How the YouTube Recommendation Algorithm Works

YouTube recommendations help viewers discover videos across surfaces such as the personalized Home page, Up Next, and Shorts. Current YouTube Help says recommendations draw on signals including watch and search history, subscriptions, likes, dislikes, “Not interested” feedback, “Don’t recommend channel” feedback, and satisfaction surveys. The relative importance of a signal differs by surface: the official help page says the currently watched video is central to Up Next, while Home primarily uses watch history. Users can manage or delete history and provide feedback. For a systems explanation, Google researchers’ 2016 paper “Deep Neural Networks for YouTube Recommendations” describes a two-stage design: a candidate-generation model retrieves a smaller set from a large video corpus, and a separate ranking model scores those candidates. Google’s current Machine Learning Crash Course teaches a broader three-stage recommender pattern with candidate generation, scoring, and re-ranking. These are public architecture descriptions and teaching models; they should not be presented as a full specification of YouTube’s current internal ranking system. Signals and model stages serve different purposes. Candidate generation narrows the search space; ranking evaluates a smaller set using more detailed features; later adjustments can account for constraints, diversity, or freshness. A recommendation is not simply a sorted list of views: the public YouTube help page describes viewer interests, feedback, and satisfaction signals. Creators can study traffic sources in Analytics, but cannot infer a guaranteed ranking formula from a public paper or one metric. Recommendations evolve, and surface-specific behavior matters.