Padding cliffs and publishing at true length
Stretching a six-minute idea to clear a runtime threshold leaves a visible seam in the retention graph. How to spot it and compare per-video return honestly.
There is a shape in retention graphs that has nothing to do with the video's content. The line holds fine, then breaks somewhere in the five-to-six-minute region, and the break appears again on the next upload, and the one after that, at roughly the same place regardless of subject. That is not a segment fault repeating itself by coincidence. That is the point where the material ran out and the runtime kept going.
Creators pad for one reason: a length threshold that unlocks mid-roll ad placements. Check the threshold itself against YouTube's current monetisation documentation rather than a number remembered from a blog post — it has moved before. What has not changed is the trade, and the trade is usually worse than it looks.
What the seam looks like
A padding cliff is distinguishable from a mid-video cliff by one property: it is tied to length, not to content.
| Mid-video cliff | Padding cliff | |
|---|---|---|
| Position across uploads | Scatters — follows the segment | Clusters at a similar absolute timestamp |
| Cause | One identifiable event | Material exhausted, runtime continuing |
| What's in the window before it | A tangent, an ad read, a format break | Nothing. Recap, montage, or a restated point |
| Repair | Cut that segment | Publish shorter |
The diagnostic is a catalogue test, not a single-video test, and it takes about fifteen minutes:
- Pull your last eight uploads of the same format.
- For each, mark the main back-half drop twice: once as a percentage of runtime, once as an absolute timestamp.
- Plot both columns.
If the drops cluster on the percentage axis and scatter on the clock, you have ordinary structural decline. If they cluster on the clock — five uploads all breaking between 5:15 and 6:00, at wildly different percentages of their runtimes — the length is the variable, and no amount of segment surgery will move it. The shape taxonomy covers the other patterns; this one only shows up across a catalogue.
The second confirmation is content-side. Watch the two minutes immediately before the seam on any two of those uploads. Padding has a small vocabulary:
- The recap. A summary of what was just said, positioned where the video should be advancing.
- The restated premise. The introduction, delivered a second time in different words, because the script needed length before the real material began.
- The information-free montage. B-roll that is pleasant and carries nothing the viewer did not already have.
- The second example. A new case that makes exactly the same point as the first one, added because it was available rather than because it was needed.
If two of those four appear in both windows, the diagnosis is confirmed.
The arithmetic of the trade
Padding to clear a threshold is a bet that a marginal ad slot is worth more than the retention it costs. Two things make that bet worse than it appears.
The marginal slot is served to the smallest cohort. Whatever placement the extra runtime unlocks sits at the far end of the video, where the fewest people are. Practitioner estimates of how much audience remains by the late minutes of a ten-minute video cluster around 40 to 50%, but that figure circulates without a traceable source. Read your own curve instead — the number is right there, and it is the only version that applies to you.
The retention cost applies to the whole video, not the padded part. This is the asymmetry that makes the bet bad. A curve that breaks at 5:30 lowers average view duration, lowers total watch time relative to what the same content would have achieved at true length, and feeds a worse signal into whatever the recommendation system does with the video afterwards. You paid across the entire asset to buy an impression at the thin end.
A third cost never shows up in analytics. Padding trains viewers to skip, which later shows up as a curve scalloped everywhere rather than broken in one place.
The comparison worth running
Most attempts go wrong here. Cut a video from ten minutes to six, watch average view duration jump, and you have learned nothing: the metric rose because the denominator moved.
The honest comparison is per-video return, held over a fixed format, across enough episodes to escape single-video variance.
- Fix everything except length. Same series, same slot, same packaging approach, same production standard. Length is the only variable.
- Publish three to five episodes at true length. True length means the runtime at which the material actually ends, decided in the script before shooting rather than in the edit afterwards.
- Compare against the three to five padded episodes immediately preceding. Not against your best ever, not against a benchmark.
- Read four columns per video, not one:
| Column | Why it's in the list |
|---|---|
| Total watch time per video | The absolute figure. Rises or falls honestly; unlike average view duration it does not move just because runtime changed |
| Impressions and click-through | Confirms packaging held constant. If CTR moved, your test is contaminated |
| Subscribers or follows per video | Catches whether shorter videos convert worse, which is the real risk of the change |
| Revenue per video, if monetised | The actual thing padding was supposed to buy. Per-video, not per-thousand — rate metrics hide the volume change |
- Pre-commit the decision rule. Write down before you publish what result would make you keep true length and what would send you back. Otherwise you will interpret the data in whichever direction the first episode points.
Two traps. Do not compare a shorter video to a longer one on average view duration alone; it is the metric most sensitive to the change you made and least informative about whether it helped. And do not run this over a period when you also changed thumbnails, slot, or topic mix. The one-variable discipline that governs any creative test applies at the programming level too.
Programming at true length
The reframe that makes this stick is that length should be an output of the idea, not an input to the schedule.
Decide length in the brief. Estimate the runtime the material supports before production, and treat overrun as a signal to split rather than a signal to expand. A brief that survives handoff is where that number belongs, because by the time you are in the edit the padding decision has already been made for you by the footage you have.
Split, don't stretch. A ten-minute slot and a six-minute idea have an alternative: two six-minute ideas across two episodes. YouTube's Shows feature lets eligible creators organise videos into seasons and episodes with their own artwork, so a two-part idea can ship as two episodes rather than one padded runtime. Treat "part two inherits part one's audience" as a working assumption, not a published ranking weight.
Spend the recovered time upstream. Paddy Galloway's public split is that top creators put roughly 30% of their effort into ideation and packaging, while smaller creators put around 5% there and 95% into filming and editing. Padding is production hours spent on filler. Those hours belong on the title, the thumbnail, and the idea.
Watch what happens to the shape. The curve you want after the switch is not flat, it is unbroken. Every retention curve declines. What you removed was the event, and the confirmation is a back half whose gradient matches its front half. Keep that read in a standing analytics review rather than checking once, and if you publish the same material across surfaces, compare shape per platform — the cross-platform view often shows the seam appearing in one place and not another.
Where padding survives the test, keep it and stop feeling bad about it. Some formats genuinely carry their length. The point of the comparison is to find out which case you are in, rather than inheriting a runtime target from someone whose format is not yours.
FAQ
How do I find my "true length"?
Write the script, cut every line that does not advance the thread or pay a promise, then time what remains. That is true length. If it lands well short of your usual runtime, the honest options are to publish short or to find a second idea — not to reinstate the lines you just cut.
Won't shorter videos mean less total watch time overall?
Per video, often yes. Across a period, frequently not, because shorter videos hold a higher proportion of their audience and get sampled more widely as a result. That is why the comparison runs over three to five episodes, and why watch time sits in the table next to subscriber conversion — you are reading the whole return, not one column.
Does the padding cliff apply to short-form?
Not in this form. Short-form has no runtime threshold to pad toward, so the equivalent failure is different: a clip that reaches its payoff at nine seconds and continues to twenty-five. The shape is a drop straight after the payoff, and the fix is the same in spirit — end where the material ends. Ad-format length decisions are their own question, covered in what the data says about video ad length.
If I cut a published video shorter, does the old retention data still count?
The graph will blend both versions and become uninterpretable for as long as the old data dominates. Treat a re-cut as a repair you confirm on the next episode rather than one you verify in place. Re-rendering the shorter version off an EDL-based timeline is cheap enough that the constraint is measurement, not production: the timeline is re-renderable, a free 480p preview pass with a short per-user cooldown lets you check the new ending before committing, and the final export is charged once however many clips are on it.