THE MACHINERY OF COHORTS
A Complete Guide to How Groups Born Together Reveal What Aggregates Conceal
Why the When of Acquisition Determines Everything That Follows
What follows is not advice.
It is not a retention playbook. Not a dashboard template. Not a guide to reducing churn with better onboarding sequences. Not a framework for segmenting customers into buckets and sending them targeted emails.
It is mechanism.
The actual machinery that determines why a business looks healthy on the aggregate dashboard while bleeding out underneath. Why two customers acquired six months apart exhibit fundamentally different economics. Why the number the operator looks at every morning is the one most likely to be lying.
Most operators live inside averages their entire careers. Average revenue per customer. Average retention rate. Average lifetime value. These numbers feel solid. They come from real data. They update in real time. They point up or down and the operator reacts.
But the averages are concealing the substrate.
The substrate is the cohort. The group born together under the same conditions. The unit that reveals what the aggregate hides.
This document is a description of that unit.
What the operator reading it does next is their business.
PART ONE: THE UNIT OF TRUTH
What a Cohort Actually Is
A cohort is a group of entities that entered the system at the same time.
Customers who signed up in January. Employees hired in Q3. Loans originated in 2024. Restaurants onboarded in week 12.
The defining feature is not similarity of behavior. It is simultaneity of entry. Everything that happened in the world at the moment they arrived. The marketing channel that was running. The pricing that was live. The onboarding flow that existed. The competitive landscape. The season. The economy.
These conditions stamped into the cohort at birth. They do not wash out. They persist for the entire life of the group.
A cohort is not a segment. Segments slice by attribute. Cohorts slice by time. The difference matters because time carries context that attributes cannot capture.
SEGMENT vs. COHORT
┌──────────────────────────────┐ ┌──────────────────────────────┐
│ SEGMENT │ │ COHORT │
│ │ │ │
│ Grouped by attribute │ │ Grouped by entry time │
│ │ │ │
│ "Enterprise customers" │ │ "January 2026 signups" │
│ "West Coast users" │ │ "Q3 2025 hires" │
│ "High-spend accounts" │ │ "Week 12 onboards" │
│ │ │ │
│ Static slice │ │ Temporal slice │
│ Context-free │ │ Context-embedded │
│ Can be reclassified │ │ Birth date is permanent │
│ │ │ │
└──────────────────────────────┘ └──────────────────────────────┘
The enterprise customer can be reclassified as SMB if they downgrade. The January cohort member is always a January cohort member. The conditions of their arrival are baked in.
This permanence is the source of the cohort’s diagnostic power.
The Timestamp as Diagnostic
Every cohort carries a timestamp. That timestamp encodes everything the operator was doing at that moment.
The January cohort arrived during the winter discount campaign. The February cohort arrived after the price increase. The March cohort arrived through the new referral channel.
When retention diverges between these groups, the timestamp tells the operator where to look. Not “retention is declining.” Rather: “retention for customers acquired through the February pricing is declining.” The timestamp localizes the problem.
Without cohort separation, the operator sees a number moving. With cohort separation, the operator sees which decision caused the movement.
THE DIAGNOSTIC POWER OF TIMESTAMPS
┌────────────────────────────────────────────────────────┐
│ │
│ AGGREGATE VIEW │
│ │
│ "Retention fell from 82% to 74%" │
│ │
│ Response: Something is wrong. What? │
│ │
└────────────────────────────────────────────────────────┘
│
▼
┌────────────────────────────────────────────────────────┐
│ │
│ COHORT VIEW │
│ │
│ Jan cohort: 84% retained (referral channel) │
│ Feb cohort: 71% retained (paid ads, new pricing) │
│ Mar cohort: 85% retained (referral channel) │
│ │
│ Response: February's pricing change broke fit. │
│ │
└────────────────────────────────────────────────────────┘
The aggregate said “something is wrong.” The cohort said “this specific decision was wrong, and the rest of the business is fine.”
One produces anxiety. The other produces action.
PART TWO: THE AGGREGATION TRAP
Simpson’s Paradox in Business
In 1951, Edward Simpson formalized a statistical phenomenon that had been observed for decades. A trend that appears in every subgroup of a dataset can reverse or disappear when the subgroups are combined.
This is not a curiosity. It is the central failure mode of aggregate metrics in business.
The classic demonstration: a hospital treats two types of patients, mild cases and severe cases. Hospital A has a lower success rate than Hospital B for mild cases. Hospital A has a lower success rate than Hospital B for severe cases. But Hospital A has a higher overall success rate. Because Hospital A treats proportionally more mild cases, which have higher baseline success rates regardless of treatment quality.
The aggregate lied. The subgroup data told the truth.
Every business dashboard that reports a blended metric across customer cohorts of different sizes, different ages, and different acquisition conditions is vulnerable to exactly this reversal.
SIMPSON'S PARADOX IN RETENTION
COHORT-LEVEL TRUTH:
┌──────────────────────────────────────────────────┐
│ │
│ Q1 cohort (100 customers): Month-6 = 78% ↓ │
│ Q2 cohort (400 customers): Month-6 = 65% ↓ │
│ Q3 cohort (200 customers): Month-6 = 72% ↓ │
│ │
│ Every cohort declining individually. │
│ │
└──────────────────────────────────────────────────┘
AGGREGATE REPORT:
┌──────────────────────────────────────────────────┐
│ │
│ Overall 6-month retention: 68% → 71% ↑ │
│ │
│ "Retention is improving!" │
│ │
│ (Because the mix shifted toward the │
│ higher-retaining Q1 and Q3 cohorts │
│ as the large Q2 cohort churned out.) │
│ │
└──────────────────────────────────────────────────┘
The aggregate showed improvement. Every cohort was deteriorating.
