Can an app with a great rating have mostly bad recent reviews?

A 4.7-star app can still have mostly bad recent reviews.

Ten of 45 qualifying apps had a rating of at least 4.5, while most of their latest 50 retrieved written reviews were 1 or 2 stars.

The theory: an app’s strong overall rating can coexist with a very different picture in its recent written reviews.

We found the mismatch in ten different apps, including Gmail, McDonald’s, Instagram and Shop. This demonstrates that a high rating does not guarantee mostly positive recent written feedback. It does not mean most of those apps’ users are unhappy.

The star rating alone does not describe the latest written feedback.

An ASO report showing “4.7 stars” can make an app’s review situation look healthy. But that number does not tell you what the newest written reviewers are saying.

Gmail is the clearest example in this test. Its US listing data showed 4.7 stars, backed by more than 2.4 million ratings. Yet 46 of the latest 50 written reviews we retrieved were 1 or 2 stars. Those 50 reviews averaged just 1.34 stars.

Both measurements can be correct. The summary rating includes a broader set of ratings. Our recent-review measure covers a small, specific group of people who left written feedback. They answer different questions.

Gmail

4.7 stars

US public listing rating

46 of 50

recent retrieved reviews were bad
92% of this written-review sample

The selected reviews were dated 24–26 October UTC. Their average was 1.34 stars.

McDonald’s

4.8 stars

US public listing rating

42 of 50

recent retrieved reviews were bad
84% of this written-review sample

The selected reviews were dated 24–26 October UTC. Their average was 1.62 stars.

“Recent” means the latest 50 dated reviews in the retrieved feed, all within the preceding 30 days. It does not mean all reviews posted during those 30 days. The feeds were collected on 27 October, and their newest review timestamps were already about 30–41 hours old in the primary app group.

Ten apps combined a high rating with mostly bad recent written reviews.

For all ten cases, the public listing’s structured data matched both the lookup rating rounded to one decimal and the exact rating count. The smaller API value under each rating shows the precision used for the study’s thresholds.

The mismatch was not universal. The other 35 qualifying apps did not have a bad-review majority in their latest 50 retrieved reviews. Google Gemini had 7 bad reviews out of 50; Vinted also had 7. The full panel, including every non-case and exclusion, is below.

These are different spans of time. Instagram’s 50 reviews covered about six and a half hours; FOX One’s covered about 15 days. Equal sample sizes make the counting rule consistent, but do not create equal time windows.

The mismatch also appeared with 20 and 100 reviews.

We set seven extra checks before calculating the comparisons. The pattern appeared under each check, although the apps that qualified changed. These overlapping checks are not independent studies.

Even when only 1-star reviews counted as bad, seven of the 45 apps still had a strict majority. Requiring every selected review to fall within the previous seven days left 34 qualifying apps, eight of which had the mismatch.

The main result remains 10 of 45 using the original rule. We did not replace it with whichever variation produced the largest percentage.

A real mismatch, without assuming the rating is wrong.

Apple lets users submit a star rating with an optional written review. The overall rating and the written-review sample therefore need not contain the same people. Apple’s explanation.

Apple also says summary ratings are specific to each territory and can be reset when a new version is released. Written reviews remain after a reset. We did not determine whether any of these ten apps had reset its rating. Apple’s ratings and reviews guidance.

This study directly supports one ASO conclusion: a strong summary rating can coexist with mostly negative recent written feedback. An analysis that includes only the star rating can miss that distinction.

It does not establish that satisfaction recently declined, that ratings are manipulated, or that reviews caused a change in rankings or downloads. Those are different claims requiring different data. We also cannot tell whether written reviewers are representative of all current users or all recent star-only raters.

One fixed panel, clear rules, every app accounted for.

We began with the same 57 app IDs used in Experiment 15. Each had appeared in the top 20 of the preserved US top-free apps chart at least once between 19 September and 25 October 2026. This selected group contains prominent apps, so it does not represent the whole App Store.

We reused the 21,750 written-review records collected for Experiment 15 earlier on 27 October. Those are reused data, not 21,750 newly collected reviews. We then obtained fresh US rating metadata, receiving usable results for 55 apps from three lookup requests.

For each app, we sorted its retrieved reviews by their recorded update time. We selected the latest 50 whose timestamps fell within the 30 days before that app’s first feed request. Metadata and review-capture times were no more than 22 minutes apart.

An app entered the main comparison only if it had:

1. A US rating of at least 4.5, using Apple’s unrounded lookup value.
2. At least 1,000 ratings in the general rating-count field.
3. At least 50 retrieved written reviews dated within the previous 30 days.

We counted a mismatch when at least 26 of those 50 reviews had 1 or 2 stars. Before acquiring the new rating metadata and running these comparisons, we set a descriptive evidence rule: finding this in at least three distinct apps would establish a repeated pattern within the panel. We found ten. This was not a statistical significance test.

The main group therefore uses 2,250 selected written reviews: 50 per app across 45 apps. Every app has equal weight in the 10-of-45 result. We did not treat all 2,250 reviews as independent evidence about how common the pattern is across apps.

Limits that affect the interpretation

• This is a selected panel. The observed 22.2% must not be presented as the share of all highly rated apps with unhappy recent reviewers.
• The review feed is a sample. We fetched at most ten pages per app in the earlier collection. We cannot guarantee completeness, and the feed lagged behind retrieval time.
• Review dates can reflect edits. These are the latest dated records we retrieved, not a verified original posting history.
• The two ratings summarize different groups and periods. The recent written-review average is not a corrected version of the App Store rating.
• This is a snapshot, not a change-over-time test. There is no equivalent earlier sample in this experiment showing that complaints increased.

The other apps are part of the result too.

Expand the table to inspect all 57 apps. “No bad-review majority” means 25 or fewer of the selected 50 reviews were 1 or 2 stars. It does not imply an absence of complaints.

Reproduce the result from the included facts.

The download contains every app result, the exact selected review IDs and stars, the full reused review-fact pool, fresh metadata, request logs, source checks, analysis rules and two independent implementations. No reviewer names or review text are included.

Extract the ZIP, then run python analyse.py and python verify.py. Both use Python’s standard library and work offline.

1. Apple: ID-based lookup API. Three new US requests returned metadata for 55 apps. Exact request URLs, capture times and raw-payload hashes are included.
2. Apple US written-review feed, Gmail example. The full original 437-request acquisition log is included. These live endpoints change after collection.
3. Apple: ratings, reviews and responses, and ratings and reviews design guidance. Sources for territory-specific ratings, optional written feedback and rating resets.
4. Apple: ratings and reviews overview. Explains how edited ratings and reviews replace previous submissions.
5. AppOrbit pinned source, by ivershuo. Source of the earlier chart-based panel selection. Derived panel database licensed under ODbL 1.0; attribution and license included.