Beyond Average Revenue: A Four-Part Decomposition for Measuring Discovery Experiments
Mean revenue is a noisy metric for A/B tests in e-commerce - it conflates purchase probability, visit frequency, and basket size into a single estimand. We decompose them, fit each with an appropriate GLM, and embed the result inside a 2² factorial design to measure two platform features simultaneously. The approach recovers clean per-feature estimates even under a confounded allocation mechanism.