How do you read a retention curve?

The x-axis shows time since a user's start event, and the y-axis shows the percentage of users who did a return event. Before reading any point, write down the start event, the return event and the time bucket, because changing any one produces a different curve. Each point divides returning users by the users who have been around long enough to reach that bucket, not by everyone who signed up.

What does it mean when a retention curve flattens?

A curve that stops falling means a group of users found enough value to keep coming back, which suggests product-market fit for some market or audience. Where it flattens shows how long it takes to reach that fit. A flat line proves fit only for the people who stayed, so the next step is to identify the characteristics of those who retained and those who did not.

What is the difference between N-day and unbounded retention?

N-day retention, called Return On in Amplitude, counts users who did the return event on exactly day X. Unbounded retention, called Return On or After, counts users who returned on day X or any later day, so it reads higher. In a worked example of 100 sign-ups, day 7 N-day retention was 10% while day 7 unbounded retention was 40%, so name the method next to every figure.

What is a good retention rate for a consumer app or SaaS?

The first goal is a curve that flattens at all, and raising the height of that flat line comes second. A flat retention curve that drives a scalable acquisition strategy is enough to keep a business alive. Timing varies by model: consumer transactional businesses can take longer than six months to flatten, and business-to-business software as a service (SaaS) companies take longer still, since they are usually measured on monthly churn.

How to read a retention curve: flattening, cohorts and the three mistakes

Q: What is a cohort retention table?

It splits the curve into its parts: each row is a group of users who started in the same period, and each column is the time since they started. Read down a column to compare cohorts at the same age, and along the diagonal to spot calendar events such as an outage or pricing change. In a worked example, an older cohort at 40% and a newer one at 20% blend to 30%, which hides that the newer cohort retains at half the rate.

Q: What is a smile curve in retention?

A smile curve dips first, then bends back up, as churned or lower-usage customers return or customers spend more. Andreessen Horowitz split revenue cohorts into acquisition (M0 to M3), retention (M3 to M6 or M9) and expansion (M9 and later). Rebasing to month 3 removes early tourist churn: in a worked example, revenue of 20 at M3 and 25 at M12 gives an M12 to M3 ratio of 125%.

Retention curve analysis means reading where the curve stops falling, not any single point. Plot the share of a cohort that returns over time; if it levels off, as in a worked example where 1,000 signups hold at 25% from month 6 to month 9, a stable group is getting value. Split by cohort, define a real return event, and ignore incomplete buckets before trusting the shape.

What retention curve analysis tells you

A retention curve answers one question: of the people who started using your product, how many keep coming back. In a standard retention chart, the x-axis represents the time since a user's initial start event and the y-axis represents the percentage of users who engage with one or more return events. You choose both events. The start event is usually signup or first purchase. The return event is the action that shows real use.

Each line on the chart is a cohort. One common definition is a group of people who did the starting event on the same date. Grouping by start date lets you compare January signups with March signups and see whether product changes moved the curve.

Shape matters more than any single point

Every curve drops early. What you are looking for is where it stops dropping. Greater retention signifies greater product market fit, and where your curve flattens out indicates how long it takes to reach your fit. A curve that keeps sliding toward zero means users try the product and leave. A curve that levels off means a stable group keeps getting value.

Here is a worked example, not a benchmark. A cohort of 1,000 signups:

Month after signup	Share still active (example)	Users active (example)
1	40%	400
3	30%	300
6	25%	250
9	25%	250

The loss between month 6 and month 9 is zero, so this curve has flattened at 25%. Those 250 users are your retained base. Be patient with the read: in consumer transactional businesses, these businesses can take longer than six months to have their graphs flatten.

One trap to avoid this week: do not include users who signed up too recently to reach the later months. They pull the late points down artificially. Some tools handle this for you, and one describes its approach as this analysis always gives all users the chance to be included in all duration buckets.

Why the curve drives lifetime value and word of mouth

Lifetime value can be framed as how long a user is active along with revenue per active user. Retention sets the first half of that formula. A higher flat line means more active users for longer.

It also feeds acquisition. For referral and viral growth, the more retained users, the more potential sharers. For paid channels, the more retained users, the higher lifetime value, the more you can spend on paid acquisition or sales and still have a comfortable payback period. See our note on payback period for that math.

