How do recommendation algorithms decide what to show me?

Mostly it's predicting which items you're most likely to engage with, learned from the behaviour of millions of people — including yours. It feels like it knows you because behaviour is far more predictive than people expect, not because it has special insight. The two core techniques: 1. **Collaborative filtering** — 'people who behaved like you also liked this'. The system finds users whose history resembles yours and recommends what they engaged with that you haven't seen. It needs to know nothing about the *content* itself. This is why a recommendation can be eerily good in a category you never searched for: a few thousand people with your exact pattern watched it next. 2. **Content-based filtering** — matching item attributes (topic, genre, creator, audio features, visual features) to a profile of what you've engaged with. Useful for new items with no engagement history. Modern systems combine both in a two-stage pipeline: a fast *candidate generation* step narrows millions of items to a few hundred, then a heavier *ranking* model scores those precisely for you, right now, on this device. What actually goes into the ranking, which is broader than most people assume: - Explicit signals: likes, subscriptions, saves, searches. - **Implicit signals, which matter more**: watch time, completion rate, rewatches, how long you paused on something while scrolling, whether you turned sound on, whether you came back. Dwell time is enormously predictive and entirely passive — you're 'telling' the system things constantly without any deliberate action. - Context: time of day, device, session length, what you just watched. - Item signals: how it's performing with similar users right now, freshness, and creator-level patterns. What's being optimised, and why it matters: usually a blend of predicted engagement and time spent, with adjustments. That objective is the source of most criticism — content that provokes strong reactions performs well on engagement metrics regardless of whether it's good for the viewer. Platforms have added counterweights (diversity, 'satisfaction' surveys, downranking borderline content) with mixed and largely unaudited results. The filter bubble effect is real but subtler than the popular version: the system doesn't decide to narrow your world. It reinforces whatever it observes, and small initial preferences compound through feedback loops. You engage slightly more with one type of thing, you see more of it, you engage more. Practical control: your watch and search history is the main input, and most platforms let you view, delete and pause it. Removing a few items you regret engaging with changes recommendations noticeably. Explicit 'not interested' signals do work, though less strongly than the passive behavioural ones — which is the honest asymmetry at the heart of it.