The state of app UGC, in numbers

Ask our library how many apps it holds and you get four answers: 32,749, 32,751, 33,538 and 34,624. All four read the same table in the same harvest, and they disagree because they are computed and cached differently.

Every other article we publish links here for its method. This page carries the counts and the caveats; the articles carry the argument.

Each number below sits next to the denominator it was computed against. A count with no denominator is not a fact, and a share with no denominator is a decoration.

Where two of our own counters return different answers for the same quantity, both are printed and neither is called the right one.

Two calls count apps and two calls count creators, and neither pair agrees. The library counter reads 32,749 apps and 204,953 creators. The coverage counter reads 34,624 apps and 207,516 creators.

The gap between the coverage counter's 34,624 apps and the library counter's 32,749 is 1,875, 5.7% of the smaller of the two. The creators gap, 207,516 against 204,953, is 2,563, 1.25%. Both gaps point the same way, with the coverage counter larger on both entities.

The two files were fetched three seconds apart, so this is not the library growing between two readings. It is two counters answering at the same moment with different numbers, and the likely difference between them is age.

The coverage payload stamps itself: generated at 04:29:38 UTC, sixteen minutes before we read it. The library counter is cached for six hours and carries no stamp at all. Its value may have been computed at any point in those six hours.

We cannot prove that from two files. Cache age and real growth over an unknown interval leave the same evidence, and only one of the two counters can be dated at all. What we can give you is the rule we apply when they conflict.

Headline totals on this page come from the library counter. Every share of apps takes its denominator from the coverage counter, 34,624, and every share of creators from 207,516. Those are the counters that produced the numerators. Mixing them would report a share of a count that was never taken.

Apps have two further counters. The category facet over in-library apps sums to 32,751, and the category facet for the apps scope sums to 33,538. Four calls, four answers, spread across 1,875 apps.

A fifth counter is worse and we print none of its output. Browse counts stop at 10,001, a ceiling that means "10,000 or more" and nothing else. Where a filter saturates that ceiling, this page omits the figure.

Every video is classified by a language model into a fixed vocabulary of formats, hooks, tones and CTAs. It is a classification, not a human tag, and it is spot-checked, not formally benchmarked. We publish it as a distribution across hundreds of thousands of posts, never as a judgement about any single video.

419,998 of the 565,898 videos in the library carry a full label set, 74.2%. Hook, call to action, content type and tone all resolve to that same 419,998 videos, which is the denominator under every table here.

The vocabularies are closed enums enforced on the model's output. Two of them still carry a small tail of free text from older, looser runs. That is 1,807 extra rows on content type and 938 on tone, and both tails are excluded from the shares below.

In the 30 days to 14 August 2026, 59.3% of the TikToks our classifier read carried no call to action. On Instagram Reels the figure was 39.0%. Those are exact counts over 16,432 and 6,529 classified rows.

That number is partly a measurement limit. Posts where we resolved a transcript show a no-CTA share 15 to 21 points lower than posts where we did not. Nothing in the pipeline reads text shown on screen.

The gap runs the opposite way to the one you would expect from the pipelines. On TikTok we read the caption and the spoken transcript, and on Instagram the caption alone. A caption is typed, so the words link in bio are there to read; speech often omits the ask the video shows on screen.

The pooled all-time figure sits between the two platforms, as a mix should: 197,558 of 419,998 classified videos, 47.0%. Among the posts that do carry an ask, link_in_bio outnumbers download_app by 2.2 times, 87,004 against 40,103.

Ads are the newer library by a distance. 231,885 of 245,937 started inside 180 days, 94.3%, and 241,928 started inside a year, 98.4%. That leaves 4,009, 1.6%, outside the year — a residual between two counters computed differently, not a count of ads with a start date older than a year.

Organic reaches further back. 498,736 of 565,898 videos were published inside the year, 88.1%, leaving 67,162 older than it. Ad campaigns end and drop out of the Ad Library; a video published two years ago stays where it was posted.

Shopping is the eighth largest category by app count, at 1,895 apps. It is the largest by content in all three other columns: 106,755 organic posts, 53,341 creators, 30,328 ads. That is 5.8% of the apps against 17.3% of the organic column.

