Chapter 1 showed that a naive comparison decomposes into a true effect plus a bias term, and that the bias term exists only because treatment was correlated with a user's pre-existing type. This chapter works through exactly how randomization severs that correlation, walks a full 3-arm coupon experiment end to end, and covers the checks that separate a trustworthy randomized result from one that quietly wasn't.
Recap: Last Chapter's Key Formula
Chapter 1 landed on one decomposition worth carrying forward: any naive comparison between a treated group and an untreated group equals the true effect on the treated, plus a bias term coming from the two groups being different kinds of users to begin with...