We measured the padding in 3,002 videos. The number has not moved.

The median measured video carries a padding share of 40% — the share of its runtime that restates, recaps or sets up rather than adds. That is the same figure the first edition published eight days ago, on a set of videos a little over half the size.

Which is the finding. A census whose headline moves every time the sample grows is not measuring the platform; it is measuring the sample. This one did not move.

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. The first edition remains at its own address, unedited.

The library grew by three quarters and the distribution stayed put

Between the two editions the set of videos carrying a padding measurement went from 1,683 to 3,002 — a growth of 78%. Here is what that did to the numbers.

The quartiles are identical. The mean moved by three tenths of a point. Even the 1,319 videos added since the first edition, taken on their own, land at a median of 39% and a mean of 40.8%.

Three tenths of a point is inside the noise of any measurement like this one, and we are not going to pretend it is a trend. What it does establish is that the first edition’s headline was not an artefact of a small library, which is the thing a second edition exists to test.

Two videos in three sit between 30% and 50% padding

The distribution is narrower than the complaint about YouTube would lead you to expect.

1,976 of 3,002 videos — 65.8% — carry between 30% and 50% padding. Below 20% there are 33 videos in the entire census, a little over one in a hundred. Above 60% there are 201, one in fifteen.

So the platform does not divide into lean videos and bloated ones. It has a dense middle at around two fifths, a thin tail of genuinely tight material, and a smaller tail where the padding has taken over. The middle is the platform.

Read the 40% carefully, because it is easy to inflate. It does not mean two of every five minutes are worthless. It means that if the same material were delivered without its structural repetition, the typical video would be substantially shorter than it is.

Fourteen and a half points separate travel from science

Subject moves the number more than anything else we measure except the title.

Travel sits at 48.9% and science at 34.4%. The order is close to the first edition’s, and it tracks one property: how far the subject’s core idea can be compressed.

Travel, design, productivity and language learning sit at the top because each is built around ideas that can be stated in a sentence. There is nowhere for a twenty-minute video on such an idea to go except sideways. Science, history and machine learning sit at the bottom because their material chains: the next minute depends on the previous one, and cutting the chain produces a list of facts rather than a shorter explanation.

Four of those bars — travel, comedy, gaming and DIY — rest on fewer than sixty videos each, because those subjects are newer to the library. Read the ends of that chart; treat the middle as an ordering.

The long videos are still the least padded

This was the finding that contradicted our own expectation in the first edition, and eight days and 1,319 videos later it is still there.

Padding falls from 45.0% under ten minutes to 27.6% over an hour, seventeen points, and the only step that goes the other way is a bump at forty to fifty minutes that rests on 140 videos.

The explanation is not complicated once the result is in front of you. A subject that genuinely requires an hour fills an hour. A subject that requires six minutes and is being delivered in eleven has to find five minutes somewhere, and restatement is the only material available. Length is a symptom of a bloated subject, not a cause of bloat.

The practical consequence is unchanged and still counterintuitive: runtime is close to worthless as a filter, and the very common habit of preferring the shorter of two videos on the same topic is, on average, backwards.

The title still predicts the padding, and less dramatically than last time

The first edition closed on a third result: group the library by how accurately each title describes its video, and the padding share separates sharply. It still does.

Where the title describes essentially a different video, padding runs at 56.7%. Where it is accurate, 34.9%. Twenty-two points, on a signal visible from the results page.

The interesting part is what happened to the extreme. In the first edition that group held ten videos and carried 76% padding — a figure that came with a warning attached, because ten videos is an illustration and not an estimate. The group now holds 222 videos, and it sits twenty points lower. The direction held; the number a small sample produced did not.

That is the whole argument for reissuing a census rather than leaving one up. The headline was robust. The most quotable line in the first edition was not, and the only way to find out which was which was to measure again.

We are not claiming causation in either direction. The reasonable reading is that both are downstream of the same decision: a video optimised for the click is optimised for the click throughout, and the title is simply the part of that you can see before you commit.

What the 40% is not

Padding is not waste. It is the share of runtime that does not add. Some of it does real work — a cue that needs saying twice, a recap that lets somebody who looked away rejoin. It becomes a problem at the point where it stops being for you.

It is not a claim about a channel. Every figure here is per video, and a subject sitting high on that chart is a statement about the subject, not about anybody making videos in it.

It is not accuracy. We measure how much a video says per minute, not whether what it says is true. Every result on this page is about how videos are built.

It is not the verdict either, though it is most of it. The videos we would not have you watch carry 56.9% padding against 37.5% for the ones we would. That is the single largest gap between the two piles in the verdict census — larger than subject, larger than length. Padding is where a video’s decision about your time becomes visible.

How this was measured, and what it does not say

For each video we measure the share of runtime spent restating, recapping or setting up rather than adding, and publish it as a percentage. The reading is ours. Method and its limits are in what value density means.

The census is 3,002 videos, not 3,243. Padding is only measured on explanatory and instructional material. The 241 videos in the library that are entertainment, news or ambient are scored on everything else and carry no padding figure, because a padding share would not mean anything for them. Every other report on this site quotes the full 3,243; this one quotes the set the measurement applies to.

The sample is not YouTube. It is English-language explanatory video across twenty subjects, assembled to build a recommendation library rather than to be representative of the platform. It skews towards material somebody thought was worth measuring.

The first edition counted differently. It reported its sample as 1,893 — every video scored at that date — while its padding figures were computed over the 1,683 of those that carried a padding measurement. The table above recomputes both editions the same way, so the two columns are comparable to each other even though the first edition’s own masthead says a larger number.

The figures are frozen at 25 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 third census will be published separately.