1. Running many experiments at once

Your grocery app has 2M weekly users and about 40 A/B tests live right now: a new search ranking, a free-delivery banner, two checkout-button redesigns, and dozens more. Every user is in many of them at once. Sooner or later the interviewer asks the follow-up: "You run all these tests at the same time on the same people. Don't they interfere?"

Concurrent experiments are tests that overlap on one population, and this tutorial is about the relationship between them. The principle is that each test assigns users independently, so every other test lands evenly in both of its arms and adds noise, not bias. Tests need to be kept apart only when they change the same element, such as two redesigns of one checkout button.

This is an optional follow-up in a data-science experimentation round, usually a few minutes after you have analyzed a single test. For a mid-level role, explain why independent overlap does not bias the result and describe salted-hash assignment. For a senior role, also show an interaction check you can compute, a clear rule for when to make tests exclusive, a global holdout, and a sensible traffic policy.

The tutorial assumes you can randomize one test, covered in Experiment Design, and compute a standard error and z-statistic for a difference in proportions, covered in Hypothesis Testing and Confidence Intervals.