In autumn 2024 I spent two weeks auditing Nordvik (fictional), a physiotherapy app with 40,000 patients in the Nordics. The founders asked for a growth plan. The data asked for something else, and it showed up as a cliff in the retention curve at week four.

The curve that looked fine

The headline retention curve looked normal for a health app: a drop in week one, then a slow decline. Week-eight retention was 31%. Nobody had split it by where patients came from.

When we did, there were two products hiding in one chart. Patients referred by a clinic were still active at week eight 52% of the time. Patients who came from paid ads were at 14%, and almost all of that loss happened in one week: week four.

Retention by source (SQL): -- share of each cohort still active, by week · SELECT source, week, · COUNT(DISTINCT patient_id) * 1.0 · / MAX(COUNT(DISTINCT patient_id)) · OVER (PARTITION BY source) AS retained · FROM weekly_sessions · GROUP BY source, week · ORDER BY source, week;

What happened in week four

The free exercise plan lasted 28 days. On day 29 the app showed an upgrade screen and nothing else. Clinic patients had a physiotherapist telling them to continue. Ad patients had a paywall, a sore knee that felt better, and no reason to stay.

In 26 interviews, nine ad patients said a version of the same sentence: 'I thought I was finished.' The product had told them the plan was complete. We had built the cliff ourselves.

I thought I was finished.

Nine of 26 interviewed patients, Nordvik audit, 2024

What we changed

The audit recommended three things. A video check-in with a human physiotherapist in week four, before the plan ended. A second plan, built from that check-in, so the next step was a plan rather than a payment. And moving EUR 1.2M of the growth budget to retention until the ad cohort's week-eight number came within ten points of the clinic cohort.

Six months later week-eight retention across all patients was up 18 percentage points.¹ Nordvik restarted paid growth in spring 2025, on a base that now kept the people it paid for.

Try this on Monday: Split your retention curve by acquisition source before you spend another euro on acquisition. If one source falls off a cliff in a specific week, look at what your product does to those users that week.

What I would do differently

I would have asked for the support inbox on day one. When I finally read it in week two, the week-four cliff was already there in plain words: dozens of messages asking whether the plan was over. Interviews confirmed it, but the inbox would have found it a week sooner.

Notes

Nordvik's product analytics, October 2024 to April 2025, all acquisition sources combined. The clinic cohort improved too, by 6 points, from the week-four check-in alone.