Cohort Analysis for Retention and LTV

Most growth dashboards are built to make you feel good. Total users up and to the right, revenue climbing, a nice green arrow next to last month. I have sat in plenty of reviews where those numbers looked healthy while the product was quietly falling apart underneath. The reason nobody noticed is simple: aggregate metrics average away the thing that actually matters, which is whether the people who show up today stick around longer than the people who showed up last year.

Cohort analysis is the fix. A cohort is just a group of users who share a starting point, usually the month they signed up or made their first purchase. Instead of asking “how many active users do we have,” you ask “of the people who joined in March, how many are still here in month one, month three, month six.” Do that for every month and you can see, plainly, whether each new group of users is behaving better or worse than the ones before it. That single shift, from the whole to the cohort, is the most useful analytical habit I have picked up as a growth PM.

This post is about how to actually use cohorts to understand retention and lifetime value, not just admire a colourful heatmap. I will cover why aggregate numbers lie, what a healthy retention curve looks like, the difference between logo and dollar retention, how to read a cohort table, how cohorts feed directly into LTV, and the mistakes that make cohort analysis useless. The goal is that you leave able to tell whether your retention is genuinely improving, because most teams cannot answer that question honestly.

What a cohort actually is

A cohort is a set of users grouped by a shared event and a shared time window. The most common grouping is the acquisition month: everyone who signed up in January is the January cohort, everyone who signed up in February is the February cohort, and so on. You could also cohort by first purchase, by the week they activated, or by the campaign that brought them in. The grouping you choose depends on the question, but the principle never changes. You lock a group of people to a starting line and then watch what they do over the periods that follow.

Why does the starting line matter so much? Because it lets you compare like with like. A user who joined three years ago has had three years to either stick around or leave, and lumping them together with someone who joined last week tells you nothing about either. When you separate them into cohorts, you are effectively running the same experiment over and over: same starting condition, different calendar month, and you get to see whether the outcome is drifting in one direction. That is what makes cohort analysis the single best way to judge whether your product and your retention are truly getting better. It removes the confounding effect of growth and lets the underlying behaviour speak.

The cohort is also the unit that maps most cleanly onto money. Every user in a cohort has a start date, a stream of activity, and a stream of revenue. Sum those streams across the cohort over its life and you have the raw material for lifetime value. I will come back to that, because it is where cohort work stops being a reporting exercise and starts driving decisions about how much you can afford to spend to acquire a customer.

Why aggregate metrics lie

Here is the trap that catches almost every team at least once. Your total monthly active users are growing. Your total revenue is growing. Leadership is happy. And yet every single cohort you have acquired this year retains worse than the cohorts from last year. How is that possible? Because you are acquiring new users faster than the old ones churn. The flood of fresh signups masks the leak. The topline rises while the product rots, and the rot is invisible until growth slows down, at which point the churn you were papering over suddenly becomes the whole story.

I have watched this happen in real time. The aggregate retention number, computed across all active users, was flat and unremarkable. It looked fine. When I broke the same users into monthly acquisition cohorts, the picture inverted. Month-three retention had dropped several points with each successive cohort. The average was being propped up entirely by a handful of loyal early cohorts who joined before some product changes and had no intention of leaving. New users were bouncing at a rate that would eventually catch up with us, and the blended metric hid it completely.

This is the core argument for cohort analysis, and it is worth stating bluntly. Aggregate metrics answer “how big are we right now.” Cohort metrics answer “is what we are building getting better or worse.” Those are different questions, and only the second one tells you whether your growth is durable. If you are trying to reason about how much a customer is worth over their life, you have to start from cohorts, which is exactly why they sit underneath any honest LTV and CAC calculation. Blended lifetime value computed off aggregate revenue and a guessed churn rate is a number that will betray you.

The retention curve and what healthy looks like

Take a single cohort and plot the percentage still active over time. Month zero is 100 percent by definition, because everyone is active in the period they joined. Then it drops. Month one might be 60 percent, month two 48 percent, month three 42 percent. The shape of that decline is the retention curve, and its shape tells you almost everything about the health of your product.

