How it works
Five steps, and the interesting parts are the constraints: what the model is allowed to say, and what you are charged when the answer is thin.
A brand name, a domain, or a trustpilot.com review URL. A name is resolved through Trustpilot's own autosuggest; a domain goes straight to that business unit, so it can't wander to an adjacent company with a similar name. Set how many reviews you want (200–600), and optionally a recency window or a single language.
Credits are reserved before anything runs — that is what lets a short wallet be refused without an upstream charge ever happening. The pull then runs in the background, sampling proportionally so the star distribution you get back mirrors the brand's real mix rather than an even spread across ratings.
Totals, average rating and the 1–5 star distribution are plain arithmetic over the reviews that arrived. Sentiment is taken from each reviewer's own star rating — 1–2 negative, 3 neutral, 4–5 positive — rather than inferred by a model, so every number traces to something a person chose.
A model groups the reviews into recurring subjects by referring to their ids. The report resolves those ids against the reviews we actually hold and builds each quote from the real row. Ids that don't resolve are dropped; a theme with nothing quotable behind it is dropped whole.
Blocks that filled are billed at 25 credits each; blocks that didn't are refunded to the pools they were drawn from. A brand with no reviews bills nothing. Reading and exporting the finished report is free.
This is the part worth understanding, because it is where a review tool usually starts lying. The model that clusters themes never writes quote text. Its whole output surface is a label, a stance, and a list of review ids. Everything a reader sees — the words, the star rating, the date, the link — is read back out of the reviews we pulled.
The stance gets checked too. If a theme is labelled positive but the reviews under it are one and two stars, the ratings win.
It is the fastest way to judge whether the themes match reality.