This is not a rare edge case. It is the default condition whenever cohort sizes vary and acquisition quality shifts over time. Which is to say, always.
The Blending Problem
The aggregate metric blends all active customers into a single pool. A customer acquired three years ago who has survived intense selection pressure sits in the same average as a customer acquired last week who has survived nothing.
The three-year survivor has a retention probability near 1.0. Not because the business is excellent at retaining that individual, but because that individual has already demonstrated they are the type who stays. The new customer has unknown retention probability.
Blending them produces a number that describes neither.
The three-year survivor does not have a 74% chance of staying. They have a near-certain chance. The new customer does not have a 74% chance of staying. They might have a 40% chance. The 74% is the average of certainty and uncertainty. It is not the truth about either.
THE BLENDING DISTORTION
┌───────────────┐ ┌───────────────┐ ┌───────────────┐
│ 3-YEAR │ │ BLENDED │ │ NEW │
│ SURVIVOR │ │ "AVERAGE" │ │ CUSTOMER │
│ │ │ │ │ │
│ True P(stay) │ │ Reported │ │ True P(stay) │
│ = ~0.95 │ │ = 0.74 │ │ = ~0.40 │
│ │ │ │ │ │
│ Described │ │ Describes │ │ Described │
│ by: nothing │ │ neither │ │ by: nothing │
│ in this avg │ │ customer │ │ in this avg │
│ │ │ │ │ │
└───────────────┘ └───────────────┘ └───────────────┘
Peter Fader at Wharton has spent two decades demonstrating this single point. The aggregate customer metric is a fiction. The cohort is the minimum unit of truth.
PART THREE: THE SHAPE OF THE CURVE
Retention Curves Reveal Substrate
Plot a single cohort’s retention over time. The x-axis is time since acquisition. The y-axis is the percentage of the original cohort still active.
The shape of this curve contains more diagnostic information than any other metric in the business.
There are exactly three shapes that matter.
THE THREE RETENTION CURVE SHAPES
Retention
%
│
100% │█
│
80% │ █ SHAPE 1: CLIFF
│ █ Steep drop, no recovery.
60% │ █ The product has no core.
│ █ No one finds durable value.
40% │ █
│ █
20% │ █
│ ██████
0% │
└──────────────────────────────────────────► Time
Retention
%
│
100% │█
│
80% │ █ SHAPE 2: FLATTENING
│ █ Steep early drop, then plateau.
60% │ █ Product has a core.
│ █ Those who survive early churn
40% │ ██ found durable value.
│ ████████████████████████████
20% │
│
0% │
└──────────────────────────────────────────► Time
Retention
%
│
100% │█
│
80% │ █ SHAPE 3: SMILE
│ █ Flattens then rises.
60% │ █ Retained users increase
│ █ usage or spend over time.
40% │ ██ Expansion revenue.
│ ████████
20% │ ████████████████
│ ████████
0% │
└──────────────────────────────────────────► Time
Shape 1 is a business with no product-market fit. Shape 2 is a business with product-market fit for a subset of its acquired customers. Shape 3 is a business where the retained cohort compounds in value.
Andrew Chen at Andreessen Horowitz identifies the flattening curve as the single most important signal of product-market fit. Not the height of the plateau. The existence of the plateau. A curve that flattens has found its core. A curve that never flattens has found nobody.
The height of the plateau tells the operator what percentage of acquired customers belong to the core. If the curve flattens at 35%, then 35% of the people the business acquires are good fits. The other 65% are acquired at full cost and contribute nothing beyond the first few periods.
What the Curve Is Actually Showing
The retention curve looks like it is showing a single process. Customers arriving, some leaving, the remainder persisting.
It is not showing a single process.
It is showing selection.
A freshly acquired cohort is a mixture of heterogeneous individuals. Some have high intrinsic retention probability. Some have low. The early periods of the curve are dominated by the rapid departure of the low-probability individuals. The later periods are dominated by the persistence of the high-probability individuals.
The curve flattens not because something changed in the business. It flattens because the people most likely to leave have already left.
WHAT THE CURVE IS ACTUALLY SHOWING
┌──────────────────────────────────────────────────┐
│ │
│ MONTH 1: MIXED POPULATION │
│ │
│ ████████████████████████████████████████ │
│ High-fit Medium-fit Low-fit │
│ (30%) (30%) (40%) │
│ │
└──────────────────────────────────────────────────┘
│
▼ Low-fit depart rapidly
┌──────────────────────────────────────────────────┐
│ │
│ MONTH 3: SELECTION IN PROGRESS │
│ │
│ ████████████████████████████ │
│ High-fit Medium-fit (Low-fit gone) │
│ (45%) (40%) │
│ │
└──────────────────────────────────────────────────┘
│
▼ Medium-fit departing
┌──────────────────────────────────────────────────┐
│ │
│ MONTH 12: SELECTION COMPLETE │
│ │
│ ████████████████ │
│ High-fit only │
│ (near-100% individual retention) │
│ │
└──────────────────────────────────────────────────┘
This is Fader and Hardie’s central insight from the beta-geometric model, published in 2007 and extended in 2018. The phenomenon of increasing cohort-level retention rates is purely due to cross-sectional heterogeneity. An individual customer’s propensity to churn does not change over time. The cohort’s apparent improvement is an artifact of composition change.
