Most channel post-mortems begin with one number and end with the wrong conclusion. You open Analytics, read an average view count that a single video from eighteen months ago has dragged upward, and decide the channel is in decline. Or you read an average upload gap of nineteen days on a channel that publishes weekly, because one hiatus sits in the middle of the sample.
Both of those are arithmetic means, and a mean is the wrong summary for anything with a long tail. Your view counts have a long tail. Your upload gaps have a long tail. What follows is what to measure instead, which comparison across your own back catalogue is worth making, how large a difference has to be before it means anything, and why a comparison that finds nothing is a result rather than a failure.
Everything below assumes your fifty most recent uploads, which is what Channel Analysis reads from a channel URL, a handle URL such as youtube.com/@name, or a link to any one of its videos.
Why the median beats the mean, twice over
Take ten uploads: nine between 800 and 2,000 views, and one that reached 240,000. The mean is roughly 25,000; the median is about 1,400. Plan against 25,000 and every upload since looks like a failure. Plan against 1,400 and you can see which uploads actually beat your typical one — the only comparison that leads anywhere.
Upload gaps break the same way: fourteen weekly gaps and one three-month break give a mean of 13 days and a median of 7.
So the report gives you a median gap between uploads and a median view count, labelled "median, not average" on the tile. The dormancy check runs on the same number: nothing published for more than three typical gaps counts as a stall. A weekly channel is flagged after about three weeks of silence, a monthly channel after three months — one threshold, scaled to your own rhythm.
The busiest publishing day it reports is counted in UTC, so a late-evening upload from a western time zone lands on the following day.
Compare your best quarter against your weakest half
The comparison worth making is not your best video against your worst — a single pair is a coin flip. It is a group against a group.
Channel Analysis sorts the fifty uploads by view count, takes the top quarter — about thirteen videos — and the bottom half — about twenty-five — and compares them on five properties. The middle is dropped deliberately: videos that landed near your median are, by definition, the ones that distinguish nothing. The comparison needs at least four videos carrying view counts before it runs at all.
| What is compared | How it is measured | Reported when |
|---|---|---|
| Title length | Mean characters across each group | 5 characters or more apart |
| Titles with a number | Share of the group | 10 points or more apart |
| Titles asking a question | Share of the group | 10 points or more apart |
| Videos carrying tags | Share of the group | 10 points or more apart |
| Custom thumbnail | Share of the group | 10 points or more apart |
Every one of these is packaging: a decision you made in Studio before publishing. That is the point — a comparison telling you "your best videos are about better topics" would be true and useless.
The thumbnail line needs a qualification: the check reads whether YouTube returns a maximum-resolution image, which stands in for a custom thumbnail. A good proxy, not a perfect one — maxres images also depend on the source being uploaded at 720p or higher.
Ten points is the floor for a real difference
Move one video across the boundary between those groups and the top quarter's share of any property shifts by roughly eight points on its own, and that single fact sets the threshold.
A difference under ten percentage points is not reported, because it cannot be told apart from which videos fell on which side of the cut. Title length uses five characters for the same reason. So 40% against 34% is nothing; 62% against 20% is something.
The threshold protects you from reading noise. It does not protect you from reading a correlation as a cause. If your top quarter carries numbers in 62% of titles and your bottom half in 20%, the honest reading is that number-led titles and strong performance travel together on this channel. Adding a number to your next title is a hypothesis to test over the following month, not a fix.
What the comparison cannot see
Four limits worth holding while you read the output:
- View count is not performance. An upload from three years ago has had three years to accumulate views. On a weekly channel fifty uploads span roughly a year and the effect is mild; on a channel posting twice a year it dominates the ranking.
- Views fold three things into one number. Impressions, click-through rate, and retention all sit inside a view count, and they are separate levers. A video with a strong thumbnail and a weak first thirty seconds can beat a better one on views.
- Retention is not in the sample. Average view duration and CTR are visible only in your own Studio, so the report stays on packaging and cadence.
- The sample is the most recent fifty, not the best fifty. A channel with 600 videos gets a read on its current year.
When nothing separates them, that is the finding
If none of the five properties clears its threshold, the report says so instead of showing an empty panel: nothing separates the top quarter from the bottom half in this sample. Which leaves the topic doing all of the work.
That result is not a dead end. It usually resolves into one of two situations:
- Packaging is uniformly good. Every title is the right length, every video has a real thumbnail, and the variance came from the topic instead — subject demand varies more than any title trick can offset.
- Packaging is uniformly weak. Fifty titles built to the same tired formula produce a flat comparison too. Flat is flat in both directions, so check your median title length before congratulating yourself.
The action is the same either way. Take the three uploads that beat your median by the widest margin, read what they are about rather than how they were dressed, and find adjacent queries with the keyword research workflow. Then vary packaging on purpose — one variable per video, held for two weeks — so the next run has something to find. The title generator helps: it returns eight titles per angle and varies the shape across them.
Reading the rest of the report
Each of the other findings is a fixed threshold rather than a judgement:
- Tag coverage below 50% is flagged. Tags carry the spellings and phrasings that do not belong in a title. What tags actually do covers how much effort it deserves.
- Titles past 70 characters. If 30% or more of your titles run past the roughly 70 characters search results display, the report says so and shows your median length beside it.
- Words this channel keeps using. The fifteen most frequent meaningful words in your titles, counted once per title so one repetitive headline cannot inflate a word — where creators find a verbal tic they did not know they had.
- Most reused tags. Your top 25 by video count, the fastest way to notice a tag list copied forward unchanged for two years.
Run a competitor through the same tool straight after your own channel. The report is identical for any public channel, and the useful reading is comparative: their median gap against yours, their median title length against yours, and whether their top quarter separates from their bottom half on something yours does not.
The short version
- Use the median for view counts and for upload gaps. One viral video and one long hiatus each destroy a mean.
- Compare your top quarter against your bottom half and ignore the middle, which by construction tells you nothing.
- Treat any gap under ten percentage points across fifty videos as noise, and under five characters for title length.
- Views hide impressions, CTR, and retention inside one number, and older uploads in the sample have had longer to collect them.
- When nothing separates your best videos from your worst, the topic is doing the work and the packaging is not. That is a finding: make more of that topic, and start testing packaging deliberately.
