Cohort Retention for Startups: Measure and Improve It

Cohort retention tracks what share of the users who started in a given week or month are still doing your product's core action in each later period. In our view, the shape of the curve matters more than any single percentage: a curve that flattens above zero suggests a real, lasting group of users, while one that keeps sliding toward zero suggests the product is not yet something people want.

When each cohort holds steady, you grow on top of the people you keep. When it doesn't, you spend every month replacing the people you lost.

Definition: Cohort retention is the percentage of a group of users who started in the same period (the cohort) that performs a defined value action again in each subsequent period, measured separately for every cohort so that new sign-ups cannot hide old users leaving.

Sign-ups can be bought. Revenue can be pulled forward. A flat cohort curve is much harder to fake, which is why we think it is one of the most honest quantitative tests of product-market fit an early startup has. This guide covers how to define it, a worked cohort table computed in Python, how to read the curve, the traps that inflate it, how to improve it, and the B2B version (net revenue retention).

Write your retention spec before you run any numbers

A lot of bad retention analysis goes wrong before the math starts. The fix is boring and it works: write down a three-line retention spec, and don't change it without rebuilding history.

1. The cohort. Group users by the week or month of first use. Once that works, slice by acquisition channel, country, device, plan or company size.
2. The value action. Pick an action that shows the user got real value, not that they opened the app. For a ride-hailing app that is a completed ride; for a photo app it might be viewing a photo full screen; for a social product it might be reading several posts. A good test: picture yourself sitting beside a happy customer and write down what you would see them do.
3. The cadence. Match the period to how often a satisfied customer would naturally use the product: daily for social and entertainment, weekly for utilities and payments, quarterly or longer for something like travel. A consumer bank, for example, might count someone as active if they made at least one transaction a week (Monzo used that definition).

Retention is rarely your headline metric, and that's fine. We like one top-line number for the company (revenue, usage such as daily or weekly actives, signed enterprise contracts or pilots, or a technical milestone), backed by three to five driver metrics. For many software companies, cohort retention is among the most important of those drivers.

One way to picture it: growth is five levers multiplied together (visits, sign-ups, first valuable action, return use and sharing). Retention is the return-use lever. When it is weak, it tends to cap everything else. Our startup KPI framework covers where retention sits at each stage.

How to build a cohort retention table, step by step

Build your first cohort table yourself, from raw event logs in a script or spreadsheet, before you trust a dashboard. It takes an afternoon, and it shows you what the numbers actually mean.

1. Export one row per qualifying action, with a user ID and a timestamp. Use your value action, not logins.
2. Assign each user a cohort, usually the month (or week) of sign-up or first value action.
3. Compute each event's age: the number of periods since the user's cohort start.
4. Count unique users per cohort per age. A user who comes back ten times in a month counts once.
5. Divide each row by its starting size to turn counts into percentages. Month 0 is 100 percent by definition.
6. Leave unreached cells blank. A cohort that started last month has no month 3 yet; filling that cell with zero makes recent cohorts look terrible.
7. Plot each row as a line and look at the shape.

A minimal version in Python (pandas), assuming a file events.csv with one row per qualifying action and the columns user_id and ts, looks like this:

import pandas as pd

ev = pd.read_csv("events.csv", parse_dates=["ts"])
ev["month"] = ev["ts"].dt.to_period("M")
ev = ev.join(ev.groupby("user_id")["month"].min().rename("cohort"), on="user_id")
ev["age"] = (ev["month"] - ev["cohort"]).apply(lambda d: d.n)
counts = ev.groupby(["cohort", "age"])["user_id"].nunique().unstack()  # NaN = not reached yet
retention = counts.div(counts[0], axis=0).round(3)
print(retention)

Worked example: a six-month cohort table

In this illustrative example, a product acquired 200 users in January, 220 in February, 250 in March, 240 in April, 260 in May and 280 in June. You counted how many of each cohort did the value action in every later month. The counts were: January 200, 90, 70, 62, 58, 57; February 220, 99, 75, 66, 62; March 250, 120, 95, 85; April 240, 118, 94; May 260, 135; June 280. Dividing by each starting size gives this table (all figures recomputed in Python):

Three ways to read it:

• Read across a row for the shape. January loses more than half its users in month 1, but from month 3 on it loses only one to four users a month (62, 58, 57). The curve is flattening at roughly 28 percent.
• Read down a column for progress. Month 1 retention rose from 45 percent for January and February to 52 percent for May, and month 2 from 35 to 39 percent. Newer cohorts are doing better, which is what product improvements should ideally produce.
• Weight the average by cohort size. The bottom row sums the retained users and divides by the users in cohorts old enough to have reached that month (for month 3: 213 of 670, or 31.8 percent). A simple average of percentages over-weights small cohorts.

