TikTok's Recommendation Engine: Anatomy of a Billion-Dollar Feed

AI · 7 min read

TikTok uses a cascade of ranking models that start with massive candidate generation and progressively filter to a ranked feed. Initial recall pulls content based on coarse signals like user history, trending clusters, and author network; lightweight scoring then prunes to a shortlist for deeper models. The ranking models optimize for predicted watch time, rewatch probability, and downstream actions such as follows and shares.

Crucially, the UX hides this complexity behind an infinite scrolling surface where immediate feedback is constant. Short video length and a low-friction next action make it easy to encode user preferences quickly, which in turn accelerates model retraining. Design decisions like autoplay, mute-by-default, and the like/follow affordance directly shape the label quality used for training.

There are also deliberate interventions for novelty and diversity: temporal decays, topical exploration buckets, and manual editorial boosts. For product teams, the main takeaways are to instrument short-term behavioral signals, provide low-friction feedback loops, and treat the feed as an ecosystem where model, UX, and editorial policy must co-evolve.