Field note
Week-4 return is a reasonable starting question for a consumer app. For a learning product sold into Thai schools and tutoring homes, it is often a measure of the calendar, not of the product.
Every February we watch the same argument. A growth lead sees a cliff. A tutor says the children are in exam weeks. Both are looking at a true chart and a false conclusion. The chart is usually “percentage of install cohort still opening the app in week 4.” The conclusion is usually “we have a retention problem,” followed by push notifications.
In Cohort Signal Lab we ask a duller question first: what would a healthy learner look like during this teaching week? If the subject has a mock exam, silence in the app can mean revision on paper. If the family is travelling after exams, the learner may return on a sibling’s phone. Counting that as churn trains the product to nag people who were never gone.
We still look at return, but we change the denominator. Return among people who submitted at least one in-app check in week 1 is more stable than return among everyone who installed. Install includes parents who opened the store listing once. Check-submit is a person who attempted the subject.
The second signal is revisit of a specific lesson after a gap of more than ten days. In language and maths apps this often marks genuine return, not a stray notification tap. It is ugly to implement if your events do not carry lesson identifiers. That ugliness is useful: it forces the dictionary work.
The third is device switch. Shared Android phones are common in the homes we see. If your identity model treats a new device as a new learner, February “churn” is partly a family logistics problem. Instrumenting a stable learner key is product work; pretending it is a campaign problem wastes money.
Session count during exam weeks. Time-in-app. Streaks. These can be interesting later. They are terrible north stars when the school calendar is loud. A streak that breaks because a child sat O-NET is not a failure of onboarding copy.
We also stop averaging all subjects together. A coding prep app and a primary English app do not share a week shape. If you only have one chart, split it before you brief the board.
Overlay the teaching calendar on the same axis as your return curve. Use actual term dates for the schools you sell into, not a generic “Q1.” When the overlay explains the dip, write that sentence in the brief. When it does not, you have a product question worth a sprint.
This is slower than a win-back campaign. It is also cheaper than running one every February. If you want to sit that desk, the syllabus is on the courses listing.