web search (or start deep research) and paste in the sub-questions:

Use web search. Research the sub-questions below.

[paste sub-questions]

Rules:
- Every factual claim needs a link to the exact page where it appears, plus the publisher and the date of the data.
- Prefer primary sources: official statistics, company filings and investor pages, official pricing pages, trade associations. Use news and blogs only when nothing primary exists, and say so.
- If two sources disagree, show both numbers. Do not average them.
- Never estimate a market size, growth rate or price yourself. If you can't find it, write "Not found" and suggest where a person could look.
- Output a table: Claim | Number | Source link | Publisher | Date of data | Primary or secondary

The "never estimate" rule matters most. Without it, a model asked for a market size will often hand you a number that sounds right instead of saying it found nothing. Also watch for scope drift, like a global figure passed off as a segment figure, or a number from four years ago passed off as current.

Step 3: Competitors and customer language (minutes 15–22)

This step works best on material you collected yourself, not on the model's memory. Open three competitors' pricing pages and copy 30–50 public reviews from the platforms your buyers actually use. Then:

Below are pricing pages and customer reviews I collected for three competitors.

[paste or upload]

Using only this material:
1. Build a comparison table: competitor, target customer, pricing model, entry price, main promise, top 3 complaints in reviews.
2. Quote the exact review phrases behind each complaint.
3. List gaps: needs that come up in reviews but that no competitor addresses on its pricing page.
Label anything you infer rather than read directly as "Inference".

For marketers, the review quotes are often the most useful thing the sprint produces. They are the words buyers actually use, and you can test them in messaging right away.

Step 4: Check the claims (minutes 22–27)

Open a new chat so the model isn't defending its own work:

Below is a research table produced by an AI assistant. Act as a skeptical reviewer.
For each row:
- Open the link and confirm the number appears on that page with the same meaning (same year, same scope, same unit).
- Label the row Confirmed, Mismatch (explain), or Link does not support the claim.
- Flag any number that comes from a press release or a report sales page rather than the original publisher.
Do not add new claims.

[paste table]

Next, pick the three numbers that would change your decision the most, open their links and read the pages yourself. Five minutes is enough for three. Our hallucination checklist for AI answers covers the full method. For each of those numbers:

• The link opens and the number is on the page
• The year, unit and scope match (global vs. segment, revenue vs. users)
• The publisher is the original source, not a site summarizing a paid report
• For forecasts, you know who made it and when
• You redid any math yourself (growth rate, share, units × price). See how accurate ChatGPT is with numbers

Step 5: The one-page brief (minutes 27–30)

Using only the rows marked Confirmed and the competitor table above, write a one-page brief:
- Answer to the research question in 3 sentences, with a confidence level (high / medium / low) and the reason for it
- What we know: confirmed facts, each with its source
- What we don't know yet, and how to find out (customer interviews, a short survey, our own sales data)
- 3 hypotheses to test with real customers, each with the interview question we would ask
Under 400 words. Do not include any claim that is not in the material.

A worked example

Here is a typical task for a B2B marketer. The company sells scheduling software to dental clinics and wants to know whether veterinary clinics are worth testing as a second segment.

In Step 1, "research the vet market" becomes "Should we build a vet-specific landing page and run a 6-week test campaign next quarter?" The sub-questions cover how many clinics there are, which scheduling tools they use now, what those tools cost, and what buyers complain about.

A first desk research pass on a question like this usually comes back mixed: an official count of establishments from a statistics agency, a few competitor pricing pages, and a market-size figure taken from a report sales page with no methodology. The table format makes this obvious at a glance, because the source column for that row is a page selling a report.

You fix this with a follow-up, not a new prompt:

Row 4 cites a report sales page, not the report itself. Find who published the original figure and whether its methodology is public. If you can't, mark the row "Unverified" and leave it out of the brief.

With that row gone, the brief rests on the establishment count, competitor prices and review complaints. That's enough to justify a small test campaign but not a product build. That's an honest answer, and it's the kind a 30-minute sprint should give you.

Where it breaks

• No search, no facts. With search off, numbers come from training data. They may be years out of date, or made up.
• Recycled numbers. Many market-size figures online come from press releases quoting other press releases. If you can't trace a number to its original publisher, treat it as unverified.
• Synthetic customers. Asking ChatGPT to "act as a clinic manager" gives you hypotheses for an interview guide. It isn't evidence.
• Confidential inputs. Client data, pipeline exports and anything under NDA don't belong in a consumer chat. Read what not to upload to AI and check your company policy first.
• Your own data. Survey exports and sales data need a different workflow. See how to analyze a spreadsheet with ChatGPT.

For consultants and marketers

If you're a consultant, run the sprint before a kickoff call. The first meeting then starts from a scoped question and a clear list of what nobody knows yet. Our AI for consultants page covers how this fits into client work. If you want ready-made versions of these prompts for competitor scans, interview guides and synthesis, the Consultant's Research & Analysis Prompt Kit has them in one place.

If you're a marketer, the review language from Step 3 and the hypotheses from Step 5 go straight into positioning work and test campaigns. More on that at AI for marketers.

This sprint is part of Session 3 of our practicum, "Documents, data, visuals." We run it live on a research question you bring. The same session covers where AI is reliable with numbers and where it gets them wrong with full confidence, and what you should never upload. You leave with your own versions of these prompts and a checking routine you'll actually keep using. The full program shows what each session covers.