Your dashboard says retention is 80%. It sounds fine. But 80% of whom? The customers who joined two years ago and never leave are propping up the customers who joined last month and are already drifting away, and a single blended figure quietly hides the difference. Cut the same base into the groups that joined at the same time, and the truth comes out: some cohorts are healthy and some are quietly dying, and you finally know which lever to pull.

The quick version

  • A cohort is a group of customers grouped by when they started. Tracking each group over time turns one misleading average into a clear picture of who actually sticks.
  • The retention curve, the share of a cohort still active each month, is the single most diagnostic chart you have. If it flattens, you have a durable base; if it keeps falling to zero, you have a leaky bucket no amount of marketing fixes.
  • Lifetime value (LTV) is the net profit a customer brings over the whole relationship. It rises directly with retention, which is why curve shape and LTV are the same story told two ways.
  • Optimise the shape of the curve, not the headline number. Lifting where the curve settles compounds; a one-off spike in month-one signups does not.

The idea in depth

Start with the definition, because it carries the whole insight. A cohort is simply a set of customers sharing a start point, everyone who signed up in January, say. Instead of asking "what is retention this month" across a churning, growing, blended base, cohort analysis follows each month's group and asks the cleaner question: of the people who joined together, what fraction are still here a month later, three months later, a year later? Plot that, and you get a retention curve.

Andreessen Horowitz, in its widely-cited 16 Startup Metrics, defines this precisely: in retention-by-cohort, "Month 1 = 100% of the installed base" and the latest month is the percentage of that original group still transacting. The discipline is in the denominator, you always divide by the people who started, never by who happens to be active now, so a flattering month of new signups can't disguise an old cohort quietly leaking away.

flowchart TD
    A(["All customers
(one blended number)"]) --> B(["Split by join month"])
    B --> C(["Jan cohort
track month 1, 2, 3 …"])
    B --> D(["Feb cohort
track month 1, 2, 3 …"])
    B --> E(["Mar cohort
track month 1, 2, 3 …"])
    C --> F(["Retention curve
= % of each cohort still active over time"])
    D --> F
    E --> F
Cohort analysis turns one average into a set of curves you can compare and act on. Leaders Loop

Why the shape of the curve is the whole game

Every retention curve falls at first, some customers always leave early. The question that matters is whether it keeps falling to zero or flattens onto a plateau. A flat tail means a stable group has found durable value and will keep paying; a curve that slides to nothing means you are filling a leaky bucket, and growth depends entirely on out-spending the leak.

That shape is, for many investors, the clearest signal of product–market fit. Andrew Chen of Andreessen Horowitz lists a flattening cohort retention curve first among his "magic metrics" for it: if a cohort's curve levels onto a plateau instead of bleeding to zero, you have found a group of users the product genuinely sticks for. How high is good enough depends entirely on the model. Lenny Rachitsky's benchmark survey, "What is good retention?" (2020), put rough six-month retention at, for example, ~40% (good) to ~70% (great) for consumer SaaS and ~70% to ~90% for enterprise SaaS, proof enough that a single universal target is meaningless.

So the move is to stop optimising the headline and start optimising the plateau. Pull up a cohort chart, find where each curve settles, and treat that height as the number to move. A spike of signups who all churn by month three does nothing to the plateau; an onboarding change that lifts the tail by five points compounds across every future cohort. Stack successive cohorts together: if newer curves sit above older ones, your product is improving; if they sit below, something you shipped, or some channel you scaled, is bringing in worse-fitting customers.

Lifetime value is the retention curve, priced

Cohort retention and lifetime value are the same fact wearing different clothes. LTV is the present value of the net profit a customer delivers across the whole relationship. The a16z framework spells out the build: revenue per customer, minus the variable cost to serve, gives a contribution margin; average lifespan is roughly one divided by the churn rate; and LTV ≈ contribution margin × average lifespan. Crucially, the article insists you base it on net profit, not revenue or even gross margin, an LTV built on revenue flatters you into over-spending on acquisition.

The lever inside that formula is retention. Halve your churn and you roughly double the lifespan term, which roughly doubles LTV, without selling anyone a single extra thing. The classic evidence is older than SaaS: Reichheld and Sasser's "Zero Defections: Quality Comes to Services" (Harvard Business Review, 1990) found that cutting the customer defection rate by five percentage points raised profits by 25% to 85% depending on the industry, developed further in Reichheld's The Loyalty Effect (Bain & Company, 1996). The mechanism is exactly the lifespan term: a retained customer keeps paying, costs less to serve over time, and often spends more.

Acquisition fills the top of the bucket. Retention decides how big the bucket is. Cohort curves show you the holes.

LTV only means something next to what a customer cost to win. The widely-used rule of thumb, popularised by David Skok of Matrix Partners on forEntrepreneurs, is an LTV-to-CAC ratio of at least 3:1, recover the cost of acquisition several times over the lifetime, or the unit economics don't work. So the move is to compute LTV and CAC per cohort and per channel, not as a single company-wide blur. A channel can show a healthy blended ratio while one of its segments quietly loses money on every customer, visible only when you split the cohorts apart. (We unpack the cost side in customer-acquisition metrics and the ratio's place in the wider picture in recurring-revenue metrics.)

