पाठशाला Pathshala · उत्पाद Utpād, The product · Lesson 05 · Build

Retention curves and the flattening test

A retention curve is the most honest chart a company owns and the easiest to draw wrong. How to build the grid, the test that says whether it flattens, and what the floor says about your market.

Pathshala, The Founder Library · 11 October 2026 · 10 min read

Growth can be bought. Press can be arranged. Sign-ups can be driven by a festival discount and a budget. The one number that cannot be purchased is the share of people who came back a month later without being asked, and then the month after that. A retention curve is that number drawn over time, and a company that cannot draw its own is guessing about whether it has a product.

The [product-market fit lesson](/library/what-product-market-fit-looks-like-in-the-numbers) says why the curve matters. This one is about the drawing itself: how to build the grid from a transactions table, the three decisions that make or ruin it, the test that says whether the curve has flattened or is only falling more slowly, and how to read the height of the floor as a statement about the market rather than the product.

Build the grid from what you already have

You need one table: a row for each time a user did the thing your product is for, with the user’s identifier and a timestamp. Payments, orders, documents created, invoices sent, rides completed. From it, three steps. First, for each user find the month of their first action; that is their cohort. Second, for each user and each later month, mark whether they did the action at least once. Third, for each cohort and each month-since-start, divide the users marked active by the users in the cohort. The result is a triangle: the January cohort has nine months of history, the September cohort has one.

Andreessen Horowitz’s 16 Startup Metrics describes this as the way investors measure churn: month one is 100 per cent of the installed base, and the latest month is the share of that original base still transacting. It is also the only form in which an investor will believe your retention, because a blended monthly churn figure hides the thing a cohort grid shows, which is whether later cohorts behave better than earlier ones.

Three decisions that decide what the curve means

The grid is arithmetic. What goes into it is judgment, and Reforge’s essay on retention as the silent killer names the three choices that most often go wrong. The unit: measure usage, not revenue. Revenue retention is the output of engaged users, and a customer still paying for a tool nobody opens is a churn event that has not been reported yet. The frequency: match the measurement to the natural rhythm of the job. A product people need daily that is measured monthly looks healthier than it is, because a user who opened it once in thirty days counts as retained. A product used quarterly, like filing a return, measured monthly looks like a disaster. The core action: define active as the thing the product exists to do, not as a login. Sequoia’s data team gives the example of a ride-hailing app counting a completed ride rather than an app open, and a16z’s warning is that companies have almost unlimited definitions of active, some of which count accidental or one-time activity, so be clear which one you are using and keep it fixed.

Write the three decisions at the top of the sheet. Every reader of the grid, including you in six months, should be able to see that active means invoiced at least once in the calendar month, or completed a ride in the week, and that the grid has meant that since it was first drawn.

Then read the grid properly. The commonest misreading of a cohort grid is to average all the cohorts into one curve. The young cohorts have only early months, the old cohorts have the late ones, and the blended curve is a mixture of different companies at different times. Read the grid two ways instead. Along a row you see one cohort’s life. Down a column, month three say, you see every cohort at the same age, and that column tells you whether the product is getting better: if the cohorts that joined after a product change hold more at month three than the ones before, the change worked, and no other chart will show you that as plainly.

The average curve is useful for one thing, which is the flattening test, and it should be computed only across cohorts old enough to have reached the months you are testing. A month-six retention figure built from two cohorts is a rumour. Wait for five.

The flattening test

Sequoia’s description of the shapes is the standard one: a curve that flattens means a percentage of users who sampled the product found value and continued to return; a curve that declines continuously will eventually reach very few or zero users; and the rare curve that smiles rises again as product improvements and network effects bring churned users back. The question for your grid is the first. Flattening is not the same as slowing, and almost every declining curve slows, so the test has to be specific.

Use this rule. Take the cohorts at least six months old. For each, compute the point drop between consecutive months from month three onwards. The curve has flattened when the drop is under about two percentage points for three consecutive months and the level at which it sits is clearly above zero, meaning well above the noise of a small cohort. A curve going 60, 48, 40, 35, 32, 30, 29 has flattened at around 30. A curve going 60, 45, 34, 26, 20, 15, 11 is slowing in absolute terms, because 15 to 11 is a smaller drop than 60 to 45, but it is losing a quarter of what remains every month and will reach zero. The second shape fools founders because the chart looks like it is levelling. Check the ratio of each month to the previous one as well as the difference; a ratio stuck near 0.75 is a leak with a smaller bucket.

The grid above is a model to read, not your data, but the three sliders are the three things a real curve is made of: what month one holds, where it settles and how fast it gets there. Set the settling point to zero and watch the readout: the curve can still look gentle at month six while the twelve-month figure says what it is.

