Grok, Perplexity and Claude lost their reviewers in a year. ChatGPT and Gemini kept theirs.

An App Store rating looks like a verdict. It behaves more like a monument. An app sitting at four-and-a-half stars has earned that number over years of installs and early goodwill, and it moves slowly. It says very little about how the people using the app this month actually feel.

This is the Nativerse lab reading underneath that number. For a whole category we separate two things the single star rating blurs together. The first is population truth: Apple's full ratings histogram across every rating an app has ever received, and it is almost immovable. The second is the mood of the people who write: a 90-day window of written reviews, set against what that same app's reviewers were saying a year earlier. One population, two points in time. Then the question that decides whether an app recovers. When users turn, does the developer answer?

This study covers 12 tracked Productivity apps on the US App Store, 35,895,616 ratings in all. Their mean lifetime rating is 4.77, and it will still read about that a year from now whatever happens next. The movement is underneath it.

The Friction Matrix

Each app sits on two forces. Left to right is the movement: how the people writing reviews now rate the app against the people who wrote about it a year or more ago. Both are written reviews, so the comparison is like for like. Top to bottom is the response: how often the developer replies. On this cohort only the movement separates the apps, so two archetypes fall out.

No app in this cohort replies to more than 20% of its recent reviewers, so the empathy axis separates nobody and these apps are told apart by movement alone.

Ghost Ships4

Fallen against their own past. Recent reviewers rate the app below its own reviewers of a year ago.

Complacent Giants6

Steady against their own past. Recent reviewers rate the app about where its own reviewers of a year ago did.

Off the matrix: Things 3, too few recent written reviews; QuickBend, too few recent written reviews. They carry no movement or reply figure here.

Against their own reviewers of a year ago 4 of these apps have fallen and 2 have risen. 4 have held, and the lifetime average shows none of it. 4 apps are Ghost Ships, fallen against their own past. 6 apps are Complacent Giants, steady against their own past.

The movement, ranked

The same measure, app by app. Bars run left of the line where today's reviewers rate an app below its own reviewers of a year ago, and right where they rate it higher. The two n values on each label are how many written reviews the earlier figure and the recent figure each rest on, so a bar built on fifty can be read against one built on five hundred.

The anatomy of the drop

Behind the movement are recurring complaints. We classify recent reviews with a rule-based taxonomy and name the dominant patterns. These are illustrative archetypes from a biased sample, not a verdict.

4 apps have fallen more than half a star against their own reviewers of a year ago. All of them are below.

Grok AI

The Velvet Rope (29.9%)The Encroaching Paywall (21.7%)The Confident Liar (6.3%)

Waste of money. And they made the cancellation process an extreme pain to do. I hope someone files a lawsuit for this. Worst part was it wouldn’t generate a single image. Kept on saying, error please try again. I keep…1★ · 2026-07 · bug_integrity, monetization, systemic_core_value

There’s been this bug going on for the last month that waste your resources that you pay for by writing double of everything.1★ · 2026-08 · bug_integrity, monetization, systemic_core_value

Perplexity

The Encroaching Paywall (21.5%)The Confident Liar (13.8%)The Velvet Rope (13%)

If you don’t have credits or want to pay 200 for a couple monthly credits this is absolutely useless. I hate saying this but ChatGPT is even better and I truly hate chat gpt. Didn’t realize I was using credits for one…1★ · 2026-09 · systemic_core_value

Total garbage. Give it a task, fails every time. Give it blueprints it states the obvious and gives you slop, no answers, no solutions. Don’t waste your time or money!!!1★ · 2026-07 · monetization, systemic_core_value

Microsoft Authenticator

Locked Out at the Door (59.2%)The Vanishing Files (4.8%)The Confident Liar (2.5%)

Literally the worst app ever. Codes don’t work. A waste of time.1★ · 2026-07 · systemic_core_value

The worst app ever developed by Microsoft. It locks me out of my email constantly and especially when I upgrade my phone. I set everything up to use my Authenticator app and when the app doesn’t send a notification…1★ · 2026-08 · login_auth

Claude by Anthropic

The Confident Liar (14.6%)The Velvet Rope (9.5%)The Encroaching Paywall (9.5%)

Since the last update, the app crashes immediately on launch. Logo flashes, then straight to "Something went wrong" , no login screen, nothing. "Try Again" does nothing. Already tried: uninstall/reinstall, restart…1★ · 2026-07 · accuracy, login_auth, performance

Terrible user experience. The model suffered a severe Chain of Thought (CoT) leak, dumping its internal deliberations and defensive profiling of me directly into the chat. When called out on the bug, instead of handling…1★ · 2026-08 · bug_integrity, systemic_core_value

The corporate response

Developer replies are a proxy for how hard a team is fighting the friction. Across this category the reply share is about 0.2%, a median of 2 days after the review.

What it means

The lifetime star is the slowest number on the page to move, and the easiest to mistake for a signal. What moves is the mood of the people who write, measured against what that same group was saying a year earlier. The category answers about 0.2% of recent reviewers, a median of 2 days later.

For Productivity, 5 of the 10 apps with enough recent reviews to measure answer nobody at all.

Method and limits

• Ratings and star distribution are population truth.
• Recent average and response rate are from a biased sample.
• Movement over time is measured within one population only: an app's written reviews from the 90 days ending at its latest captured review (minimum 50) against its own written reviews from at least 365 days earlier (minimum 50).
• The histogram average and the written-review average are never subtracted from one another to claim a decline. The difference between them is a standing population offset, reported as written_vs_population and nothing more.
• Taxonomy is rule-based keyword/n-gram matching (v1, heuristic); buckets can overlap and some reviews are unclassified.
• No version-tied analysis: app_version is sparse and snapshots are not version-segmented; no claim links sentiment to a release.
• The reviewer-sentiment series (where shown) is sample-based and self-selection-biased; deep-backfilled apps only.
• Developer-reply latency uses the response last-edit date as a proxy for first reply.
• Reply quality (substantive vs boilerplate) is heuristic: boilerplate = a reply sharing a templated opening 8-gram with another reply. It catches macros, not a bespoke but empty non-answer.
• Review capture runs nightly and missed some nights; for busy apps the store may hold fewer reviews than were written in the window.
• Re-run at 60 and 120 days, no count changes.
• Quotes are short illustrative excerpts selected by polarity and length, not a representative sample.
• The free/deep split is structural; no payment gating exists yet.

Grounded in prior art on app-review mining and review selection bias:

The cohort

ChatGPT4.23 recent +0.12 vs a year ago n=3729 recentMicrosoft Outlook3.91 recent +0.56 vs a year ago n=1056 recentGoogle Drive3.79 recent -0.04 vs a year ago n=760 recentGmail2.33 recent -0.19 vs a year ago n=518 recentGoogle Gemini4.18 recent +0.01 vs a year ago n=3385 recentGrok AI3.7 recent -1.07 vs a year ago n=3291 recentDropbox3.39 recent +0.64 vs a year ago n=241 recentMicrosoft Authenticator1.55 recent -0.91 vs a year ago n=418 recentPerplexity3.51 recent -1.02 vs a year ago n=683 recentClaude by Anthropic3.03 recent -0.59 vs a year ago n=1791 recentThings 34.47 recent n=34 recentQuickBend4.5 recent n=2 recent

Independent research from the Nativerse lab. Population data from Apple's public ratings histogram. Movement is measured inside written reviews only: an app's reviews from the 90 days ending at its latest captured review, against its own reviews from a year or more earlier. Figures are cited, not invented.