N-day retention

Definition

N-day retention is the percentage of a cohort that is active on the Nth day after their starting event, where day zero is the day of the starting event itself.

D1, D7 and D30 are the shorthand almost every app team uses, and the label is doing more work than it looks. Two decisions sit underneath it and neither is usually stated: what "day N" is measured against, and what counts as active. Get either wrong and your retention curve is internally consistent and externally meaningless.

Day zero, and the off-by-one that follows it

The near-universal convention is that day 0 is install day and D1 means the day after. It is not the only convention in use, and the alternative — D1 meaning install day itself — shifts the whole curve by one position, which makes every cross-tool comparison wrong in the same direction.

The first and third rows are the interesting pair. Calendar-day bucketing counts a twenty-minute gap as retention; 24-hour bucketing does not. Neither is wrong, but calendar bucketing systematically inflates D1 for apps whose users install in the evening — which is most consumer apps — and that bias does not appear anywhere in the output.

What counts as active

The second undeclared decision. "Active" can mean the app was opened, a session exceeded some duration, a screen was viewed, or a meaningful action occurred. Each produces a different curve from identical logs, and the gap between the loosest and strictest definitions is routinely larger than the difference between a good month and a bad one.

A background wake or a silent push must never count. It is the easiest way to manufacture retention, it is invisible in the aggregate, and it corrupts the one metric everything else is judged against. If your SDK reports a session on a background refresh, exclude it explicitly rather than assuming it does not.

The whole curve in one query

N-day retention is worth computing as a curve rather than as three isolated numbers. The shape between D1 and D30 tells you whether you have a leaky onboarding, a missing habit loop, or a healthy plateau, and none of those are visible from the endpoints alone.

-- Day 0 is install day. Activity is a session past the splash screen.
-- Both decisions are stated here rather than left to the reader.
WITH cohort AS (
  SELECT user_id, DATE(installed_at) AS cohort_date
  FROM installs
  WHERE installed_at >= CURRENT_TIMESTAMP() - INTERVAL 90 DAY
),
active_days AS (
  SELECT DISTINCT
    c.user_id,
    c.cohort_date,
    DATE_DIFF(DATE(s.started_at), c.cohort_date, DAY) AS day_n
  FROM cohort c
  JOIN sessions s USING (user_id)
  WHERE s.reached_first_screen              -- not a background wake
    AND DATE(s.started_at) >= c.cohort_date
),
sizes AS (
  SELECT cohort_date, COUNT(*) AS cohort_size FROM cohort GROUP BY cohort_date
)
SELECT
  a.day_n,
  SUM(z.cohort_size)                                   AS users_in_scope,
  COUNT(DISTINCT a.user_id)                            AS retained,
  ROUND(100.0 * COUNT(DISTINCT a.user_id) / SUM(z.cohort_size), 2) AS retention_pct
FROM active_days a
JOIN sizes z USING (cohort_date)
-- A cohort only contributes to day N once it has had N full days.
WHERE a.cohort_date <= CURRENT_DATE() - a.day_n - 1
  AND a.day_n BETWEEN 0 AND 30
GROUP BY a.day_n
ORDER BY a.day_n;

That correlated maturity filter is the important line. It lets a 60-day-old cohort contribute to D30 while a 5-day-old cohort contributes only up to D4, so the tail of the curve is not diluted by cohorts that have not lived long enough to reach it.

Reading the curve

• A steep D0→D1 drop is normal everywhere; the size of the drop is a first-session problem, not a product-market-fit verdict.
• A curve that flattens has found its habitual core. The plateau height, not D1, predicts long-term value.
• A curve that never flattens projects to zero users. No acquisition budget fixes that.
• Weekly spikes on days 7, 14, 21 indicate a weekly habit and are a reason to prefer classic retention over smoothed alternatives.
• Divergence by source is the fraud and quality signal — see retention rate for what each shape implies.

N-day is the label; the measurement rule underneath it is a separate choice. Classic counts day N exactly, rolling counts day N or later, and both are legitimately reported as "D7". Stating which one you used costs six words and prevents most retention disputes outright.