The business did not get better at retaining people. The people who were going to leave, left.
This distinction matters enormously for forecasting. If the operator believes retention improved because of something the business did, the operator will project that improvement forward. If the operator understands it is selection, the operator knows the improvement has a ceiling and knows what drives that ceiling: the proportion of high-fit customers in the initial acquisition mix.
PART FOUR: THE VINTAGE IMPRINT
Conditions at Birth Persist
In credit risk, the concept is called vintage analysis. Loans originated in the same period are grouped into a vintage. The performance of that vintage is tracked over its entire life.
Decades of lending data show the same pattern. The conditions at origination predict the lifetime performance of the vintage more reliably than any factor that occurs afterward.
Loans originated in easy-credit environments have higher lifetime default rates than loans originated in tight-credit environments. Not because the later performance differs. Because the initial pool composition differs. Easy money attracts marginal borrowers. Tight money selects for strong borrowers. The selection happened at entry. Everything after is the echo.
THE VINTAGE IMPRINT
Origination Lifetime Performance
Conditions (Default Rate)
┌──────────────────┐
│ LOOSE CREDIT │
│ Easy approval │────────► Default: 12-18%
│ High volume │ Marginal borrowers entered
│ Low standards │
└──────────────────┘
┌──────────────────┐
│ TIGHT CREDIT │
│ Hard approval │────────► Default: 3-6%
│ Lower volume │ Only strong borrowers entered
│ High standards │
└──────────────────┘
The lifetime curve was determined at the gate.
Nothing the servicer does later changes the composition.
The same mechanism operates in every business that acquires customers over time.
The January cohort arrived through a Facebook campaign targeting broad audiences. The March cohort arrived through a referral program. These two groups will have different lifetime economics not because of anything that happens to them after arrival, but because of who they were when they arrived. The channel selected for different types of people. The selection is the vintage.
The Operator’s Implication
When cohort performance degrades over time across successive vintages, the instinct is to look at what changed in the product, the service, the operations.
Often the answer is simpler. What changed is the acquisition channel, the pricing, the competitive environment, or the marketing message. The pool composition shifted. The vintage imprint shifted with it.
The most common pattern: a business that starts with organic and referral-driven acquisition produces excellent early cohorts. As the business scales and adds paid acquisition channels to grow faster, later cohorts perform worse. The operator blames the product or the operations team. The real cause is that paid channels attract a different population than organic word of mouth.
The vintage imprint means the operator cannot fix later-cohort performance by improving the product alone. If the problem is who walks through the door, the fix is at the door. Not on the factory floor.
PART FIVE: THE ECONOMICS LAYER
Unit Economics Are Cohort Economics
Customer lifetime value is the single most referenced metric in business strategy. It is also the most commonly miscalculated.
The standard formula takes average revenue per customer, multiplies by average retention rate, and projects forward. This produces a single number that supposedly describes the value of acquiring a customer.
It describes nothing. Because it uses averages that blend cohorts of different ages, sizes, and acquisition conditions. The number is a composite of the near-certain three-year survivor and the about-to-leave new signup.
Real unit economics require cohort-level calculation.
| Metric | Aggregate View | Cohort View |
|---|---|---|
| LTV | $2,400 (blended) | Q1: $3,800 / Q2: $1,200 / Q3: $2,900 |
| CAC | $180 (blended) | Q1: $90 (organic) / Q2: $340 (paid) / Q3: $120 (referral) |
| LTV:CAC | 13.3x (looks great) | Q1: 42x / Q2: 3.5x / Q3: 24x |
| Payback | 2.2 months (blended) | Q1: 0.7mo / Q2: 8.5mo / Q3: 1.2mo |
The aggregate says the business has a 13.3x LTV:CAC ratio. Excellent. Scale the acquisition spend.
The cohort view says the Q2 paid channel is barely viable at 3.5x and takes 8.5 months to pay back. Scaling that channel means deploying 8.5 months of working capital per customer for a marginal return.
The aggregate invites a scaling decision. The cohort view reveals that the decision would drain cash to acquire low-value customers.
AGGREGATE vs. COHORT ECONOMICS
┌────────────────────────────────────────────────────────┐
│ │
│ AGGREGATE: "LTV:CAC = 13.3x, scale everything" │
│ │
│ ████████████████████████████████████████████████████ │
│ │
│ One number. One decision. Wrong. │
│ │
└────────────────────────────────────────────────────────┘
┌────────────────────────────────────────────────────────┐
│ │
│ COHORT VIEW: │
│ │
│ Q1 (organic) ████████████████████████████ 42x │
│ Q2 (paid) ███ 3.5x │
│ Q3 (referral) ████████████████████████ 24x │
│ │
│ Three numbers. Three decisions. Clear. │
│ │
└────────────────────────────────────────────────────────┘
The Payback Period Constraint
The CAC payback period is the number of months it takes for a customer’s cumulative gross profit to recover the cost of acquiring them.
This metric only makes sense per cohort. Because different cohorts have different CACs, different revenue trajectories, and different churn rates. The blended payback period obscures the working capital requirement of the worst-performing cohort.
