Here is a situation that plays out quietly in a lot of businesses. The dashboard says repeat-purchase rate is holding steady at around 27%, roughly where it has been all year. Nobody is alarmed. Meanwhile, every group of customers acquired in the last six months is performing noticeably worse than the groups before them, and the only reason the headline number looks calm is that a loyal base acquired two years ago keeps buying and props the average up. By the time the blended figure finally moves, the damage is several months old and far more expensive to fix.
That is the problem cohort analysis exists to solve. It is the one view that shows direction rather than a snapshot, because it compares each new group of customers against the previous one at the same age. The reframe worth holding onto: your retention isn't a number, it's a shape — and the most important part of that shape is not how steeply it falls at the start, but the level where it stops falling.
What a cohort actually is
A cohort is just a group of customers who share a starting point, almost always the month they first bought or signed up. Instead of one pooled retention figure, you track each group separately as it ages. Arrange those groups as rows, with periods since first purchase as columns, and you get a grid that turns raw orders into a readable picture of whether people stay.
The barrier to entry is lower than most people assume. You need three things: a customer identifier, a transaction or activity date, and ideally the amount. That is enough to build a first version in a spreadsheet, and it is worth doing manually once before buying a tool, because building it teaches you how to read it. The one prerequisite is trustworthy data, and it has to be data you own rather than borrowed from a platform, which is where a solid first-party data strategy earns its keep. Duplicate customer records will also split one person across two cohorts and distort everything, which is why CRM data hygiene is the unglamorous prerequisite for any of this.
Building the table in four steps
The mechanics are simple arithmetic, and the whole thing works as percentages so that cohorts of different sizes can be compared fairly.
1. Assign each customer to a cohort — the month of their first purchase. A customer who first bought on 12 January belongs to the January cohort forever, no matter when they buy again.
2. Calculate elapsed periods — for every later order, how many months after that customer's first purchase did it happen? Month 0 is the first purchase itself.
3. Count unique active customers per cohort, per period. Unique matters: three orders from one person is one active customer, not three.
4. Normalise to percentages — divide by the original cohort size. If 1,000 people first bought in January and 90 bought again in month 1, that cell reads 9%.
Worked example: January cohort = 1,000 customers → month 1: 90 active = 9% → month 2: 60 = 6% → month 3: 45 = 4.5% → month 4: 42 = 4.2% → month 5: 41 = 4.1%. The fall is slowing, and the curve is flattening at roughly 4%.
That last detail is the whole point of the exercise. The early drop looks brutal, and it always does. What matters is that it stops.
Goflow
Multichannel operating system unifying orders, inventory, shipping and listings across marketplaces and storefronts.
Best for: High-volume multichannel e-commerce sellers managing orders & inventory
Reading the table in three directions
A cohort grid answers three different questions depending on which way you read it, and most people only ever use the first.
| Direction | What it shows | The question it answers |
|---|---|---|
| Across a row | One cohort ageing over time | What is our retention shape? |
| Down a column | Different cohorts at the same age | Are we getting better or worse? |
| Diagonally | All cohorts in one calendar month | What happened that month? |
Reading across gives you the retention signature: a steep initial drop, then ideally a plateau. Reading down is the view marketers under-use and should live in, because it is your early-warning system: newer cohorts sitting above older ones at the same age means your changes are working, and newer cohorts sitting below means something is degrading months before any average will admit it. Reading diagonally catches events that hit everyone at once, like a stockout, a delivery failure, a price rise or a checkout bug, and that is the diagnostic habit behind any honest end-to-end funnel audit.
The reframe Retention isn't a number, it's a shape. Everyone fixates on how steeply the curve falls at first. What matters is the height at which it stops falling, because that plateau is your actual business.
The plateau is the business
A healthy curve does not decline toward zero; it flattens. The flattening point is where casual buyers have finished leaving and the people who genuinely found value remain. That plateau, not the first-month drop, is what compounds: every new cohort adds another layer of loyal customers on top of the ones already there, which is what makes growth feel easier over time rather than harder.
Two practical implications follow. First, judge yourself against your own baseline rather than a published benchmark, because retention depends enormously on how often people naturally need your product. Take cohorts from six to twelve months ago, see where their curves settled, and treat that as your reference line. Second, watch the plateau's level over successive cohorts. A sinking plateau is the clearest sign that something upstream has changed, often in the mix of customers you are acquiring rather than in the product itself.
Cohorts by channel, not just by month
Time-based cohorts tell you whether retention is improving. Cohorts built on other shared traits tell you why, and this is where cohort analysis stops being a report and starts changing decisions.
Group customers by acquisition channel or campaign and the differences are often dramatic: one channel delivers cheap first orders that never repeat, while another costs more upfront and produces customers who stay for years. That single comparison can justify reallocating budget in a way that a cost-per-acquisition column never will, and it is a far more reliable guide than last-click reporting in an era where attribution keeps getting harder. Group by first product purchased and you often find that certain entry products predict loyalty, which should shape what you promote and how you build product pages for those items. Group by behaviour, such as people who opened the onboarding sequence or joined the newsletter, and you learn which early actions correlate with staying, which is exactly the signal that good behaviour-based scoring is built on.
