Method
This page exists because the hard part of this measurement is not the arithmetic. It is knowing which apparent results are artefacts of your own panel. We produced three findings that did not survive scrutiny before producing none that we would publish, and every one of them is documented below. Every numeric figure on this page is a retracted artefact, published as a worked example of how this measurement fails. None of them is our finding and none should be quoted as one.
A panel of 302 YouTube channels across 12 categories was frozen before 2026-08-24 and has not been altered since. Public view counters on 18,671 videos are read hourly. Counts are cumulative, so only the difference between consecutive readings of the same video can speak to a change on a particular day; the unit of analysis is a per-video accrual rate, not a view total.
3 full days of pre-change baseline were banked before the change took effect. That is the part that cannot be reconstructed afterwards. Once a platform changes a definition, campaign data contains the change and ordinary audience movement in the same week, and no amount of later analysis separates them.
The headline cohort is videos published on or before 2026-07-25 — at least 30 days old at the change — because their view velocity decays smoothly and predictably. Newer uploads are tracked but excluded from the headline; section 03 explains why that turned out to matter more than expected.
Our first published figure was 0.56× — a false result, an apparent 44% fall in public view counts that was an artefact of our own method. It was wrong, it is retracted, and it was live on this site for about an hour.
The post-change window at that moment covered 00:00–02:00 UTC, two of the lowest-traffic hours of the day, and it was being compared against a 72-hour average that included peak hours. Matched hour-for-hour against the same hours before the change, the same data gave 0.94 and 1.09.
Fix: every comparison is matched on hour of day, and the site publishes no figure until a full daily cycle has accrued.
The next result looked excellent. Split by cohort, the effect appeared concentrated exactly where the mechanism predicts: newer videos up 28%, newer long-form up 2.02× — also a false result, retracted, older videos flat. Long-form and live carried the 30-second threshold that was removed; Shorts never did. It was mechanistically coherent and commercially attractive.
It was also entirely present on 23 August, a date when nothing happened. Running the identical analysis against that date gave newer long-form 2.02× and newer videos 1.32× — the same apparent effect, on a date with no change to detect, which is what proves both were artefacts.
The cause was our own instrument. The panel gains newly published videos over time, and a brand-new video accrues views far faster than one a few weeks old — so any window later in time contains fresher videos and looks inflated, whatever the platform did. Hour-matching controls for time of day. It does not control for video age or panel churn.
Fix: the newer-video cohort is treated as unusable for this measurement, and no claim is published without a placebo test.
Run the whole measurement against a date when nothing happened. Anything that shows up there is your method, not the platform. Only the gap between the real date and the placebo is attributable to the change.
This is the single cheapest safeguard available to anyone measuring a platform change, and it is the one that killed our best-looking finding. We publish it because a figure that has not survived a placebo is not a measurement — it is a coincidence with a decimal point.
With the first two confounds handled, a clean-looking result appeared on the headline cohort — statistically strong, and negative. It is not usable either.
Our baseline covers Friday to Sunday. The change landed on a Monday. Measured on identical hours using pre-change data only, day-to-day accrual swings by up to 24% — larger than any effect we have measured. With no pre-change Monday to compare against, the counting change and the calendar cannot be separated.
A detail worth stealing: the placebo dates all fell on a weekend, so their agreement with each other looked like a tight noise floor. A tight spread from a homogeneous sample is not a small error bar — it is a sample that has not been asked the hard question.
Fix: the answer waits for matched weekdays — post-change Saturday against pre-change Saturday, Sunday against Sunday. That is why this site still publishes no figure.
A figure appears here when it survives all four checks: a full daily cycle, a placebo date, matched weekdays, and a stated confidence interval. If it does not survive them, we will say that instead, in the same place and with the same prominence.
A null result is a result. If the change did not move public view counts by an amount distinguishable from ordinary variation on 30-day-old videos, that is worth knowing — it means anyone re-pricing contracts against a specific multiplier right now is acting on something nobody has measured.
The limitations that will remain even then: there is no control group, because the change was global and simultaneous; the estimate rests on a decay curve fitted before the change and extrapolated across it; counters are revised downward when the platform purges invalid views, and 4,178 such revisions were observed and excluded from accrual rather than hidden. None of those go away with more data.