You size a test and the answer comes back: 8 weeks. The interviewer asks: "How would you get an answer sooner without cutting corners?" The honest answer is usually less noise, not more traffic. The variance of your metric, how much it spreads from user to user, sets the standard error of the per-arm estimate, and the number of users a test needs grows in step with that variance. Variance reduction is the set of techniques that shrink it without changing the effect you are trying to measure, so the same lift becomes detectable with fewer users.

These techniques offer little benefit when the test already fits its window. Adjusting for user history also helps little when most users are new and have no history.

Before you shrink a variance, you have to compute it correctly. Metrics like click-through rate are counted per search, while users are what you randomized, and the textbook formula quietly understates their noise. The first job is getting it right with the delta method. Then come the levers: adjusting with pre-experiment data, known as CUPED (Controlled-experiment Using Pre-Experiment Data), stratification, capping heavy tails, and a more sensitive metric. All of it runs on one test, a home-goods marketplace trying a new search ranking model where revenue per user is dominated by a few sofa orders.

This question comes up after sizing in product data-science rounds and usually gets 5 to 10 minutes. For a mid-level role, the bar is spotting the per-search variance mistake, explaining CUPED and how much it removes, and capping a revenue metric. For a senior role, interviewers listen for the cost of each lever, when stratification is worth the trouble, and the order you would pull them in. The tutorial assumes you know that sample size grows with variance, and the standard error and z-test from Hypothesis Testing and Confidence Intervals.