Acting on the drop
The curve tells you not just whether you have a problem but precisely when it happens, and the steepest drop is where a small improvement pays back across every future cohort. For most e-commerce businesses that drop sits between the first and second purchase, which makes post-purchase education, usage guidance, replenishment timing and a well-judged incentive the highest-leverage work available. Those nudges are exactly what lifecycle email segmentation and well-timed SMS flows are for.
Two disciplines make the difference between an interesting chart and a compounding improvement. Change one thing per cohort, so the next curve actually attributes the result to something. And re-measure rather than declaring victory, since the only proof an intervention worked is that the following cohort's curve sits higher at the same age. Customers who have already gone quiet are a separate job, handled by re-engagement automation; cohort analysis is about stopping the leak before it happens.
Start this week
Export twelve months of orders with three columns: customer ID, order date, order value. Build the grid in a spreadsheet, monthly cohorts as rows, months-since-first-purchase as columns. Look at three things in order: where does the curve flatten, what level does it flatten at, and are newer cohorts above or below older ones at the same age?
That single afternoon usually produces a clearer picture of the business than a quarter of dashboard watching. Once the manual version has taught you what to look for, most analytics platforms and many CRM platforms will generate cohort tables automatically, so the habit is cheap to maintain once the thinking is in place.
The short version
Blended retention metrics hide direction, because loyal older customers mask the performance of recent ones, and by the time the average moves the problem is months old. Cohort analysis fixes that by grouping customers by when they first bought and tracking each group separately, which lets you compare every new cohort against the last at the same age. You need only customer ID, date and amount to build the first version in a spreadsheet: assign cohorts, calculate elapsed periods, count unique active customers, normalise to percentages. Read the grid across a row for the retention shape, down a column to see whether you are improving, and diagonally to catch month-specific events. The key number is not the first drop but the height at which the curve flattens, because that plateau is the loyal core that compounds. Build cohorts by channel, first product and early behaviour to learn why some customers stick. Then act at the steepest drop, usually the gap between the first and second purchase, change one thing per cohort, and confirm the improvement in the next group's curve.
Want to know if your customers are actually sticking?
We help brands read retention properly and fix the drop-offs that cost the most.
Explore E-commerce Marketing →Frequently asked questions
What is cohort analysis in marketing?
Cohort analysis groups customers by when they first bought or signed up, then tracks each group separately as it ages. Instead of one blended retention number that mixes brand-new customers with loyal ones acquired years ago, you get a grid: each row is a group, such as everyone who first purchased in January, and each column is a period since that first purchase. The cells show what share of the group is still active or still buying. The reason it matters is that blended averages hide direction. A business can report a stable repeat-purchase rate while every recently acquired group performs worse than the last, because the older, loyal customers prop the average up. Cohort analysis makes that visible immediately, because you can compare each new group against the previous one at the same age. You need only three data points to start: a customer identifier, a transaction or activity date, and ideally the amount. That is enough to build a first version in a spreadsheet before buying any tool.
How do you read a cohort retention table?
Read it in three directions, and each one answers a different question. Reading across a row follows a single group as it ages, showing the shape of its retention: a steep early drop followed by a flattening plateau is normal and healthy, because the plateau is the loyal core who found lasting value. Reading down a column compares different groups at the same age, which is the most valuable view for a marketer, because it tells you whether your changes are working. If newer groups sit above older ones at the same age, retention is improving. If they sit below, you have a problem that blended metrics would not reveal for months. Reading diagonally shows what happened to every group in the same calendar month, which surfaces seasonality or the impact of a specific event, such as a stockout, price change or site issue. The single most important number is not the size of the first drop, which is always large, but the height at which the curve flattens.
What is a good cohort retention rate?
There is no universal figure, because retention depends heavily on your industry and how often people naturally need to buy. A coffee subscription and an annual insurance product cannot be compared on the same scale, and a business selling mattresses should not expect monthly repeat purchases. This is why the useful benchmark is the shape of your own curve and its direction over time rather than someone else's percentage. Establish your baseline from older groups: look at cohorts acquired at least six to twelve months ago, see the level at which their curve stabilised, and treat that plateau as your reference point. Then judge every new cohort against that same-age benchmark. Healthy signs are a curve that flattens rather than falling toward zero, a plateau that holds, and newer cohorts tracking at or above older ones. Warning signs are a plateau that keeps sinking with each new group, or a fresh drop appearing at a period where older cohorts held steady, which usually points to a specific change you can trace.
What should you do after running a cohort analysis?
Act at the point where the curve drops most steeply, because that is where a small improvement compounds across every future cohort. For most e-commerce businesses that is the gap between the first and second purchase, which is why post-purchase education, usage guidance, replenishment reminders and a well-timed incentive are the highest-leverage interventions. For subscription and software businesses it is usually early activation, meaning the first days when someone either experiences the product's value or quietly stops using it. Change one thing at a time so the next cohort's curve tells you whether it worked, then re-measure rather than assuming. Cohorts built on acquisition channel are just as actionable: if customers from one channel or campaign consistently retain better, that is a strong argument for shifting budget toward it, even when its upfront cost per acquisition looks worse. The goal is a loop, where you read the curve, intervene at the drop, and confirm the improvement in the next group.
