A short-video feed decides what you see next with a simple loop: a huge pool of clips is cut down to a short list, a ranker scores that list, and your reaction to each clip feeds back into the next score. This post draws that loop and explains each part. It describes a generic recommendation […]
A short-video feed decides what you see next with a simple loop: a huge pool of clips is cut down to a short list, a ranker scores that list, and your reaction to each clip feeds back into the next score. This post draws that loop and explains each part. It describes a generic recommendation loop, not any one platform’s real system.
The loop in four boxes
Two things follow from this shape. The pool is far too large to score clip by clip, so a cheap filter runs first and the expensive scoring happens on a short list. And the loop never stops. A new account has little feedback, so early clips are broad guesses that tighten quickly.
What happens in the first fifteen seconds
The signals, side by side
| Signal | What it tells the ranker | Weak spot |
|---|---|---|
| Watch time | Whether the clip held attention, scaled by clip length. | Favors clips that are only long, not good. |
| Replay | The viewer wanted to see it again. | Looping content can inflate it. |
| Like or share | An active choice, so a stronger vote than passive viewing. | Rare, so each one carries a lot of noise. |
| Fast skip | A clear no, and cheap to collect. | Says nothing about why. |
This is a general explanation written for readers who build or study apps. Platforms publish few details about their real systems, so treat the diagram as a mental model and check each platform’s own documentation for specifics.
