Classic retention

Definition

Classic retention is the percentage of a cohort that is active on day N exactly, counting only activity on that specific day and ignoring activity on any other day.

Also called day-N retention or bracket-free retention, it is the strictest of the three common definitions and the lowest number any of them will produce. Its strictness is the point: because each day is measured independently, the resulting curve carries information about usage rhythm that smoothed definitions deliberately average away.

One day, counted alone

A user counts toward classic D7 only if they were active on day 7. Being active on days 1 through 6 and on day 8 does not help. This makes classic retention a snapshot rather than a cumulative measure, and it is why the number is so much lower than teams expect the first time they compute it.

User B is the case that provokes the most objection: six consecutive active days and classic D7 records nothing. That is correct behaviour for the metric, not a flaw. Classic asks whether the cohort is present on a given day, which is the question you need when comparing one day against the same day for another cohort — and B's absence on day 7 is real information.

User C shows the complementary case, where classic misses a returning user that rolling retention catches. Which of those blind spots you can tolerate is the actual basis for choosing between them.

The jagged curve is the signal

Plot classic retention day by day for almost any consumer app and the line has teeth. Days 7, 14, 21 and 28 sit above their neighbours; weekend days sit above or below depending on the category. This is usage rhythm, and it is one of the few metrics that shows habit formation directly.

A peak at D7 that survives to D14 and D21 means a genuine weekly loop exists. A D7 peak that vanishes by D14 usually means a scheduled notification produced a one-off return rather than a habit — a distinction worth knowing before crediting the lifecycle campaign that produced it.

Computing it

-- Active on day N exactly. No BETWEEN, no >=.
WITH cohort AS (
  SELECT user_id, DATE(installed_at) AS cohort_date
  FROM installs
),
day_index AS (
  SELECT DISTINCT
    c.cohort_date,
    c.user_id,
    DATE_DIFF(DATE(s.started_at), c.cohort_date, DAY) AS day_n
  FROM cohort c
  JOIN sessions s USING (user_id)
  WHERE s.is_foreground
),
sizes AS (
  SELECT cohort_date, COUNT(*) AS cohort_size FROM cohort GROUP BY cohort_date
)
SELECT
  d.day_n,
  ROUND(100.0 * COUNT(DISTINCT d.user_id) / SUM(z.cohort_size), 2) AS classic_retention_pct
FROM day_index d
JOIN sizes z USING (cohort_date)
WHERE d.day_n BETWEEN 1 AND 30
  -- A cohort counts toward day N only after N full days have elapsed.
  AND d.cohort_date <= CURRENT_DATE() - d.day_n - 1
GROUP BY d.day_n
ORDER BY d.day_n;

The equality on day_n is the entire difference between this and the other definitions. Change it to BETWEEN 1 AND 7 and you have range retention; change it to >= 7 and you have rolling. Keeping the three in one file, with the predicate as the only difference, is the most effective documentation available.

One practical note: because classic counts a single day, small cohorts produce noisy numbers. A cohort of 200 installs at 6% D7 is twelve users, and a swing of three users moves the figure by more than a point. Aggregate cohorts by week before drawing conclusions at low volume, and never compare single-day cohorts across a holiday.

When classic is the right choice

• Comparing cohorts. Same-day-index comparisons are exactly what classic is built for.
• Detecting release regressions. A cohort-wide drop at a specific day index localises the change.
• Diagnosing habit loops. The peaks and troughs are the diagnostic.
• Comparing acquisition sources. Strictness makes low-quality traffic obvious sooner.

It is the wrong choice for apps with a naturally irregular cadence — a travel booking app, a tax tool, a delivery service used fortnightly. There, a user active on day 6 and day 8 but not day 7 is a perfectly healthy user that classic scores as churned. Range or rolling retention describes those products better, and the retention rate overview compares all three on one cohort.