The honest limitation: LTV is a forecast dressed as a fact

Here is where the tool breaks down if you trust it too literally. LTV is a prediction, and it inherits every weakness of the assumptions behind it. The "1 ÷ churn rate" lifespan assumes churn stays constant forever, but retention curves bend, prices change, and a young company simply hasn't lived long enough to know its real long-run churn. Quote a confident "£4,200 LTV" for a product 14 months old and you are extrapolating a decade of behaviour from barely a year of data. The a16z authors make the same point in their own way: measure 12- and 24-month LTV from actual historical cohorts, conservatively, rather than projecting a lifetime you cannot yet observe.

There is a subtler trap too. LTV is an average, and averages hide concentration: a blended LTV can look strong while a long tail of customers is unprofitable and a few whales carry the number, the same masking problem that net and gross revenue retention exists to expose. So treat LTV as a decision aid with a confidence interval, not a precise asset on the balance sheet. Use early-cohort LTV to compare channels and segments against each other, a relative judgement it is good at, and resist banking a far-future absolute number you haven't earned the data to claim.

A worked example

The figures below are illustrative, chosen to show the mechanics rather than to report a real company.

Two acquisition channels look identical on the surface. Both bring in customers at a CAC of £300, and both sell a subscription with a contribution margin of £40 per customer per month. The marketing dashboard, blending everything together, reports a single happy retention number and a single LTV. But split the cohorts by channel and the curves diverge sharply.

Channel A (a content-led, intent-driven channel) loses customers at about 2.5% per month. Average lifespan ≈ 1 ÷ 0.025 = 40 months, so LTV ≈ £40 × 40 = £1,600. Against a £300 CAC that's an LTV:CAC of about 5.3:1, comfortably above the 3:1 bar.

Channel B (a discount-led paid channel) brings in worse-fitting customers who churn at 8% per month. Lifespan ≈ 1 ÷ 0.08 = 12.5 months, so LTV ≈ £40 × 12.5 = £500. Same £300 CAC, but now the ratio is only about 1.7:1, below 3:1, and barely covering the cost to serve once overheads are counted.

flowchart LR
    A(["Channel A
churn 2.5%/mo"]) --> A2(["Lifespan ≈ 40 mo
LTV ≈ £1,600"])
    A2 --> A3(["LTV:CAC ≈ 5.3:1
healthy"])
    B(["Channel B
churn 8%/mo"]) --> B2(["Lifespan ≈ 12.5 mo
LTV ≈ £500"])
    B2 --> B3(["LTV:CAC ≈ 1.7:1
below the 3:1 bar"])
Same CAC, same margin, same blended dashboard, opposite economics once you split the cohorts. Illustrative figures. Leaders Loop

Read on the blended average, the two channels were indistinguishable, and you might happily pour more budget into the cheaper-looking one. Read by cohort, the move is obvious and would otherwise have been invisible: shift spend toward Channel A, and either fix Channel B's onboarding to flatten its curve or stop feeding it. Notice where the real lever sits, you don't need to touch CAC or price at all. Drag Channel B's monthly churn from 8% down to 4% and its lifespan doubles to 25 months, LTV rises to £1,000, and the ratio crosses the 3:1 line. Fixing the curve, not the spend, is frequently the cheapest growth in the building.

Frequently asked questions

What is the difference between cohort retention and plain churn rate?

A single churn rate is a blended snapshot across everyone active right now, so it mixes loyal old customers with shaky new ones and hides the trend. Cohort retention follows one group from its start date forward, so you see when people leave and whether the curve flattens. Churn tells you the rate today; cohorts tell you the story over time, and the story is what you can actually act on.

How long should I wait before trusting an LTV number?

Long enough to see the curve bend. Because LTV leans on the lifespan term (≈ 1 ÷ churn), an LTV computed in a product's first few months is extrapolating heavily from thin data. A practical discipline, echoing the a16z guidance, is to report 12- and 24-month LTV from real historical cohorts and treat any "full lifetime" figure as a directional estimate, not a banked asset, especially for comparing channels rather than declaring an absolute value.

Is a higher LTV:CAC ratio always better?

No, and this surprises people. The 3:1 rule of thumb is a floor, not a target to maximise. A very high ratio (say 8:1) often means you are under-investing in acquisition and leaving growth on the table; you could afford to spend more to win more customers and still clear the bar. The ratio is a balance test between retention quality and acquisition spend, not a score to push as high as possible.

What actually flattens a retention curve?

Usually the early experience, not a loyalty programme bolted on later. Most of the drop happens in the first weeks, so the work that pays back most is onboarding and getting a new customer to the moment the product becomes useful, the "aha" point. Fixing why month-one and month-two customers leave lifts the entire plateau for every future cohort, which is why it compounds far better than chasing back lapsed long-timers one at a time.

Does this only apply to subscription or SaaS businesses?

No. Any business with repeat custom, e-commerce, marketplaces, apps, services, has cohorts and a retention curve, even if "active" means "purchased again" rather than "still subscribed." The a16z definition explicitly frames retention by cohort around customers still transacting. Subscription businesses just make the maths cleaner because the recurring charge is an obvious heartbeat to measure.

Related in the Toolkit

Where to go next