If the Q2 paid cohort takes 8.5 months to pay back and the operator is acquiring 400 customers per month through that channel at $340 each, the operator needs $340 x 400 x 8.5 = $1.156 million in working capital floating against that single channel before it breaks even.
The aggregate payback period of 2.2 months would suggest the capital requirement is $340 x 400 x 2.2 = $299,200. Off by a factor of four.
This is how businesses run out of cash while their dashboards show healthy metrics. The aggregate lied about the capital requirement. The cohort told the truth.
PART SIX: THE SURVIVORSHIP PROBLEM
What the Data Does Not Show
Survivorship bias is not a quirk. It is the structural default of any dataset that only contains entities still in the system.
When the operator pulls a report on customer satisfaction, the report surveys customers who still exist. The customers who left are not in the dataset. Their dissatisfaction, their reasons for leaving, their experience of the product at the moment they decided it was not worth continuing. All gone.
The remaining customers are, by definition, the ones who found enough value to stay. Their satisfaction score reflects their selection. Not the business’s performance.
THE SURVIVORSHIP FILTER
┌────────────────────────────────────────────────────────┐
│ │
│ ORIGINAL COHORT (Month 0) │
│ │
│ ○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○ (100 customers) │
│ Mixed satisfaction: avg 6.2 / 10 │
│ │
└────────────────────────────────────────────────────────┘
│
▼ 12 months pass
┌────────────────────────────────────────────────────────┐
│ │
│ SURVIVING COHORT (Month 12) │
│ │
│ ●●●●●●●●●●●● (35 customers remain) │
│ Measured satisfaction: avg 8.4 / 10 │
│ │
│ The 65 who left took their low scores with them. │
│ The average improved. The product did not. │
│ │
└────────────────────────────────────────────────────────┘
The operator looking at this trajectory sees “satisfaction improved from 6.2 to 8.4.” The operator concludes the product got better.
The product did not get better. The dissatisfied people left. The remaining pool is composed of the people who were satisfied all along.
This is the same mechanism as the retention curve flattening. Selection, not improvement. The distinction determines whether the operator invests in product improvement (because the real satisfaction is low) or maintains course (because the apparent satisfaction is high). Survivorship bias systematically pushes operators toward the second, more dangerous conclusion.
The Churned Cohort
The most valuable cohort for diagnosis is the one the operator never looks at: the people who left.
Exit surveys capture a fraction. Most people who leave a product, a service, a restaurant, a subscription simply stop showing up. They do not announce their departure. They do not explain their reasoning. They produce absence, not signal.
The churned cohort’s silence is itself data. The timing of their departure (month 1, month 3, month 7) tells the operator where the value proposition broke. The characteristics they shared at entry (acquisition channel, plan type, geography) tell the operator who the product is not for.
But this requires the operator to study the dead, not the living. Most reporting systems are built to study the living. The default dashboards show active users, current revenue, existing customers. The dead are in a database table marked “inactive” that no one queries.
PART SEVEN: THE READING DISCIPLINE
The Cohort Table
The cohort retention table is the oldest and most reliable diagnostic instrument in business analytics. It predates digital products by decades. Insurance actuaries built the first versions in the 19th century. Life tables. Mortality tables. The same structure.
Rows are cohorts, defined by entry period. Columns are time periods since entry. Each cell contains the percentage of the original cohort that is still active at that age.
THE COHORT RETENTION TABLE
┌──────────┬───────┬───────┬───────┬───────┬───────┐
│ Cohort │ M0 │ M1 │ M3 │ M6 │ M12 │
├──────────┼───────┼───────┼───────┼───────┼───────┤
│ Jan │ 100% │ 82% │ 64% │ 51% │ 42% │
│ Feb │ 100% │ 78% │ 58% │ 44% │ -- │
│ Mar │ 100% │ 85% │ 71% │ -- │ -- │
│ Apr │ 100% │ 74% │ -- │ -- │ -- │
└──────────┴───────┴───────┴───────┴───────┴───────┘
Read DOWN a column: "How does month-3 retention
compare across acquisition cohorts?"
Read ACROSS a row: "How does this cohort
age over its lifetime?"
The diagonal is the present moment.
Each cell below-left of the diagonal is the future.
There are two ways to read this table. Both are essential. Neither alone is sufficient.
Reading across a row tracks a single cohort’s aging. The January cohort went from 100% to 82% to 64% to 51% to 42%. This is the lifetime decay curve for that specific vintage.
Reading down a column compares the same age across different vintages. Month-1 retention: Jan 82%, Feb 78%, Mar 85%, Apr 74%. This reveals whether acquisition quality is improving or degrading over time, independent of lifecycle effects.
The operator who only reads across sees how customers age but misses whether the incoming quality is changing. The operator who only reads down sees quality shifts but misses lifecycle patterns. Both axes together form the complete diagnostic.
The Diagonal
The diagonal of a cohort table represents the most recent data point for each cohort. January at month 12. February at month 6. March at month 3. April at month 1.
The aggregate retention number that appears on the executive dashboard is, effectively, the weighted average of this diagonal. It blends the oldest cohort’s most mature data with the newest cohort’s most immature data.
This is why aggregate retention can improve while every cohort is worsening. If the business acquired many customers early (large old cohorts with high mature-stage retention) and few customers recently (small new cohorts with low early-stage retention), the diagonal is dominated by the old survivors.
