Ranking systems in social media feeds learn from what you actually do, not just what you click. Dwell time, meaning how long you linger on a post even without liking or commenting, is treated as a quiet signal of interest. A related idea is the "pause," when a viewer stops on a piece of content instead of swiping past it. These pauses suggest attention and processing, and in many systems they predict interest better than quick scrolling. Negative feedback is gathered when you actively push back: hiding a post, unfollowing a creator, or tapping "not interested." Each of these is an explicit training example that tells the system to show less like it in the future. Because "not interested" is so direct, it retrains a profile faster than passive scrolling ever could.
Underneath, recommendation systems often use what is called a two-tower model. In simplified terms, the system learns separate numerical representations of users and of items, and matches them by similarity to retrieve candidates that might interest you. After retrieval, a more expensive ranking stage decides what to actually show. That ranker is typically multi-objective: instead of optimizing only for clicks or watch time, it juggles retention, satisfaction, diversity, and safety. This is why you'll see "integrity" objectives added to ranking, so that content which might perform well on pure engagement but is spammy, scammy, or harmful gets suppressed.
Two specific short-video metrics capture most of what rankers watch. Watch time per impression is the average watch time earned each time a video is shown, and it combines whether viewers start the video, how long they stay, and whether they replay. Hold rate captures what happens in the first few seconds: the percentage of viewers who do not swipe away immediately. A low hold rate is a strong negative signal because it means the first impression failed, no matter how good the later content might be. Together, these signals drive what you see next.