Every post on this site starts from a pile of reader comments: a script reads thousands of them, ranks the questions inside, and we answer the one with the most signal. This post is that loop in a form you can run on your own customers, where the pile is reviews, survey answers, support email, and comment threads instead. AI does the reading; the two prompts below and the checks around them are what keep the summary honest.
Start with the words customers already wrote
Before you write a survey question, collect the feedback your customers have already produced. Product reviews on your listings and your competitors', open-ended survey answers, support email threads, comment sections, sales call notes. Collecting the pile is mostly an export job, and it's worth doing before writing any new survey, because existing feedback tells you which questions to ask. Gather it into one document or spreadsheet with a column for the source; where a comment came from matters when you read the results.
Ask for themes, counts, and quotes
ChatGPT reads uploaded files, including spreadsheet exports, PDFs, and plain text. Hand it the pile and ask for structure rather than a summary in prose:
Here are 214 customer comments from [product reviews /
support email]. Group them into themes. For each theme,
give it a short name, count how many comments touch it,
and quote one comment word for word. If a comment doesn't
fit a theme, put it in 'other' instead of forcing it.
Don't give advice yet.
Each part of the request does a specific job. Counts keep the summary honest: a theme touching 8 comments out of 200 reads differently than one touching 120. Quotes keep the customer's own language, which stops you paraphrasing their complaint into marketing words. "Don't give advice yet" keeps the first pass descriptive, so you see the pile before deciding anything about it.
Code survey answers and interview transcripts the same way
Open-ended survey answers and interview transcripts get the same treatment with one change: instead of themes, ask for a code per answer.
Here are open-ended answers to the question: "[paste it]".
Code each answer by its main topic. Return a table with
columns: answer number, topic, and whether the answer is
positive, negative, or mixed. Quote any answer you find
ambiguous instead of guessing its topic.
For an interview transcript, the request changes again: "list every moment the customer describes a problem in their own words, with the quote." The transcript version matters because interviews answer why, and the why often lives in the customer's phrasing, not in a paraphrase of it.
Strip the identifying details first
Customer feedback contains things the customer didn't agree to share with a new tool: names, email addresses, order numbers, sometimes health or financial details. Delete the direct identifiers before pasting anything, then check what's left, because an unusual purchase combined with a city can still point back to one person. Check your survey tool's terms before exporting, and for business feedback, look at the AI provider's controls for retention, training on your data, and who can see the workspace. None of this changes the analysis; themes don't need names to form.
Where the summary goes wrong
An AI summary of customer feedback fails in predictable ways, and each failure has a check.
- The pile is who wrote in. A theme built from 400 support emails reflects the people who email support, which skews toward problems, not your whole customer base. Say where the pile came from when you present the results.
- AI can't ask a follow-up. It reads what's on the page; the question that would have unlocked the reason goes unasked. Read a sample of the source comments yourself, enough to know whether the themes match what you're seeing.
- An empty pile produces invented answers. Ask a chatbot what customers want without giving it their words, and it writes something plausible from training data that hasn't heard of your customers. Research runs on what customers wrote, not what a model assumes, and the same reason says not to generate fake survey answers to fill out a sample.
- The analysis can slip on its own method. When ChatGPT runs calculations on a spreadsheet, its documentation says to review the code and assumptions before relying on the result. A spot-check of one theme's count can reveal whether the method is behaving (Data analysis with ChatGPT, OpenAI Help Center).
A production example: this site
The topics here come from this loop at scale: a script reads reader comments, 3,734 of them as of this morning, and clusters them into 13 ranked themes, and the day's post answers the highest-signal question we haven't covered. Some days the research changes the plan: this morning's rank showed every high-signal theme already covered, so we wrote from our backlog instead. The smallest version of this loop is one spreadsheet and one AI pass, and the deciding part stays a person's job at either size.
Point it at a decision
Research earns its keep when it changes a decision: which feature to build next, which objection a sales page has to answer, which complaint keeps coming back. Naming the decision first also tells you when to stop analyzing. If the summary can't change the decision, you don't need a deeper analysis, you need a different pile.
FAQ
Can I use ChatGPT for customer research?
For the reading and grouping work, yes, and a free tier handles it: upload a spreadsheet of comments and get themes, counts, and tables back. It can't distribute a survey, sit in an interview, or watch what customers do, so collection and judgment stay with you.
How do you analyze open-ended survey responses with AI?
Export the answers to a spreadsheet, upload it, and ask for each answer coded by topic and sentiment, with ambiguous ones quoted back instead of guessed. The per-answer table is the part worth keeping: you can audit any row against its source, then read a sample of the raw answers before trusting the whole table.
Is it safe to paste customer feedback into ChatGPT?
The sensitive part is less the analysis than the copies: once feedback leaves your survey tool, it lives in another provider's hands. Check your survey tool's terms before exporting, strip direct identifiers, and for business data use a plan whose retention and training controls you've read.
Can AI replace customer interviews?
No. An interview earns its keep on the follow-up question, and follow-ups are where reasons surface; a transcript analysis can't ask them. What AI takes over is part of the digest work afterward: record the interview, get a transcript, paste it, and ask for the customer's own words on each problem, quoted. Same interviews, faster summaries.
What's the difference between customer research and market research?
Customer research reads the words of people who already use you; market research sizes up people who don't yet. The same three moves apply to both, but market research pulls from a wider pile: competitor reviews, forum threads, industry surveys. Start with your own customers, because the pile is smaller and the decisions closer.
Collect the pile, ask for themes with counts, check a sample, point it at a decision. The pieces, when you need them:
- How to Analyze Data with ChatGPT: Beginner's Guide: the file-upload mechanics, step by step.
- How to Fact-Check AI Output: A Practical Workflow: the verification pass, in full.
- How to Use AI for Small Business Operations: where customer messages fit in the wider operations job.
- How to Use AI for Research Without Getting Burned: when the research is about a market instead of your own customers.