Rolling Retention

The simplest way to count who came back is to pick one day and check whether they showed up. Present on day 7 counts as retained, absent counts as gone. N-day retention is that method.

The trouble starts with products nobody opens daily. Pagedeck's 7-day retention read 12 percent for a while. Nine out of ten users vanishing in the first week made for heavy meetings. Then the records got opened up, and plenty of those users had worked three days that first week and come back on day 8 and day 11. A collaborative document tool gets opened on days when there's work to do, so the odds of any one specific day landing on a visit were low to begin with. When Mia counted activity from day 8 onward, 42 percent of the same group had come back.

Rolling retention closes that gap by changing how the counting works. Instead of taking attendance on a fixed day, it counts anyone who showed up on any day after day N. Asking which of 12 percent and 42 percent is the real number has no answer. The first asks whether a daily habit formed; the second asks whether the person has given up on the product.

Put a signup group in the denominator, put everyone from that group with at least one activity record on or after day N in the numerator, and divide. Four things get settled first.

1. Define the group that forms the denominator. Everyone who arrived in the same window becomes one block, and that block is called a cohort — the 500 people who signed up in the first week of September, say. Mix new and existing users and nobody can explain what the resulting number means.
2. Define what counts as coming back. Whether a login is enough, whether a document has to be opened, or whether it has to be edited can move the number by a factor of two. Pagedeck counts opening or editing a document.
3. Pick N. Match it to how often the product is meant to be used. Seven or fourteen days for something opened on working days; ninety for something used once every few months.
4. Pick the day observation ends. This step exists only for this method. Since any visit after day N counts, a number that never stops collecting keeps climbing. Most teams cut it at a fixed multiple of N.

If 210 of the 500 September signups opened or edited a document between day 8 and the cutoff, the answer is 42 percent. Implementing it needs only the last-activity date: everyone whose most recent visit falls more than N days after signup is the numerator.

Measure one cohort both ways and rolling always reads higher, because everyone present on day N is also someone who came back at least once after day N. Which is why the two numbers can't be lined up to argue that something improved or got worse.

Comparing the absolute figure against another company's is close to useless. A messenger opened several times a day and a tax filing service used once a year don't belong on one scale. What gets read instead is the shape of the curve as N grows.

Chart one cohort at 7, 14, 30, 60, and 90 days and a series like 60, 45, 33, 31, 30 falls and then stops. That last 30 percent is the share who found a reason to keep using the product. A series like 45, 28, 14, 6, 2 runs to the floor instead, which means nobody found that reason, and in that state new signups drain out as fast as they arrive.

There's one more trap. The longer the observation window, the higher the same cohort reads on its own. Measure the September cohort in October and it's 42 percent; measure it again in December and it's 47. Nobody came back extra — there were simply more days to count. So comparing cohorts means holding the window at one length: 90 days for the June cohort, 90 days for the September cohort.

A jump in the days after a feature launches deserves the same caution. The illusion of improvement that comes from newness is the novelty effect, and two weeks later the number usually sits where it was. So improvement gets judged by comparing the same stretch across several cohorts rather than by one cohort's swings. The broader practice of grouping users by arrival window to compare them is cohort analysis.

"7-day retention is 12 percent and rolling is 42. Both are correct."

The same group counted two ways, so neither figure is wrong.

"To compare against last quarter's rolling retention, cut both at 90 days."

Windows of different lengths don't produce a valid comparison.

"High rolling doesn't mean people are using it. One visit in 90 days still counts."

A single visit and steady use are worth separating.

Pagedeck keeps both numbers in one table. The 7-day figure is the column for whether a habit formed; the 30-day rolling figure is the column for who hasn't left. Mia reads the two together at the monthly review.

Last quarter the first column moved from 12 to 14 percent while the second held at 42 and then 41. New users were opening the product more often in week one without any change in how many stayed, so the team pointed the next quarter at reasons to come back a month later rather than at first-week onboarding.

The most common reading is that a higher number makes it the better metric. The same product does look better through it, so the reaction follows naturally. But it reads high because the counting condition is loose, not because the product is strong. Leave that in place and a team picks whichever method flatters the report, and a table whose calculation changes each quarter stops showing whether anything improved. Pick one method and keep it, even in the quarters where the stricter reading looks bad.

The other common belief is that one return visit makes someone retained. The definition does say exactly that. But a person who logged in once in 90 days and a person who logs in three times a week both count as 1. Judge product health on this figure alone and an account that nobody uses and nobody deleted gets read as retention. So this number gets read alongside a measure of how often people actually use the product.