You do not need a high plateau to survive. One benchmark study put it this way: a flat retention curve that drives a scalable acquisition strategy is enough to keep your business alive. In our reading, the first goal is a curve that flattens at all, and raising the height of that flat line comes second.

How do you read the axes and each point on the curve

The first section covered what the two axes show. Before you read any single point, write down three settings from your analytics tool: the start event that puts a user on the curve, the return event that counts them as retained, and the time bucket (day, week or month). Change any one of them and the same users produce a different curve.

What one point is made of

Amplitude's documentation defines each data point as "the number of unique users who triggered the return event during a specified timeframe (day, week, or month) divided by the total number of unique users who reached that point in time."

The denominator is not everyone who signed up. It is only the users who have been around long enough to reach that bucket.

Here is an example with round inputs, not a benchmark. Say 1000 users signed up this quarter, but only 100 of them signed up early enough to reach week 4. If 25 of those 100 returned in week 4, the week-4 point is 25 percent. Dividing the same 25 returners by all 1000 signups gives 2.5 percent, a number that describes nothing.

Two ways to count a return

Amplitude offers two definitions, and they divide by different groups. For Return On, the numerator counts users who returned in that exact bucket. For Return On or After, the numerator counts users who returned in that bucket or later.

Measure	Who is in the numerator	Who is in the denominator
Return On	Users who did the return event in that bucket	Users who did the start event in the first bucket
Return On or After	Users who did the return event in that bucket or later	Users who completed that bucket

Return On or After also counts a user in every earlier point before their return. In our reading, that makes it the gentler of the two, so name the measure whenever you share a curve.

Four checks before you trust a point

Look for incomplete buckets. In Amplitude, results for days with incomplete data carry an asterisk. Ignore them until the bucket closes.
Expect a lift at the right edge on live data. Amplitude notes that ongoing data can trend upwards because users only enter a point after enough time has passed. A rising tail on a young cohort is often a small denominator, not better retention.
Make the start event a first-time event. Mixpanel's First Time Filter narrows an event to a user's first instance, so repeat signups do not restart the clock.

N-day, unbounded and bracket retention compared

Amplitude offers three ways to count a return, and each answers a different question. Return On was formerly known as N-Day. Return On or After was formerly known as Unbounded. Return On (Custom) lets you create custom brackets for your Return On retention instead of using predefined units of time.

Mixpanel uses the same two-event setup. In its docs, a Retention Behavior is doing an event and then coming back to do another event. You pick a start event, such as sign-up, and a return event, such as any event.

Method	What it counts on day X	Denominator	Fits (our reading)
Return On (N-day)	Users who did the return event on exactly day X	The start cohort, as covered earlier	Daily products, where a skipped day is a signal
Return On or After (unbounded)	Users who returned on day X or any later day	Users who have reached day X, as covered earlier	Occasional products, such as a tax or travel tool
Return On (Custom) (bracket)	Return On, grouped into brackets you define, such as days 7 to 13	Same as Return On, in our reading	Weekly products, where any visit in the week counts

Why unbounded reads higher

Amplitude states the reason plainly: because "on or after" counts a user as retained if they return on that day or any later day. In its sample chart, the unbounded line falls to about 69% by Day 7, holds around 60% at Day 14, and settles near 43% by Day 30. That is a measure of who ever came back, not who comes back on a schedule.

Mixpanel says On or After retention is a better fit for most businesses, and it is the default calculation for the Retention report. It suggests On when your product relies on users returning on every time unit.

A worked example

These are example inputs, not a benchmark. Take a cohort of 100 sign-ups:

10 of the 100 return exactly on day 7, so day 7 Return On is 10%.
40 of the 100 return on day 7 or later, so day 7 Return On or After is 40%.
25 of the 100 return at least once in a days 7 to 13 bracket, so that bracket reads 25%.

Same users, three numbers. When you report retention, name the method next to the figure every time.

24-hour windows or calendar days

Amplitude lets you define how Amplitude measures days when it calculates retention (24-hour windows or strict calendar dates). In our reading, the choice moves late-night sign-ups. Someone who signs up at 11 p.m. and returns at 9 a.m. is on day 1 by calendar date, but still inside their first 24-hour window. Pick one setting this week, write it in the chart notes, and keep it fixed across every comparison.

