Retention

For a long stretch, the proof that a product was working was how many new people showed up that month. Three thousand signups beat last month's twenty-five hundred, so the company was growing. For software sold once and installed, that arithmetic held. The money arrived at the moment of sale, and whether the buyer ever opened the thing again changed nothing on the books.

Subscriptions broke the arithmetic. A signup is only the start, and revenue exists only if that person is still there next month. Spend thirty million won on ads to bring in three thousand signups, watch twenty-seven hundred of them disappear within a month, and the only number that grew is the signup count while the only number that shrank is the bank balance.

The same meeting kept repeating at Pagedeck. Ravi would report three thousand new signups for the month, Grace would ask how many of last month's three thousand were still around, and nobody in the room could answer. The company was watching the number of people arriving, not the number staying. Everyone knew how much water was being poured in each month; nobody had counted the holes in the bucket.

Retention exists to count those holes, and it does it by turning the measurement around. Instead of asking how many arrived this month, it asks how many of the people who arrived last month are still here. Look at the inflow and you get a report card for marketing. Look at what stays and you get one for the product. People coming back means they found a reason in the product itself to open it again, and that is not a result any ad copy can manufacture.

Metrics on the product side line up in the order a user passes through them. Acquisition, where a new person first arrives, is the first box. Activation, where that person first feels what the product is good for, is the next. Retention — how much they keep coming back after that — is the third. The order matters. A user who never activated never met a reason to return, so when retention is low the thing to fix is often not retention but the box in front of it.

Five terms sit under retention, in two groups: three ways of counting the same phenomenon, and two terms that look at it from the leaving side.

Whichever counting method you use, the skeleton of the calculation is the same. Four decisions have to be made before a number exists.

1. Fix the group in the denominator. Everyone who arrived in the same window gets treated as one block, and that block is called a cohort — the 500 people who signed up in the first week of September, for example. Mix new and existing users in one denominator and nobody can explain what the result means.
2. Define what "came back" means. Whether a login counts, or opening a document is required, or only an edit counts, can double the number on identical behavior. Pagedeck counts opening or editing a document.
3. Pick the interval. Match it to how often the product is meant to be used. A collaborative document tool gets opened every working day, so daily or weekly windows fit. Measure a flight booking service — used once or twice a year — in daily windows and nearly every day reads zero.
4. Divide. If 160 of the 500 came back on day 7, that's 32 percent.

The three counting methods are variations on steps 3 and 4.

• N-day retention — the share who came back on exactly the Nth day after signing up. Day 7 alone is what counts. Someone who used the product on three days in between but didn't show up on day 7 isn't counted. It suits products meant to become a daily habit, and because it's the strictest method it produces the lowest numbers.
• Rolling retention — counts anyone who came back at least once on or after day N. Missing day 7 but showing up on day 12 still counts as retained. It suits products used occasionally, and on the same cohort it always reads higher than N-day retention. That gap is why the two numbers must never be compared side by side.
• Cohort retention — builds several cohorts by signup period and lines them up against each other. Put the week-4 figure for June signups next to the week-4 figure for September signups and you can see whether the product got better in between. The first two produce one number for one group; this one shows change over time.

The two terms on the leaving side describe the same events from the other direction.

• Churn — users or customers going away. It's the mirror image: 30 percent retention is 70 percent churn. Both words survive because different people need them. Product teams use retention because they want to look at what the remaining users do; finance and revenue teams use churn because they have to total up the money that walked out this month.
• Resurrected user — someone who left and came back. These people fall outside the original denominator, so they get counted separately. A user who hadn't logged in for 90 days and reopened the product after a feature announcement email belongs here.

Once a number exists, judgment comes next, and comparing the absolute value against someone else's is nearly useless. A messenger opened several times a day and a tax filing service used once a year cannot be held to the same bar. What gets read instead is the shape of the curve over time.

Plot one cohort at weeks 1, 2, 4, 8 and 12 and you might see 60, 45, 33, 31, 30 — falling, then leveling off. That flat 30 percent is the share of people who found a lasting reason to use the product. A curve that reads 45, 28, 14, 6, 2 went all the way to the floor, which means nobody found such a reason, and in that state any amount of new traffic drains out at the same rate. This is why practitioners check whether the curve has flattened before anything else.

A jump in the days right after a feature ships deserves suspicion. The illusion of something looking good simply because it is new is called the novelty effect, and it usually settles back within two weeks. So improvement is judged not by one cohort rising and falling but by comparing the same week number across several cohorts. The general practice of grouping people by when they arrived and comparing the groups is cohort analysis, and retention is where it gets used most.

Products sold to companies often measure the same thing in money rather than headcount. Net revenue retention takes what existing customers paid, adds expansion and subtracts cancellations, and reports what remains a year later. If two of ten accounts left but the remaining eight added seats, headcount retention reads 80 percent while the revenue figure can read 110 percent.

Growth teams watch the same stretch for a different purpose. There, the period in which users keep coming back is itself called the retention stage, and the question is which change makes the number move. The measurement is identical; the use is not. On the product side retention is a diagnosis of the current state. On the growth side the same value is the verdict on whether an experiment worked. Two teams can sit in front of the same table and ask completely different questions of it.

"Signups keep climbing but week-4 retention is 12 percent. We should stop pouring in at the top and look at the bottom."

A proposal to pause acquisition work and find out why nobody is staying.

"Was that measured rolling or N-day? We can't mix the two in one report."

A correction: different methods produce different values, so the basis has to be agreed first.

"Last week's bump might be the novelty effect. Let's wait for the September cohort's week 4 before calling it."

Don't read a short-term rise as improvement; compare the same week number across cohorts.

Pagedeck reviews a weekly retention table every Monday. Columns are weeks since signup, rows are signup weeks, so lower rows are more recent arrivals. Mia doesn't follow a single row across; she reads one column up and down. In the week-4 column alone she can see that 27 percent of June arrivals stayed and 34 percent of September arrivals did.

One finding came out of that table last quarter. Users who invited at least one colleague and users who worked alone split at 41 percent and 9 percent by week 4. When Mia brought those two rows into the meeting, the next quarter's work settled itself: instead of adding features, the team would rework the early screens to get a colleague invited during the first week. The number itself never says what to do. Split the group and the work appears inside it.

The most common mistake is reading low retention as a retention problem. The name invites it. But this number is usually a result of what happened in an earlier box. If the wrong people were brought in, or if the people who arrived never once experienced what the product is good for, there is no reason for them to be there in week 4. When users from a discount promotion sit at 5 percent in week 4 while users who arrived through search sit at 30 percent, the product didn't fail to hold anyone — the wrong people were recruited. Raising the number by sending more "come back" notifications works briefly, and the visits bought that way drain out again the following month.

Judging a product by one blended average is the other frequent error. The dashboard shows a single figure, so it looks like a report card. But an average flattens unlike groups together: behind 30 percent there may be team accounts staying at 70 percent and individual accounts falling to 4 percent. Set a goal of moving the average from 30 to 32 in that state and nobody knows which side to touch. Until the number is split by acquisition source, plan, and company size, it doesn't tell anyone what to do.