Retention Analysis

Numbers used to be read only at a single point in time. "9,100 active users this month" states the current size and hides everything happening inside it.

The same 9,100 can be two completely different situations. In one, most of the people who were there last month are still there. In the other, half of last month's people left and new arrivals filled the seats. The first is growing; the second is pouring water into a leaking bucket, and at a single point in time the two are indistinguishable in a report. A company whose user count halves the month after advertising stops is the second kind.

Retention analysis exists to expose that difference, by grouping people by when they arrived and standing those groups along a time axis. Seeing how much of each arrival group remains after a few weeks separates the problem of bringing people in from the problem of keeping them, and those two problems get solved by different teams.

Group users by when they arrived and count how much of each group remains as the weeks pass. Five things to settle.

1. Pin down what counts as remaining. Whether a sign-in is enough or someone has to open a document changes the figure substantially. Picking the one action through which the product delivers its value gives a number that can be acted on, even though it always looks worse than the sign-in version.
2. Choose the grouping unit. Everyone who signed up in the same week counted as one group is the default, and grouping by arrival period in order to compare is what cohort analysis means. Products used on a longer rhythm group by month instead.
3. Choose the time axis. Daily for a product used daily, weekly for a work tool. Get it wrong and you spend the analysis explaining why the line dips every weekend instead of looking at the trend.
4. Build the table and draw the curve. Elapsed periods across, arrival periods down, is the standard shape. Start with 100 people and keep 45 in week one, 30 in week two, 24 in week four, and that group's line falls steeply and then eases. That shape is a retention curve.
5. Watch whether it flattens. Where curve flattening appears, the line settling at some height instead of continuing down, that share is the proportion of arrivals who make a home in the product. A line that runs all the way to zero means nobody is being held, and adding traffic will not change that.

That much covers the whole population; the next step is splitting it. Layering on segment analysis, which divides groups to compare them, produces separate curves by plan, by team size, by acquisition channel. A flat curve for one segment inside an overall curve that never flattens is a common finding, and that segment is usually who the product is actually for.

It fits products used repeatedly. Where revenue comes from subscriptions, or the tool gets opened every week, the share that remains is what sets the size of the business. When the decision is whether to staff acquisition or staff keeping people, this analysis answers it.

It pairs often with funnel analysis, which counts drop-off at each step. A funnel shows the losses in the first days after signup well; retention analysis covers the following weeks and months. Half the arrivals disappearing in week one surfaces in the funnel; people quietly leaving in month four surfaces here.

The bad fits are clear. For a product used once, the analysis doesn't apply — measuring week-four retention on a service for getting a moving quote means nothing. In enterprise products sold on annual contracts, an individual's weekly return doesn't connect to whether the contract renews; renewal rate and seats in use get watched separately there.

The limit is time. Knowing the week-four figure takes four weeks, and the three-month figure takes three months. So the first job is drawing past groups from records you already have rather than waiting for new ones to mature. Small groups also make the values jump: in a product with twelve signups a week, 25 percent week-four retention means three people, and one person changing their mind makes it 33.

"The overall curve doesn't flatten, but for teams of five or more it does."

The signal appeared only after the population was split by segment.

"Monthly actives are flat while week-four retention has fallen two months running."

Worrying about next quarter's number rather than this month's.

"Let's change remaining from a sign-in to opening a document."

Aligning the counting rule with the place the product's value actually lives.

At Pagedeck this analysis changed a decision in a quarter when active users were rising. Monthly actives had climbed two quarters in a row, and a proposal to increase the advertising budget came to the review.

What Mia found after grouping by arrival period was different. The curve for solo users ran down to 6 percent by week eight and hit the floor, while users on teams of five or more flattened at 26 percent. Almost all of the growth in actives had come from the first group. The decision to spend more on advertising was held, and reworking the first experience so that people arrive as teams became the quarter's work instead.

The most common mistake is reducing it to one retention number. "Week-four retention is 24 percent" finishes a report, which makes it tempting. But the answer this analysis gives is the shape of the curve, not a figure. The same 24 percent means one thing on a line still falling and another on a line that flattened in the low twenties weeks ago. The first becomes 10 percent shortly; the second stays put. Report one cell and that difference disappears.

Reading a low number as a bad product is the other one. A low figure is certainly a signal, but what it's low against comes first. Week-four retention on a messaging app opened daily and on a tool used once a quarter cannot be held to one standard. What the figure gets read against is the rhythm the product is naturally used on, and whether it rose or fell compared with this product's own earlier groups.