Return cost shows up on this month’s P&L. The benefit of a good return shows up months later, as an order from someone who might have left. Measure only the first, and every policy decision drifts toward cutting cost and quietly losing customers.

The four-way split

Take every customer who made a first purchase in a given quarter, and split them by what happened to their first return: none, refunded, exchanged, or store credit (add returnless refunds if you use them). Track each group’s repeat rate at 90 days and twelve months, and their contribution. Appendix A has the query; The Second Order covers cohorts in depth.

Expect the exchange group to look best, and don’t read too much into it: customers who exchange were already more committed. The split is a map of where the value sits, not proof of what caused it.

The split tells you where to look. Only a holdout tells you what a policy did.

Test policy changes against a holdout

Almost every change in this guide can be randomized by customer: bonus credit on or off, refund at scan or at receipt, fee or no fee. Keep a random 10% to 50% of customers on the current policy and compare repeat purchase and contribution per customer. The Honest Test covers reading a test without fooling yourself.

Repeat rate is slow and noisy, so size the test first. To see twelve-month repeat move from 30% to 33%, you need roughly 3,700 customers per group Derived, from the rule of thumb 16 × p(1 − p) ÷ d² (80% power, 5% significance). Read a signal at 90 days; decide at twelve months.

What a change must do to pay for itself

A more generous policy has a cost you can count: labels you now pay for, a few more returns. The tool below turns that cost into the repeat-rate lift it has to buy. For a fee you’re considering, enter its expected income as the cost of not charging it: the answer is how much repeat purchase the fee can lose before it loses money.

With the defaults, free labels cost $49,500 a year, which repeat purchase repays if the twelve-month repeat rate rises 3.3 points, from 28% to 31.3%. If only the 20% of customers who return something respond, their repeat rate must rise 16.5 points. A holdout needs about 3,000 customers in each group to see the overall lift. Whether that’s plausible is exactly what Bower and Maxham’s numbers suggest and don’t prove.

Do this

• Build the four-way split for customers acquired 12 to 24 months ago, and add it to a quarterly report.
• Before any policy change, run it through the tool and write down the lift it needs. If the needed holdout is bigger than a year of first-time customers, make the change bolder or don’t make it on this evidence.