N-day Retention

Teams have always wanted to know whether the people who signed up are still around, but turning that question into a number kept stalling in the same place. Nobody had agreed on when "came back" gets checked. One analyst counted a user as retained if they showed up at any point during the first week. Another counted them only if they were still active once that week had passed. The same 500 users produced 45 percent and 22 percent at the same time.

So meetings turned into arguments about where a number came from rather than conversations about the product. Someone would say retention was up from last month, someone else would ask how exactly it had been counted, and half the meeting went into reconciling definitions.

N-day retention removes the ambiguity by nailing down a single day. Not any day inside a window — day 7, day 30, one specific day after signup. Show up that day and you count as 1, otherwise 0. The trade for that clarity is strictness, so this measure always produces the lowest number of the retention family. That's the job it was built for: catching where a product that wants to be opened daily first breaks down.

Group everyone who signed up in the same period, then divide the number of people who showed up on exactly the Nth day after signup by the size of that group.

A group of users who arrived in the same period is called a cohort — the 500 people who signed up in the first week of September, for example. Mix new and existing users into one denominator and nobody will be able to explain what the number means later.

The calculation has four steps.

1. Pick the cohort that forms the denominator. Everyone with the same start date, like the 500 people who signed up on September 1.
2. Decide what "came back" means. 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 inside the same product. Pagedeck counts opening or editing a document.
3. Pick N. The common values are 1, 7 and 30 days, which show how many remain after a day, after a week, and after a month.
4. Divide the day's returning users by the cohort size. If 160 of 500 show up on day 7, that's 32 percent.

Step 4 is where people slip. Day-7 retention is not a running total from day 1 through day 7 — it's day 7 alone. Somebody who showed up on days 2, 3 and 5 but not on day 7 counts as a zero. That's why the same cohort always scores lower on N-day retention than on any other method.

Match the interval to how often the product is meant to be used. A collaboration tool that gets opened every working day fits a daily interval. Measure a flight booking service that people touch once or twice a year on a daily interval and nearly every day reads zero. Products like that either stretch N out to 30 or 90 days or count a different way entirely.

A single figure can't be called high or low on its own. Products have different natural rhythms and each company defines "came back" differently, so somebody else's 30 percent and your 30 percent don't point at the same thing. What's worth reading is the shape of the curve you get by plotting one cohort across several values of N.

Plot days 1, 7 and 30 at 60, 33 and 30, and the fall stops somewhere — that 30 percent is the slice of users who found a reason to keep using the product. A run like 45, 14 and 2 drops all the way to the floor, which means nobody found a reason to stay, and in that state new signups drain out as fast as they arrive. In practice teams look at whether the curve has flattened before they look at how large the values are.

Whether things improved is a question for several cohorts at the same N, not for one cohort moving around. If June signups held 24 percent on day 7 and September signups hold 31 percent, something got better in between. Comparing groups that arrived in the same period this way is called cohort analysis.

Treat a jump right after a feature launch with suspicion. The illusion of something looking good simply because it's new is the novelty effect, and it usually settles back within about two weeks.

"D7 dropped from 32 percent to 24 percent, starting with the August cohort."

D7 is shorthand for day-7 retention, and the point is that something changed for people who arrived in August, so look at what shipped that month.

"D1 is fine, but it falls apart at D7. There's nothing holding people through the first week."

Users come back the next day but not a week later, so look at the first-week experience rather than the signup flow.

"Our product isn't a daily thing, so N-day reads zero across the board."

The measurement interval doesn't match the product's rhythm, so count a different way.

At Pagedeck, Mia opens the cohort table every Monday and reads three rows: D1, D7 and D30. When one of them moves noticeably, she starts by retracing what changed during the week that cohort signed up — copy on the signup screen, a new ad campaign, an outage.

In the weeks where D7 fell, what Mia found was usually on the acquisition side rather than inside the product. People arriving from a new ad had signed up for personal notes rather than collaboration, wrote one document alone, and never came back a week later. Weeks like that get fixed by rewriting the ad, not the product.

The most common mistake is reading a zero on day 7 as a user who left. It sits in the same column as churned users in the table, so it reads that way easily. But this metric counts attendance on one specific day, which means someone who showed up on day 6 and day 8 also lands as a zero. Read that as departure and win-back emails go out to people who are actively using the product, while the count of who genuinely left stays permanently inflated. If the question is who left, look at a measure taken from that side, like churn.

Lining the number up against figures another company published is just as common. The name matches, so a comparison feels fair — except they may have counted a bare login while you require a document edit. Comparing two numbers built on different definitions makes a healthy product look like a failing one, and next quarter's priorities get set on top of that misreading. Compare your own cohorts to each other and nothing else.