What actually performs on short-form: 4,404 TikTok posts, measured
Most advice about short-form video is asserted rather than measured. This page is the opposite: 4,404 public TikTok posts, harvested with their real engagement, across nine business verticals. It exists to be quoted, checked and argued with — so the method and the sampling limits are on it, not in a footnote. Every figure here is recomputed from the source files by a script, not typed.
Key takeaways
- The median post gets 30,000 plays. Not the average — the median, which is the number that actually describes a typical post once the outliers are set aside.
- The distribution is brutally top-heavy. The 90th percentile is 1.2 million plays, forty times the median, and the top post in the set has 84.1 million.
- 490 of 4,403 posts cleared a million plays — about eleven in a hundred. A million-view post is rare but not exotic, which is the fact most "go viral" advice gets wrong in both directions.
- Likes track plays much more steeply. Median likes are 1,509, so the typical post converts roughly one viewer in twenty into a like.
- The sample is hashtag-drawn and TikTok-only. That is a real bias and it is described below rather than glossed over. Treat these as benchmarks for hashtag-discoverable business content, not for all of short-form.
On this page
What are the headline engagement numbers?
Across 4,403 public TikTok posts carrying metrics, median plays are 30,000, the 90th percentile is 1.2 million, and the highest single post recorded 84.1 million plays.
| Measure | Median | 90th percentile | Maximum |
|---|---|---|---|
| Plays | 30,000 | 1.2 million | 84.1 million |
| Likes | 1,509 | 97,700 | 6.8 million |
Two things are worth drawing out of that table, because they are the ones that change what you do on Monday.
The gap between the median and the 90th percentile is the whole story. 1.2 million is forty times 30,000. In a distribution shaped like that, the arithmetic mean is close to meaningless — it describes neither a typical post nor a good one — which is why every figure on this page is a median or a percentile. Anyone quoting an "average views" number for short-form is quoting a number that describes no real post.
A million plays is roughly a one-in-nine event, not a lottery win. 490 posts out of 4,403 cleared it. That reframes the strategy in a specific way: if a bit more than one post in ten in this sample breaks a million, then the number of posts you make is a larger lever than the polish of any one of them. It also means the reverse of the usual advice is true — the goal is not to make a viral video, it is to take enough shots that the distribution does the work.
How was the corpus collected, exactly?
Posts were sampled by hashtag across nine business verticals on TikTok, with engagement metrics captured at harvest time rather than estimated, giving 4,404 rows of which 4,403 carry usable metrics.
The method matters more than the numbers, so here it is in enough detail to be criticised.
- Platform. TikTok only. Nothing here is drawn from Reels, Shorts or anywhere else.
- Sampling frame. Hashtags associated with each of nine business verticals. Posts were collected as they appeared under those tags.
- Metrics. Play count, like count, comment count, share count and collect count as reported by the platform at the moment of harvest. They are a snapshot, not a final tally — a post collected while still circulating will have grown since.
- Size. 4,404 rows. 4,403 carry metrics; one row is an empty object and is excluded from every percentile on this page.
- Percentile method. Values sorted ascending, index taken as
floor(fraction × (n − 1)). Stated because a percentile without its method is not reproducible, and the choice moves the answer.
An earlier Surge post published the 90th percentile for likes as 98,200. The correct figure against the rows that carry metrics is 97,700. Taken over all 4,404 corpus rows. One row is an empty object; sorted, its undefined metric lands last and pushes the percentile index up by one. Against the 4,403 rows that carry metrics the value is 97,700.
It is a difference of about half a percent and nobody would have caught it. It is published because a page whose argument is "check our numbers" earns nothing by quietly fixing the one that was wrong.
What does the data look like by industry?
The nine verticals are unevenly sampled — SaaS accounts for roughly half the corpus at 2,193 posts while e-commerce has 60 — so per-vertical figures are only meaningful for the larger groups.
| Vertical | Posts | Share of corpus | Usable on its own? |
|---|---|---|---|
| SaaS and software | 2,193 | 49.8% | Yes |
| General / mixed | 671 | 15.2% | Yes |
| Home services | 620 | 14.1% | Yes |
| Beauty and skincare | 320 | 7.3% | Yes |
| Dental | 160 | 3.6% | Yes |
| Food and hospitality | 160 | 3.6% | Yes |
| Real estate | 140 | 3.2% | Too thin — read with the whole corpus |
| Fitness | 80 | 1.8% | Too thin — read with the whole corpus |
| E-commerce | 60 | 1.4% | Too thin — read with the whole corpus |
The imbalance is the honest headline here. SaaS at 2,193 posts is close to half of everything, and e-commerce at 60 is not a sample you should draw a conclusion from. Publishing a confident "e-commerce benchmark" off sixty posts would be the single easiest way to make this page untrustworthy, so it is not published.
Six of the nine clear a 150-post floor and can carry a claim of their own. Three do not, and are marked as such. If you are reading this because you want a benchmark for your own vertical, the useful advice is: use the whole-corpus medians as your baseline, and treat your vertical's row as a directional hint rather than a target.
What can this data not tell you?
It cannot tell you about non-hashtag discovery, other platforms, paid distribution, or whether any of these posts converted — it measures reach and reaction only, on one platform, among posts that used hashtags.
Every dataset has a shape, and citing one without its shape is how a reasonable number becomes a wrong conclusion. Four limits, stated plainly.
It is hashtag-drawn, so it over-represents hashtag-discoverable content
Posts that reach an audience purely through the For You page without meaningful tagging are under-represented by construction. If your strategy does not lean on hashtags, these distributions describe a slightly different population from yours.
