Ask ChatGPT who your customers are, what they want, and why they buy. You’ll probably get an answer that sounds reasonable.
That’s the problem. Reasonable is a very low bar for something you’re about to build a marketing budget around.
Can ChatGPT do your market research? It can help with substantial parts of it: organizing questions, comparing public information, and summarizing evidence you provide. But an answer about customers is only as useful as the evidence behind it. A beautifully formatted guess still spends real money.
Start with a decision, not “research my market”
“Tell me about the accounting services market” invites a report you can admire and then ignore.
“Should our bookkeeping firm test a monthly service for independent restaurant owners in Phoenix?” gives the research a job.
Now you can ask useful questions. How many businesses fit? What alternatives do they use? What problems prompt them to switch? Who approves the purchase? What would make the service worth its price?
Define the audience, geography, time period, and decision before asking for answers. Otherwise, you’re ordering information by the pound.
Desk research gets you oriented
Desk research uses information that already exists. Think government business counts, competitors’ published services and prices, industry reports with visible methods, and public customer reviews.
The Small Business Administration’s planning guidance includes market research to help you understand your customers and competitive environment. Existing sources help establish context. Direct research helps answer questions specific to your customers and business.
Use ChatGPT to help organize that first pass. Request a comparison of competitors, links to original sources, publication dates, and a list of unanswered questions. When current information matters, use a workflow that can retrieve current sources and check what it actually found. Asking for freshness does not make an unsourced answer current.
Keep the categories straight. A competitor’s website tells you what it promises. Reviews tell you what those reviewers reported. Neither tells you exactly why your next prospect will buy.
Your imaginary focus group cannot buy anything
“Act as ten restaurant owners and tell me whether you’d buy this” can produce ideas for questions. It does not produce ten restaurant owners.
Synthetic respondents are generated responses. Don’t count them as interviews, quote them as customer feedback, or use their fictional enthusiasm to justify a launch.
Talk to actual buyers, including people who chose another provider or decided to do nothing. Ask about a recent decision: what happened, what they tried, what it cost, and what they did next. “Would you buy this?” is easy to answer politely. An actual purchase requires a little more commitment.
Surveys need similar care. Record who responded, how they were recruited, and who is missing. Ten friendly customers can reveal useful problems. They cannot establish what an entire market believes.
Make every important claim show its work
A citation gives you somewhere to check. It does not finish the checking.
OpenAI’s guidance on research synthesis emphasizes checking citations against original sources and preserving the findings’ scope and uncertainty. Apply that discipline before a claim reaches your strategy.
- Open the original source. Does it support the specific statement?
- Check the fieldwork date as well as the publication date.
- Confirm the geography, business size, and customer category match yours.
- Look for sample size, recruitment method, and relevant exclusions.
- Separate reported facts, your interpretation, and unanswered questions.
If the original evidence is inaccessible, mark the claim unverified. Repetition across five articles does not create five independent studies.
Here’s what this looks like in practice
Suppose that hypothetical Phoenix bookkeeping firm wants to lead with “save time.”
First, review local competitors’ published offers and gather relevant business-count data. Then interview a small, deliberately selected group of restaurant owners about their last bookkeeping decision. Treat those interviews as exploratory, not representative.
Suppose several interviewees describe unexpected cash shortages before they mention administrative time. That suggests a hypothesis: clearer cash visibility might be a stronger opening message.
Give ChatGPT deidentified interview notes and ask it to group recurring problems, preserve contradictory accounts, and link each theme to its supporting notes. Don’t upload sensitive customer records. Remove names, contact details, account information, and identifying context. Use only material you’re authorized to process in an approved tool and account.
Then test the competing messages with actual prospects. Track qualified inquiries, sales conversations, and purchases. The interviews informed the test. They did not prove the outcome.
A prompt you can reuse
Start with this, then provide the approved sources and context the task requires:
Help me investigate [decision] for [audience] in [region] during [period]. Use the approved materials I provide and identify additional public sources needed. Separate verified findings, interpretations, and hypotheses. For each material claim, provide its original source, date, relevant population, and limitations. Flag missing or contradictory evidence. Do not invent statistics, quotations, respondents, or citations. Finish with customer interview questions and a practical test that could change our decision.
A prompt is an instruction, not a quality guarantee. Review the output against the sources before using it.
Use the findings to connect your marketing spending to revenue. You can browse the marketing resources or work with me on the decisions behind your plan.
Let ChatGPT help you do the homework. Your customers still get a vote.
