Sounds like they optimize on the full population and not for individual. Because if they did it for OP well the metrics would improve.

It is just the curse of naive AB testing. Same stuff happens with apps like Duolingo where it is optimized e2e, not personalized, and then they behave as if their approach is sane.