How to read a retention curve: find the part that breaks, then fix it

A retention curve answers three questions: the early drop tests activation, the slope tests ongoing value, and the floor tests product-market fit. Picture Monday's metrics review: the founder opens the cohort chart and asks whether to fix the product or buy more users.

Most teams read the curve as one number, such as week-4 retention, and argue about whether it is good. Read it as three parts instead. Name the part that breaks first, fix it, and only then scale spend. The method below works in Mixpanel, Amplitude, PostHog or a spreadsheet.

How to build a retention curve you can trust

A retention curve shows the share of a starting group that is still active after each interval. Four choices decide whether the line means anything: the cohort, the definition of active, the interval and the counting rule.

Group users by the date they started, not by the date you ran the report. A weekly or monthly sign-up cohort is the usual unit. Each line then follows one group of people as they age. A change shipped in March shows up first in the cohorts that start after it. That separation is the core of cohort retention analysis.

Define active by the core action, not by a login. Amplitude calls this the critical event: "a specific action you most want users to perform." Sequoia makes the same point: "it's also important to define what constitutes meaningful activity." For a note app, the core action is writing a note. For an invoicing tool, it is sending an invoice.

Choose the interval from natural usage frequency. Sequoia notes that "a daily, weekly or monthly time frame may be most meaningful," depending on the product's vertical. Amplitude gives a rule of thumb: "Your product usage interval is when 80% of users have completed the action a second time." A payroll tool used once a month will look broken on a daily chart.

Check the counting rule last. Mixpanel's Retention report docs describe two. "On" counts users who come back in one specific day, week or month. "On or After" also counts users who return in "any time unit afterward," and Mixpanel calls it "the default calculation." An On or After curve never sits below an On curve for the same data.

Before Friday, write your core action, interval and counting rule at the top of the retention report. Then rebuild the chart with them.

The three parts of a retention curve

Every retention curve has three parts, and each answers a different question. Read them in order, because a break early in the curve distorts everything after it.

Early drop: did new users activate?

The early drop runs from day 0 to day 1 or week 1. Users who leave here never reached the value they signed up for. Treat a steep early drop as an activation problem, not a retention problem.

The early drop is usually the steepest part of the line. Andrew Chen and Ankit Jain's 2015 Android data found that "the average app loses 77% of its DAUs within the first 3 days after the install." Their conclusion: "the best way to bend the retention curve is to target the first few days of usage." The data is from 2015 and covers Android only, so use the principle, not the number.

The first week also says something about later months. Amplitude's 2025 analysis of more than 2,600 companies found that "69% of top day seven performers were also top three-month performers." The sample is Amplitude's own customers across all categories. Read the finding as a sign that early weeks carry signal, not as a law.

Masa, a consumer app for social event discovery, shows what an early-drop fix looks like. Its case page names the mandatory sign-up wall as the biggest drop-off point. Kreto deferred registration until after the core experience and added a Day 1 activation sequence. Day 1 retention stabilized at 20%.

Our guide to SaaS onboarding best practices covers fixes for this stage step by step. Build a funnel of your onboarding steps for the last four weekly cohorts and find the step where most new users stop. Remove or defer one step before it.

Slope: do users keep finding value?

The slope is the middle of the curve, after activation and before the line settles. Users here got some value and still left. A slope that keeps falling points to weak ongoing value: the product solved a first problem but gave no reason to return.

Slope fixes differ from onboarding fixes. Look at what retained users do in their second and third weeks that churned users do not. Common answers are a repeat use case, saved work worth coming back to, or an invited teammate. Build the next prompt, email or feature around that behavior.

Pick 20 users who activated and then left between week 2 and week 6. Read their event history and note the last action each one took before stopping.

Floor: is there product-market fit?

The floor is where the curve flattens, if it does. A flat floor means a group of users found lasting value. Lenny Rachitsky puts it plainly: "If this curve flattens above 0%, that means there is some part of the market that wants your product."

Flattening is the start, not the finish. Rachitsky adds that "you'll need this curve to flatten a lot higher." A low floor shows that some demand exists. A higher floor makes each acquired user worth more, and that value sets what you can pay to acquire the next one.

Extend your oldest cohorts as far as your data allows and mark where each line stops falling. If no line stops, work on the slope before you spend more on acquisition.

Three curve shapes and what each one means

Step back from the three parts and look at the whole line. Sequoia's retention guide describes three shapes: declining, flattening and smiling.

A declining curve never levels out. Sequoia is direct: "When a product has not achieved product-market fit, the retention curve will continuously decline." More acquisition spend on a declining curve buys users who will eventually leave.

A flattening curve levels out at some share of the cohort. In Sequoia's words, "The higher the level at which the curve flattens, the higher the long-term retention and the healthier the product."

A smiling curve dips and then rises. Sequoia describes a product whose "retention curve will actually rise as product development and network effects propel churned users to return." a16z sees the same shape in some AI products and writes that "ChatGPT's retention curve is the perfect example of this behavior."

The two sources measure different things. a16z tracks revenue retention by monthly cohort, while Sequoia's examples use active users. The shape carries over, but the units do not.

