Every title and thumbnail change, with what it was before.

When a creator rewrites a title, the previous one is gone from YouTube. When they replace a thumbnail, the old image is gone too. The archive holds both sides: the title as it was, struck through beside the title as it is, and the previous thumbnail image next to the replacement.

That is what makes packaging studyable rather than anecdotal. You can see that a video was renamed twice in a week, what it was called each time, and which version is live now — across every channel in your niche, filterable by channel, by axis, by outcome and by date.

YouTube runs packaging experiments by showing different titles or thumbnails to different viewers. One browser session cannot see that: it sees one variant and has no way to know a second exists.

So a probe is several independent browser identities, launched cold, arriving separately. If they come back with different packaging for the same video, that is an experiment, and every variant they saw is recorded. Those identities are never batched into one session and their number is never reduced to save capacity — batching them would defeat the measurement, so when the system is busy it probes less often rather than probing with fewer eyes.

Thumbnails need one more piece of care. The same image re-encoded by a CDN is not byte-identical, so comparing files would report a change every time YouTube re-compressed something. Images are compared by perceptual hash instead, and two hashes within a distance of six are treated as the same picture — a threshold that is frozen across the whole system so that hashes recorded a year apart still compare correctly.

Probing is bounded to the first three weeks after upload, which is the window in which YouTube actually runs these experiments, and an experiment that has not been seen for six hours reads as ended. When it ends, the winner is whichever recorded variant matches what the video is showing now — exact text for titles, the perceptual-hash rule for thumbnails. If neither matches, the creator picked a third option we never saw, and the record says the winner is unknown rather than naming one.

Above the archive is the summary of everything recorded for your niche: how many experiment runs and packaging changes are in view, how often titles and thumbnails settled on a variant that was tested, how long a typical experiment ran, and how many times a run was seen.

Every figure carries its own n, because each one counts a different population. And when a population is too small to express as a rate, the panel prints the count and a note saying so rather than a percentage — “not enough data yet, the count, not a rate”. A percentage over four observations is a number that looks like evidence and is not.

The same rule governs the effect of a change. Where there are enough view samples on both sides of a swap, the record says whether the view rate improved, declined, or did not clearly move. Where the video was too new, or the samples too sparse, it says that instead. “Settled on a tested variant” is likewise stated carefully: it means the video now shows one of the variants that was recorded, and whether that was the original packaging or the new one is not something YouTube exposes, so it is not claimed.