[gwern]

I think changes over time in device should be like any other unobserved variable like age or country or gender which might affect traffic: averaged out by randomization. The point of randomization is that on average, the two groups will be balanced. (If the intervention itself changes the mix and the changed mix then yields more traffic, then that changes what one might speculate about how it works but the simple result of ads->less-traffic is still true.)