Google Play Release Regression Monitor

Catch broken Android releases before ratings collapse.

Turn public Google Play reviews into an early-warning system for mobile releases. Schedule one Actor to track app versions and Google Play request contexts, surface only new feedback, and flag rating drops, 1–2 star spikes, crash/login/payment complaints, and multi-signal release regressions.

Built for release decisions—not bulk review scraping. Get deterministic incidents, version analytics, explicit review-scope metadata, and stateful change detection without a browser or paid AI API.

See the value in one run

Keep the prefilled Spotify input and click Start. The first successful run creates a baseline; scheduled runs then return only new reviews and newly active incidents. One app across US, GB, and DE request markets uses three app_market_checked events.

After scope validation, a review-derived incident is explicitly unscoped while retaining the gl request context:

{
  "recordType": "INCIDENT",
  "eventType": "VERSION_REGRESSION",
  "severity": "HIGH",
  "appId": "com.spotify.music",
  "appName": "Spotify",
  "country": null,
  "requestedCountry": "us",
  "reviewScope": "UNSCOPED_PUBLIC_FEED",
  "language": "en",
  "version": "9.1.78.2218",
  "previousMetric": 0.2222222222222222,
  "currentMetric": 0.3333333333333333,
  "delta": 0.1111111111111111,
  "increasePercent": 50,
  "evidence": {
    "previousVersion": "9.1.76.2055"
  },
  "detectedAt": "2026-08-29T10:47:31.325Z"
}

Who it is for

• Mobile developers and QA teams checking whether a production release escaped with defects.
• Product managers monitoring post-release feedback across Google Play request contexts without assuming reviewer residence.
• ASO agencies and mobile publishers monitoring app portfolios.
• Competitive-intelligence teams tracking the release quality of other public Android apps.

What it detects

Release monitoring

VERSION_REGRESSION, NEGATIVE_REVIEW_SPIKE, RATING_REGRESSION, and the combined RELEASE_REGRESSION identify meaningful deterioration with minimum-sample guards.

Explicit review scope

Google Play's public review RPC accepts gl, but gl alone did not reliably isolate review feeds in validation. Review-derived records therefore use country: null, preserve the request as requestedCountry, and report reviewScope: UNSCOPED_PUBLIC_FEED. COUNTRY_COMPARISON and COUNTRY_REGRESSION are disabled. Identical feeds and incident keys returned through multiple requested countries are emitted only once, and REVIEW_SCOPE_NOTICE explains the limitation.

Crash, login, payment, and update complaints

English reviews are classified with transparent phrase dictionaries into:

• CRASH_COMPLAINT
• LOGIN_COMPLAINT
• PAYMENT_COMPLAINT
• PERFORMANCE_COMPLAINT
• UPDATE_REGRESSION
• FEATURE_BROKEN
• UI_UX_COMPLAINT
• ADS_COMPLAINT
• NOTIFICATION_COMPLAINT

Resolution phrases such as “fixed”, “resolved”, and “works now” suppress obvious false positives. No LLM, embedding service, or paid AI key is involved.

Example input

{
  "apps": [
    { "appId": "com.spotify.music", "name": "Spotify" }
  ],
  "countries": ["us", "gb", "de"],
  "languages": ["en"],
  "maxReviewsPerAppCountry": 100,
  "onlyChanges": true,
  "detectRegressions": true,
  "compareVersions": true,
  "minimumReviewsForRegression": 5,
  "negativeRatingThreshold": 2,
  "regressionThresholdPercent": 50,
  "includeReviewText": true
}

appId is the Android package ID from the Google Play URL. The Actor validates malformed IDs, two-letter country/language codes, thresholds, and impossible sample-size settings before making requests.

First run and scheduled runs

The first successful run for each app × country × language creates a baseline in the named google-play-regression-monitor-state-v1 Key-Value Store, which is retained across scheduled Actor runs. Its MARKET_SUMMARY contains baselineRun: true; historical reviews are not called new regressions. With onlyChanges: true, the baseline emits summary/comparison records but does not dump the initial review window.

On later runs, the Actor:

1. Loads the last good market state.
2. Fetches newest reviews first and deduplicates native Google review IDs.
3. Stops pagination after 25 consecutive known IDs or the configured review limit.
4. Merges the result into a bounded recent history window.
5. Calculates rating, negative-rate, version, country, and complaint metrics.
6. Writes new reviews, newly active incidents, and comparison records.
7. Commits state only after extraction, validation, analytics, and Dataset writes succeed.

If a regression stays active, it is not repeated every run. If it resolves and later returns, it can be emitted again.

Output records

The default Dataset contains typed records:

• REVIEW — a newly discovered review (or fetched review when onlyChanges: false). A new rating at or below the threshold has eventType: NEW_NEGATIVE_REVIEW.
• INCIDENT — a deterministic regression or release-health event.
• MARKET_SUMMARY — baseline flag, fetched/new counts, recent rating metrics, missing-version counts, and store rating.
• VERSION_COMPARISON — per-version sample size, average rating, negative rate, and signal rates.
• REVIEW_SCOPE_NOTICE — requested countries, duplicate-feed evidence, and why country comparison is unavailable.
• MARKET_ERROR — an isolated, actionable failure that did not modify that market’s state.

