Runtime is a useless filter: longer videos carry less padding, not more

Across seven runtime bands, from under ten minutes to over an hour, padding falls and measured value rises, in a straight line, all the way through. The common habit of choosing the shorter of two videos on the same subject is, on this library, backwards.

Runtime is the easiest thing to see on a results page, and it turns out to be one of the worst reasons to choose between two videos on the same subject. We have published this pattern once before, inside the padding census; this report is the pattern on its own, with the runtime axis as the whole point rather than one line in a table.

A note on the numbers. Unlike the rest of this site, the figures on this page are frozen at the date above rather than recalculated on each visit. A report people cite should not move under them. When we run this census again it will be published as a new edition with a new date.

Padding drops every step of the way

Seven bands, ordered by runtime, and padding falls in six of the seven steps.

Under ten minutes, padding averages 45.0% of runtime. Over an hour, it averages 27.6% — a drop of seventeen and a half points across the scale. The one band that breaks the pattern, 40 to 50 minutes at 38.2%, sits a single point above its neighbour on either side and rests on 140 videos; read it as noise around a falling line, not a reversal of it.

The direction is not subtle and it holds at every scale we can measure it at. A video’s runtime tells you almost nothing about its subject, but it tells you a fair amount, on average, about how much of that runtime is going to be spent restating things you already know.

Value rises the same way padding falls

The same seven bands, the other measurement, and the line points the other direction.

Measured value climbs from 55.9 under ten minutes to 72.0 over an hour — the highest score of any band, on the smallest sample. Between those two ends the climb is not perfectly smooth; the 40–50 and 50–60 bands dip half a point below the 30–40 band ahead of them, the same small wobble the padding chart shows in the same place. It is one underlying pattern showing up twice, not two independent findings that happen to agree.

The verdict moves the furthest of the three

Padding and value are our own internal measurements. The verdict — worth the runtime or not — is the one built to answer a reader’s actual question, and it moves the most dramatically of the three.

30.2% of videos under ten minutes were not worth their runtime. None of the fifty videos over an hour was. We published this half of the pattern separately, in the verdict census; seeing it lined up against padding and value in one table is what makes the mechanism legible. All three numbers move together because they are, largely, measuring the same underlying thing from three different angles.

Why the relationship runs this way

The mechanism is not exotic once it is stated plainly. A channel decides how long a video should be based on how much its subject actually supports, and then the runtime is filled to match that decision — sometimes with material the subject genuinely has, sometimes with material invented to reach a length the channel had already committed to.

Short runtimes are where the second kind concentrates. A four-minute idea stretched to eleven minutes has seven minutes to fill from somewhere, and restatement, recapping and setup are what is available. An idea that genuinely takes forty minutes does not have that problem — there was always forty minutes’ worth of material, so nothing needed inventing.

This also explains why the relationship is not perfectly linear. The dip in the 40–60 minute range likely reflects a second population entering the mix at that length: some genuinely dense forty-minute videos, and some videos padded well past where their subject runs out, arriving at the same runtime from opposite directions.

What this changes about choosing between two videos

Stop treating runtime as a proxy for effort spent watching, and start treating it as almost irrelevant to quality. Two videos on the same subject, one twice the length of the other, are not more and less demanding versions of the same thing on this measurement — the longer one is, on average, the less padded of the pair.

Weight subject and title over length. Which subject you are in and how accurately the title describes the video both separate the worthwhile from the wasteful harder than runtime does. Runtime is a weak, if directionally useful, third signal.

Do not overcorrect into always picking the longest option. The pattern is a population-level average, not a per-video guarantee, and the longest band here is fifty videos. A twenty-minute video that is dense start to finish beats a padded ninety-minute one; runtime is a hint, not a verdict.

How this was measured, and what it does not say

Each video carries a measured padding share, a measured value score, and a keep-or-skip verdict; this report groups all three by the video’s own runtime, in seven bands. Method and its limits are in what value density means.

Padding rests on a smaller sample than the other two columns. It is measured only on explanatory and instructional material, 3,002 of the 3,243 scored videos; value and the verdict cover all 3,243. Each chart’s own note states which n applies.

The longest band is thin. Fifty videos sit over an hour, and every figure for that band should be read as an illustration of the direction rather than a precise estimate. The direction itself — falling padding, rising value — is consistent across the six bands ahead of it, which is what makes the pattern trustworthy even where the tail is not.

This is a correlation across a library, not a causal law. We are not claiming that making a video longer would improve it. The evidence points the other way: a subject that supports length gets made longer, and a subject that does not gets padded to reach a target instead.

The sample is not YouTube. It is 3,243 English-language explanatory videos across twenty subjects, assembled to build a recommendation library rather than to be representative of the platform.

The figures are frozen at 26 August 2026. Scores are recalculated as the underlying measurement improves, so the live numbers on the shelves will already have moved slightly from these. This edition will not be edited; the next census will be published separately.