Retention Curve

While retention was reported as a single number, there was no way to judge whether that number was good. A report saying 24 percent were still there in week four doesn't reveal whether the line is still falling or has settled at that level. The same 24 percent might be 6 percent next month or still 23 percent five months from now, and those two situations imply completely different futures for the company.

Companies watching only their signup chart fell into the same trap over and over. The cumulative signup line keeps climbing and monthly new users keep rising, so everything looks fine. Then about six months in, the total user count stops at some level, and the moment ad spending drops the level sinks with it. That people were leaving as fast as they arrived only becomes visible after the budget is gone.

Put one group on a time axis and connect the points, and that situation shows up as a single picture. The line either bends and settles or it keeps going down, and that distinction is legible before anybody argues about a number. What the curve does is answer the question of whether anyone stays, in a form you can see.

Take everyone who signed up in the same period as one group, calculate what share of that group is still active at each week mark, plot the points, and connect them. Getting a single point requires settling four things first.

1. Define the group in the denominator. Everyone who arrived in the same period gets bundled together, and that bundle is called a cohort — the 500 people who signed up in the first week of September, for instance. Mix new and existing users in the count and nobody can explain what the number means.
2. Define what counts as still active. Whether a login is enough or the core action is required can swing the value by a factor of two in the same product. Pagedeck counts opening or editing a document.
3. Pick the interval. Match it to how often the product is meant to get used. A collaborative document tool people open on working days is read weekly; read a service used once or twice a year in daily buckets and nearly every cell comes out at zero.
4. Divide. If 225 of the 500 are still there in week one, that's 45 percent.

Repeat that calculation for the same cohort week after week and connect the points. Starting at 100 people and landing at 45 in week one, 30 in week two, 24 in week four, 23 in week eight, and 23 in week twelve produces a shape that drops steeply and then flattens.

Two drawing conventions help. Week zero is 100 percent by definition, so it adds nothing to the shape and the axis starts at week one instead. And a single cohort can carry whatever happened to that particular group, so cohorts are plotted as several lines side by side, one per signup period. Comparing cohorts that way has its own name, cohort retention.

Comparing the absolute values against someone else's is close to useless. A messaging app opened several times a day and a tax filing service used once a year cannot be held to one standard. A week-four number that is healthy in one category is alarming in the other, and neither fact travels between them.

What matters is the shape. A curve that stops descending and lies down at some level is flattening, and that is the real signal. Flattening at 24 percent means a quarter of the people who arrive find a place in the product, and it supports the arithmetic that the same share will apply to the people you bring in next. Bring in 1,000 a month and keep 240 of them and the base accumulates by that much every month.

A curve that falls to the floor instead — 45, 28, 14, 6, 2 — means nobody is being held. In that state the total user count stops at some level no matter how much traffic you add. Bring in 1,000 a month and lose 1,000 the next month and nothing accumulates. The thing to do here is not more advertising.

The comparison that works is against yourself. Put the week-four value for June signups next to the one for September signups and you can see whether the product got better in between. Watch out for the days right after a new feature opens, though. The illusion of something looking good briefly because it's new is the novelty effect, and it usually returns to baseline within two weeks. So improvement gets judged by comparing the same week across cohorts, not by one cohort's ups and downs.

"The curve hasn't flattened yet. Turning up advertising in this state just spends the budget."

Arrivals aren't sticking, so this isn't the moment to add traffic.

"June's cohort was 18 percent at week four. September's is 24."

Whatever got changed in between actually worked.

"You drew that chart from week zero. Can you pull it again starting at week one?"

A column that's always 100 percent is sitting in front and squashing the shape.

At Pagedeck, Mia pulls this chart again every week. She splits the rows by signup week and reads the week-four column vertically; when that column climbs through 18, 19, and 24 percent over three months, she takes it as evidence the onboarding changes landed. She keeps a year of rows on the chart so that one bad week never gets read as a trend. The picture Grace carries into investor materials is this curve rather than the cumulative signup line.

The most common mistake is reading a higher curve as a better product. A line sitting above another one naturally looks better. But height depends heavily on how often the product is meant to be used. Forty percent at week four is unremarkable for a tool people open daily, while flattening at 8 percent is normal for one used once a quarter. Rank by height and any low-frequency product is judged a failure, while nobody checks the thing that matters — whether it holds the level it flattened at.

Reading any flattening as product-market fit is the other one. "Flattening is the signal" is easy to take as an absolute rule. But a curve that flattens at 2 percent has still flattened. When one in fifty arrivals stays, bringing in 1,000 people a month accumulates 20, and no company runs at that rate. A curve gets read for where it flattens and for whether that height can support a business. Running something like the Sean Ellis test, which asks users directly, alongside it serves the same purpose.