Dwell Time
A click tells you someone was curious about a thumbnail. It does not tell you whether they stayed. Dwell time — how long a user remains with an item after choosing it — is the cheapest available correction to that gap.
What gets measured
The clock starts when the item opens and stops when the user leaves, returns to the feed, or goes idle. Because the raw distribution is heavy-tailed and dominated by article length or video duration, it is rarely used as-is:
- thresholded into a binary long dwell label — a few seconds, or a share of the item's length;
- log-transformed or bucketed before serving as a regression target;
- normalised by content length, so a fully-watched short clip is not beaten by an abandoned feature film.
Where it fits, and what it fixes
Dwell time works as an online metric read from a live system and as a training label for the ranking stage; feeds routinely optimise a weighted blend of click probability and long-dwell probability. Clickbait is the failure it exists to catch — CTR rises, satisfaction falls, and offline accuracy sees nothing wrong because the click was real. It carries its own noise: an open tab, a phone put face-up and a slow-loading page all look like engagement, so it is read next to conversion and retention rather than on its own.
Dwell time is how long a user stays with an item after selecting it, used as an implicit-feedback signal of satisfaction rather than mere curiosity. It is normally thresholded into a long dwell label or normalised by content length, since the raw value tracks item duration. Ranking stages optimise it alongside click probability because a CTR-only objective rewards misleading thumbnails; dwell time is the signal that catches them.
Clicking on something only proves the headline worked. Staying with it suggests the thing was actually good. Dwell time is the measurement of that staying — how long you keep reading an article, or watching a video, after you tapped on it.
Recommendation systems care because the two can disagree completely. A picture that promises more than it delivers collects plenty of taps, and nobody sticks around. By counting the seconds afterwards, the software learns to favour things people stay with over things people merely click. It is an imperfect clue — a phone left face-up looks exactly like rapt attention — so it gets used together with other signs rather than trusted on its own.