The diagonal is the place where time and composition conspire to mislead.
THE DIAGONAL TRAP
M0 M1 M3 M6 M12
Jan │ 100 │ 82 │ 64 │ 51 │ [42] │ ← old, mature
Feb │ 100 │ 78 │ 58 │ [44] │ │
Mar │ 100 │ 85 │ [71] │ │ │
Apr │ 100 │ [74] │ │ │ │ ← new, immature
[ ] = diagonal = present-moment data
Aggregate "current retention" = weighted average
of the diagonal values.
The diagonal mixes lifecycle stage with vintage quality.
It answers no clean question.
PART EIGHT: THE ASYMPTOTE
What Flattening Means
When a retention curve flattens, it means the selection process is complete. The individuals most likely to leave have left. The remaining individuals have demonstrated, through their continued presence, that their intrinsic retention probability is high.
The level at which the curve flattens is the asymptote.
The asymptote tells the operator three things:
- What fraction of acquired customers are true fits
- How much recurring revenue to expect from each cohort in perpetuity
- Whether the acquisition machine is bringing in the right people
An asymptote at 40% means that for every 100 customers acquired, 40 will persist indefinitely (barring external shocks). The other 60 were acquired at full cost and contributed revenue only during their brief tenure.
THE ASYMPTOTE AS EFFICIENCY METRIC
Acquisition
Cost Per
Customer ┌──────────────────────────────┐
│ │ Paid for 100 customers. │
$200 │ ████████████████████████████████████████ │
│ │ 60 left within 6 months. │
│ │ 40 persist indefinitely. │
│ │ │
│ │ Effective CAC per survivor: │
│ │ ($200 x 100) / 40 = $500 │
│ │ │
│ │ 2.5x the headline CAC. │
│ └──────────────────────────────┘
│
└──────────────────────────────────────────────
The headline CAC is $200. The effective CAC per surviving customer is $500. The asymptote reveals the true acquisition cost by dividing total spend by the number who actually stay.
An operator trying to improve unit economics has two levers: reduce the headline CAC (spend less per acquired customer) or raise the asymptote (acquire a higher fraction of customers who fit). The second lever is almost always higher leverage because it compounds. Every percentage point of improvement in the asymptote reduces effective CAC and increases lifetime revenue simultaneously.
The Product-Market Fit Signal
Andrew Chen at Andreessen Horowitz identifies the flattening retention curve as the primary quantitative signal of product-market fit.
The logic is structural. If no subset of acquired users finds durable value in the product, the curve never flattens. It bleeds to zero. Every user eventually leaves. There is no core.
If some subset finds durable value, the curve flattens at the size of that subset. The product has fit. Not with everyone it acquires. With a specific population whose needs it serves durably.
The curve does not say “the product is good.” It says “the product is good for this specific fraction of the people who try it.” The fraction is the asymptote.
| Asymptote Level | Interpretation |
|---|---|
| 0-10% | No product-market fit. Near-universal churn. |
| 10-25% | Weak fit. Small core. Acquisition-heavy economics. |
| 25-40% | Moderate fit. Viable with efficient acquisition. |
| 40-60% | Strong fit. Healthy cohort economics. |
| 60%+ | Exceptional fit. The product serves most who try it. |
For SaaS businesses, a month-12 retention asymptote above 40% is generally considered strong. For consumer products with low switching costs, above 25% is notable. For ghost kitchens and food delivery, reorder rates above 30% at month-6 indicate the concept has found its regulars.
The numbers vary by industry. The mechanism does not.
PART NINE: THE DECAY FUNCTION
How Cohorts Age
The shape of cohort decay follows a pattern that has been documented across industries, geographies, and decades.
The initial period shows rapid loss. The rate of loss decelerates over time. The curve approaches an asymptote.
Fader and Hardie’s beta-geometric (BG) model captures this mathematically. The model assumes each individual has a constant probability of churning in any given period, but that probability varies across individuals according to a beta distribution. When you aggregate individuals with different constant churn probabilities, you get a cohort-level curve that shows decelerating churn. The curve decelerates not because anyone’s individual churn rate decreased, but because the high-churn individuals have been removed from the pool.
INDIVIDUAL vs. COHORT CHURN RATES
INDIVIDUAL LEVEL:
┌──────────────────────────────────────────────────┐
│ │
│ Customer A: P(churn) = 0.05/period (constant) │
│ Customer B: P(churn) = 0.15/period (constant) │
│ Customer C: P(churn) = 0.40/period (constant) │
│ Customer D: P(churn) = 0.03/period (constant) │
│ │
│ No individual changes. Ever. │
│ │
└──────────────────────────────────────────────────┘
COHORT LEVEL:
┌──────────────────────────────────────────────────┐
│ │
│ Period 1 churn rate: 15.8% │
│ Period 2 churn rate: 12.1% │
│ Period 3 churn rate: 9.4% │
│ Period 4 churn rate: 7.2% │
│ Period 5 churn rate: 5.8% │
│ │
│ Appears to improve. Composition is changing. │
│ │
└──────────────────────────────────────────────────┘
The cohort-level churn rate declines every period. But no individual became less likely to churn. Customer C with a 40% churn probability just left early, removing high-churn weight from the pool. The survivors are increasingly dominated by Customers A and D, whose low individual churn rates pull the cohort average down.
