App Review Insights: Release Drops, Complaint Terms, Reply Rates

You scraped thousands of app reviews. Now what? This Actor reads a review dataset from your own Apify account and returns the answers as a small table: which release lost stars, which words only unhappy users write, how often and how fast the developer replies, and how the rating moves month by month.

It does no scraping and uses no AI model, so there is no extra source to break and the same dataset always gives the same result. Review texts and author names never appear in the output.

Typical uses: spot the release that caused a rating drop · see complaints per language without reading them · compare a competitor's reply rate with yours · a monthly app health report for a client.

Input example

{ "datasetIds": ["YOUR_REVIEW_DATASET_ID"], "minReviewsPerVersion": 30 }

Pick the dataset of an earlier run in the input form. Several datasets are merged, and the same review is counted once. Leave the field empty to see the output on a small fictional sample, free of charge.

Works with the output of Google Play Reviews Scraper and App Store Reviews Scraper, and with other review datasets that have a 1-5 rating field (rating, stars or score). Text, version, language, date and reply fields are recognised by their usual names; a missing field only removes the rows that need it.

Output

Row type	What it says
app-summary	Reviews, exact star counts, average, 1-2 star share, reply rate, median reply time, top languages
version	Per version: reviews, average, 1-2 star share, change against the previous version, regression flag
complaint-term	Words and two-word terms typical of 1-2 star reviews, per language, with exact review counts
month	Reviews, average and 1-2 star share per month

{ "type": "version", "app": "Example Notes", "version": "3.3.0", "reviews": 80, "avgRating": 3.138, "lowShare": 0.45,
  "previousVersion": "3.2.0", "lowShareChange": 0.325, "zScore": 4.54, "regression": true }

How the numbers work

• Counts are exact. Averages and shares are rounded to 3 decimals, scores to 2.
• regression is true when the 1-2 star share rose by at least 5 points and a two-proportion test gives z ≥ 2.58 against the previous version. Versions below minReviewsPerVersion are not compared.
• score on complaint terms is a log-odds z-score (Monroe et al. 2008): how much more typical a term is of 1-2 star reviews than of 4-5 star reviews. 3-star reviews are not used for terms.

Pricing

Pay per event: one charge per insight row. Your review count does not matter, and the fictional sample is free. maxRows caps every run, and app summaries and regressions are written first.

Honest limits

• The result describes your dataset, not the whole app. A dataset filtered by stars or cut by a limit gives a skewed star split. For fair comparisons, collect all ratings with sort "newest".
• People review with delay and may keep an old version number, so a release is a strong hint, not proof of cause.
• Stopwords exist for English, German, Spanish, French, Portuguese and Italian. In other languages some common words can appear as terms.
• Terms come from words, not meaning: "crash" and "crashes" are counted separately.
• Datasets must be selected in the input: this Actor cannot see anything else in your account.