Hours watched differences may predict retention differences farther in the future better than short-term measurements of retention, which may measure reaction to changes more than it measures the steady-state effect of the change.
If you want to act on A/B tests faster than directly measuring long-term retention would allow, having a proxy measure of long-term retention that mitigates the risk of optimizing for transitional rather than steady-state retention has value.
