YouTube Brain

Recommendation System

3 min read

Recommendation System

YouTube describes its recommendation system as automated word of mouth: it pulls each video toward the viewers most likely to be satisfied by it, rather than pushing one video at a crowd. The system has two stated objectives, to help viewers find content they want and to maximize long term satisfaction, and it learns from more than 80 billion signals. The practical takeaway for a creator is that you do not optimize the algorithm directly. You optimize for a specific audience, and the recommendations follow. See Discovery Surfaces for where this plays out and Chasing Views Over Satisfaction for the failure mode.

The eight personalization signals

The official help page names eight inputs the system uses to personalize what a viewer sees:

  • Watch history
  • Search history
  • Channel subscriptions (see Subscriptions and Subscribers)
  • Likes
  • Dislikes
  • “Not interested” feedback
  • “Don’t recommend channel” feedback
  • Satisfaction surveys

These are viewer signals, not video signals. YouTube puts it bluntly: the system pays attention to viewers, not videos. Your job is to make a video that a clearly defined viewer will choose, finish, and feel good about. That feeds Click-Through Rate, Watch Time and AVD, and Audience Retention in turn.

Satisfaction over raw watch time

The defining shift of the 2024 to 2026 period is that the system weights watch time by satisfaction. YouTube’s Todd Beaupre describes it as measuring how viewers feel about the time they spend, not just the minutes. The blog post that introduced watch time as a signal in 2012 also introduced “valued watchtime,” measured through 1 to 5 star surveys where only 4 and 5 star ratings count. The system predicts survey responses with machine learning where a survey is not shown.

This is why a viewer who finishes a short video and rates it highly can send a stronger signal than one who watches 40 percent of a long video and leaves. It is also why session continuation, whether the viewer keeps watching YouTube afterward, matters. Satisfaction, not duration alone, is the target. See Chasing Views Over Satisfaction.

How a new video is tested

The system tests a new video on a small, relevant audience first. Strong early signals, longer watching, clicks on the next recommendation, and subscriptions, expand reach, sometimes within days. For a viewer who watches a niche, the system surfaces small but promising channels that match that interest, which is why consistent topic focus helps a smaller channel. An individual video underperforming does not penalize the channel overall, and experimenting with Shorts, long form, or live does not inherently confuse the system.

What a creator controls

Sources

Graph View

Backlinks