Skip to content

Script X-ray

Where viewers went back to watch a video again. YouTube publishes a Most replayed graph for its videos, and Prepublish keeps that graph beside the caption timings for every video it reads. Each video has a transcript page showing the graph and the moments viewers returned to, with what was said at each. This page shows what the corpus holds and how the model that ranks passages is measured.

The graph counts replays. It does not show where viewers stopped watching. Every association below is measured across public videos, and an association is not a cause.

What viewers replay

The strongest associations the corpus separates from zero. They are measured on the same videos the model trains on, so they describe the corpus rather than test the model.

Flat transition

Replayed 0.26 standard deviations more than other passages in the same video.

Measured on 400 passages where it is present against 51,421 where it is not. 95 percent interval 0.16 to 0.35.

Call to action

Replayed 0.76 standard deviations less than other passages in the same video.

Measured on 1,931 passages where it is present against 49,890 where it is not. 95 percent interval -0.82 to -0.69.

These are associations across public videos, not causes. A passage carrying one of these features is not guaranteed to be replayed, and many passages without it are.

How well the model ranks passages

On 244 videos from 134 channels the model never saw in training, its ranking of a video's passages matched what viewers replayed at a rank correlation of 0.33. Ranking the same passages by their position in the video alone reaches 0.21.

Of the three passages it ranks highest in each held-out video, 22 percent hold a replay peak. Picking three passages at random would hold a peak 13 percent of the time.

The words add measurable signal beyond position: the gain over the position-only baseline has a 95 percent interval of 0.09 to 0.14, which sits above zero. That is the bar an audit requires before it shows replay evidence.

Training runs

Every run, newest first. A run is promoted only when it ranks passages on channels it never saw at least as well as the model already in use.

DateVideosPassagesCorrelation vs positionPeak precision vs chancePromoted
Oct 4, 20261,50351,8210.33 vs 0.210.21 vs 0.13Promoted
Oct 3, 20261,41448,6960.34 vs 0.230.22 vs 0.13Kept
Oct 2, 20261,40348,2750.34 vs 0.230.21 vs 0.13Promoted
Oct 1, 20261,30342,5300.34 vs 0.230.22 vs 0.13Promoted
Sep 30, 20261,13337,3020.35 vs 0.240.23 vs 0.13Promoted
Sep 29, 202698231,8030.35 vs 0.240.23 vs 0.13Promoted
Sep 28, 202682126,2150.32 vs 0.200.21 vs 0.13Promoted
Sep 27, 202669922,3350.29 vs 0.160.19 vs 0.13Promoted
Sep 26, 202659018,7670.30 vs 0.180.19 vs 0.13Kept
Sep 25, 202644914,3730.28 vs 0.200.19 vs 0.12Kept
Sep 24, 20262507,5820.30 vs 0.230.10 vs 0.11Promoted

The corpus

One row per passage of a video whose replay graph was read and whose caption timings could be aligned to its own words.

Videos

1,620

Channels

748

Passages

57,671

Holding a peak

4,893

Recent X-rays

The videos read most recently, newest first. Each one has its own page at /youtube-transcript/<videoId>.

How it works

  1. 1

    Every transcript read keeps YouTube's public Most replayed graph and the caption cue timings beside the words.

  2. 2

    Each passage of a video with a graph is given the replay level over its own span, and the answers to a fixed set of questions about what the passage does.

  3. 3

    A nightly run fits a model on the corpus and measures it on channels kept out of training, against a model that knows only where in a video a passage sits.

  4. 4

    A new model replaces the one in use only when it ranks passages on channels it never saw at least as well as the model already in use.

Questions

Does this predict retention?

No. The graph behind Script X-ray counts replays: how often a moment was watched again. It is not a record of how many viewers stayed with the video, and a model trained on it cannot tell you where viewers stopped watching. What it estimates is replay potential, which is how closely a passage resembles passages viewers chose to watch twice.

Why do some videos have no X-ray?

YouTube publishes a Most replayed graph only once a video has enough views. Below that there is nothing to read, and the transcript page says so in one line rather than showing an empty chart.

Where do the numbers come from?

From YouTube’s public Most replayed graph and the public caption timings of videos that have both. Nothing comes from a creator’s private analytics, and no creator’s studio data is read.

How is the model measured?

On channels kept out of training entirely. Every figure on this page is measured there, and it is measured against the same model with the position of a passage alone, which is the honest baseline: passages near the start of a video are replayed more often than passages near the end.

When does replay evidence appear on an audit?

Once the model in use clears the bar: its ranking of held-out passages has to beat position alone with a confidence interval above zero. Until then an audit does not show it, because an estimate that is no better than knowing where a passage sits is not worth acting on.

Find the moments in your own script

Paste a draft and the check reads it before you record: hook strength, pacing, and the passages most likely to lose viewers, with copy-paste rewrites.

Check a script freeRead a video transcript