Evaluate
AI clip acceptance rate: how to work out your own
Work out your own AI clip acceptance rate: a usable-clip definition, reject codes, a cost-per-clip example and how big a gap must be before it counts.
AI clip acceptance rate is the share of a tool’s suggested clips that you would actually post. No independent acceptance-rate figure for gaming footage was found in the sources checked on 8 November 2026, so the useful number is the one you get from your own count on your own VODs. This page shows how to define it, count it and judge whether a difference between two tools is real.
Sections on this page (7)
How to define a usable clip
An acceptance rate only means something if “usable” is defined before you look at any clip. A workable definition has two passing levels and one failing level, and each clip gets exactly one.
| Result | Anchor | Examples of what fits |
|---|---|---|
| Accepted as is | You would post it with no change | Good moment, clean start and end, captions and crop acceptable |
| Accepted with light edits | You would post it after trimming or fixing a caption, within two minutes of work | A few seconds off the start, one misheard name |
| Rejected | Needs over two minutes, or is not worth posting at any effort | Wrong moment, or the layout must be rebuilt |
The two-minute line is a rule of thumb. Pick your own, but fix it first.
Reporting both passing levels matters, because a tool that needs light edits on everything is very different from one that needs none on a few clips. That gives two numbers:
- Strict rate: accepted as is, divided by clips scored
- Total rate: accepted as is plus light edits, divided by clips scored
Also record why each rejected clip failed, with one reason code: wrong moment, bad start, bad end, framing, captions, duplicate of another clip, or policy or copyright risk. A tool rejected mostly for bad starts can be rescued by an editor. One rejected for wrong moments cannot.
As reasoning, not a finding: expect the rate to fall from a tool’s top picks to its later ones, and expect it to differ more between genres than between tools. You can check the first by comparing clips ranked 1 to 5 with those ranked 6 to 10, and the second by running two different VOD types.
Why published figures are hard to compare
Keep-rate claims are hard to compare because they tend to leave out the footage, the rater, the number of clips and the definition of usable. Treat any figure without those details as a claim, whether it comes from a vendor or from an individual creator’s estimate.
Measuring your own rate
You can get a first reading in an evening with one VOD and a free tier.
- Decide the two-minute rule and your reason codes before you look at any clip.
- Take one VOD you have not watched back, of a kind you usually stream.
- Take the tool’s top 10 suggestions, in its order. Do not regenerate or reorder them.
- If you are comparing tools, hide which tool made which clip.
- Mark each clip as post, post after a trim, or discard, and time your review and any edits.
- Repeat on a second, different VOD type, because one result will not tell you much.
Limiting the count to the top 10 tests the tool’s own ranking, which is what you rely on when you do not have time to watch everything. A spreadsheet layout for the counts is on the clip evaluation log template, and the clipper picker helps you decide which tools are worth trying.
Worked example: from rate to real cost
An acceptance rate turns a price into a cost per usable clip, and it can reverse a ranking once editing time is counted. Illustration only, with invented inputs for two unnamed tools: each scored on 20 clips from two VODs totalling 360 source minutes on a one-credit-per-minute plan, with one minute of review per clip.
| Tool A | Tool B | |
|---|---|---|
| Accepted as is | 6 | 2 |
| Accepted with light edits | 6 | 12 |
| Rejected | 8 | 6 |
| Strict rate | 30% | 10% |
| Total rate | 60% | 70% |
| Credits per accepted clip | 360 / 12 = 30 | 360 / 14 = 25.7 |
| Edit minutes (light edits × average) | 6 × 1.5 = 9 | 12 × 1.8 = 21.6 |
| Your minutes per accepted clip (20 review + edits) | 29 / 12 = 2.4 | 41.6 / 14 = 3.0 |
Tool B has the better total rate and the cheaper credit cost per clip, and costs you more time per clip. Which one wins depends on whether your time or your credits are scarcer. The cost per VOD hour page covers the input side of that sum, and your acceptance rate supplies the output side.
How big a gap has to be
A rate from ten clips is a very rough estimate. The Wilson score interval, which the NIST statistics handbook recommends for a proportion and notes cannot give a negative lower limit, shows how rough. NIST also says normal-approximation limits may be inaccurate when the sample is very small and gives an exact binomial method for that case. As reasoning, exact limits at n = 10 would be somewhat wider than these.
| Accepted | 95% Wilson interval (arithmetic, not a finding) |
|---|---|
| 3 of 10 | 11% to 60% |
| 5 of 10 | 24% to 76% |
| 6 of 20 | 15% to 52% |
| 12 of 40 | 18% to 45% |
| 30 of 100 | 22% to 40% |
Fisher’s exact test compares two counts of accepted and rejected clips. The SciPy documentation describes it for a 2x2 table, with a two-sided p-value that covers a difference in either direction. The smallest gap that clears p below 0.05, starting from a tool at 30%, shrinks as the sample grows. Invented inputs again:
| Clips per tool | Tool at 30% | Other tool needs | Gap |
|---|---|---|---|
| 10 | 3 | 9 | 60 points |
| 20 | 6 | 14 | 40 points |
| 40 | 12 | 22 | 25 points |
| 100 | 30 | 45 | 15 points |
In the worked example above, 12 of 20 against 14 of 20 gives a p-value of about 0.74, which is a tie.
Treat these as minimums. As reasoning: clips from one VOD share a streamer, a game and a mood, so they are not fully independent, and real uncertainty will be wider than the tables suggest. For a home comparison, call a tie unless the gap is at least as big as the row for your sample size, and use tie-breaks such as editing time and cost instead of forcing a winner.
What can distort an acceptance rate
Your own habits move this number as much as the tools do.
- Rater drift. Scoring many clips in a row makes standards slide. Score in short sittings and shuffle the order.
- Regenerating. Rerunning until the list looks good measures your persistence, not the tool’s ranking.
- Different standards by tool. Once you know which tool made a clip, your expectations change. Hide the names.
- Acceptance is not performance. A clip you would post may still get no views. The rate measures editorial judgement, not audience response.
- Cherry-picked footage. A great stream raises the rate for every tool, so a high number from one VOD says little.
Reading your result
Use the two rates together, along with the reason codes.
- High strict rate: the tool can run close to unattended on this kind of footage.
- Low strict, high total: it finds moments but needs finishing, so budget editing time per clip.
- Mostly wrong-moment rejections: switch tools, because more editing will not fix ranking.
- Rate collapses after rank 5: review only the top few, and do not pay for more clips.
- A tie after two VODs: choose on price, caption quality and how quickly you can fix a clip.
Other measures to run alongside this one are on the guides hub.