I run a visa agency in Bali, and the rest of my week goes into scrapers on Apify. One of them reads Apple's App Store: search results, app pages, developer catalogues and the top charts, one row per app with ratings, price, version history, size, languages and the developer's website.

There is no browser involved. Apple publishes a free JSON endpoint — the iTunes Search API at itunes.apple.com/search and /lookup — plus a chart feed at rss.marketingtools.apple.com. No key, no login, no Apple ID. That makes App Store metadata the easiest app data on the internet to collect, and it is exactly why people trust it more than they should.

Everything below comes from direct calls to those endpoints today, 20 November 2026, plus the last real runs of the Actor.

limit=20 returns 18

The search endpoint takes a limit parameter, documented up to 200. Here is what US search for vpn actually returns:

limit resultCount 20 18 30 27 50 46 200 189 

Five repeated calls at limit=30 gave 27 every time, same apps, same order — so this is not flakiness, it is the endpoint. Other terms behave the same way:

term limit=10 limit=50 limit=200 meditation 9 47 191 calculator 9 45 190 bank 9 49 198 duolingo 9 39 174 

None of the missing places are duplicates or broken rows: every result I got had a trackId, wrapperType: "software" and was unique. Apple seems to fill limit slots first and then drop entries it will not serve in that storefront. The shortfall is usually 5–10%, and for narrow terms like duolingo it reaches 22%.

Two consequences:

  1. You cannot get 200 apps for a search. The maximum is whatever survives Apple's filter — 189 for vpn. If your pipeline asserts len(results) == limit, it fails on every term.
  2. "Give me 20" needs a bigger ask. My Actor requests the wanted number plus at least 10 (or +30%) and trims back, which is why a recent run with "Apps per search: 30" delivered exactly 30 apps in 4.5 seconds. The run-summary row said charged app ×30 = $0.027.

The rank depends on the limit you asked for

This one surprised me. I took the vpn results at three limits and looked up where each app sits in the limit=200 list:

limit=10 → positions in the 200 list: 1 2 3 4 5 6 7 9 8 10 limit=20 → 1 2 3 4 5 6 7 9 8 10 12 14 16 15 19 18 17 21 limit=30 → 1 2 3 4 5 6 7 9 8 10 11 12 14 13 16 15 19 21 23 24 25 20 27 28 22 30 29 

At limit=10, "VPN - Free VPN Potato" is 8th and NordVPN 9th. Ask for 200 and they swap. Further down the list, the order shuffles more: at limit=30 the app in 22nd place of the big list shows up 25th. And at limit=20, places 11, 13 and 20 of the big list are simply not in the small one.

If you are doing app store optimisation and tracking "my position for keyword X", position is only comparable between reads that used the same limit. Pick one and never change it. The Actor writes the position it actually saw and the sourceValue (the search term) into every row, so at least the numbers carry their context — but they are Apple's positions for that request size, not the position a person sees on a phone.

Gibberish returns apps

I searched zzqxv, a string that matches nothing:

limit=10 → 6 apps (ZOX., جيل زد-GZ, ZBX robot, زوينه - Zweina, Zip-unzip Files opener, Zipym Lite) limit=50 → 14 apps limit=200 → 14 apps 

Apple does fuzzy matching on everything, so a typo in a keyword list does not produce an empty result — it produces plausible-looking rows about Arabic social apps and ZIP openers. Since the Actor charges per app row, a typo costs a few tenths of a cent and pollutes your dataset. If your keyword list comes from a spreadsheet, check the title column against the term before you aggregate anything.

The same app is a different row in every country

The country parameter does not just change the currency. Here is WhatsApp (id 310633997) in six storefronts:

country trackName rating ratings description chars us WhatsApp 4.68 18,986,745 1,924 gb WhatsApp Messenger 4.70 4,250,438 2,382 de WhatsApp Messenger 4.60 3,649,211 2,907 id WhatsApp Messenger 4.72 3,846,086 2,382 jp WhatsApp Messenger 4.54 266,854 1,376 br WhatsApp Messenger 4.78 8,697,806 2,831 

Same version (26.45.75) everywhere, but a different name in the US, a different rating in each country, and rating counts from 267 thousand to 19 million. App Store ratings are per storefront. There is no global rating, and adding the six counts together gives you a number Apple never shows anywhere. If you join apps across countries, join on appId or bundleId, never on title.

