Breakout before the click

What the replay curve is, and what it is not

The public most-replayed heatmap is the only per-video shape YouTube shows the outside world. It is a curve of replays normalised to a video’s own mean, re-normalised again before it is used here. It is not a count of how many people stayed or left, and no such count is public for a channel launchtube does not own.

That limitation shapes everything below. Videos that flopped usually carry no heatmap at all, because they lack the watch volume to generate one, so the shapes available for study are mostly the shapes of videos that already did well. Ninety-two videos across eleven channels had a usable one. The rest of what these measures miss is set out in how growth was measured.

Where a video’s replay curve first drops

The first sustained drop is the second at which replay mass first falls below the video’s own mean. It moves with the length of the file: six seconds on a Short, nine at three to eight minutes, sixteen at eight to twelve, twenty-three at twelve to twenty, and 136 seconds on files over twenty minutes.

The bands themselves are the subject of the page on video length, which is where the question of which runtime grew is answered. The figures here say only where in a file the replay shape turns down, on the videos that had a shape to read.

Why the first drop does not decide the breakout

A winners-against-losers comparison was possible on only two of the eleven channels with reconstructed histories. On one of them the winning and losing curves were 0.722 and 0.720 — a gap of two thousandths across 39 videos. The replay shape does not predict which video breaks out.

That is a narrow finding and it should be read narrowly. It argues against spending review effort on mid-video structure in the hope of buying a breakout. It does not say structure is irrelevant to a viewer who has already clicked, and the ledger records that limit alongside the number.

How early a breakout arrives, and on whose traffic

On every channel whose history could be reconstructed, the first real break came within the first five uploads rather than later. The research tags that as an inference rather than a measurement, because it rests on reconstructed histories and one catalogue turned out to be missing roughly fifty early uploads.

The dated numbers behind it are unambiguous. One channel’s first and third uploads took 151,064 and 383,127 views while it held 53 subscribers or fewer. On two other channels the second upload took 667,660 views at about 2,700 subscribers, and the third took 966,442 at about 12,800.

The reading the research puts on that is also an inference: early breakouts arrive on browse and suggested traffic, which is independent of how many subscribers a channel has. The counter-case is the one the sweep can show plainly — the channels that needed twenty to fifty uploads to reach a first 100,000-view video were all still small at the time of measurement. What the fastest-growing channels were publishing while that happened is in what grows a faceless channel.

The title patterns the sweep counted

Titles were counted in one field only: the forty most-viewed videos pooled from a 200-video sample of the AI-automation niche, on 15 September 2026. A number in the title appeared in 31 of the 40, the single figure that two separate counts of the same pool agree on.

Every one of these is correlational. They are the titles that happened to be on the winning videos of one field on one date, and the field’s own winning content is screen-recorded software, which a generated pipeline cannot make.

Titles that name a thing rather than a category

Across the other classes swept, the pattern that repeats is naming: a specific entity, a specific place and period, or a definition, in place of a category word. The multipliers below are each a video’s views against its own channel’s rolling median, so they compare a video with its channel and not with YouTube.

Which of those subjects a generated pipeline can actually render, and which classes returned nothing at all, is the subject of the page on niches.

What the strongest thumbnails in the genre have in common

This one is an inference from the files rather than a measurement. Across three separate reports, the best-performing thumbnails inspected carried heavy baked-in text and several characters in frame, which is the opposite of what a one-character-per-shot generated pipeline produces.

It was thumbnail inspection on small samples, not a systematic study, and it describes what the field does rather than proving that text-heavy thumbnails cause the views. Treated as a description of the competitive shelf, it is still the most concrete thing the sweep recorded about the picture a reader sees before deciding.

The opening seconds, for a viewer who has already clicked

On the reference channels measured, the subject of the video is named in the first two to fifteen seconds and the payoff is promised before the body begins; on one set of references the definition itself is answered by about second 31. That is directional structure drawn from auto-captions of three to five top videos per channel, not a tested rule.

On a Short there is less room than that. The first sustained drop arrives at second six, and on a 25 to 40 second file all of the replay mass sits inside the first minute, so the subject has to be on screen before anything else is.

What this page does not claim

No figure here states how many people clicked, how long they stayed, or what any channel made. None of that is public for a channel launchtube does not own, and the public replay heatmap is the ceiling of what can be seen from outside. Whether the way a video is made affects its reach is a separate question, answered from YouTube’s own published pages in does YouTube penalise AI channels.

The title levers carry one more caution, and the research tags it as speculative: every one of them in the two fields with no generated channel in them was measured on presenter-led channels, and nobody has run them on a faceless generated channel, because none exists in either field. The levers are real where they were measured. Whether they transfer is untested, and this page does not pretend otherwise. The rest of the research sets out the same limits for the other five questions.