The critical thing to look for is whether the curve flattens. An unhealthy curve keeps sliding toward zero, meaning that given enough time every user eventually leaves and you have no durable base. A healthy curve drops at first, because some people were always going to bounce, and then it levels off into a flat plateau. That plateau is the fraction of users who found real, lasting value. When people talk about the retention “smile,” they mean a curve that flattens or even ticks back up as dormant users return and expansion kicks in. A flattening curve is the most reliable signal of product-market fit I know of, far more trustworthy than any survey or NPS score, because it is revealed behaviour rather than stated intent.

Where the curve flattens matters as much as whether it flattens. A product that stabilises at 40 percent retention has a fundamentally different business than one that stabilises at 8 percent, even if both curves technically level off. The plateau height sets the ceiling on how large you can grow for a given acquisition rate, because in steady state your user base settles roughly where new arrivals balance the churn of the flat tail. When I am evaluating whether a retention initiative worked, the plateau is the number I care about most. Moving the early drop-off is nice, but lifting and flattening the tail is what compounds.

Logo retention versus dollar retention

Counting users is only half the picture, and often the less important half. There are two ways to measure whether a cohort is sticking, and confusing them leads to bad conclusions. Logo retention, sometimes called user or customer retention, counts how many accounts are still active regardless of how much they spend. Dollar retention, or revenue retention, measures how much of the cohort’s original revenue you are still collecting. These can point in opposite directions, and the gap between them is where a lot of the truth lives.

Imagine a cohort where half the accounts have churned, so logo retention is 50 percent, but the accounts that stayed upgraded to bigger plans and now spend more in total than the entire cohort did at the start. Your dollar retention is above 100 percent even though you lost half your logos. This is net revenue retention, and when it exceeds 100 percent it means your existing customers grow in value faster than churn erodes them. A cohort like that will generate revenue that expands over time without a single new signup, which is close to the best position a subscription business can be in. It is common in products with usage-based pricing or clear expansion paths, and rare enough elsewhere that hitting it is worth real effort.

The practical rule is to always look at both. Logo retention tells you whether people find enough value to keep showing up. Dollar retention tells you what that base is worth and whether expansion is offsetting churn. A team that only watches logo retention can miss that its remaining users are downgrading and quietly bleeding revenue; a team that only watches dollars can miss that it is propped up by a few whales while the broad base leaves. For lifetime value, dollar retention is the input that matters, but you need logo retention to understand why the dollars move.

Reading a cohort table

The standard way cohorts get presented is a triangular table, often coloured as a heatmap. Rows are acquisition cohorts, one per month. Columns are the periods since acquisition: month zero, month one, month two, and so on. Each cell shows the retention, in percent or in revenue, for that cohort at that age. The result is a staircase, because the January cohort has more months of history than the June cohort, so its row extends further to the right.

There are two directions to read it, and you need both. Reading across a row shows you the retention curve for one cohort as it ages, which is the shape I described earlier. Reading down a column is the more powerful move: it holds the age constant and compares cohort to cohort. Look down the month-three column and you are asking “at the same point in their life, is each new cohort retaining better or worse than the one before it.” That vertical comparison is how you detect whether a product change helped, whether a new acquisition channel brings worse users, or whether retention is slowly decaying. If the numbers in a column get darker as you move down, newer cohorts are worse and you have a problem the aggregate would never show you.

A colour heatmap makes patterns jump out, but do not let the colour do all the thinking. I always sanity-check the raw cohort sizes, because a cohort of forty users will produce noisy percentages that swing wildly and mean very little. A cell showing 25 percent retention off a base of eight people is not a signal, it is a rounding accident. Read the table for direction and magnitude, confirm the cohorts are large enough to trust, and be suspicious of any dramatic movement in a small or very recent cohort.

