I run a visa agency in Bali. The other half of my week goes into Apify Actors, and one of them reads Apple's public App Store review feed: stars, title, text, author, app version, helpful votes, date, in any of Apple's storefronts, plus one summary row per app.

My health check has run it against the same target since 17 October — Duolingo, US App Store, 100 most recent reviews. Thirty-one successful runs to 14 November, 2,427 unique reviews. I read the whole pile this weekend because a buyer asked a fair question: "can I watch my rating after a release?" The answer is no, and three other things I believed about this feed were wrong too.

Apple's "most helpful" list is not sorted by helpful votes

Every entry in the feed carries im:voteSum and im:voteCount — helpful votes and total votes. My rows carry them as helpfulVotes and totalVotes. So the "most helpful" list should be those numbers in descending order.

It is not. Here are the first fifteen entries of Duolingo's US most-helpful list, read today, as (helpfulVotes, totalVotes, date):

(0,0,2026-06-08) (0,0,2026-05-23) (0,0,2024-05-24) (0,0,2022-03-02) (0,0,2025-12-09) (0,0,2025-10-23) (0,0,2026-05-29) (0,0,2026-04-12) (0,0,2022-01-15) (0,0,2024-08-03) (0,0,2026-08-08) (0,0,2023-11-19) (0,0,2025-11-22) (0,0,2026-05-27) (8,12,2024-11-19) 

Fourteen reviews with no votes at all, then one with eight. The most-voted review in the entire list — 1,078 helpful out of 1,897 votes, written 28 September 2019 — is at position 58. Across all 500 entries, only 110 of 499 consecutive steps increase the vote count; the rest go down or stay flat, which is what you would expect from a list that is roughly ordered by something and not by this.

And 369 of those 500 "most helpful" reviews have zero helpful votes. In the most-recent list it is worse: 498 of 500.

The consequence is not that the sort is bad. It is that you cannot reproduce or verify Apple's helpfulness order from the data Apple gives you. Sort the rows by helpfulVotes yourself and you have replaced Apple's ordering with a column that is zero 95% of the time. Keep the arrival order — rows are pushed in feed order, and sortedBy says which list each came from.

One good surprise: the two lists do not overlap at all. 500 most-recent and 500 most-helpful for the same app, same storefront, read minutes apart — zero shared review ids, 1,000 unique reviews. That is why sortBy: "both" is worth the second pass.

"Most recent" is a day and a half old

In all 31 runs, the newest review in the dataset is older than the run itself by a lot:

hours between newest review and the run
minimum 26.5
median 32.0
maximum 37.4

Not once in four weeks did a run return a review written that day. The feed is a cache, and for this app it trails about a day and a half behind. If you are watching for the first angry reviews after a release, you are reading the day before yesterday.

The second number matters more. Those 100 "most recent" reviews span a median of 13.9 hours — in one run only 2.4 hours. Duolingo US produces roughly 173 text reviews a day, so 100 rows is about half a day of the app. The full 500-review list, read today, spans 13 to 15 November: 60 hours. That is the entire recent history Apple publishes for an app this size.

The gap between two daily runs is invisible and permanent

Here is the part that would have bitten me silently. In 24 of 30 consecutive run pairs, the two runs shared zero review ids. Not "mostly fresh" — no overlap whatsoever.

Zero overlap sounds like perfect efficiency. It is the opposite: it means the window moved further than it was wide, and the reviews in between were never read by anybody. Measuring each gap — the time between the oldest review of a run and the newest review of the run before it — gives 21 gaps totalling 400 hours, median 11.5 hours each.

2026-10-22 -> 2026-10-23: gap of 25.2h with no reviews read 2026-10-25 -> 2026-10-26: gap of 16.4h 2026-10-28 -> 2026-10-30: gap of 34.5h 2026-11-05 -> 2026-11-10: gap of 106.3h 

Some of that is my health check skipping days. Most of it is not: a daily run of 100 rows on an app with 173 reviews a day loses the difference every day, and because the ids never repeat, nothing in the output looks wrong. The rule is arithmetic: your row cap must cover more hours than your schedule interval. For an app this size, daily means 300 rows, not 100.