Label your curve with one of the three shapes before your next metrics review. If the line is declining, say so in the meeting.

How to compare cohorts over time

A single curve shows where you are. A stack of cohort curves shows whether the product is getting better. Plot each monthly cohort on the same axes, aligned by months since sign-up, not by calendar month.

Newer cohorts above older ones at the same age suggest recent changes are working. Lines that fall below older ones mean something got worse: the product, the onboarding or the mix of users coming in.

Keep one caveat in view. Sequoia warns: "For non-social products, the retention of newer cohorts may not be correlated with how well older cohorts are doing." An old cohort's floor is not a forecast for a new one. Compare cohorts at the same age, and let each new cohort show its own floor.

Plot your last six monthly cohorts on one chart. Mark the month in which you shipped each major change, then check whether the next cohort moved.

How to use retention benchmarks without fooling yourself

Benchmarks come second. Each one rests on its own definition. A benchmark only compares with your number when the unit, window and counting rule match.

One reference set comes from Lenny Rachitsky with Casey Winters. They define retention as "the percentage of users who signed up and are still active," measured at month 6. The figures come from a survey of about 20 growth practitioners plus public company data. Treat them as expert consensus, not a measured dataset.

Enterprise SaaS is left out of the chart. The post gives two different "good" values for it, about 70% in one place and 75% in another.

A second benchmark uses a different definition. The same Amplitude analysis says that if you "get just 7% of your original cohort of users to return on day seven," the product sits in the top 25%. That figure counts a return on day 7 itself, across Amplitude customers in all categories.

The two numbers look far apart because they measure different things. One is activity at month 6 among sign-ups, the other a return on one day in the first week. Chen's 2015 figure is a third definition again: daily actives on Android. A Mixpanel chart on the default On or After setting is a fourth, and never reads lower than an On chart of the same data.

Write the definition next to every benchmark in your deck: unit, window, counting rule and population. Delete any comparison where one of the four differs.

Five mistakes that distort a retention curve

Most bad retention decisions come from a chart that was built wrong, not from a misread shape. Check for these five before you present a curve.

1. Reading blended retention. One total number mixes old and new users. Sequoia notes that "total retention is the weighted average of your cohort retentions," so one large cohort can hide a change in the others.
2. Trusting small cohorts. In a cohort of 40 users, each person is 2.5 points, so four people returning or leaving move the line by 10 points. Merge weeks into months until each cohort holds steady.
3. Counting logins. A login is not value. Users who open the app and leave inflate the curve and hide the real slope. Count the core action from the first section.
4. Mixing paid and organic users. Paid users often retain differently from organic ones, and a budget change shifts the mix. Split cohorts by acquisition source before you judge a product change.
5. Mixing definitions. A day-7 chart, a month-6 benchmark and an On or After report are three different metrics. Compare like with like.

Rebuild your main retention chart split by acquisition source, counting only the core action. Put it next to the blended version you use today and note where the two disagree.

How retention decides whether to spend on acquisition

Retention sets the value of each acquired user, so it caps what you can pay to get one. a16z writes: "When projecting 5-year LTV / CAC returns, retention is the most important input." Sequoia calls retention "by far the best lever for product growth."

A stable curve also makes spend easier to plan. The same a16z piece says: "If your curve stabilizes, you can predict unit economics more reliably and invest more aggressively in GTM." The authors add that "you can afford longer payback periods when you're confident your customers will stick around."

The data behind those lines is revenue retention at AI companies. The reasoning applies to user curves as well: a predictable floor makes payback predictable.

Turn the three parts into a rule for the planning meeting:

• Steep early drop: fix activation before you add budget. New users will leave at the same rate.
• Falling slope with no floor: fix ongoing value first. Spend now buys users who will churn.
• Flat floor, stable across recent cohorts: run the LTV/CAC math on that floor. Then scale spend in steps and recheck the curve by channel.

Kreto orders its growth marketing work the same way: find the broken part of the funnel, fix it, then scale.

Before you approve the next budget increase, write down which part of your curve is broken. If the answer is the early drop or the slope, move that budget to the fix.

Frequently asked questions

What does a good retention curve look like?

A good curve flattens, and a higher floor is better. Compare it first with your own older cohorts at the same age. Use an outside benchmark only when its definition matches yours.

When is a cohort too small to read?

At 100 users, each person is one point of retention, so a few returns or losses move the line visibly. Merge weekly cohorts into monthly ones until the curve stops jumping between neighbors. Treat points from small cohorts as direction only.

Should I track day 1, week 1 or month 1 retention?

Track the interval that matches how often people naturally use the product. Use the first point for activation and the later points for the floor. A monthly tool on a daily chart will always look broken.

Read the curve, then decide

A retention curve is three diagnoses on one line. Build it by start-date cohort with the core action. Read the early drop, the slope and the floor in order, and fix the first part that breaks. Scale acquisition once the curve flattens at a level your unit economics can carry.

If you want a second pair of eyes on your cohort chart, talk to Kreto. On a discovery call we can read your curve with you and see whether working together fits your stage.