One store cut 47% of its apps, the other added at record speed. Half of what we scraped is simply silent.

The three numbers

Google Play shrank from over 3.4 million apps in early 2024 to about 1.89 million by July 2026 — a cut of roughly 47% (Android Police). Over the same stretch, Apple moved the other way: the App Store added nearly as many new apps in the first half of 2026 as in all of 2025 (9to5Mac). Both stores describe what they are doing as a quality decision.

And here is a third number, measured on 9 September 2026 from our own review archive rather than from stored summaries: of the 27,392 apps we selected and scraped, 14,477 — 52.9% — could not clear a 100-review bar. 5,886 of them had zero usable reviews at all; another 8,591 had between one and 99. Only 12,915 apps, 47.1%, had enough written about them to analyse.

Those 14,477 are not broken apps. Most of them work. They are simply silent — and silence is invisible to any quality rule built on crashes and content checks.

What each store means by quality

Google's removals target apps with limited functionality and content — text-only or PDF-wrapper apps, single-wallpaper apps — along with apps that crash, freeze, or fail to offer an engaging experience (TechRepublic). Separately, Google says it blocked 1.75 million policy-violating apps from Play in 2025. Every one of those tests is a property of the binary: what the app contains, whether it runs, whether it behaves.

Apple's growth is the mirror image of the same definition. An app that installs, runs and meets review guidelines is a good app, so more of them is a better store. Neither definition asks whether a single person ever had anything to say about the app.

What our funnel measures instead

We only see an app through what its users wrote. That makes the 100-review bar a measurement of attention, not of engineering. An app can pass every technical check a store applies, ship on time, never crash — and still be unanalysable, because almost nobody wrote anything about it. 52.9% is not a defect rate. It is how much of the shelf is quiet.

That is the part neither store's rule can see. A purge based on crashes and static content does not touch the silent half; a store adding apps at record speed is adding into it.

The review side has its own funnel, and it is worth being exact about what it removes: of 8,204,639 reviews collected, 2,097,244 — 25.6% — were set aside by our quality rules before any analysis ran. Some of that is what you would expect: text under 15 characters, fewer than four words once emoji are stripped, keyboard-mashing, and word-for-word duplicates repeated across one app's review set. But the largest share is not junk at all — it is reviews written in a language we cannot score. We read English with VADER and five more languages with a multilingual model; everything outside that set is held back rather than guessed at. In a sample of the backlog, only 19.3% fell inside the languages we support and 80.7% did not. Producing a sentiment number with a model never tuned for the language would be a figure nobody could stand behind, so those reviews wait for a model rather than getting one.

What it means if you ship an app

If you are deciding where to launch, the two stores are asking you different questions. Play is currently pruning for functional substance, so a thin wrapper is a liability there in a way it was not two years ago. The App Store is not pruning at that rate, which means the shelf you are landing on is more crowded, and being findable matters more than being permitted.

Either way, the constraint that actually decides whether anyone can tell how your app is doing is review volume. Below roughly a hundred reviews there is no trend to read, no sentiment worth quoting, and no way to distinguish a bad month from noise — for you, for us, and for anyone comparing you to a competitor. Getting a real review count is not vanity work; it is the difference between having evidence and having a rating.

The limitation, plainly

Our 27,392 apps are apps we chose to scrape, and we select for apps that looked worth analysing. That is not a random sample of either store, and it is certainly not the whole of either store — we have harvested 66,722 App Store apps and 11,195 Play Store apps to date. So the honest reading of 52.9% is "of the apps we scraped", never "of the App Store". If the true population skews anywhere, it skews quieter than our sample, not louder: a random draw would pull in more obscure apps than a deliberate one does.