What is a cohort retention table and how is it different from a blended curve

A cohort retention table splits the curve into its parts. Each row is one group of users who started in the same period. Each column is the time since they started. Adobe's retention tool places a table below the chart, with the option to show individual cohorts, and its interval can be set to Daily, Weekly, and Monthly. In our reading, for most startups, weekly rows hold enough users to read and still show change quickly.

Read it three ways

Across a row: one cohort aging. Amplitude's music app example showed that 11,528 users signed up for your music app on July 16, and Day 5 retention was 49.7%.
Down a column: cohorts compared at the same age. This is the test that matters. If week 4 falls row after row, recent sign-ups are retaining worse.
Along the diagonal: cells that share a calendar week. A dip running diagonally points to an outage, a pricing change or a holiday, not to the cohort itself.

The table forms a triangle because recent cohorts have not yet reached the later columns.

Why the blended curve hides a change

Here is an example, not a benchmark. One older cohort of 1000 users retains 40% at week 4, so 400 return. A newer cohort of 1000 users retains 20%, so 200 return. The blended curve shows 400 plus 200, or 600 of 2000 users, which is 30%. In our reading, a blended number like that looks like a mild decline. The column view shows the newer cohort retaining at half the older rate. As older cohorts pile up, they dilute each new row further.

Three kinds of cohort

Amplitude separates three types of cohort data for your retention analysis: acquisition, behavioral, and predictive:

Acquisition groups users by The date users were acquired or signed up. This is the triangle below.
Behavioral groups users by Past actions and user profile properties, for example users who finished setup in week 1.
Predictive uses Behavioral data used to forecast future actions.

Copyable template

Fill each cell with the returning share of the row's users. Leave a cell as "open" until that cohort has reached that week.

Sign-up week	Users	W0	W1	W2	W3	W4	W5	W6	W7	W8	W9	W10	W11	W12
Week A	__	100%	__	__	__	__	__	__	__	__	__	__	__	__
Week B	__	100%	__	__	__	__	__	__	__	__	__	__	__	open
Week C	__	100%	__	__	__	__	__	__	__	__	__	__	open	open
Latest week	__	100%	open	open	open	open	open	open	open	open	open	open	open	open

Add a behavioral split as a second copy of the table, one per segment. Then compare the W4 columns side by side.

What it means when a retention curve flattens

A curve that stops falling means a group of users found enough value to keep coming back. Brian Balfour states the test plainly: if the curve flattens off at some point, you have probably found product-market fit for some market or audience. Note the qualifier. A flat line proves fit for the people who stayed. It does not tell you how many more people like them exist.

Mixpanel gives a simple shape to look for. In its example, 100% of users come in on day one and the line drops to 50% active within five days. If it still shows 50% at 20 days, that is a flat curve.

When it flattens matters as much as where

Adobe's documentation for its retention analysis makes the timing point directly: where your curve flattens indicates how long it takes to reach your fit. Here is a worked example with round numbers, not a benchmark. Product A levels off at 25 percent by day 10. Product B levels off at 40 percent, but only by day 30. On a cohort of 1000 sign-ups, Product A keeps 25 percent of 1000, which is 250 users, and you know it twenty days sooner.

We think the early, lower flattening is often the more useful signal this quarter. It gives you a stable base to run experiments against while the slower curve is still moving. Read the time unit against your model, too: Balfour notes that B2B SaaS companies take the longest, since they are usually measured on monthly churn.

Survey first, behavior as the test

Check	What it measures	Our reading of its weight
Sean Ellis survey	What users say, 40% very disappointed bar	Fast, cheap, early signal
Flattened retention curve	What users do over time	Stronger test, slower to collect

Use the survey to decide where to look, then let the curve decide. Once it flattens, Balfour's next step is to identify the characteristics of those who retained and those who did not. That is how you find the audience the flat line belongs to.

What is a smile curve in retention

A smile curve dips first, then bends back up. The early drop is normal churn. The later rise comes from users who return or from customers who spend more. A flat curve means you kept a core group. A smile means that group grew.

Andreessen Horowitz described a smiling curve as one that indicates that churned or lower-usage customers come back or use the product more often as the product’s capabilities improve. In our reading, it was writing about revenue cohorts at AI companies, but the shape reads the same way for any product.

