Your customers have already told you exactly what to fix and exactly what to say in your next campaign. The answers are sitting in a spreadsheet nobody reads.
Stop paying for market research you already own. Export your reviews, survey comments, and support tickets, remove names and emails, and paste batches into ChatGPT or Claude with the three prompts in this guide: one extracts ranked themes with verbatim quotes, one scores sentiment by topic, one mines exact customer language for ad and landing page copy. Verify the top themes against real counts, assign each one an owner, and close the loop monthly.
Marketing teams will pay an agency five figures for customer research while 2,000 unread survey responses sit in a spreadsheet from last quarter. I have seen it more times than I can count. The most honest, specific, unprompted account of what your customers think already exists. It is in your reviews, your support inbox, and that NPS export nobody opened.
The reason it goes unread is volume, not value. Nobody has three days to read a thousand comments and tally themes by hand. That bottleneck is now gone. Reading a pile of messy text and sorting it into patterns is one of the few tasks current AI does almost perfectly, because the source material is right there to check.
And the companies doing this are seeing real numbers. Online fashion retailer Motel Rocks used AI-powered sentiment analysis in its support operation and reported a 9.44% lift in customer satisfaction alongside a 50% drop in ticket volume. [2] Zendesk’s CX Trends research finds CX leaders increasingly betting on exactly this kind of AI-assisted analysis to personalise and prioritise. [3] You do not need their budget to start. You need an export button and three prompts.
Before any AI touches anything, gather the raw material. Most businesses have five sources and forget three of them:
One rule before anything gets pasted into an AI tool: strip the personal data. Remove names, emails, order numbers, and anything identifying. A find-and-replace pass takes ten minutes. If your company has rules about AI tools, this is the moment to check them (our guide to using AI without leaking company data covers the safe setup).
AI gets you from 1,000 comments to 8 themes in minutes. The judgment about which theme matters commercially is still yours, and it should be.
“Here are customer comments from [source, date range]. Identify the recurring themes. For each theme: a short name, an estimated share of comments, three verbatim quotes, whether it is getting better or worse within this period, and which team should own it (product, support, marketing, ops). Rank by frequency. Do not invent themes that appear fewer than three times; list those separately as one-offs.”
“Classify each comment by topic (shipping, quality, price, support, usability, other) and sentiment (positive, negative, mixed). Output a table of topic vs sentiment with counts, then list the five most intense negative comments and the five most enthusiastic positive ones, quoted exactly.”
“From these comments, extract the exact phrases customers use to describe: the problem they had before buying, the moment the product won them over, and the hesitation they almost did not overcome. Quote them word for word. I will use these in ad copy and landing pages, so flag the ten most vivid phrases.”
That third prompt quietly pays for the whole exercise. Plugging real customer phrasing into your ads and landing pages is the closest thing to a cheat code in copywriting; it pairs perfectly with the workflow in how to use ChatGPT to write ad copy.
The paste-into-chat approach has a ceiling: it is manual, it samples rather than monitors, and it does not run while you sleep. When feedback volume passes roughly a few hundred items a week, look at the tools that do this continuously.
If you run support through a platform like Zendesk, sentiment and intent analysis is increasingly built in: AI reads tone and urgency across chat, email, and messaging and routes accordingly. [1,4] Survey platforms have shipped similar text-analysis layers. The buying question is not “is the AI clever?” but “does it sit where my feedback already lives?” A mediocre analyzer inside your help desk beats a brilliant one that requires a weekly CSV ritual you will abandon by August.
Until you hit that volume, honestly, the chat workflow is not a compromise. It is the better teacher: you see the raw comments, you argue with the themes, and you build instinct about your customers that a dashboard never gives you.
One sizing note from experience: if you are evaluating paid tools, run the chat workflow for two months first and keep your theme spreadsheet. It becomes your benchmark. When a vendor demos their dashboard, you can check their themes against the ones you found by hand-plus-AI. If the expensive tool cannot beat your free workflow on your own data, that is your answer, and you learned it before the contract, not after.
Analysis that does not change anything is a hobby. Close the loop with three habits:
Here is what closing the loop looks like in practice. A client of mine ran this workflow on a quarter of post-purchase surveys and found a fast-growing theme: customers loved the product but could not tell which size to order. The fix was a size comparison photo on every product page, shipped in a week. The phrase customers kept using, “I wish I had known it runs small”, became the FAQ heading on the page. Complaints on that theme dropped the next month, and the team found it without reading a single survey end to end.
This week’s move: export your last 90 days of reviews, strip the personal data, and run Prompt 1 on the first 150. You will know more about your customers by Friday than most of your competitors learn in a quarter.
Yes. Paste batches of 100 to 200 reviews with personal details removed, and prompt it to extract ranked themes with verbatim quotes, classify sentiment by topic, or pull exact customer phrases for marketing copy. Verify theme frequencies against the raw file before acting, since AI approximates counts.
Start with ChatGPT or Claude: free, flexible, and good enough for most teams below a few hundred feedback items a week. Above that volume, use AI built into where feedback already lives, such as Zendesk for support tickets or your survey platform for NPS comments, so analysis runs continuously.
Patterns start appearing at around 50 comments, and 150 to 300 gives you reliable theme rankings. Below 50, read them yourself: the value of AI is compressing volume, and at small volume the verbatim detail matters more than any summary.
Only after removing personal data: names, emails, order numbers, and anything identifying. Use a business plan with model training disabled where possible, and check your company AI policy first. The themes survive anonymisation; the risk does not need to.
Good on obvious positive and negative, weaker on sarcasm, mixed feelings, and culturally specific phrasing. Treat sentiment scores as directional, ask for quoted examples behind every classification, and spot-check a sample by hand before reporting numbers to anyone.
This workflow comes from a decade of running customer research for B2B and consumer brands, plus Zendesk’s published research and case studies on AI-assisted feedback analysis. The prompts are the ones I run on real client exports.
Weekly tips. Real examples. Practical help for busy professionals.
We care about your data, check out our privacy policy.