This is the mechanical heart of cohort analysis. The aggregate improves because the composition shifts. Not because the individuals change.
Duration Dependence
Fader and Hardie’s 2018 extension addressed a subtlety. In some businesses, individual churn probability does change over time. A customer who has been subscribed for two years may genuinely be less likely to churn in month 25 than they were in month 2. Not just because of selection, but because of habit formation, sunk cost accumulation, integration depth, or social ties within the product.
Their beta-discrete-Weibull (BdW) model separates the two effects: cross-sectional heterogeneity (different people have different baseline rates) and duration dependence (individual rates change with tenure).
For most businesses, heterogeneity dominates. Duration dependence is small. The curve shape is primarily selection.
But for businesses with deep integration (enterprise software, platforms with data lock-in, financial products), duration dependence can be significant. The longer a customer stays, the higher their individual switching cost becomes, which genuinely reduces their individual churn probability over time.
The operator’s diagnostic question: “Is my improving cohort retention driven by selection or by genuine stickiness?” The answer determines whether the operator should invest in acquisition quality (if selection) or in deepening integration for existing customers (if duration dependence).
PART TEN: THE CONSTRAINTS
Constraint 1: The Observation Window
Cohort analysis requires time. A six-month-old business cannot produce a twelve-month retention curve. The data does not exist yet.
This creates a fundamental tension with the speed at which operators need to make decisions. The operator wants to know if the February cohort will retain at month 12. The February cohort is only four months old. The answer does not exist.
Projection models like Fader-Hardie’s BG can extrapolate from early-period data. But extrapolation degrades with distance from observed data. The model can project month-6 from month-3 with reasonable accuracy. Projecting month-24 from month-3 is speculation wearing a confidence interval.
Constraint 2: Cohort Size
Small cohorts produce noisy data. If the January cohort contains 20 customers, the difference between 40% retention and 45% retention is one person. Statistical significance evaporates.
The operator running cohort analysis on small populations will see wild variance between cohorts that is entirely attributable to sample noise, not business dynamics. The temptation is to interpret every fluctuation as meaningful. Most of it is not.
COHORT SIZE AND SIGNAL RELIABILITY
Signal
Clarity
│
│ ████████████████
HIGH │ ██████
│ ██████
│ ██████
MED │ █████
│ ██
│ ██
LOW │ ██
│█
│
└──────────────────────────────────────────────►
10 50 100 200 500 1000 5000
Cohort Size
The minimum cohort size for reliable month-over-month retention comparison is roughly 100 to 200 customers. Below that, the operator is reading noise. Above 500, the signal becomes clean enough for decision-making. Above 1,000, the operator can begin detecting small effects.
Constraint 3: The Definition Problem
What counts as “retained”?
For a subscription business, it is clear. The customer is paying or not. Binary.
For a marketplace, it is ambiguous. Did the customer place an order this month? This quarter? Does browsing count? Does opening the app count?
The definition of retention determines the shape of every curve, the height of every asymptote, the economics of every calculation. Two operators looking at the same business with different retention definitions will reach different conclusions.
This is not a problem to solve once. It is a problem to solve consciously and consistently. The operator who switches retention definitions between board meetings is comparing curves that cannot be compared.
Constraint 4: External Shocks
Cohort analysis assumes that the world external to the business remains roughly stable between cohorts. When it does not, interpretation breaks.
A recession changes customer behavior across all cohorts simultaneously. A competitor launches and pulls customers from every vintage at once. A platform changes its algorithm and alters acquisition mix for all future cohorts.
These external shocks create period effects that cross-cut cohort effects. Separating them requires more data and more sophisticated methods than most operators have access to.
The practical implication: when something external changes dramatically, treat pre-shock and post-shock cohorts as belonging to different regimes. Do not blend them. Do not project one into the other.
COHORT EFFECTS vs. PERIOD EFFECTS
┌────────────────────────────────┐
│ COHORT EFFECT │
│ │
│ Different cohorts behave │
│ differently because of │
│ conditions at their birth. │
│ │
│ Visible: reading DOWN │
│ a column in the table. │
│ │
└────────────────────────────────┘
┌────────────────────────────────┐
│ PERIOD EFFECT │
│ │
│ All cohorts change at │
│ the same calendar time │
│ because the world changed. │
│ │
│ Visible: a sudden shift │
│ across the diagonal. │
│ │
└────────────────────────────────┘
PART ELEVEN: OPERATOR NOTES
For the Multi-Unit Operator
A ghost kitchen or multi-location food business acquires customers per location, per platform, per concept. Each of these dimensions creates a natural cohort boundary.
Customers who discovered the brand through DoorDash in January on concept A are a different vintage than customers who discovered through UberEats in March on concept B. Blending them into “our customers” destroys every signal that matters.
The pattern seen across food delivery: first-order to second-order conversion is the cliff. Roughly 60-70% of first-time orderers never order a second time. Of those who order a second time, roughly 50-60% order a third. By the fifth order, the customer has self-selected into the core. Reorder rates stabilize.
The asymptote for delivery concepts typically sits between 15-30% at month-6, heavily dependent on concept differentiation and order experience. Commodity concepts (pizza, wings) have lower asymptotes because switching costs are near zero. Distinctive concepts have higher asymptotes because the customer cannot get the same thing elsewhere.
