
What is Cohort Analysis and How Do Founders Use It
You pull up your retention number before a first meeting, it looks healthy and then an investor asks to see it broken out by signup month. Split that way, one group of customers has been leaving steadily since the week they arrived, and the blended figure was averaging them against a loyal older base.
Cohort analysis is the practice of building that breakout on purpose. This guide covers what a cohort is, how to build and read one and what investors look for in your cohort data.
What is Cohort Analysis
Cohort analysis groups customers who share a starting point, then tracks one metric for each group across equal time intervals. Everyone who signed up in March forms one cohort and everyone who signed up in April forms another. You follow each group forward at one month, two months and three months, then compare them at the same age.
How Cohort Analysis Differs From Segmentation
Segmentation is a snapshot and a cohort is a time series. A segment is everyone on the enterprise plan right now, a group whose membership changes each time someone upgrades or cancels. Everyone who started in March, followed forward for as long as you keep measuring, is a cohort, and nobody joins or leaves after the fact. Customers can move between segments for years and still belong to the same cohort.
"Enterprise customers who signed up in March, over the next 90 days" is a segmented cohort question, and combining the two lenses produces sharper answers than either alone. The segment narrows who you are looking at, the cohort fixes when they arrived and the pair turns a vague worry into something testable.
The Types of Cohorts You Can Build
The starting point you choose decides which question the cohort can answer. Signup date is the obvious choice and where most companies stop, but an action, an attribute or a revenue figure can anchor a cohort too.
Acquisition Cohorts
Acquisition cohorts group customers by the week or month they signed up, though for some products the entry event is a first purchase or a first install. They answer whether newer arrivals do better than older ones, the first thing worth knowing after an onboarding change or a pricing change. Diagnosis is a different job: a March cohort that retains worse than February points at March without naming a cause.
Behavioral Cohorts
Behavioral cohorts drop signup date entirely and group customers by an action they took or skipped. One cohort reaches an activation milestone in its first week, another gets there later and a third never arrives at all. Grouping customers this way gives you a leading indicator early enough to act on, and a change you can test by guiding more of them to the milestone sooner.
Segment-Based Cohorts
Here an attribute does the dividing rather than a date or an action, usually acquisition channel, plan tier or company size. These cohorts answer which source produces customers who last, a spending question as much as a product question. A company whose organic search cohorts retain better than its paid search cohorts on the same plan has a reason to shift budget, then measure whether blended retention improves.
Revenue Cohorts
Revenue cohorts use the same grouping rules, then track recurring revenue in place of customer count, which separates keeping customers from keeping dollars. Cohorts can lose customers and still grow revenue if the ones who stay expand, or keep most of their customers and shrink because nobody upgrades. Net revenue retention (NRR) counts expansion, so it can stay above 100 percent in a period when a real share of the cohort's original revenue disappeared.
How to Run a Cohort Analysis
The hard part of a cohort analysis happens before you open a spreadsheet. Vague questions and loose definitions produce tables that look rigorous and mean nothing.
1. Start With a Specific Question
"Why are customers leaving" cannot produce a cohort table, because it names neither a group nor a period to measure it over. A workable version ties one change to one observation window: compare customers who hit May's new setup flow against earlier signups, and measure first-month cancellation. That framing tells you which cohorts to build and how long to wait before the answer means anything.
2. Group Customers by a Shared Starting Point
The entry event is the one decision you can't revise midway without throwing away comparability. Signup date in one analysis and first payment date in the next produce two tables that look alike, but don't line up. The second group already cleared a hurdle the first never faced. Most teams start from signup date and move to a behavioral entry point such as "invited a teammate" once they suspect which action predicts retention.
3. Track One Metric Across Equal Periods
One metric per analysis keeps the finding legible, and the interval has to match how often someone should use the product. Daily-use products need daily intervals, and something tied to a weekly workflow or a monthly reporting cycle needs a longer one. Measuring a weekly-use product on a daily interval produces a curve far worse than the product deserves. Periods also have to be equal, so that month three covers the same third 30-day stretch for every cohort.
4. Compare Cohorts and Test a Change
Cohort tables can show that customers who did something stayed longer, but never that the doing caused the staying. Customers who create a shared workspace may retain better because it helps, or because the kind of customer who sets one up was always going to stay. Separating those takes an experiment: offer one randomly chosen group extra guidance on creating a workspace, give another the standard onboarding, then compare the cohorts.
How to Read a Cohort Table
A retention cohort table puts cohorts in rows and elapsed time in columns. Each cell shows the share of the original group still active, and the table comes out triangular because newer cohorts haven't filled the later columns. Working through it in a fixed order keeps one finding from being mistaken for another:
- One row across time: Following one cohort from left to right shows how it decays, from 100 percent at period zero down to whatever floor it reaches. Steep falls in the first two or three periods usually trace back to the first-run experience or to a gap between marketing's promise and the product.