The last three columns are link counts and not entity counts. A video matched to two apps in different categories is counted once in each. The organic column sums to 618,636 against 565,898 videos, and the creator column to 259,695 against 204,953 creators.

The ads column runs the other way: it sums to 230,607, below the 245,937 ads we hold. Two effects pull against each other. Multi-matching inflates a column, and content with no matched app deflates it, because an ad with no matched app appears in no category at all.

For organic, almost everything is matched, so inflation wins and the column overshoots. For ads, enough ads have no matched app that deflation wins.

The 15,330 shortfall is a lower bound on unmatched ads and not a count of them. The two effects cannot be separated from two numbers. Read a cell as ads linked to apps in that category, never as a slice of 245,937, and never subtract one from the other.

No ad spend. We do not have it, and for ordinary commercial ads nobody does. Meta publishes spend only for political and issue advertising: 12 of 1,164 ads in one production sample. Any spend figure attached to an ordinary Meta creative is modelled, because the source figure does not exist.

No usage data of any kind. No daily or monthly actives, no retention, no session length, no time in app. No source in our pipeline carries any of it, so nothing on this page or in the product is derived from it.

No ratings history and no chart rank. Ratings and review counts are the latest scraped values, and each refresh writes over the last one, so there is no rating trend to plot. We store which category an app is in and not where it ranks inside that category.

TikTok ads and YouTube are not collected. Every ad figure on this page is Meta. Every organic figure is TikTok or Instagram. Neither gap is a sampling decision — no collector exists for either, so an article about TikTok ad creative is one we cannot write.

Instagram is mostly classified without a transcript. On TikTok we read the caption and the spoken transcript. On Instagram we read the caption alone, apart from one rescue path. 51 of the 586 classified Instagram rows in our sample did carry a resolved transcript. All of them came through that one path, and on the rest a call to action spoken aloud is invisible to us.

Nothing in the pipeline reads text shown on screen. One text-only pass runs over the caption and the transcript. An end card reading download now is invisible on both platforms, which puts a floor under every no-CTA count on this page.

A quarter of the organic corpus carries no labels. 419,998 of 565,898 videos have a hook and a call-to-action label. The other 145,900, 25.8%, were never analysed and the backfill script has never been run. Every facet share here is a share of the analysed subset and says nothing about the rest.

The newest content is the least labelled. A probe of the 1,995 newest videos across both platforms found 13 carrying an analysis row. Labelling lags collection, so a share computed today describes work finished weeks ago and not what was posted this week.

Category figures only see content we matched to an app. Category filters resolve through the matched app. A post we could not tie to an app is invisible to every category number on this page. Only 8,938 of 34,624 apps have any organic content at all.

Content is linked to apps by store links where we can find them and by name matching otherwise. Name matching is deliberately generous and it is not perfect. Our own log records videos about a sunset and a yawning rabbit matched to a meditation app on the shared word "calm".

The library is a sample, and not a random one. It grew by our customers asking about their competitors. A category that looks large here is a category we collected heavily. Comparing two categories on this page compares our collection before it compares the market.

Download and revenue figures are third-party estimates, not figures reported by the developer. They are rounded at source, so month-to-month movements of a million are rounding and not growth. 14,498 of 34,624 apps carry any estimate, 41.9%, and June 2026 was the most recent month available on the day of this count.

Ad reach is mostly missing. Impression buckets appear on roughly 4% of ads, the political, issue and EU-delivered inventory Meta reports on. Ad country and audience demographics are not stored at all. What the ad data does carry is the creative, the copy, the destination, the run length and the variant count.

Virality is our own score. It is Apptonic's 0–100 number, weighted toward view-to-follower breakout and recency. It measures what is travelling now, not lifetime performance, and it is a ranking signal, not a platform-reported metric. Its classes are percentiles of a moving distribution, so an "exploding" share from July and one from August are answering different questions.

Video view counts are as-of-last-scrape. Most videos were collected once and never collected again, so there is no view history behind any figure here. Instagram reports no meaningful share count, which is why share and save numbers are never pooled across the two platforms.

Creator figures are library figures. Follower counts come from a profile fetch that has not reached the long tail, and rows without one carry a null. The same person posting on TikTok and on Instagram is two rows on this page, because we do not stitch identity across platforms.