youtube-outlier-research

YouTube outlier research

A video's view count says more about the channel it sits on than about the video. 500,000 views is unremarkable on a channel that always gets a million, and 60,000 views is a discovery on a channel that usually gets 4,000. The second is evidence that an idea worked without an audience carrying it, which is the only evidence worth copying.

So the unit of analysis is the multiple, meaning a video's views divided by its own channel's baseline. Report multiples. Never rank videos from different channels by raw views.

The measurement

Baseline. The median views of the channel's 30 most recent settled uploads of the same format. Median rather than mean, because one runaway hit drags a mean upward and makes every other upload look like a failure against it.

Settled means published at least 7 days ago. Younger uploads are still accumulating views, so including them drags the baseline down and understates every multiple.

Format splits at 180 seconds. Shorts and long-form get separate baselines and are never compared to each other, since a 40-second video judged against 20-minute numbers is flattered or punished by nothing but its length.

Typical range is the 25th to 75th percentile of the same pool. A video inside that range is the channel doing what it always does, whatever its raw views.

Not enough data. Fewer than 5 settled uploads in a format means no baseline for that format. Say so rather than computing one from three videos.

Reading a result

• Below 1x is under the channel's own normal, whatever the view count looks like.
• 3x is the common bar for calling something an outlier.
• 10x and above reshapes a niche, and is rare enough to be worth checking for a cause, a collaboration, an external link, or a video that got picked up somewhere.

Two things change the reading and are the most common analysis errors.

Age. A 12x from two years ago has already been absorbed, every serious competitor made their version. A 12x from last month is still open. Always report how old an outlier is next to its multiple.