What you’ll be able to do

You’ll be able to take a pile of messy reviews, survey answers or support tickets, and turn them into a short list of themes with counts attached, the kind of thing you can actually put in front of a team and say “this is what we fix first”. No more scrolling through 200 reviews trying to remember if the delivery complaints outweigh the packaging ones. You just ask, and the AI does the sorting.

Before you start

You need feedback in one place first. That’s the bit people skip. If your reviews live on Google, your survey answers live in Typeform, and your support tickets live in Zendesk, gather them into one document or spreadsheet before you do anything else. A CSV is ideal, one row per piece of feedback, one column for the actual text.

Strip out anything identifying before you upload it. Names, email addresses, phone numbers, order numbers, all of it. Not because the AI tools are especially untrustworthy with data, but because you don’t need to hand over personal details to get themes and quotes. Keep the feedback text and maybe a date or a star rating. That’s plenty.

Small samples mislead you, and this is worth saying properly. Ten reviews will give you noise, not themes. One angry customer who mentions the car park twice can look like a pattern when it’s really just one person having a bad day. Aim for at least 30 to 50 pieces of feedback before you trust what comes out the other end, and more if you can get it.

Get set up

For most small businesses, ChatGPT, Claude or Gemini will do this job well. All three let you upload a file, usually a CSV or a spreadsheet, and ask questions about what’s in it. You don’t need a paid plan to try a small sample, though file upload limits and how much data you can paste in vary, so check what your current plan allows.

If your feedback already lives in a spreadsheet, Google Sheets with Gemini and Excel with Copilot can do this analysis without you leaving the sheet, on the plans where those features are switched on. Handy if you’re the sort of person who lives in spreadsheets anyway.

Where does the feedback come from in the first place, though? That’s the question a lot of small businesses haven’t really answered. If you’re not systematically collecting reviews after every job, you’re working with whatever happens to land on Google or Facebook, which is a biased sample skewed towards people who were furious or delighted, nothing in between.

This is where ReviewNudge, which we make, is worth a mention: it sends an SMS or email after a job asking for a review, and if the rating’s low, it routes that feedback to a private form instead of a public one. That second bit matters more than it sounds, because a lot of the useful, specific, actionable feedback that would help you improve never makes it to a review site at all. It ends up in an inbox, or nowhere. Collecting it is step one. Analysing it, which is what this guide covers, is a separate step after that.

For a fair comparison, Google Forms and Typeform also export your survey answers as CSV, which you can then feed into any of the AI tools above. There’s no single “must use this” tool here. It depends whether you need to collect feedback from scratch or you’ve already got a folder full of it.

One more thing worth flagging: MarketFrame, also ours, does a similar kind of analysis but on your competitors’ public reviews rather than your own customers, if you’re trying to work out what people complain about elsewhere.

Try it yourself

Here’s a prompt for finding broad themes across a batch of feedback. Paste your feedback in as text, or upload the CSV first and then send this.

I've uploaded a file (or pasted below) containing customer feedback from reviews, surveys and support tickets. Please read through all of it and identify the main recurring themes. For each theme, give me:
- A short name for the theme
- How many times it comes up (a rough count is fine)
- Three verbatim quotes from the feedback that best illustrate it

Don't invent themes that only appear once or twice. Focus on what's actually recurring.

Next, once you’ve got your themes, this one narrows things down to what’s actually worth fixing first.

Based on the themes you identified, what is the single change we could make that would address the most complaints? Explain your reasoning using the counts and quotes from the feedback, and tell me what you'd expect to happen if we made that change.

And a third for when you just want the quotes pulled out cleanly, useful if you’re writing a report for a manager or a founder who wants the receipts, not the summary.

For each of the themes you found, pull out five verbatim quotes from the feedback (not paraphrased) that show that theme clearly. Group them under the theme name. Keep the quotes exactly as written, including any spelling or grammar in the original.

That last instruction, keeping quotes exactly as written, matters more than it looks. AI tools have a habit of tidying up quotes as they go, which is fine for readability but useless if you’re trying to show someone the actual words a customer used.

Check the result

Once you’ve got themes back, read a handful of the original feedback entries yourself and check whether they’d have landed in the theme the AI put them in. Pick five or six at random. If you keep disagreeing with the categorisation, the themes are probably too broad or too vague, and it’s worth asking the AI to split a theme into two or narrow the wording.

Check the counts add up roughly to your total. If you uploaded 80 pieces of feedback and the themes only account for 40 mentions, ask what happened to the rest, there may be feedback that’s too vague or off topic to categorise, and that’s worth seeing rather than silently dropping.

Ask the AI directly: “which of these themes are you least confident about, and why?” A decent response here should mention things like low counts, vague wording in the original feedback, or overlap with another theme. If it just repeats the same theme back to you with more confidence, push again with a smaller, more specific slice of the data.

Try running the same broad-themes prompt twice on the same data. If the results are wildly different each time, something’s off, either the data’s too messy or too small a sample to produce a stable pattern.

If it doesn’t work

Before uploading anything, check it for personal data one more time. Names, emails, phone numbers, addresses, order numbers, anything that identifies a specific customer. None of this is needed for theme analysis, so there’s no reason to hand it over. If your export includes it automatically, strip the columns out in a spreadsheet before you paste or upload.

Don’t hand over anything containing payment details, even partial ones. Don’t hand over internal notes that mention other customers by name if that’s mixed in with your support ticket data. And if your feedback includes anything sensitive, health information, safeguarding concerns, that sort of thing, don’t put it anywhere near a general-purpose AI tool. That’s not what these tools are built to handle responsibly.

If the AI misses obvious themes, the most common cause is sample size. Below 20 or 30 entries, patterns just don’t show up reliably; there isn’t enough repetition for anything to count as a theme. The fix is more data, not a better prompt.

Messy formatting causes problems too. If your CSV has feedback split weirdly across multiple columns, or HTML tags left in from an export, or half the rows are blank, clean that up first. A quick pass in a spreadsheet, deleting empty rows and merging split text into one column, saves a lot of confusion.

And sometimes the feedback itself is just too vague. “Good service” and “would recommend” don’t tell you anything actionable, no amount of clever prompting turns nothing into something. If a big chunk of your feedback is like that, the honest answer is you need more detailed feedback going forward, which loops back to how you’re collecting it in the first place.

Keep exploring

If you’re trying to get found and picked, have a look at “How to get AI to recommend my business”. If this whole process has convinced you that spreadsheets are where you want to live, “How to get AI to make a spreadsheet” walks through building one from scratch with AI help. And if the feedback analysis is feeding into bigger decisions about where the business goes next, “How to get AI to write a business plan” is the natural next step.

Sources and review notes

Review date: 15 September 2026 — every source above was opened and checked against the text on that date.

Disclosure: ReviewNudge and MarketFrame are our products — this site is made by the same studio, Adapt Progress Evolve. We have tried to compare fairly, and the alternatives named are real and worth a look.