How cohorts feed lifetime value

This is where cohort analysis pays for itself. Lifetime value is not a single guessed multiple of first-month revenue. Done properly, it is the retention curve multiplied by the revenue per user in each period, summed over the life of the cohort. Take the fraction of the cohort still active in each month, multiply by what an active user pays that month, and add up the periods. The area under that curve is the value a cohort delivers, and because it is built from observed retention rather than an assumption, it is honest.

This construction makes the link between retention and value concrete. A flatter, higher retention curve means more area under the curve, which means higher LTV, full stop. It also means you can model the effect of a retention improvement directly: lift month-six retention by five points across the plateau and you can read off exactly how much LTV increases, and therefore how much more you can afford to spend acquiring a customer. That is the entire game, because LTV is only meaningful next to acquisition cost. I have laid out how the two sides fit together in the piece on LTV and CAC, and the short version is that a durable business needs LTV comfortably above the fully loaded cost of acquisition, with the payback happening fast enough that you are not betting the company on a distant plateau.

One caution on the tail. Newer cohorts have not lived long enough to reveal their full curve, so you are forced to project the plateau from limited data. Be conservative here. It is tempting to extrapolate an optimistic flat line and book a big LTV, but if the real curve is still sliding you will have overpaid for every customer. I would rather anchor the projection on the observed plateau of older, mature cohorts and only revise upward once newer cohorts prove they behave better.

Segmenting cohorts to find what drives retention

A single blended retention curve tells you the average, but averages hide the levers. The real value of cohort analysis shows up when you slice the same cohorts by some attribute and compare the curves. Segment by acquisition channel and you often find that paid social brings users who churn fast while organic or referral users retain for years. Segment by plan and you see whether your pricing tiers actually correspond to different levels of commitment. Segment by behaviour and things get genuinely useful.

The most valuable behavioural cut is around activation. Split a cohort into the users who reached the activation moment, the specific action that correlates with sticking, and those who did not. Almost always the activated group has a dramatically higher, flatter retention curve, and the unactivated group falls off a cliff. That comparison does two things. It confirms your activation metric is real, and it quantifies exactly how much retention you would gain by pushing more of each cohort past that moment. If activated users retain at 55 percent and unactivated at 10 percent, the entire retention problem might just be an activation problem in disguise. I have written about how to find and define that moment in the post on activation metrics, and cohort segmentation is how you prove the moment matters rather than just asserting it.

Segmentation is also how you diagnose bad growth. When a channel scales and blended retention dips, the cohort-by-channel view tells you immediately whether the new users are structurally worse or whether it is a temporary mix shift. Without that cut you are guessing, and guessing about channel quality is an expensive way to run acquisition. Cohort a cheap channel against an expensive one over six months and the true cost per retained user is often the opposite of what the cost per signup suggested.

Using cohorts to prove a change actually worked

Most retention initiatives are declared successful on vibes. Someone ships an onboarding revamp or a lifecycle email sequence, the next month’s numbers look okay, and everyone moves on. Cohort analysis is the discipline that turns that hand-waving into evidence. Because a cohort is anchored to a start date, you can cleanly separate the users who joined before a change from those who joined after, and compare their curves at the same age. If the post-change cohorts retain better at month one, month three, and into the plateau, the change worked. If they do not, it did not, regardless of how the blended number moved.

This is the honest way to evaluate any retention or lifecycle effort. When I test a change aimed at reducing early drop-off, I hold the pre-change cohorts as the baseline and watch whether successive post-change cohorts bend the curve upward at the same age. The tactics themselves, the specific interventions that stop people leaving, are their own subject, and I have collected the ones that reliably move the curve in the post on churn reduction tactics. On the revenue side, cohorts are how you tell whether a re-engagement or expansion programme is genuinely lifting dollar retention over time, which is the measurement backbone of any lifecycle CRM built for repeat revenue. The programme either shows up as a fatter tail in the post-launch cohorts or it does not exist.