The table also explains growth. Add up the last cell of each row and you get the June active users: 57 + 62 + 85 + 94 + 135 + 280 = 713. Of those, 433, or 60.7 percent, came from cohorts acquired before June. Stack those cohorts on top of each other in a chart (often called a layer cake) and you can see it happen. When the old layers stay thick, total actives compound instead of resetting every month.

How to read a retention curve: flattening is the signal

Judge the shape first and the level second.

Picture two products. Product A keeps many more users in the first months but keeps sliding toward zero. Product B loses most of its users early, then holds steady around 20 percent. Our read: Product B is the better business, because each new cohort adds lastingly to the base. Product A is a treadmill. You buy users to replace the ones walking out the back door.

YC's David Lieb, who helped build Google Photos, has made this flattening test the center of how he evaluates retention. His own evidence is worth noting. Google Photos' weekly curves dropped fast and then went flat somewhere between 20 and 40 percent, depending on country and device. Within about six weeks of launch, that flatness gave the team confidence in a product that went on to pass a billion users.

The level still matters, just less. As a rule of thumb for consumer products, we'd want the plateau at 20 to 30 percent or better; a weekly cohort with only 6.25 percent still active in week four has not found its core yet. Category benchmarks help calibrate. Lenny Rachitsky and Casey Winters (Lenny's Newsletter, June 2020) surveyed 20 experienced growth practitioners and put good six-month user retention at about 25 percent for consumer social products, rising to about 70 percent for enterprise SaaS. Our product-market fit guide walks through the full benchmark table.

Two other shapes are worth knowing on sight:

• The shark fin. A product goes viral, acquires users fast, then collapses because the curve never flattened. Mixpanel co-founder Suhail Doshi has described watching about 15 companies grow virally and then die this way. Reactivation campaigns to win those users back often go badly. Fixing the curve usually matters more.
• The smile. The curve flattens and then rises, because surviving users use the product more over time. Many investors consider this the most encouraging shape you can see. Andreessen Horowitz partners Santiago Rodriguez and Alex Immerman (a16z, "Retention Is All You Need," September 10, 2025) report that some leading AI products show smiling curves as users who drifted away return when capabilities improve.

Six ways founders fool themselves with retention data

Most retention mistakes aren't math mistakes. They're definition mistakes. These are the ones we'd check first.

1. Too easy an action. Opens triggered by notifications inflate actives. Google+ is the classic cautionary case: clicking a notification bell inside other Google products counted as using Google+. The value action from your spec is the guard here.
2. Too long a period. Monthly or quarterly windows can make almost anything look retained. If weekly curves look bad for a product meant to be used weekly, switching to monthly before an investor meeting hides the problem. It doesn't solve it.
3. Paying as the action. People often stop using a product before they stop paying for it, so track paying and active together.
4. A single data point. A figure like 80 percent week-over-week retention says little without the week, the numerator and the denominator. One strong week-three number can sit on a curve that is still falling.
5. The tool's default definition. Analytics tools may compute rolling or returned-at-any-point-after retention rather than activity within each period. Amplitude's documentation, for instance, distinguishes N-day (return on) retention from unbounded (return on or after) retention, which produces smoother, flatter-looking curves. Check that your tool's chart matches the table you built by hand.
6. Blending. A single averaged curve can hide one segment that is flat and another heading to zero. Slice by channel, country and customer type before concluding anything.

How to improve cohort retention: five levers in the order we would pull them

When curves don't flatten, work through these in order. The first ones are usually the cheapest.

