The problem is that optimizing for long-term satisfaction metrics is extremely tricky and hard. Such metrics lead to the better user retention, potentially even

The problem is that optimizing for long-term satisfaction metrics is extremely tricky and hard. Such metrics lead to the better user retention, potentially even with less watch time. Paradoxically higher watch time may lead to lower user satisfaction (and decreased retention). In ML world these metrics require combining DNNs with methods like Q-learning. YouTube learned these lessons and now probably has the most advanced ML among competition in its’ recommendation system.