Prices are local too — not converted, set by the developer per tier. Procreate today:

us 12.99 USD gb 12.99 GBP de 14.99 EUR id 195000 IDR jp 1800 JPY in 999 INR br 79.90 BRL 

price is a bare number. 12.99 in the US and 12.99 in the UK are different amounts of money. Always keep currency next to it, which is why every row carries price, formattedPrice and currency together.

"Current version" ratings are a copy

The API exposes both userRatingCount and userRatingCountForCurrentVersion, and the same pair for the average. In theory, the second one tells you how the latest release landed. I looked up all 100 apps of the US top-free chart and all 100 of the Indonesian one:

us: ratingCount == currentVersionCount → 100 of 100 id: ratingCount == currentVersionCount → 100 of 100 

Apple fills the "current version" fields with the all-time numbers. My rows still carry ratingCurrentVersion and ratingCountCurrentVersion because that is what Apple returns, but do not build a "rating after release" chart on them. The only per-version signal in public data is the version field on individual reviews.

Charts are the clean part

The chart feed comes in exactly four sizes — 10, 25, 50, 100 — and all four returned the expected number of entries for the US and Indonesia today. Looking up the 100 chart ids in one /lookup call found 100 of 100 in both storefronts, so a chart read is two requests. The Actor fetches the next size up and trims, so "top 5" returns and charges 5 apps rather than 10.

The smallest rating count in the US top-free 100 was 143; in Indonesia, 716. New apps do reach the chart with a tiny rating base, so a 4.9 there can mean much less than it looks.

Reviews come from a different, moodier feed

Reviews are not in the Search API at all. They come from Apple's RSS customer-review feed: 50 per page, up to 10 pages, so at most 500 per app per country. Earlier this week that feed started returning empty pages to requests coming from Apify's own container addresses — no error, just a valid JSON document with no entries — and its cache sometimes served empty copies too. The fix in the Actor was to re-ask an empty page with a cache-busting query string, and after that through a proxy, before concluding the app has no reviews.

A real run from 18 November: search duolingo, two apps, 60 reviews each, most recent first. It returned Duolingo and Babbel plus 120 reviews in 11.7 seconds, and the summary row said $0.0078. All 60 Duolingo reviews were tagged version 7.149.0, while the app row said the live version was 7.149.1 — the newest release had not collected a single review yet. Of the 120, 58 were five stars and 28 were one star; App Store reviews are bimodal, and the average hides that.

Calling it

{ "searchQueries": ["vpn"], "apps": ["https://apps.apple.com/us/app/whatsapp-messenger/id310633997", "com.burbn.instagram"], "topChart": "top-free", "topChartSize": 25, "country": "id", "maxAppsPerSearch": 30, "includeReviews": false } 
import requests rows = requests.post( "https://api.apify.com/v2/acts/lergassy~app-store-scraper/run-sync-get-dataset-items", params={"token": "apify_api_..."}, json={"searchQueries": ["vpn", "meditation"], "country": "us", "maxAppsPerSearch": 30}, timeout=300, ).json() for r in (x for x in rows if x.get("type") == "app"): print(r["sourceValue"], r["position"], r["title"], r["rating"], r["ratingCount"], r["formattedPrice"], r["updatedAt"][:10]) 

Filter out type == "notice" (the free run-summary row) and type == "error" (an id or bundle id that does not exist in that storefront — also free, with the reason in the text).

The Actor is App Store Scraper: $0.001 per app, $0.00005 per review, no start fee.

When it is the wrong tool

  • Downloads and revenue. Apple does not publish them anywhere public. Anything that shows you "estimated downloads" is a model on top of chart positions, not scraped data.
  • Keyword rankings as a phone shows them. Search results on a device are personalised and include ads; the API returns its own ordering, which, as shown above, depends on the limit.
  • Full review history. 500 most recent per country is the ceiling. For Duolingo in the US that was about 60 hours of reviews when I measured it last month.
  • Rating after a release. The current-version fields are copies; you would have to rebuild it from review rows, and reviews are a biased sample of who bothers to write.
  • Global numbers. Every rating, price and even the app name is per storefront. If you need "worldwide", you are summing countries yourself, and you should say so in whatever chart you make.