Insights
Sentiment and themes
How ReviewDesk classifies the mood of a review and extracts the topics people keep raising
Every review is classified twice after it arrives: once for sentiment, and once to extract the themes it mentions. Both run in the background and feed the dashboard, the sentiment filter, and theme trends.
Sentiment
Each review is labelled Positive, Neutral, or Negative based on what it says, not what it scored. The distinction matters: a three-star review describing a crash is negative, and an occasional one-star review is really a neutral question.
Reviews with an empty or very short body are classified from their rating alone, without an AI call — there is nothing to read, and it would be a waste of quota.
Themes
Themes are short topic tags extracted from review text — the recurring subjects in your feedback, such as pricing, a sign-in problem, or a specific feature. Similar wordings are merged so that "log in", "login", and "signing in" become one theme rather than three, and generic filler that would tell you nothing is discarded.
Theme trends on the dashboard show which are growing against the previous period. A theme climbing sharply is usually the earliest signal that a release has broken something.
When classification runs
| Aspect | Behaviour |
|---|---|
| Cadence | A classification pass runs every few minutes. |
| Batch size | Up to 200 reviews per app per cycle. |
| New reviews | Show as unclassified until the next pass reaches them — usually minutes. |
| Backlog | A newly connected app with many reviews is worked through over several cycles. |
The dashboard surfaces the state of this pipeline, including whether it is running and when the next cycle is due, so an unclassified backlog is visible rather than mysterious.
Coverage
The coverage indicator beside theme trends is the share of reviews in the period that have been classified. Low coverage means the trends are drawn from a partial picture — worth waiting for it to rise before acting on a theme that appears to be spiking.