Retention curve - the shape of who is left, second by second

A retention curve plots the share of viewers still watching against elapsed time. Nearly every video ad has the same overall shape: a steep fall in the first seconds, then a shallower slope. The curve is therefore read by its deviations, not by that shape: the cliff at the moment the hook's promise expires, the stretch where nobody leaves at all, the small bump where a viewer scrubs back. Check the denominator before reading one: a curve drawn against everyone the ad was served to starts well below 100% at second one, while a curve drawn against the viewers who reached second zero starts at 100%. It answers a question no summary metric can: not how many people left, but exactly what they were watching when they did.

How to Read One

• A cliff in the first three seconds is a hook problem. The opening promised something the next frame did not continue.
• A cliff five to eight seconds in usually means the ad revealed it was an ad: a logo, a price, a sales register change.
• A long flat stretch is the good part: the viewers who are left are watching it through. Read it as the ad's baseline and measure the cliffs against it. What you are looking for is the second where the line breaks, not the minutes where it does not.
• A drop at the CTA is normal and mostly harmless. Those viewers already got the message.
• As craft advice rather than a curve reading: a flat stretch running longer than the ad's own attention span is still where to look next, because the viewers stayed but nothing new was asked of them, and the cost of that usually lands at the following cut, not inside the flat.

Worked Example

A 30-second ad holds 42% of its impressions at 3 seconds, 38% at 8 seconds, 21% at 9 seconds and 19% at 30 seconds. That first point is the ad's hook rate, and everything after it is what the hook handed over. The 17-point fall across a single second at 0:09 is a specific edit, not general fatigue: at 0:09 the creator's face cuts to a packshot. Re-cutting that transition so the claim continues over the product shot is one afternoon of work, and it is worth more than any change to the other 29 seconds.

Reading a Curve You Cannot See

Retention curves are private to whoever ran the ad, so a competitor's curve is never available to you. What is available is the structure that produced it. Admetra segments a public ad at its real cuts and labels every beat with the job it does, which reconstructs the decision points a curve would have exposed: where each escalation lands, which stretch carries no new information, and where the ad changes what it is asking for.