• Acquire better-fit users. Often the easiest fix, and the most overlooked. A campaign that brings in people the product was not built for produces cohorts that fall to zero, however good the product. Google Photos saw exactly that from a push toward younger users, who had fewer years of memories to revisit. Consider measuring acquisition cost per retained, active user, not per sign-up; Monzo shut down a cheap channel once its customers turned out to be unprofitable.
• Fix onboarding and activation. Find the magic moment that predicts long-term retention by comparing your best users with everyone else, then redesign sign-up so more people reach it sooner. Facebook found that users who added seven friends in their first ten days usually stayed, and Monzo users who added three friends at sign-up retained about 20 percentage points better. We wouldn't spend long debating the exact threshold. The direction tends to matter more.
• Remove friction early. Email and text confirmation steps can cause heavy drop-off, and many early products may not need fraud defenses yet.
• Improve the product itself. New use cases, speed and simpler flows show up as newer cohorts sitting higher and flattening sooner, exactly the column pattern in the worked example.
• Build network effects. Products that get better as more people join (messaging, sharing, marketplaces) tend to see cohorts improve as the network gets denser.

None of this replaces conversations. A cohort curve tells you whether you're on track. Talking to customers, especially the ones who left, is usually what tells you what to change. Our guide on how to talk to users covers that side, and we've written more on reducing customer churn on VC Unfiltered.

B2B cohort retention: logos, gross revenue retention and net revenue retention

For B2B companies, it usually makes sense to track retention in dollars as well as users, per annual (or monthly) customer cohort.

• Logo retention: the share of customers from a cohort still paying.
• Gross revenue retention (GRR): starting recurring revenue minus churn and downgrades, divided by starting revenue. It cannot exceed 100 percent.
• Net revenue retention (NRR, also called net dollar retention): the same, plus expansion from upsells, added seats, cross-sells and price increases. It can exceed 100 percent.

Illustrative worked example. A cohort of customers begins the year at $1,000,000 of annual recurring revenue. Over twelve months, customers worth $80,000 cancel, others downgrade by $30,000, and the remaining accounts expand by $220,000. GRR is ($1,000,000 - $80,000 - $30,000) / $1,000,000 = 89 percent. NRR is ($1,000,000 - $80,000 - $30,000 + $220,000) / $1,000,000 = 111 percent. That company grows 11 percent a year from existing customers before closing a single new deal.

Small monthly churn compounds. A 7 percent monthly revenue churn sounds like keeping 93 percent, but over twelve months it loses about 58 percent of the starting revenue (1 - 0.93^12 = 0.581). At 3 percent a month the annual loss is about 31 percent, and at 1 percent about 11 percent. Usage is the early warning: Mixpanel found a strong link between customers' daily usage and lower revenue churn, which is why usage cohorts tend to move before dollar cohorts do.

For benchmarks, SaaS Capital's 2025 retention survey of private B2B SaaS companies (published September 18, 2025) found a median NRR of 102 percent for companies with $25,000 to $50,000 average contract values, with the top quartile at 111 percent. Companies with NRR of 110 percent or more grew faster than the survey median. Rachitsky and Winters (June 2020) put good twelve-month NRR for enterprise SaaS at about 110 percent and great at about 130 percent, against about 100 and 120 percent for bottom-up SaaS. Our sibling guide on startup business models and unit economics shows how retention feeds lifetime value and payback.

Where views differ in cohort retention

Practitioners broadly agree on the core idea and split on three details. Here's our lean on each, though reasonable teams choose differently.

• Where to start the curve. One view is to measure every cohort from its first period, as the most complete picture. But AI products attract many tourists who try once, and the a16z team suggests judging commitment from month 3 (the ratio of month-12 to month-3 retention). We tend to look at both: the full curve for completeness, the post-tourist curve for the quality of your core.
• Which active-user metric. Some practitioners see monthly actives as close to a vanity metric for anything people should use often, and argue a strict daily definition is more revealing. For products with a naturally slower rhythm, such as travel, tie the period to natural usage instead. Our rule of thumb: be strict about the action and realistic about the period.
• Tools vs. hand-built tables. At an early stage, the mainstream analytics tools (Mixpanel, Amplitude) will give you the same answer, and Amplitude's retention report is a reasonable default. We'd still build the table by hand once and reconcile it with the tool. That is often how you catch a definition problem.

