Start on the envelope, then meet its flattery. The simple model: margin per order × orders per year × expected years − CAC. A D2C brand with ₹1,000 orders at 30% margin, 2.5 orders a year, 'about three years' of life and ₹500 acquisition: 300 × 2.5 × 3 − 500 = ₹1,750. Fast, communicable, and quietly dishonest twice over: 'three years' is an assumption wearing a memory, and rupees in year three are counted as if they were rupees today. The envelope is fine for a first conversation and dangerous for a budget.
The retention-perpetuity model replaces the guessed lifespan with observed behaviour. If each year a customer survives with probability r, and future money is discounted at rate d, the infinite sum collapses to CLV = m × r ÷ (1 + d − r), the 'margin multiple' formula. Our brand at m = ₹750, r = 60%, d = 15%: 750 × 0.6 ÷ 0.55 ≈ ₹818, well under the envelope's number, because the envelope's 'three years' smuggled in immortality the data doesn't support. Now move the dial: r = 70% gives 750 × 0.7 ÷ 0.45 ≈ ₹1,167. The denominator (1 + d − r) is the insight: as retention approaches the discount rate's ceiling, value doesn't grow linearly, it compounds. This is why retention initiatives out-argue acquisition spend in mature businesses, and why a CLV model without a discount rate always overpays for growth.
Cohort analysis is the model that can't be argued with, because it's just your history arranged honestly. Take everyone acquired in a given month, track their cumulative contribution margin over the following months, and lay the cohorts in a triangle: each row a vintage, each column an age. The triangle answers what averages hide: are newer cohorts better or worse than older ones (is product improving or is acquisition scraping weaker segments?), where does the repeat curve flatten (your practical horizon, beyond which you project nothing), and what does a customer actually return by month twelve against what CAC assumed? Blended CLV rises while every new cohort worsens: the classic growth-story autopsy, visible only in the triangle.
The contractual/non-contractual split decides how far the maths can honestly go. Contractual businesses observe the death certificate (a cancelled plan), so survival rates are data and the perpetuity model fits naturally. Non-contractual businesses observe only silence: the customer who hasn't ordered in 90 days may be churned or merely between weddings. Here the honest tools are probabilistic. Fader and Hardie's BG/NBD family models each customer's purchase rhythm and dropout probability from recency and frequency, producing 'probability alive' instead of fake churn dates. You don't need to run the estimation by hand to be literate. You need to know that applying subscription-style churn maths to a retailer manufactures precision from a category error, and to ask which family your dashboard's number came from.