Review Theme Miner

A free tool for founders asking what buyers praise and complain about, in their own reviews and their rivals'. Paste reviews of your product and one or more rivals, one per row with the product, the star rating and the text. The miner reports each theme's share of reviews with a 95% interval, the themes where a rival draws complaints and you don't, the theme to fix first, switching sentences, feature asks and the phrases your happy reviewers use. It says how much of the text its coding reached. It runs in your browser; nothing you paste is sent anywhere.

How it codes the reviews

It runs the engine of the AAJ review-mining agent skill, unchanged. It counts reviews, not quotes: a theme raised three times in one review is one review. Sentiment comes from the star rating only: four or five stars is praise, one or two is a complaint, three is mixed, and a review with no rating is a mention only. Themes come from your own coding when a review has it, and otherwise from a word list of twelve themes common to software reviews, which you can extend. The result counts the reviews coded by hand, by the word list and not at all.

The engine's rules

Each theme's share carries a 95% Wilson interval. With fewer than 10 reviews of a product, the intervals are too wide to compare it with another.
A rival's weak spot is a theme where the rival's complaint share is 20% or more and yours is at least 10 points lower, with 10 or more reviews of each. It is a lead to prove, not a claim the engine has checked.
Fix first is your theme with the most complaints, needing at least two.
Below 60% coverage from the word list, code the reviews by hand or add your own themes.

What it can't see

Reviews describe reviewers, not the market. Use more than one review site before generalising.
A complaint inside a four- or five-star review counts as praise for that theme.
The word list matches words, not meaning.

Why I built it

Rivals' customers have already written down what annoys them, in public, and almost nobody counts it. I wanted counting that a founder can check: reviews, not quotes, sentiment from the stars the reviewer chose, and a plain statement of how much the coding missed.

The same engine runs as an agent skill, review-mining. To turn a weak spot into a sales card, use the Battlecard Builder. To build a persona from what reviewers say, use the Three-Lens Persona Brief.

Run this in your agent

The free AAJ agent skill Review Mining does this job inside a coding agent, on your own data. Install it with npx skills add sarojkjha/aaj-marketing-skills --skill review-mining.

Using Claude or ChatGPT instead? Add the AAJ engines as a custom connector, with no install: https://aaj-engines-mcp.sarojkjha10.workers.dev/mcp

Frequently asked questions

Where do I get the reviews?

From the review sites, app stores and forums where your buyers write, copied into a sheet with one review per row: the product, the star rating and the text. Check each site's terms before you copy reviews from it.

Why doesn't it read sentiment from the words?

Because words mislead: "the support is unbelievable" can be praise or a complaint. The star rating is what the reviewer chose, so the miner uses that and nothing else.

What does coverage mean?

How many reviews got at least one theme, by your own coding or by the word list. Reviews that matched no theme are counted and listed as uncoded, so you can read them for themes the list doesn't know.

Can I use a rival's weak spot in my copy?

Only as a lead. The miner shows where a rival draws complaints and you don't; it hasn't checked you are better. In AAJ's view, make the claim only with your own proof.

Is my data stored or shared?

Nothing is sent anywhere. The miner runs in your browser. Your last reviews and settings are kept in this browser only, so a refresh doesn't lose them. Start over clears them.