The operator’s highest-leverage cohort question: “What is the month-6 reorder rate by acquisition platform and concept, and how has it changed across quarterly cohorts?” If the answer is declining, the fix is almost never operational. It is either a channel-mix problem (scaling the wrong platform) or a concept-fit problem (the menu no longer matches the audience the platform delivers).
The Cohort Tells the Operator Where They Are
The shape and level of the retention curve is a stage indicator.
Early-stage businesses have volatile, noisy cohort curves because cohorts are small and the product is changing underneath them. The operator reads the curve for direction, not precision.
Growth-stage businesses have stabilizing curves with clear asymptotes. The operator reads the curve for asymptote height and vintage-over-vintage comparison. Is the asymptote rising (improving acquisition quality)? Flat (stable)? Falling (degrading)?
Mature businesses have well-established curves with tight confidence intervals. The operator reads the curve for marginal shifts. A one-percentage-point drop in the asymptote, spread across thousands of acquired customers per month, is a material change.
The same metric at different stages requires different reading protocols.
The Customer Cohort Chart (C3)
Fader’s Customer Cohort Chart (C3) is the single most useful visualization for understanding where revenue comes from.
The chart shows total revenue in each period, broken down by the cohort that generated it. The operator sees, visually, how much of current revenue comes from the January vintage, the February vintage, the March vintage.
This reveals a truth that aggregate revenue conceals: whether the business is building on a compounding base of retained customers, or running on a treadmill where new acquisition must replace churned revenue every period.
THE CUSTOMER COHORT CHART (C3)
Revenue
│
│ ┌──────────────────────────────────────────┐
$50K │ │▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓│
│ │▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓░░░░░░░░│
$40K │ │▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓░░░░░░░░▒▒▒▒▒▒▒▒│
│ │▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓░░░░░░▒▒▒▒▒▒▒▒████████│
$30K │ │▓▓▓▓▓▓▓▓▓▓░░░░░░▒▒▒▒▒▒████████ │
│ │▓▓▓▓░░░░░░▒▒▒▒▒▒████████ │
$20K │ │░░░░▒▒▒▒▒▒████████ │
│ │▒▒▒▒████████ │
$10K │ │████████ │
│ └──────────────────────────────────────────┘
└──────────────────────────────────────────────►
Q1 Q2 Q3 Q4 Q5 Q6 Q7 Q8
████ = Oldest cohort ░░░░ = Middle cohorts
▒▒▒▒ = Newer cohorts ▓▓▓▓ = Newest cohort
Healthy: old cohorts form a thick, stable base.
Unhealthy: old cohorts thin out, new cohorts
must replace them just to stay flat.
In a healthy business, the bottom layers (oldest cohorts) remain thick over time. Revenue compounds because retained customers continue generating revenue while new cohorts add on top.
In an unhealthy business, the bottom layers thin rapidly. Each new period requires a larger and larger new cohort just to replace the revenue lost from prior cohorts. The business is on a treadmill. Growing the top line while the base erodes.
The C3 chart makes this visible in a way that a single revenue number never can.
PART TWELVE: THE COMPLETE PICTURE
The Unified Framework
Everything connects.
THE COMPLETE COHORT FRAMEWORK
┌────────────────────────────────────────────────────────┐
│ │
│ ACQUISITION │
│ │
│ Channel, pricing, message, timing, and competitive │
│ context determine WHO enters. The vintage imprint. │
│ │
└────────────────────────────────────────────────────────┘
│
▼
┌────────────────────────────────────────────────────────┐
│ │
│ THE COHORT │
│ │
│ A mixed population with heterogeneous fit. │
│ Some will stay. Some will not. The ratio │
│ was determined at the gate. │
│ │
└────────────────────────────────────────────────────────┘
│
┌─────────────┼─────────────┐
│ │ │
▼ ▼ ▼
┌──────────────┐ ┌──────────────┐ ┌──────────────┐
│ SELECTION │ │ DECAY │ │ ECONOMICS │
│ │ │ │ │ │
│ Low-fit │ │ Curve │ │ True CAC, │
│ members │ │ decelerates │ │ true LTV, │
│ depart │ │ as pool │ │ true payback│
│ rapidly │ │ purifies │ │ visible │
│ │ │ │ │ only here │
└──────────────┘ └──────────────┘ └──────────────┘
│ │ │
└─────────────┼─────────────┘
│
▼
┌────────────────────────────────────────────────────────┐
│ │
│ THE ASYMPTOTE │
│ │
│ The fraction that persists. The true core. │
│ The compounding base. The real business. │
│ │
└────────────────────────────────────────────────────────┘
The cohort is the unit of truth.
The aggregate is the unit of concealment.
The retention curve is the diagnostic instrument.
The asymptote is the product-market fit signal.
The vintage comparison is the acquisition quality signal.
The C3 chart is the compounding signal.
Together, they form the only honest picture of whether a business is building on solid ground or running on a treadmill painted to look like a road.
Final Synthesis
The cohort is not a metric. It is a lens.
Every metric that matters. Retention, lifetime value, acquisition cost, payback period, satisfaction, engagement. All of them become honest only when viewed through this lens. Without it, they are averages. Averages that blend the certain with the uncertain, the old with the new, the selected with the unselected.
The operator who looks at aggregates is looking at a composite photograph. Fifty faces merged into one. The result is smooth, symmetrical, and describes no one.
