TikTok completion floors by clip length
Completion rate means nothing without a duration band. Floor targets per band, how to build your own from 60 posts, and when shortening beats fixing.
"Our completion rate is 38%" is not a sentence that carries information. It is two numbers pretending to be one: a measure of how well the clip held people, and a measure of how long the clip was. On a 12-second post, 38% is a disaster. On a 75-second post it is comfortably above the working floor. Reported as a single channel-wide figure, the two cancel out into something you cannot act on.
This matters more than it sounds, because completion rate is doing real work in TikTok's ranking. Watch time and completion together are commonly estimated to account for something like 40–50% of the ranking decision — a practitioner estimate rather than a disclosed weight, so hold it loosely, but nobody serious argues the signal is minor. If it is that load-bearing and you are measuring it wrong, you are optimising against a corrupted read.
Why a channel-wide average lies
Completion is mechanically length-dependent. Every additional second is another opportunity to lose someone, so the same quality of content produces a lower completion figure at 60 seconds than at 15. That is not a finding about your content; it is arithmetic.
Which means a channel-wide completion average moves whenever your content mix moves, independent of quality. Publish a month of 15-second clips and the average rises. Publish a month of 70-second explainers and it falls. If you are tracking that average month over month, you will read mix shift as a performance change and act on it. Accounts do this constantly: a genuinely successful pivot to longer, denser content shows up as a "declining completion rate" and gets reversed.
The fix is to stop reporting one number. Bucket by duration band and report a completion figure per band. Nothing else in your analytics changes; you have just stopped averaging across a variable that dominates the metric.
Starting floors, by band
These are the retention floors that circulate among people who work on TikTok performance. They are expressed as average watch time as a share of clip length, which is the closest thing to a comparable figure across bands.
| Clip length | Working floor | What it means |
|---|---|---|
| Under 30s | Above 50% | Below this the opening is failing, not the body — there is not enough clip for anything else to be responsible. |
| 30–60s | Above 40% | The band where the body starts to carry real load. Below the floor, look mid-clip. |
| Over 60s | Above 30% | Long enough that some viewers leave because they got the answer, which is not the same as a failed opening. |
Source honesty matters here. These come from vendor-published benchmark work, not from TikTok, and the underlying sample composition is not something you can inspect. Use them the way you would use any borrowed prior: as a starting line to tell you roughly where "bad" begins, and as something you replace with your own numbers as soon as you have enough posts to compute them.
The band boundaries are worth more than the percentages, honestly. Even if the floors are off for your niche, the act of splitting the report into three buckets fixes the mix-shift problem on its own.
Build your own floors in one afternoon
Sixty posts is enough. The procedure:
- Export your last 60 posts with clip length, average watch time, and completion for each. One platform only.
- Bucket them into under-30s, 30–60s, and over-60s.
- Check the bucket counts. You need at least 15 in a bucket for its median to mean anything. If one bucket has four posts, you do not have a floor for that band — you have an anecdote, and you should keep using the borrowed prior there until you do.
- Take the median, not the mean, within each bucket. One post that broke out will drag a mean upward and give you a floor no ordinary post can clear.
- Set the floor at the median. Not at your best post. The floor's job is to flag the bottom half for attention, not to define excellence.
- Re-compute quarterly. Your own floors drift as your content changes, and a floor from a year ago is a borrowed prior again.
What you now have is a triage list: every post below its band's median, with the band recorded so you know which failure to look for.
The shorten-versus-fix rule
Here is the decision the floors are actually for. A post is under its band floor. Do you shorten it, or fix it?
The answer comes from the retention curve shape, and the rule is short enough to memorise:
Shorten when the curve bleeds. Fix when the curve cliffs.
A slow bleed — a gentle, steady decline with no single drop — means the clip outlived its idea. Viewers are leaving continuously because nothing new keeps arriving. Shortening genuinely repairs this, because you are removing the seconds where nothing was happening. A 55-second clip with one idea in it becomes a 25-second clip with one idea in it, and it moves into a band where the floor is higher but its curve now clears it.
A cliff — a steep early drop, then a flat survivor line — means the opening failed. Shortening does nothing here except produce a shorter clip with the same broken opening. Worse, it moves the post into a stricter band, so the same failure now shows up as a bigger miss. The fix is the opening, full stop.
There is a third case people forget. A mid-clip cliff, where the curve holds and then falls off a shelf somewhere in the middle, means one segment lost them: a tangent, a format break, a section that repeats what you already said. The fix is cutting that section out, which shortens the clip as a side effect but is not the same operation as trimming the front or the tail. Cut the offending segment, not an equivalent number of seconds from wherever is convenient.
What shortening actually costs
Shortening is not free, and treating it as the default fix has two failure modes.
You lose the band you were building for. If your content is genuinely mid-length and works there, chasing a higher completion figure by compressing everything to 20 seconds trades a real audience for a better-looking number. Completion is an input to distribution, not the goal; a 70-second post at 32% completion delivers more total watch time than a 20-second post at 60%, and total watch time is the thing that compounds.
You break the cut rhythm. Compressing a clip without re-cutting it produces a clip whose pacing no longer matches the band. TikTok's native feel sits around a cut every 1.5–3 seconds, and the per-platform cut rate has to be re-checked whenever you change length materially. A 60-second edit trimmed to 30 without touching the cut points reads as rushed at the front and slack at the back.
The generation-side version of this is worth planning for: decide the target band before you write the shot list, then generate clips at the lengths that band implies. That is cheaper than generating long and trimming, and it happens to be where models are most stable anyway. Model shortlists differ by platform — the best video models for TikTok is a different set from the one you would pick for widescreen, largely because of aspect ratio and duration ceilings.
If you are cross-posting the same asset, the band question changes per destination. The format cheat sheet for TikTok, Reels and Shorts covers what has to change between them; the completion floors above are TikTok's and should not be carried over unmodified.
FAQ
Should I report completion or average watch time?
Track both, decide on average watch time. Completion is the cleaner communication number and the easier one to compare against a floor. Average watch time in seconds is the one that tells you whether a change actually delivered more attention, because it does not move when clip length moves. If the two disagree after an edit, average watch time is the one to believe.
Do these bands apply to Reels and Shorts?
The principle does — completion is length-dependent everywhere, so banding is always right. The floors are not transferable. Shorts exposes Viewed vs. Swiped Away, which is a different measurement entirely and has its own conventions. Reels sits closer to TikTok but its feed composition differs enough that you should compute your own floors rather than importing these.
What about posts that go viral — do they break the floors?
They break the medians, which is why you take the median rather than the mean. Leave breakouts in the dataset but expect them to sit far above the floor and carry no diagnostic information. A post that got fifty times your normal reach tells you about distribution, not about whether its opening held people relative to its band.
Is a low completion rate always a problem on long clips?
No, and this is where the over-60s band needs judgement. A viewer who leaves after getting the answer they came for is not a failed retention outcome, which is why raw average-percentage-viewed is a weaker proxy than it looks. On long clips, read curve shape before you read the percentage: a clip that holds a healthy plateau and then declines from the point where the answer landed is working as intended.