One trap to avoid: do not confuse a cohort dip with the effect of your change when it might just be seasonality. A cohort acquired in December may retain differently from one acquired in September for reasons that have nothing to do with your product, especially if your business has a holiday pattern. Compare like seasons where you can, look at the same age across multiple cohorts rather than a single before-and-after pair, and give the post-change cohorts enough time to reveal their plateau before you claim victory.

Tooling, from spreadsheets to product analytics

You do not need fancy tooling to start. A spreadsheet with acquisition cohorts down the rows and periods across the columns will get you a working cohort table, and building one by hand at least once is the best way to understand what the numbers mean. It is tedious to maintain, but it demystifies the whole exercise, and for a small product it is entirely adequate.

Once you outgrow the spreadsheet, you have two serious options. SQL over your own event and revenue tables gives you total control: you define the cohort, the active event, and the revenue logic exactly as your business works, which matters because retention is never quite generic. The cost is that you own the queries and the maintenance. The other option is a product analytics tool that builds cohort and retention reports for you. I lean on PostHog for exactly this, because it turns cohort and retention analysis into a few clicks and lets me segment by any event or property without writing a query every time, and I have gone deeper on that setup in the post on PostHog for product analytics. In practice most teams end up using both: the product analytics tool for fast exploration and shared dashboards, and SQL for the definitive revenue cohorts that feed LTV, where the exact logic has to be beyond dispute.

Whatever the tool, the requirement is the same. You need a reliable event that marks a user as active, a clean acquisition date, and accurate per-period revenue. Get those three inputs right and any tool will give you honest cohorts. Get them wrong and the prettiest heatmap in the world is just confidently misleading colour.

Common mistakes that make cohort analysis useless

The failures I see are consistent, so it is worth naming them. The first and biggest is never leaving the aggregate. Teams look at blended active users and blended retention, feel informed, and never break the data into cohorts at all, so they miss decaying retention until it is a crisis. If you take one thing from this post, take the habit of asking “what does this look like by cohort.”

The second is using too short a window. If you only ever look at month-one retention, you cannot see the plateau, and the plateau is where the durable value lives. A product can have great month-one numbers and a curve that keeps sliding to nothing. Give cohorts enough time to reveal their tail before you judge them. The third is ignoring dollar retention entirely and reasoning only about logos, which hides both downgrade bleed and expansion strength. The fourth is failing to segment, so you never learn which channels, plans, or behaviours actually drive the retention you are seeing, and you cannot act on the average. And the fifth is the seasonality confusion I mentioned: reading a calendar effect as a product effect, or the reverse, because you compared two cohorts acquired in very different conditions.

None of these are exotic. They are the default state of a team that has a dashboard but not the cohort habit. The fix is not a tool, it is a discipline: group by start date, read down the columns, watch the plateau, look at both users and dollars, and always segment before you conclude.

The short version

• A cohort is a group of users sharing a start point, usually signup or first-purchase month, and cohort analysis is the clearest way to see whether your product and retention are truly improving.
• Aggregate metrics lie: total users and revenue can rise while every new cohort retains worse, because new signups hide the leak. Cohorts expose it.
• The retention curve should flatten into a plateau rather than slide to zero. A flattening curve, the “smile,” is the most reliable signal of product-market fit.
• Track both logo and dollar retention. Net revenue retention above 100 percent means expansion outpaces churn, and dollar retention is the input LTV cares about.
• Read a cohort table down the columns to compare cohorts at the same age, not just across the rows, and sanity-check cohort sizes before trusting a cell.
• LTV is the retention curve times revenue per period, summed over the cohort’s life. A flatter, higher curve means more LTV and more you can spend to acquire.
• Segment cohorts by channel, plan, and behaviour, especially activation, to find what actually drives retention.
• Use pre- and post-change cohorts to prove a retention or lifecycle change worked, and do not mistake seasonality for a product effect.

---

I am Deepanshu Grover, a Growth Product Manager in Paris. If your topline looks fine but you suspect retention is quietly rotting, connect on LinkedIn or get in touch.