Journal · 28 Feb 2026
Building a segment that survives iOS ATT
App Tracking Transparency did not only change ads. It changed how late, sparse, or duplicated a user identifier can be when you try to stitch behavior. Teams that still define segments as “users whose id did X in the last 14 days” quietly drop a large iOS slice, then wonder why activation looks healthier on Android.
In Cohort Signal Lab we ask students to write two versions of the same cut: one that requires a stable account id, and one that can run on a device-scoped event stream. If the second version cannot be stated, the first version is not a product segment. It is a CRM export.
Probabilistic stitching is not a personality
Some stacks offer “probable” identity graphs. Those graphs can be useful for measurement studies. They are a poor foundation for a segment that will trigger an in-app surface. When the graph revises itself, your “loyal” bucket gains and loses people who never changed behavior. Faculty will fail that design on critique night.
What still works
First-party events after sign-in remain the cleanest. Before sign-in, segment on sequences inside a single install: search then save, browse then exit, paywall then settings. Those sequences do not need to know who the human is across devices. They still tell product what the current surface is doing.
Thailand adds a wrinkle: many households share an Android device even when iOS ids are thin. A device-scoped segment can still mix two people. The honest move is to name that risk in the cohort frame, not to pretend ATT was the only identity problem.
If your roadmap depends on cross-device loyalty before login, fix the login prompt or accept that you are guessing. The studio will not help you decorate the guess.