Could cohort based on geo-location instead of completely at random. Makes analysis a bit more challenging though and doesn't completely solve the problem of course.