- One column across cohorts: Comparing every cohort at the same age answers whether the product is improving, so rising values in the month-two cell mean the onboarding or product changes are working. Falling values tend to implicate traffic quality before the product itself.
- Cliffs between adjacent periods: A sharp drop marks a specific moment in the lifecycle, such as the end of a free trial or a confusing step on day three. Payment failure at first renewal produces the same shape.
- Stripes that land on one date: When unrelated cohorts weaken during the same calendar window, the cause is business-wide and not anything in a cohort's lifecycle. Mapping each cell back to its real date separates an outage or a broken release from a retention problem.
- The point where a row levels off: That plateau is the durable base, the group that found something worth coming back for. A curve that never flattens hasn't produced one yet.
Going through the whole sequence before drawing a conclusion keeps a one-day outage from getting filed as a retention problem.
What Investors Look for in Your Cohort Data
Investors spend more time on the shape of the curve than on the headline number. A curve that drops early and then levels off above zero says a group of customers found something durable in the product. The question at seed and Series A is whether each new cohort retains at least as well as the one before. That puts cohort retention next to headline NRR in most startup metrics reviews.
CRV led DoorDash's first financing round and backed the company again during its Series A and B. Its 2016 cohort spent 57 percent more in its fourth year than its first, and existing customers drove 85 percent of gross order value by the third quarter of 2020. Few companies have numbers like that, and the failure that shows up in diligence is more ordinary. NRR clears 100 percent in the deck, then a cohort waterfall built in the data room reveals weak gross retention underneath. Expansion from a handful of enterprise accounts had been covering for smaller customers churning fast.
The Cohort Analysis Mistakes That Skew Your Numbers
Bad cohort analyses usually fail on construction and not on arithmetic, and the same failures repeat from company to company. These are the ones worth checking for before you act on anything a table tells you:
- Cohorts too small to trust: With a handful of customers, every movement looks like a trend, and a single departure can move the rate far enough to reverse what it appears to show.
- Time windows too wide: Monthly buckets that span a product change put the old flow and the new one in the same row. The contrast you built the table to find disappears into the average.
- Mixed populations in one bucket: Self-serve signups and enterprise contracts grouped together produce a curve that describes neither group.
- Periods that haven't closed yet: The newest column covers a stretch of time that is still running, so the most recent figure understates where the cohort will land.
Any table built on these is wrong at the source, so the check belongs at construction time rather than after the numbers reach a board deck.
What Cohort Analysis Tells You That Growth Numbers Hide
The blended number reports the past and the cohort table reports the direction. Total customers and cumulative signups can rise for a year even as every cohort underneath them erodes, and trailing churn can stay flat through all of it. A founder who can show a flattening curve under one consistent definition has answered the question a growth chart cannot: whether the business gets stronger with each customer it adds.
Founders who get the most out of cohort work treat the table as a question about which customers to go find next. That habit shows up long before a fundraise, in the metrics they track week to week. If you're an early stage founder looking for a partner who will work through your cohort table with you before the round instead of auditing it during diligence, reach out to us to see if we'd be a good fit.
Frequently Asked Questions About Cohort Analysis
How often should you run a cohort analysis?
The cadence should match the speed at which your cohorts accumulate enough activity to support a decision, and the tables belong in whatever management meeting already reviews acquisition numbers. After a product change or a new campaign, more frequent reviews make sense until the affected cohorts have aged past the point where the change should appear. The measurement interval is a separate choice from the review cadence, and cohorts measured in weekly periods can still go to a monthly meeting.
What tools do you need for cohort analysis?
A spreadsheet and clean billing data will get a seed stage company through its first several analyses. Once event volume grows, a product analytics or subscription analytics tool can automate retention tables and make behavioral comparisons easier to run. The right tool is whichever one lets a non-engineer get an answer without creating a reporting bottleneck.
What counts as a good cohort retention rate?
Business software benchmarks vary by company size, annual recurring revenue and contract value, and crossing 100 percent NRR has grown harder across every segment. Consumer apps work on a different scale entirely and usually measure retention in days instead of annual revenue. Whatever the level, a curve that flattens at a modest number beats one that starts higher and keeps sliding, especially for a company still measuring product-market fit.
Does cohort analysis work outside software as a service?
The mechanics don't change outside software, only the metric inside the cell. Marketplaces track gross merchandise value retention by cohort and usually run buyers and sellers as separate cohorts, the pattern the DoorDash figures above show on the buyer side. Ecommerce brands track repeat purchase rate by cohort, and usage-based or artificial intelligence products track whether a cohort keeps consuming at or above its starting level.