Review Mining With AI: Finding What Customers Really Say

A 4.2-star average hides almost everything

A product with a 4.2 average might be loved by most buyers and hated by people with wide feet. Another 4.2 might be fine but uninspiring to everyone. The number is the same and the fixes are completely different.

Reviews contain the answer, written by customers in plain language. The trouble is volume. A retailer with 15,000 reviews across 800 products cannot read them all, so they get skimmed when a product manager remembers, and most of the information is lost. Review mining with AI makes that information countable.

Why sentiment scores are not enough

The first generation of review analysis gave each review a positive, negative or neutral score. That tells you little the star rating did not. What businesses need is the reason. We cover general sentiment analysis elsewhere; review mining is more specific.

Aspect-based analysis, which language models now do well, extracts each thing the customer commented on and how they felt about it. One review can say delivery was fast, the fabric is lovely, and it runs a size small. Those are three separate facts for three separate teams.

Building a theme list that means something

The most important design choice is the list of themes. Too broad, such as quality, and nobody can act on it. Too narrow, and every review creates a new category.

1. Have a model read a sample of a few hundred reviews and propose themes with example quotes
2. Merge and rename them with people from product, operations and customer service in the room
3. Assign each theme an owner, the team that would act on it
4. Tag the full review set against that fixed list, allowing a miscellaneous bucket
5. Review the miscellaneous bucket monthly and add themes when a pattern appears

Turning themes into decisions

Counting themes is a report. Tracking them against products, suppliers and time is a tool. The useful views tend to be:

• Theme share by product, sorted by the return rate or revenue at stake
• Theme trend over time, so a new supplier batch with a durability issue shows up in weeks, not quarters
• Positive themes by product, which feed product page copy and ad messaging in customers' own words
• Theme comparison between similar products, to see why one outsells the other

An illustrative example: an online footwear retailer notices a sizing theme climbing on one boot style over six weeks. The size guide for that style is updated, a line added to the product page, and the supplier contacted about the last. Returns for that style are then tracked. None of that required an AI model to make a decision; it needed the AI model to notice.

Mining competitor reviews: useful, with care

Reviews of competing products on public marketplaces can show gaps in the market, such as a common complaint about a rival's product your version already solves. Many businesses read these by hand already.

Collecting them at scale is a different matter. Check the platform's terms of service before scraping, respect rate limits, and avoid storing reviewer personal data you do not need. Official APIs or licensed data providers are safer routes. When third-party data comes up in our data science projects, we check the source and terms before writing any code; if the only way to get it breaks the terms, we will say so.

Accuracy and the things to check

Language models tag reviews well, but not perfectly. Sarcasm, mixed reviews and very short text trip them up. A few checks keep the numbers honest:

• Hand-label a few hundred reviews and measure agreement with the model per theme
• Keep the quote that justified each tag, so anyone can click through and check
• Watch for themes that are always rare; they may be under-detected rather than uncommon
• Re-test when you change the model or the prompt, because tagging drifts

Also remember who writes reviews. Customers with strong feelings are overrepresented, and review requests sent at different times capture different experiences. Treat theme counts as signals to investigate, not as survey results.

What it costs and when to skip it

A one-off analysis of an export is small and a good way to find out whether the reviews contain anything new. A recurring pipeline that ingests reviews, tags them, and feeds a dashboard with alerts is a modest project, usually a few weeks, mostly integration and dashboard work.

If you get a handful of reviews a week, read them. If your reviews are mostly about delivery couriers you cannot change, the insight will be limited. For businesses with feedback spread across reviews, surveys and support tickets, it usually makes sense to analyse them together, which we discuss in voice of customer analysis. SpiderHunts typically starts with a one-off export so the client sees real themes before committing to anything recurring.

Frequently asked questions