Example incident:

{
  "recordType": "INCIDENT",
  "eventType": "RELEASE_REGRESSION",
  "severity": "CRITICAL",
  "appId": "com.example.app",
  "country": "us",
  "language": "en",
  "version": "8.4.1",
  "evidence": {
    "previousVersion": "8.4.0",
    "previousNegativeRate": 0.08,
    "currentNegativeRate": 0.42,
    "previousAverageRating": 4.4,
    "currentAverageRating": 2.9,
    "topComplaintSpikes": [
      { "signal": "CRASH_COMPLAINT", "previousRate": 0.02, "currentRate": 0.27 }
    ]
  }
}

The OUTPUT Key-Value Store record summarizes successes/failures, fetched/new reviews, missing versions, incident counts, top new-review signals, requests, runtime, and memory.

Regression heuristic

Rate regressions require all of the following:

• both compared groups meet minimumReviewsForRegression;
• the relative increase meets regressionThresholdPercent (50% by default);
• the absolute increase is at least 10 percentage points.

A rating regression additionally requires a drop of at least 0.5 stars. A RELEASE_REGRESSION requires at least two version indicators: negative-rate regression, rating deterioration, or a complaint spike. These rules are deterministic and intentionally conservative.

Extraction approach

The Actor wraps @mradex77/google-play-scraper behind its own Google Play adapter. The maintained library uses Google Play’s public HTTP batchexecute review endpoint and public app details page; it does not launch Chromium or require a Google account. Its responses are runtime-validated and pagination-token cycles are detected. The Actor adds its own normalization, stable fallback hashing, capped retry/backoff with jitter, per-market isolation, response-integrity guards, deduplication, early stopping, and state commits.

Requests use a shared two-requests-per-second limiter. HTTP 408/429/5xx, rate-limit, network, blocked, and timeout failures receive at most four attempts with capped exponential backoff. Exhausted failures create MARKET_ERROR and preserve the prior state.

Country and language semantics

country is the Google Play gl request market. It is not verified reviewer residence. Google decides which localized review selection a signed-out caller sees, and caller IP can also affect regional availability. Country result sets can overlap, differ, or occasionally be identical. The Actor preserves market context and never claims it proves where a reviewer lives.

In the August 29, 2026 live validation from one Turkish egress IP, hl=en returned 99.5–100% overlap between US, GB, and DE for Spotify and 100% for Signal. This proves that gl alone did not provide reliable country isolation in that environment. A country-matched proxy/egress may be needed when strict geographic differentiation is a business requirement.

language is the Google Play hl request language. It influences the returned/localized selection but is not a guaranteed detected language for every review. Star/rating analytics work in every language. Semantic complaint classification is currently English-only; other review records have signalClassificationSupported: false.

Version limitations

Google Play frequently exposes the app version associated with a review, but not for every review. Missing versions remain null; they are never inferred. Overall and country analytics include those reviews, while version analytics use only reviews with a public version. reviewsWithVersion and reviewsWithoutVersion make coverage visible.

The store’s displayed “current version” may be “Varies with device”. In that case it is stored as null, and NEW_APP_VERSION is not generated.

Completeness and pagination limitations

Google Play review feeds are ranked/localized public snapshots, not an official complete historical archive. New reviews can change ordering between requests, continuation tokens may expire, and the same review can appear in overlapping pages. The Actor aggressively deduplicates and optimizes recurring monitoring, but does not promise complete historical coverage.

An app page that exists but produces an empty or malformed review response is treated as a failure—not a successful zero—so broken extraction cannot silently replace a good baseline.

Privacy and compliance

Only public, signed-out Google Play data is requested. The Actor does not access private accounts, bypass authentication, or attempt to identify reviewers. Reviewer names can appear in output because Google publishes them, but persistent monitoring history removes review text, reviewer names, and developer-response text; state is for release metrics and deduplication, not user profiling.

Cost and Pay Per Event readiness

The commercial unit is app_market_checked: one successful app × country × language check. The source includes an opt-in Actor.charge() integration using that event name, disabled until the Store publisher configures an exact event price and sets ENABLE_PPE_APP_MARKET_CHECKED=true. This avoids inventing pricing or breaking free/local runs. Failed markets are not charged.

No external secret or proxy is required for ordinary runs. Direct public HTTP is the default; Google may rate-limit large/high-frequency jobs, so keep schedules and market counts reasonable.

Development

Requires Node.js 22.12 or newer.

npm install
npm run typecheck
npm test
npm run build
npm run live:smoke

Local Apify input belongs in storage/key_value_stores/default/INPUT.json. The smoke script tests Spotify and Signal against US, GB, and DE and performs an identical repeated request to report ID overlap.

The Apify SDK purges local default storage at process start unless configured otherwise. To validate state across two manual npm start invocations, set CRAWLEE_PURGE_ON_START=0 for the second run. Hosted scheduled runs retain the default Key-Value Store normally.

Store metadata

Name: Google Play Release Regression Monitor

Positioning: Release-health monitoring for Android teams—not a bulk historical review scraper.

Description: Catch broken Android releases before ratings collapse. Monitor Google Play reviews by app version and request market for rating drops, complaint spikes, and regressions.