A duplicate review id across two runs is a good sign. It is the only proof you did not skip anything.

ratingCurrentVersion is a copy of the all-time rating

This is the one that cost a buyer their use case. Apple's lookup API returns four numbers in the app summary:

"rating": 4.72, "ratingsCount": 5589849, "ratingCurrentVersion": 4.72, "ratingsCountCurrentVersion": 5589849 

In 31 runs out of 31, across nine app versions (7.144.0 through 7.148.1, six of them shipping during the month), ratingCurrentVersion equalled rating and ratingsCountCurrentVersion equalled ratingsCount — to the last digit, every time. Same for WhatsApp in the UK (4.7) and in Germany (4.6), the other two apps in my logs.

Apple no longer resets its per-version rating; the field is a mirror. My README had a tip saying "schedule a daily run and compare ratingCurrentVersion". That tip was wrong and it is out of the README as of today.

What is left to watch is ratingsCount, which does move, and only in one direction:

5,508,161 on 17 October → 5,589,849 on 14 November: +81,688 ratings in 28.8 days, about 2,835 a day. The average stayed at 4.72 for the entire month — on a 5.5M denominator, a bad week cannot shift the second decimal. A rating alert on a big app is a thing that will never fire.

Those two numbers side by side are the real headline: 2,835 ratings a day, ~173 reviews with text a day. Roughly 6% of the people rating this app write anything, and the feed only publishes the written ones. Never describe a review sample as "what users think" when 94% of the signal is a star tap you cannot see.

Who writes is not who taps

The 2,427 reviews I collected average 3.96 stars. The store says 4.72. That is not a sampling error in my collection — it is the difference between people who write and people who rate:

stars share of 2,427 written reviews
5 58.8%
4 13.1%
3 8.0%
2 5.9%
1 14.3%

One in five written reviews is a 1 or 2 — for a 4.72-star app, that is the complaint channel, and maxRating: 2 returns it while charging nothing for the rows it filters out. The split between the two lists is sharper still: today's most-recent 500 is 64% five-star, the most-helpful 500 is 38% with a fat 3–4 star middle. Pull only "most helpful" and your sentiment analysis reads a different app.

Calling it

curl -X POST "https://api.apify.com/v2/acts/lergassy~app-store-reviews-scraper/run-sync-get-dataset-items?token=$APIFY_TOKEN" \ -H 'Content-Type: application/json' \ -d '{"apps":["570060128"],"countries":["us"],"sortBy":"both","maxReviewsPerApp":300,"maxRating":2}' 

apps takes app ids, App Store URLs or plain names; a name is only accepted when the App Store result contains every word you typed, otherwise you get a free error row instead of a wrong app. The same app given twice in three spellings is read and billed once. A real row, untouched:

{ "type": "review", "appId": "570060128", "appName": "Duolingo: Learn Languages", "country": "us", "reviewId": "14784212403", "rating": 2, "title": "Falta variedad", "text": "Falta variedad en las preguntas y conversaciones. También hay cosas que no me gustan y es obligatorio participar.", "author": "shergrim", "authorUrl": "https://itunes.apple.com/us/reviews/id148809972", "appVersion": "7.148.1", "helpfulVotes": 0, "totalVotes": 0, "date": "2026-11-13T05:04:39.000Z", "sortedBy": "mostrecent", "scrapedAt": "2026-11-14T20:08:48.762Z" } 

Note the language. The US storefront is not the English storefront — the storefront is where the account pays, not where the user lives. A hundred rows takes about 4 seconds; no login, no browser, no API key. It is App Store Reviews Scraper; the Android half of the job is Google Play Scraper.

Two smaller field notes. appVersion is the version the reviewer had: 16 of my 2,427 rows were written on 7.56–7.101, long before the month I sampled. And authors barely repeat — 2,418 distinct names in 2,427 reviews, eight people wrote twice.

When this is the wrong tool

  • You need the full review history. You get two lists of 500 per app per country. For a top-100 app that is a few days; for a small app it can be years. There is no pagination past it without a developer account on the app itself.
  • You need same-day reviews. The feed trails ~32 hours.
  • You need per-version ratings. Apple does not publish them any more, whatever the field name says.
  • You need to rank by helpfulness. The votes are zero on 95% of rows and Apple's own order ignores them.
  • You need a global number. Reviews are per storefront. The US list and the German list share nothing; "the app's reviews" is not a thing that exists.

The habit I am taking from this: before trusting any "latest" or "top" list from a public feed, read it twice on a schedule and check whether the two reads overlap. No overlap is not freshness. It is a hole in the middle of your data, and it will never show up as an error.