Short video metrics: what to track and what to ignore

Views are the number everyone watches and the number that explains the least. They report how widely the algorithm chose to distribute a video, which is downstream of everything you control. Here is what to look at instead.

Fix them in this order

1. Three-second retention. Nothing below matters until this holds
2. Completion rate. Once people stay past the hook, the middle has to earn it
3. Saves. Practical value, and the strongest signal of a video worth making more of
4. Profile visits. Whether the viewer understood who you are

Order matters more than the list. Improving the middle of a video that loses half its audience at second three is wasted effort, because that middle is never seen.

Reading the retention curve

• Cliff in the opening seconds: the hook promises the wrong thing, or there is none
• Steady slide: pacing sags, and the fix is cuts and visual events, not more words
• Sharp drop at one point: something specific repels people there, go watch that moment
• Uptick at the end: viewers are rewatching the payoff, so move it earlier next time

Metrics that mean little

Comparing periods without fooling yourself

1. Use the median, not the mean: one viral video distorts an average into a false trend
2. Compare windows of at least two weeks, since variance between individual videos exceeds most edits
3. Change one variable at a time, or the result tells you nothing
4. Write down what you changed, because next month you will not remember

Numbers do not transfer across platforms

A view means different things in different places. In a subscription feed it is proof of delivery; in a recommendation feed it is the outcome of selection. Comparing the two produces conclusions about nothing.

The test for whether a metric is worth watching

If a number would not change what you do next week, it is not a metric, it is decoration. Each of the four above points at a specific action, which is exactly what views do not.

Bottom line

Track three-second retention, completion, saves and profile visits, in that order. Keep views as context, compare medians over two-week windows, change one thing at a time. Everything else produces the feeling of working rather than the result of it.