Maybe they were too busy micro-optimizing engagement rates using a very large amount of A/B experiments and got stuck at a local maximum?