It is one platform
TikTok only. Instagram Reels and YouTube Shorts have different distribution mechanics, different audience behaviour and different typical view counts. Nothing here transfers to them without argument, and this page will not make that argument for you.
It measures attention, not money
There is no conversion, revenue or attribution data in the corpus, because a public post does not expose any. A post with 84.1 million plays may have sold nothing. Anyone telling you a view count is a business outcome is skipping the step that matters.
The metrics are a snapshot
Engagement was captured at harvest. Posts still circulating at that moment have grown since, which biases the figures slightly downward. The direction of that bias is known; its size is not.
None of these undermine the headline distributions — a forty-fold gap between median and 90th percentile is not an artefact of hashtag sampling. They do mean the numbers are benchmarks for a defined population rather than universal constants, and that is the correct way to use them.
How does Surge use this corpus?
Surge learns structural patterns from high-performing posts — how the opening seconds set up tension and how the payoff lands — and rewrites those shapes onto your offer rather than copying any individual post.
The corpus is not a swipe file and Surge does not reproduce posts from it. What a large sample of measured posts is genuinely good for is separating the structural features that recur among high performers from the ones that merely feel important.
Concretely, it informs how Surge builds the opening beat, how long a thought runs before the payoff, and which of the three formats — meme ad, slideshow, or creator post — suits a given business. Those decisions used to be taste. With 4,403 measured posts behind them they are still partly taste, but they are taste with a denominator.
It also sets what Surge deliberately does not chase. A sound that peaks on a Tuesday is dead by Friday, and building a product around riding trends late is building a product around being slightly too slow. The durable signal in a distribution this top-heavy is structural, not topical.
What we measured about our own output
One number about Surge itself belongs on a page about evidence: across 144 recorded human keep-or-kill judgements on generated ads, 39.6% were kept. That is not a flattering number and it is not presented as one. It says that roughly two in five generated ads are worth posting in a human's opinion, which is precisely why the product is built as a swipe queue rather than a single-shot generator: the workflow assumes you will reject most of them, and gives you 10 a day to reject from.
See the patterns applied to your business
Paste your URL. Surge reads your site and builds 10 finished ads on it, free, no card.
Get 10 free ads →Can you cite these figures?
Yes — the figures on this page may be quoted freely with a link back, and they carry the date they were computed so a future reader can tell whether they have moved.
Quote them. Link back, name the date, and state the sample size — "30,000 median plays across 4,403 public TikTok posts (Surge, 23 August 2026)" is a citation that will still be defensible in a year, because it carries everything a reader needs to judge it.
If you find a figure here you cannot reproduce, or the corpus moves under this page, tell us. The correction above is on the page because that is the standard being claimed, and it applies to anything else found the same way.
Frequently asked questions
What is the average number of views on a TikTok post?
Across 4,403 public TikTok posts measured in August 2026, the median was 30,000 plays. The median is the right figure to use rather than the mean, because the distribution is extremely top-heavy: the 90th percentile is 1.2 million plays, roughly forty times the median, so an arithmetic average describes neither a typical post nor a successful one.
How many TikTok posts get a million views?
In this sample, 490 posts out of 4,403 cleared one million plays — roughly one in nine. That makes a million-view post rare but not exceptional, and it is the single most useful reframing in the data: at those odds, the number of posts you publish is a bigger lever than the polish of any individual one.
How many posts is the Surge corpus based on?
It contains 4,404 harvested rows, of which 4,403 carry engagement metrics. One row is an empty object and is excluded from every percentile published here. The posts span nine business verticals, sampled by hashtag on TikTok.
Which industries are in the data?
Nine verticals: SaaS and software with 2,193 posts, general or mixed with 671, home services with 620, beauty and skincare with 320, dental with 160, food and hospitality with 160, real estate with 140, fitness with 80, and e-commerce with 60. The sampling is deliberately uneven and the three smallest are marked as too thin to support a standalone claim.
What are the limitations of this data?
Four, stated on the page. It is hashtag-sampled, so it over-represents content that is discoverable by tag. It covers TikTok only and does not transfer to Reels or Shorts without argument. It measures reach and reaction but contains no conversion or revenue data, so a view count here is not a business outcome. And metrics were captured at harvest time, so posts still circulating have grown since, biasing figures slightly downward.
Can I cite or reuse these statistics?
Yes. Quote them freely with a link back, and include the date and the sample size so the claim stays checkable — for example, 30,000 median plays across 4,403 public TikTok posts, Surge, 23 August 2026. If a figure cannot be reproduced or the underlying corpus moves, email support@clipspeed.ai and it will be corrected on the page.
Why was the 90th-percentile likes figure corrected?
An earlier Surge post published it as 98,200, computed across all 4,404 corpus rows. One of those rows is an empty object, and when the values are sorted its missing metric lands last and pushes the percentile index up by one position. Against the 4,403 rows that actually carry metrics the figure is 97,700. The difference is about half a percent and would not have been noticed, which is exactly why it is published rather than quietly amended.
What percentage of AI-generated ads are worth keeping?
Across 144 keep-or-kill decisions on Surge-generated ads, 39.6% were kept — roughly two in five. The denominator matters and is easy to overstate, so plainly: these are our own decisions made during internal testing, recorded across five days, and zero customer swipes have been recorded to date. It is a small internal sample, not a benchmark. It is published because it is the reason Surge is built as a swipe queue you reject from rather than a single-shot generator.
Keep comparing
Published 2026-08-23. Last checked against every cited source on 2026-08-23.
Figures about Surge are recomputed from source by surge-seo/verify-facts.js.
Found something out of date? Tell us and we will correct it.