Google says Play Store's machine learning-based system removed millions of fake reviews and ratings and identified thousands of bad apps during a recent week
Posted by Fei Ye, Software Engineer and Kazushi Nagayama, Ninja Spamologist — Google Play ratings and reviews are extremely important …
Context & Ripple Effects
This post is an early data point in a decade-long escalation of Google's automated Play Store enforcement. Earlier in 2018, Google credited new machine learning techniques for removing 700,000+ policy-violating apps in a year, and its Android security review reported that 60.3% of potentially harmful apps were caught by ML rather than human review.
What changed here is granularity: instead of annual totals, Google is reporting what the system does in a single week — millions of fake reviews and ratings stripped and thousands of bad apps flagged. That weekly cadence foreshadows the scale the pipeline later reached, including 2.28M apps rejected and ~333K developer accounts blocked in 2023, before settling at 1.75M rejections in 2025.
First-order effects
- Apps propped up by purchased or bot-generated reviews lose their inflated ratings within days, directly changing their visibility in Play Store search and charts.
- Developers running legitimate apps see a cleaner ratings signal, while the thousands of identified bad apps move into Google's removal pipeline.
Second-order effects
- Review-farming and app-store-optimization services that sell fake ratings face a shrinking market on Play, pushing manipulation toward stores with weaker automated moderation.
- Rival app marketplaces come under pressure to match Google's ML-driven review filtering, since star ratings are the primary purchase signal users compare across stores.
Third-order effects
- Store integrity becomes an arms race between classifier pipelines and manipulation networks, with enforcement economics favoring platforms that can automate detection at weekly scale rather than rely on human review.
- As automated rejection grows from hundreds of thousands of apps a year to millions, developer recourse and appeal processes become the structural bottleneck — a governance question regulators may eventually probe.
The trend: App-store trust enforcement is consolidating around machine-learning pipelines whose throughput has scaled from hundreds of thousands of yearly removals to millions of weekly actions.