The operator who looks at cohorts is looking at individuals. Not literally. But at the closest approximation the data allows. Groups born under the same conditions, tracked through the same lifecycle, compared on equal terms.
The machinery is simple.
Group by entry time. Track over lifecycle. Compare across vintages. Read the shape. Find the asymptote. Calculate true economics.
The machinery is simple but almost no one runs it.
Because the aggregate is easier. Because the dashboard defaults to blended. Because the board deck template has one number for retention, not a table. Because reading a cohort table requires a different kind of attention than reading a single metric.
The cohort does not make the business better. It makes the operator’s vision accurate. What happens after that is a function of what the operator does with accurate vision.
The aggregate told the operator the business was healthy. The cohort told the operator which part was healthy, which part was bleeding, and exactly when the bleeding started.
One produced comfort. The other produced clarity.
The machinery does not care which one the operator chooses to look at.
It runs regardless.
CITATIONS
Cohort Analysis and Retention Modeling
Fader, P.S. & Hardie, B.G.S. (2007). “How to Project Customer Retention.” Journal of Interactive Marketing, 21(1):76-90. The foundational beta-geometric (BG) model for projecting cohort-level retention from early-period data. https://faculty.wharton.upenn.edu/wp-content/uploads/2012/04/Fader_hardie_jim_07.pdf
Fader, P.S., Hardie, B.G.S., Liu, Y., Davin, J., & Steenburgh, T. (2018). “‘How to Project Customer Retention’ Revisited: The Role of Duration Dependence.” Journal of Interactive Marketing, 43:1-16. Extension of the BG model to the beta-discrete-Weibull (BdW), separating heterogeneity from genuine duration dependence. https://journals.sagepub.com/doi/abs/10.1016/j.intmar.2018.01.002
Hardie, B.G.S. “A Spreadsheet-Literate Non-Statistician’s Guide to the Beta-Geometric Model.” https://www.brucehardie.com/notes/032/BG_intro.pdf
Customer Centricity and Heterogeneity
Fader, P.S. (2020). Customer Centricity: Focus on the Right Customers for Strategic Advantage. Wharton School Press. The argument that customer heterogeneity is the central fact of business strategy and that cohort-level analysis is the minimum viable lens. https://marketing.wharton.upenn.edu/profile/faderp/
Fader, P.S., Hardie, B.G.S., & Ross, M. (2022). The Customer-Base Audit. Wharton School Press. Framework for the Customer Cohort Chart (C3) and cohort-based revenue decomposition.
Fader, P.S. & Toms, S. (2018). The Customer Centricity Playbook. Wharton School Press.
Product-Market Fit and Retention Curves
Chen, A. (2019). “Magic metrics indicating a startup probably has product/market fit.” Includes cohort retention curves that flatten (stickiness) as a primary PMF signal, power user curves showing a “smile” pattern. https://x.com/andrewchen/status/1184170125525577728
Headline VC. “Cohorts Retention 101 for Startups.” The retention asymptote as PMF indicator and forecasting tool. https://deepdive.headline.com/learn/resources/cohorts-retention-101-for-startups
a16z (Andreessen Horowitz). “Retention Is All You Need.” AI product retention benchmarks and cohort methodology. https://a16z.com/ai-retention-benchmarks/
Simpson’s Paradox and Aggregation Bias
Simpson, E.H. (1951). “The Interpretation of Interaction in Contingency Tables.” Journal of the Royal Statistical Society, Series B, 13(2):238-241. The original formalization of the reversal paradox.
Statistics By Jim. “Simpson’s Paradox Explained.” Accessible treatment with business examples. https://statisticsbyjim.com/basics/simpsons-paradox/
Alipourfard, N., Fennell, P.G., & Lerman, K. (2018). “Using Simpson’s Paradox to Discover Interesting Patterns in Behavioral Data.” https://arxiv.org/pdf/1805.03094
Vintage Analysis and Credit Risk
Listen Data. “Credit Risk: Vintage Analysis.” Vintage curve methodology and loss-curve interpretation for cohort-based lending portfolios. https://www.listendata.com/2019/09/credit-risk-vintage-analysis.html
Finley Technologies. “What Are Vintage Curves?” Practical treatment of vintage analysis in lending contexts. https://www.finleycms.com/blog/what-are-vintage-curves
Network Science and Distribution Mechanics
Barabási, A.L. & Albert, R. (1999). “Emergence of Scaling in Random Networks.” Science, 286(5439):509-512. Scale-free networks and preferential attachment as the structural substrate of distribution.
Survivorship Bias
Brown, S.J., Goetzmann, W., Ibbotson, R.G., & Ross, S.A. (1992). “Survivorship Bias in Performance Studies.” Review of Financial Studies, 5(4):553-580. The original quantification of survivorship bias in fund performance data. https://terpconnect.umd.edu/~wermers/ftpsite/FAME/Brown_Goetzmann_Ibbotson_Ross.pdf
Unit Economics
Ordergroove. “How to Calculate CAC Payback Period for Subscription Businesses Accurately Using Cohort Analysis.” Cohort-level payback calculation methodology. https://www.ordergroove.com/blog/how-to-calculate-cac-payback-period-for-subscription-businesses-accurately-using-cohort-analysis/
Document compiled from peer-reviewed statistical modeling, Wharton customer analytics research, venture capital retention benchmarks, credit risk vintage analysis, and network science literature.