Journal · 12 March 2026
Retention windows after first purchase
Teams copy a 14-day window from a slide they liked. Fashion looks brilliant. Grocery looks broken. The window was never a measurement decision — it was interior design.
In Cohort Signal Architecture we ask a blunt question before anyone draws a curve: when does the first-purchase clock start, and when does it stop being “retention” and start being “a new campaign to a cold-ish buyer”? If you cannot answer in one sentence, your holdout is already leaking.
The clock is part of the cohort contract
Entry event is not “became a customer sometime.” It is the timestamp of the paid order that first crossed your threshold — including COD confirmation in markets where that matters. In Thailand retail, we have seen teams start the clock on checkout click and then wonder why repurchase looks weak: a slice of those checkouts never became paid orders.
Write the stop rule next. A grocery app with weekly baskets cannot share a fashion boutique’s 14-day romance. We usually ask grocery desks to try 7 and 21 days on the same holdout, then pick the window where the control group still has enough repurchase events to talk about. If the control is empty, you do not have a window; you have a wish.
Do not let category mix choose for you
A marketplace that sells both lipstick and rice will produce a “blended” 14-day number that flatters lipstick. Split the contract. If you cannot split yet, say so on the signal board. Pretending the blend is a strategy is how vanity metrics sneak back in.
Holdouts have to fit inside the window
A 7-day window with a 10-day freeze is a contradiction. Module three of the flagship course spends an uncomfortable hour on this. The guest critic from finance will ask whether anyone exposed to a brand burst during the window was excluded. If the answer is a shrug, the lift figure does not ship.
Alumni from the March fashion case on our reviews page cut their claimed lift sharply once the window and the burst exclusion agreed with each other. That is the job of Lifecycle Messaging Analytics: fewer numbers, harder to dismiss.
A practical default, not a universal law
If you need a starting point while you wait for data: grocery 7–10 days, fashion 21–28, travel 45 with a huge contamination warning around public holidays. Then throw the default away the moment your control group can speak. The Quiet-Hour Calibration Lab will not fix a wrong window; it only tells you when not to shout inside it.