Valued watch time

YouTube's recommendation system does not treat every minute watched the same. Alongside raw watch time it uses satisfaction signals gathered from in-product surveys asking viewers how much they liked what they watched, plus likes, dislikes and "not interested" feedback. Time on videos viewers rated highly counts as valued watch time, time on videos they rated poorly is discounted, and machine learning fills in a predicted rating where no survey response exists.

The practical consequence catches people out. Stretching a video to hoard minutes can raise its raw watch time while lowering its satisfaction scores, and the flat retention curve that padding produces tends to get the video pushed less over time.