Readingaretentioncurve

The retention graph is the most useful object in short-form and the most ignored. Most teams look at the headline number — average watch time — and stop. The number tells you that something went wrong. The shape tells you what.

The cliff

Retention falls off a wall in the first second and then flattens. This is a hook failure, full stop. The people who left never engaged with the content at all; they made a scroll decision on the first frame and the first syllable. Nothing about the body of the clip is implicated, and rewriting the middle is wasted work.

Fix the first frame and the first line. Nothing else.

The slide

A steady, gentle decline with no sharp features. The hook worked, and then the clip simply was not interesting enough to hold anyone in particular. This is the hardest shape to fix because there is no single failure point — the pacing is uniformly slack.

Usually the answer is compression. Cut the same content into two thirds of the runtime and the slide steepens into something with structure.

The step

Retention holds, then drops hard at a specific timestamp, then flattens again. Something happens at that moment that costs you the room. Go and watch it. Nine times out of ten it is the moment the clip reveals it is an ad, or the moment the hook gets answered.

If it is the reveal, move it later and earn more first. If it is the answer, withhold it and layer a second gap on top.

The lift

Retention rises somewhere in the middle, or exceeds 100% at a point. That is rewatching, and it is the single best signal you can get. Some segment is being scrubbed back — a detail, a number, a moment people want to see twice.

A lift is not a result to celebrate. It is an instruction. Whatever caused it should be the opening of the next five clips.

What we do with this

Every clip the fleet posts gets its curve pulled the same night and classified into one of these four shapes. The classification, not the raw watch time, is what feeds the next batch. A cliff sends the concept back to hook generation with the body intact. A step sends it to the editor with a timestamp attached. A lift gets promoted and cloned.

This is unremarkable as analysis. It is only interesting because doing it consistently across a fleet, every night, is not something a human content team has the hours for — and it is exactly the kind of loop that does not get tired.