Reading a cohort table
Read a cohort table along its rows, columns and diagonals, name its curve shape, and compute LTV and payback from the curve rather than one churn rate.
On this page
A single churn rate is an average over customers who behave very differently. A cohort table shows them apart: one row per month of new customers, one column per month since they started paying, and in each cell the share still paying. Six rows of the invoicing app's table are enough to replace the 5% assumption the last two lessons used, and to find the one month of acquisition that will never pay back.
The table
Each row below is a month's paid cohort, about the 96 customers the funnel produces from $50,000 of spend, with its size beside the month.
The blank corner is months that have not happened yet. June's cohort has one cell because it has existed for one month. A blank is not a zero, and a spreadsheet that fills it with zero reports a collapse that has not occurred; cohorts and funnels reads the same blank corner on a product's weekly cohorts.
Three ways to read it
Along a row, one cohort ages. January runs 100, 82, 72, 65, 62, 60. The monthly losses are 18%, 12%, 10%, 5% and 3%: each smaller than the last, so the curve is flattening towards a group of customers who stay.
Down a column, cohorts are compared at the same age, which shows whether the product or the acquisition is getting better. The month-2 column reads 82, 83, 74, 82, 84. May's 84 looks like an improvement, but a cohort of 95 at 84% varies by about √(95 × 0.84 × 0.16) ≈ 3.6 customers from month to month, roughly 4 points, so 84 against 82 is noise. March's 74 is not: eight points below its neighbours, twice that spread.
Along a diagonal, cohorts are compared in the same calendar month. January's month 4, February's month 3, March's month 2 and April's month 1 all happened in April. An outage, a price change or a failed billing run shows as a dip along a diagonal rather than a row. No diagonal here stands out, which rules out a calendar cause for March and points at the cohort itself.
Three shapes a row can take
Flattening is January. Losses shrink to about 3% a month by month 6, leaving a core of freelancers for whom the app is part of the week. It is the shape of a product that fits the segment it was sold to.
Decaying to zero is March. The spend that bought the March cohort, three months before it started paying, went to a broad campaign on searches for free invoice templates, and it reached people who needed one invoice, not an invoicing habit. The row loses a quarter of what is left every month, 74, 55, 41, with no sign of a floor. The cohort was the largest of the six at 140, and its CAC the lowest: $50,000 ÷ 140 = $357, the only month under the course's $368 payback ceiling. Its expected paying months, if the quarter-a-month loss continues, are 1 + 0.74 + 0.55 + 0.41 + 0.41 × 0.75 ÷ 0.25 = 3.93, worth 3.93 × $40 ≈ $157. A $357 customer worth $157 never pays back. The month with the best CAC on the dashboard was the worst month of acquisition in the half-year.
The smile appears only in a revenue table. Agency accounts that stay add seats, so a revenue cohort can dip as small accounts leave and climb back past 100% as the survivors grow: 100, 88, 80, 76, 78, 90, 104 by month 12. A customer count can never exceed the cohort it started with. When revenue smiles, LTV comes from the revenue curve, and expansion becomes a lever of its own, the one PLG versus sales-led uses for its hybrid motion.
LTV from the curve
LTV is monthly margin times the expected number of paying months, and the expected number of paying months is the sum of a row, read as fractions. One churn rate turns that sum into 1 ÷ churn. The curve lets us add it up instead.
January's first six cells sum to 1 + 0.82 + 0.72 + 0.65 + 0.62 + 0.60 = 4.41 paying months, against 5.30 for the same months under a flat 5%. Beyond month 6 the row needs an assumption; take the last observed loss, about 3% a month, so the tail adds 0.60 × 0.97 ÷ 0.03 = 19.4 months. Three LTVs come out of the same row:
one rate fitted to six months 0.60^(1/5) = 0.903, so 9.7% churn $40 ÷ 0.097 = $412
the 5% assumption $40 ÷ 0.05 = $800
the curve, with a 3% tail $40 × (4.41 + 19.4) = $40 × 23.8 = $952
The fitted rate is the common analyst's error: it spreads the heavy early losses across a tail that loses far less, and halves the value. The curve gives the highest LTV of the three.
It also gives the latest payback. Cumulate $40 × each cell against the $521 CAC from CAC, LTV and payback: $176 by month 6, $521 in month 26, five months after the flat 5% model's month 21. The early months do the paying back, and the curve has fewer customers in them than the flat rate assumed; the floor adds its value after payback is over. A higher LTV and a later payback, from one row, is the normal result for a flattening curve whose churn is front-loaded.
The tail is the weakest number in the sum: 19.4 of the 23.8 months are an assumption about months nobody has observed. So report value at a 36-month cap beside the lifetime figure. Capped at 36 months the curve is worth $641, less than the flat model's $674; everything the curve adds lies beyond month 36. The module's code exercise, payback month from a cohort, uses the same 36-month horizon.
In practice
To finance: replace the single churn rate in the model with the cohort curve, report payback from it (month 26 for January, not 21), and show LTV at a 36-month cap with the lifetime figure beside it and the tail assumption labelled.
To whoever runs acquisition: judge a month of spend by its cohort's month-3 cell, not by its customer count. A count and a CAC are known in the month; a row takes three months to say anything, and it is the row that pays.
The mistake that costs money is scaling a month of acquisition on its count and CAC before its row has aged. March delivered 140 customers at $357, the best month on the dashboard, and each is worth about $157.
Check yourself
1. July's cohort of about 100 reaches month 2 at 76%, after a new campaign aimed at freelancers who invoice every month. Is it March again, and do you cut the campaign?
Not yet known. The spread for 100 customers at 76% is √(100 × 0.76 × 0.24) ≈ 4 points, and 76 is 6 below the usual 82: one and a half spreads, not two. Hold the campaign's budget flat for a month. At month 3, March read 55 and the others about 72; a July reading near 72 says it is not March, and one near 55 says cut.
2. Finance proposes using the curve's $952 instead of $800, which lifts LTV/CAC from 1.54 to 1.83, and the founder wants to raise paid spend on the strength of it. Do you agree?
No. The same curve moves payback from month 21 to month 26, so the cash need grows, and 19.4 of its 23.8 paying months are assumed. Capped at 36 months it is worth $641, less than the $674 the flat model gave. Show the board the month-26 payback; the higher LTV argues for patience, not for more spend.