The research you have already paid for
Every review, support ticket, cancellation email and survey comment is a piece of market research you have already paid for, usually painfully. Most businesses read each one as it arrives, react to the angry ones, and never look at the pile as a whole. Yet the pile is where the value sits: one comment is an anecdote, but three hundred of them contain patterns that no single reading will ever catch.
Until recently, mining that pile meant either a research agency's invoice or a heroic weekend with a spreadsheet and highlighters. A capable large language model such as Claude or ChatGPT now does the first 80 per cent of the job in minutes, provided you feed it properly and check its work. This is a practical walkthrough of exactly how.
Gathering and preparing the data
- Pull reviews from wherever they live: Google Business Profile, Trustpilot, Amazon, app stores. Business accounts on most platforms offer bulk exports; at smaller volumes, copy and paste is perfectly workable.
- Export support tickets from your helpdesk. Zendesk, Freshdesk, Gorgias and HubSpot all offer CSV exports; include the customer's first message and the final resolution.
- Add cancellation emails, NPS or survey comments, and anything typed into a 'reason for leaving' box.
Before anything goes near an AI tool, strip the personal data: names, email addresses, phone numbers, order numbers. A find-and-replace pass works, or run a first LLM pass whose only job is redaction. Then use a paid business tier of your AI tool with a no-training commitment, and the exercise stays comfortably on the right side of UK GDPR.
Modern models accept very large pastes, but analysis quality is better in batches of one to two hundred items. Label each batch with its source and date range so that trends over time stay visible when you compare quarters.
Need a hand with this?
Our team delivers AI & Machine Learning for UK businesses — with a free initial consultation, transparent fixed quotes and no lock-in contracts. Tell us what you're working on →
Copy-paste prompts that do the heavy lifting
The difference between vague AI waffle and a usable report is a prompt that demands structure and evidence. Adapt these three:
Theme extraction
"Here are 150 customer reviews. Identify the 8-12 most common themes, positive and negative. For each theme give: a short name, a count of reviews mentioning it, two verbatim quotes, and whether mentions rise or fall across the date range. Do not invent quotes; only use text that appears in the data."
Prioritising problems
"From the negative themes above, rank the top five by likely revenue impact for a business like ours [describe your business in one line]. For each, state the probable underlying operational cause and one cheap experiment that would test a fix."
Voice-of-customer copy
"List the exact phrases customers use when praising us, grouped by the benefit they describe. Flag the ten phrases that would work best as website or advert copy."
Finding churn reasons in support tickets
Reviews tell you what the public thinks; tickets tell you why customers actually leave. Export the last six months and run something like:
"Here are 200 support tickets. Classify each by root cause: product fault, delivery, billing, expectation mismatch, user error, other. Give me the count per category, the average number of back-and-forth messages before resolution per category, and which categories are most often followed by a refund or cancellation. Quote three tickets that best illustrate each of the top three categories."
The finding many owners find uncomfortable is that a large share of tickets traces back to expectation mismatch, which is a marketing and product-page problem rather than a support problem. Customers were promised, or inferred, something the product does not do. That single reframe is often worth the entire exercise, because the fix is a paragraph of honest copy rather than a support hire.
Turning themes into decisions
- Recurring complaints are an operations to-do list: fix the top one, then measure whether mentions fall in the next quarter's batch.
- Recurring questions are missing content: turn the five most common into FAQ entries and product-page copy so they stop arriving as tickets.
- Recurring requests are a roadmap vote: weight them by customer value, not just by volume, before building anything.
- Praise themes are positioning: lead your marketing with what customers actually love, in the words they use, rather than what you assumed they valued.
Rerun the whole analysis quarterly with fresh exports and keep the results in one running document. The trend lines between quarters are frequently more informative than any single quarter's snapshot.
Key Takeaway
Export the last 12 months of reviews and support tickets, strip out names and contact details, and run them through an LLM with prompts that demand theme counts and verbatim quotes as evidence. Verify a sample of quotes against the source before acting, because models occasionally invent them. Repeat quarterly: recurring complaints are your operations to-do list, recurring requests are your roadmap, and the exact words customers use are your best marketing copy.
Pitfalls to respect
- Hallucinated quotes: models occasionally invent plausible-sounding quotes even when told not to. Spot-check every quote you plan to act on against the source file before it reaches a slide or a decision.
- Sample bias: reviewers skew towards the delighted and the furious, and tickets skew towards problems. Read findings as signals from the vocal, not a census of your customer base.
- Overcounting: ask the model to show its counts, then verify a couple yourself with a search of the raw file.
- Privacy: redact before pasting, and never feed personal data into consumer-tier AI tools that may train on inputs.
Handled with that discipline, review mining is one of the highest-return AI use cases a small business has, because the raw material is free and already yours. If you would like help setting up the pipeline or acting on what it finds, our team can help.