YouTube's recommendation engine drives % of watch time:
Two-stage approach: Candidate generation: Narrow millions of videos to hundreds using user history and collaborative filtering Ranking: Score candidates by predicted engagement (clicks, watch time, likes)
Features used:
- Watch history and search queries
- Demographics and device type
- Video freshness and quality signals
- Creator authority
Cold start: New users get popular content. New videos get exploration budget.
Real-time signals: Recently watched videos immediately influence recommendations.