What the floor says about your market

The height at which a curve flattens is the floor, and Sequoia’s sentence about it is the one to remember: the higher the level at which the curve flattens, the higher the long-term retention and the healthier the product. But the floor is a statement about more than the product. It is the share of everyone who tried the thing for whom it became a habit, which means it is also a measurement of the market you have been acquiring from.

Three readings. A high floor with a steep early drop says the product is right for a well-defined group and wrong for the rest of who you are acquiring; the fix is in targeting and onboarding, not the product. A low floor with a gentle drop says the product is mildly useful to many and essential to few, which is the hardest position, because there is no segment to double down on. No floor says you are in a market of samplers: people who will try anything once, which in India describes much of consumer app acquisition during a festive campaign. Andrew Chen’s analysis of Quettra’s data from 125 million Android devices found that the average app lost 77 per cent of its daily users within three days of install, 90 per cent within thirty and over 95 per cent within ninety, while the top ten apps still held about 51 per cent at day ninety. The floor is where the categories separate.

The benchmarks for where a floor should sit come from Lenny Rachitsky’s 2020 survey of twenty growth practitioners, combined with public company data, and they are six-month user figures: consumer social around 25 per cent good and 45 great; consumer transactional 30 and 50; consumer subscription 40 and 70; SMB software 60 and 80; enterprise software 70 and 90. Note the denominators differ by category, registered users for social and paying accounts for software, which is one more reason to write your own definition at the top of the sheet. For Indian consumer apps the AppsFlyer festive report for the Diwali 2024 window puts even the best shopping apps at 6.5 per cent day-thirty retention on iOS and the best finance apps at 13.2 per cent; a floor above 10 per cent at month three in that setting is an achievement, and a founder comparing it to a Western SaaS benchmark is comparing two different markets.

The floor is not how good the product is. It is how many of the people you reached needed it.

Why two points of retention are worth more than double the acquisition

Reforge’s worked example makes the arithmetic plain. Company A adds a million users a month and keeps 85 per cent of them each month; Company B adds two million and keeps 65 per cent. At six months B is ahead, 5.3 million monthly actives to 4.2. At twelve months A passes it. At three years A has 6.6 million and B has 5.7, on half the acquisition. The floor compounds and the leak compounds, and the difference between them is invisible for the first two quarters, which is exactly when most companies decide to spend on growth instead.

The same arithmetic sets the ceiling on what a customer is worth. Lifetime, in months, is roughly one divided by the monthly loss once the curve has settled; a floor at which 3 per cent of the remaining users leave each month implies a lifetime around 33 months for those who reach it. Multiply by monthly margin and you have the LTV the [unit economics lesson](/library/cac-ltv-and-payback-the-three-numbers) needs, and you have it from the grid rather than from a hope.

A worked example

A Bengaluru company sells inventory software to pharmacies at ₹1,200 a month and has signed about 200 a month since January. Active is defined as at least one stock entry in the calendar month. In October the grid shows the January to April cohorts, all at least six months old, holding 71, 69, 73 and 70 per cent at month one, then 58, 55, 60, 57 at month three, then 50, 49, 52, 51 at month six, with the drops from month four onwards at two points or under. The curve has flattened at around 50. The floor is half of every pharmacy that tried the product. Monthly loss after month six is running at about 2 per cent, so a pharmacy that stays past the sixth month is worth roughly fifty months of ₹1,200, or about ₹60,000, before margin. The month-one drop, 30 per cent gone before they have really started, is the onboarding problem the [activation lesson](/library/activation-first-session-that-decides) addresses, and it is a separate problem from the floor, which is healthy for SMB software in a market where paying stops with a tap.

Now the same grid with the July cohort at month three holding 66 rather than 58. The product changed in June. The column says it worked. That is the sentence the board wants and the grid is the only place it can be read.

Seven mistakes make a curve lie. Counting logins as active, which measures curiosity. Measuring revenue when the question is usage, which reports the leak a quarter late. Blending cohorts into one average, which mixes a young company with an old one. Reading month six from two cohorts, which is a sample of two. Dropping a bad cohort because a campaign brought the wrong users, which is survivorship with a spreadsheet. Choosing a frequency that flatters, monthly for a daily product. Declaring a flattening when the ratio month to month is still well below 0.95, which is a leak that has learned to look calm.

A monthly ritual, and what to check

On the first working day of the month, add last month’s column to every cohort in the grid. Read down the column for the newest age that has five or more cohorts and write the range. Run the flattening test on the cohorts six months or older and write one of three words: flat, slowing, or leaking, with the height if flat. Compare the month-three column for cohorts before and after your last significant product change. Then write one sentence that a board member could read without the grid: the curve flattens at about 50 per cent by month five, month one is the problem, and the July cohort suggests the June change helped. If you cannot write the sentence, the grid is not finished.

The benchmarks here are ranges other people have published for other markets and are cited so you can read them in full; your definition of active, your frequency and your market will move every one of them.