Platforms moved from counting plays to counting minutes watched, and almost every visible behaviour of recommendation follows from that single change of unit.
A view is a weak measurement
A play registers after a short threshold, so a video that is opened and abandoned counts alongside one watched completely. The unit records an intention rather than an experience.
That makes views easy to obtain through the title and thumbnail alone, which is a promise made before the content is seen and can be made independently of whether it is kept.
Optimising for plays therefore rewards the packaging and ignores everything after it, which is a straightforward description of the behaviour the change was designed to stop.
Minutes align the platform with the viewer
Time watched can only be accumulated by content people continue watching, so it cannot be captured by packaging alone and requires the video to hold attention.
It also aligns with the platform's own interest, since advertising inventory is generated by time spent rather than by pages opened, making the metric commercially coherent.
The result is a system that will promote a video with fewer plays and longer sessions over one with more plays and immediate departures.
The metric has its own distortions
Rewarding total minutes advantages length, and creators responded by extending videos, which produces padding that raises the number without improving the experience.
Platforms countered by weighting the proportion of a video watched alongside the absolute time, so a short piece watched through can outperform a long one abandoned midway.
Both measures still favour formats that hold viewers regardless of value, which is why watch time rewards serialised and episodic structures so strongly.
Session behaviour extends the idea
What happens after a video matters as well. A video that leads viewers into further watching contributes more than its own duration, and one that ends a session contributes less.
This is why recommendation favours content that connects to other content, and why a satisfying but terminal video can be promoted less than a middling one that leads onward.
What creators can read from it
Average view duration on its own is difficult to interpret, because it moves with length. Proportion watched is comparable between videos and is the more useful figure.
The retention curve is more useful still, since it identifies where attention was lost, which is information no aggregate number preserves.