What app-store reviews reveal about product risk

A rating averages away the thing you need to see

An app with 40,000 ratings and a 4.3 average can be quietly falling apart. The average is weighted by years of history, so a month of one-star reviews about a broken checkout moves it by a rounding error. Meanwhile the reviews themselves are screaming.

Our audited catalog splits like this: 402 apps rate under 3.0, 1,315 between 3.0 and 3.9, 1,993 between 4.0 and 4.4, and 2,064 at 4.5 or above. The interesting finding is that recurring complaint clusters appear across all four bands. A high rating is not evidence of no problems — it is often evidence of a long history.

Three patterns worth acting on

One angry review is noise. The same complaint from unrelated people across several app versions is a signal, and it is the pattern that predicts churn.

• Complaints that survive an update — the strongest signal there is, because it means the fix that shipped did not address what users are experiencing
• A cluster that is new since the last release — usually a regression, and the cheapest possible moment to catch one
• Payment, subscription or refund language — 28.6% of audited apps show a payment cluster, the largest single cluster in the catalog, and these convert directly into chargebacks and store complaints
• Device compatibility reports — 28.4% of audited apps, a statistical tie with payment at the top of the ranking — which usually means a device or OS tier stopped being tested
• Silence from the developer — 47.4% of audited apps have never replied to a review at all

Replying is a ranking behaviour, not just customer service

Both stores surface developer replies publicly, and a visible reply changes what the next prospective user reads. The average reply rate across the audited catalog is 28.0%, which means most negative reviews sit there unanswered as permanent copy on the listing.

The practical version: reply to every one- and two-star review, briefly and without argument, and say what changed when it changes. Users routinely raise their own score after a reply. Nobody raises it after being ignored.

Use evidence, not sentiment theatre

Sentiment percentages are easy to generate and hard to trust. A useful analysis names the cluster, counts the matching reviews, shows when the most recent one arrived, and quotes a couple of them verbatim so the classification can be checked.

That is deliberately unglamorous. It is also the only version an owner can act on, because "sentiment fell 8%" tells you nothing about what to fix, while "nine people since the March release cannot complete checkout on Android 14" tells you exactly where to look.

Every catalog figure on this page is a snapshot of the audited catalog taken on 3 August 2026. The catalog grows continuously, so the live totals are higher than the counts quoted here, while the shares move only slowly.