The reassignment maybe shouldn't be random but it should look at the current "score" and decide which verison (a or b) should be shown so that the significance is increased. For super large experiments this will probably not matter, but I guess since you are building for indie founders, this would help to avoid "bad luck" scores