A weekly cohort review checklist you can copy

Many teams refresh their cohort charts every week or two, put the core metrics on a shared dashboard, and send a short metrics note to advisors and investors. This list can guide the review:

• [ ] Value action and period written down and unchanged since last review (if changed, note it and rebuild history).
• [ ] Newest cohort added; unreached cells blank, not zero.
• [ ] Shape check: is the oldest cohort's curve flat over its last three periods?
• [ ] Column check: is period-1 and period-3 retention higher for recent cohorts than for older ones?
• [ ] Segment check: curves split by channel, country or plan, and customer type or company size.
• [ ] Activation check: share of new users reaching the magic moment in their first week.
• [ ] B2B only: logo retention, GRR and NRR for each annual cohort.
• [ ] One hypothesis for the worst segment, and three customer conversations booked to test it.
• [ ] Tool chart reconciled with the hand-built table at least once a quarter.

Where 1752vc fits

If you're still building or iterating on an MVP and your first cohorts are only a few dozen users, Ignite, 1752vc's startup academy for first-time founders, may be a useful next step. It runs 12 weeks, live and remote at your own pace, and is built for the stage where choosing the right value action and fixing activation matter most; applications are reviewed on a rolling basis. If you haven't shipped anything yet, start with our guide on how to build an MVP, then come back once you have a few weeks of cohorts.

The bottom line

A retention curve won't tell you what to build. It will tell you, earlier and more honestly than most metrics, whether the thing you already built is working.

Growth fills the bucket.

Retention decides whether it holds water.

Key takeaways

• Cohort retention follows each group of new users separately, so growth in sign-ups cannot hide users leaving.
• It helps to define three things first: the cohort, a value action that proves real use, and a period that matches natural usage.
• In our view the shape matters more than the level: curves that flatten above zero suggest a product people want; curves sliding to zero suggest a treadmill.
• Many retention mistakes are definitional: periods too long, actions too easy, single data points, blended segments and tool defaults.
• In B2B, logo retention, GRR and NRR per cohort are worth tracking; small monthly churn compounds into large annual losses.
• Common ways to improve retention include acquiring better-fit users, getting more of them to the magic moment quickly, and talking to the ones who leave.

Frequently asked questions

What is cohort retention?

Cohort retention is the share of users who started in the same week or month and are still doing your product's core action in each later period. Each cohort is tracked separately, so a surge of new sign-ups cannot mask older users leaving. Founders plot each cohort as a curve and look for it to flatten above zero, one of the clearer quantitative signs that people want the product.

How do you calculate cohort retention?

Group users by the week or month they started, choose a value action and a period, then count how many from each cohort did that action in each later period, counting each person once. Divide each count by the cohort's starting size. It helps to leave periods a cohort has not reached yet blank, and weight any average by cohort size rather than averaging percentages.

What is the difference between cohort retention and churn rate?

Churn rate is usually a single figure, such as the share of customers or revenue lost in a month across the whole base. Cohort retention shows the full curve for each starting group over time, so you can see whether losses slow down and whether newer cohorts do better. A blended churn rate can look stable while every cohort is still declining toward zero.

What is a good net revenue retention rate?

It depends on customer size. SaaS Capital's 2025 survey found a median of 102 percent for private SaaS companies with $25,000 to $50,000 contract values, and faster growth among companies at 110 percent or more. Lenny Rachitsky and Casey Winters put good enterprise SaaS NRR at about 110 percent and great at about 130 percent. Anything below 100 percent means revenue from existing customers is shrinking year over year.

How often should a startup review cohort retention?

Every week or two is a reasonable rhythm for many teams: often enough to catch a problem quickly without reacting to noise. Early on, when cohorts are small, pair each review with customer conversations, because a few users can swing the percentages. It is worth rebuilding the whole history whenever you change the value action or the time period, so old and new cohorts stay comparable.

Sources

• Y Combinator: How To Improve Cohort Retention (David Lieb)
• Y Combinator: How to measure your product (Suhail Doshi)
• Y Combinator: Consumer Startup Metrics (Tom Blomfield)
• Y Combinator: How to choose a metric and set your KPI (Michael Seibel)
• Y Combinator: Analytics for startups (Ilya Volodarsky)
• Lenny's Newsletter: What Is Good Retention (Lenny Rachitsky and Casey Winters, June 2020)
• SaaS Capital: What Is a Good Retention Rate for a Private SaaS Company? (2025)
• Andreessen Horowitz: Retention Is All You Need (September 2025)
• Amplitude Docs: Interpret your retention analysis