The three phases of a revenue cohort

The same analysis split cohorted revenue retention into acquisition (M0, M3), retention (M3, 6-9), and expansion (M9+). Here, M0 means month 0, the month of sign-up.

Phase	Months	What to check (our reading)
Acquisition	M0 to M3	How steep the drop is and where it stops
Retention	M3 to M6 or M9	Whether the line holds flat
Expansion	M9 and later	Whether revenue per cohort climbs back up

In that analysis, most often, the curve begins to flatten around M3, and after ~6, 12 months, retained customers may start expanding.

Rebase to month 3 when tourists inflate month 0

Some sign-ups are tourists. The same piece pointed to AI tourists who sign up and churn after a couple months. It proposed rebasing retention and customer acquisition calculations from Month 0 (M0) to Month 3 (M3), and a ratio where M12 divided by M3 measures how well customers who survive the tourist churn (M3) perform over their first full year.

A worked example, not a benchmark: a cohort brings in 100 in revenue at M0, 20 at M3 and 25 at M12. Measured from M0, M12 retention is 25 / 100 = 25%, which looks weak. Rebased, M12 / M3 is 25 / 20 = 125%, a smile. This week, rebuild your cohort table with M3 as the denominator and see which cohorts curve up. The same piece also tracked cost per retained customer at M3, which you can calculate from the same table.

The usage metric changes the shape

For a usage curve, the first choice is the core metric. For most consumer social apps, this is daily active users (DAUs), and for apps with a less frequent use case, weekly active users (WAUs) is also acceptable as a starting metric.

In our reading, a daily metric produces a noisier curve with a sharper early dip. A weekly metric smooths both. So pick the unit that matches how often people should use the product, then judge the smile in that unit.

Do not confuse this with the L-ness curve. That chart looks at the distribution of users by number of days active over a certain time period. A smile there describes how often users show up. It does not describe whether they come back over months.

In our reading, on annual contracts, billing hides churn, so watch usage.

The three mistakes that make a curve mislead, with a worked example and checklist

Most bad retention calls come from three habits. Each one is easy to spot once you know where to look.

Mistake 1: blending cohorts of different ages or sources

A blended curve averages every cohort in the window, so its shape moves whenever the mix of cohorts moves. In our reading, andreessen Horowitz describes the same effect on revenue: as a business "mixes into its older cohorts", the reported figure changes even if no single cohort does.

Here is a worked example with invented numbers, not a benchmark. In month one, a referral channel brings 1000 users who keep 40 percent at week 8, so 400 are still active. A paid channel brings 1000 users who keep 20 percent, so 200 are still active. The blend is 600 active out of 2000 users, which is 30 percent.

In month two, referral brings only 100 users, so 40 percent of them is 40 active. Paid still brings 1000 users at 20 percent, which is 200 active. The blend is 240 active out of 1100 users, about 21.8 percent. Neither channel got worse. The mix did.

Mistake 2: a return event that is not value

Tools make "came back" easy to define loosely. Mixpanel's quick-start example chooses "Any Event" in the second event slot. A login, a settings change or an email click all count, which flatters the curve. Pick the action that delivers the product's value, the same one you use for activation. Mixpanel frames the question as what percent of users are "still sending messages after seven days", which is the right shape.

Mistake 3: reading the curve before the cohort is old enough

On ongoing data, Amplitude notes, you can see your retention curve "trend upwards", because it only includes users "after they've had enough time to convert." Late points then reflect only the early users who stayed. In Mixpanel, boxes with an asterisk are "still in flux". Treat any starred cell as unfinished.

Checklist to run this week

Split the curve by sign-up month and by acquisition channel before quoting one number.
Write down your return event. If it is any event or a login, swap it for the value action.
Drop every starred or incomplete cell from the read.
Note how many users sit behind each late point. A thin denominator is a guess.
Recompute week 8 for each cohort separately and compare with the blend.

The decision you have to make

Once the curve is split and cleaned, decide which cohort is your real signal: the channel and month you plan to spend on next quarter. Report that cohort's curve, not the blend. If it still falls without flattening, your next move is a product test, not more acquisition. Plan it with how to run a growth experiment and track time to value alongside it.

What this page does not cover

This page does not give retention benchmarks by industry or step-by-step setup instructions for any analytics tool. It also does not cover how to model revenue or set acquisition budgets from your curve, and it is not financial advice.