What are a firm’s current and future customers worth?

A growing firm often reports a loss: its earnings say little about its value, and an earnings multiple becomes unusable ^[C10]. Gupta, Lehmann and Stuart propose valuing its customers, today's and those it will acquire, from four quantities: margin per customer, retention, acquisition cost and the pace of acquisition ^{[C02, C03, C15]}. This is the value of the customer base (customer equity); it approximates the value of the firm to the extent that customers make up most of it ^{[C01, C77]}.

The page's program first reproduces the published case: from the authors' inputs alone, with a calendar convention that had to be chosen, it recovers the value of the customers of Amazon, Ameritrade, eBay and E∗Trade in March 2002 and their sensitivity to retention, margin and acquisition cost; it does not recover Capital One. Then, on a synthetic online retailer whose data are drawn from the model itself, with the true retention and no departures before the first quarter, the method values the customer base at €223.73M, against €214.33M computed with the true curve: +4.4%, a gap within chance variation. Counting current customers only understates it by 27.7%; the "average lifetime" shortcut overstates it by 34.3%; an acquisition curve fitted too early, before its peak, stays within 20% of the reference in only 30 draws out of 100. Raising retention from 80% to 80.8%, a 1% relative improvement, adds 3.75% to the value of the base here; cutting the acquisition cost by 1% adds 0.12% — before counting what either improvement costs.

An online shop selling hiking gear publishes its number of active customers every quarter. It acquires customers fast: 35,972 at the start, 1,070,589 sixteen quarters later. Each active customer brings a margin of €12 per quarter; acquiring a customer costs €40 of marketing; 80% of one year's customers are still there the following year. Yet the shop reports a loss: its overheads, added to its acquisition spending booked as expenses, still exceed its margin.

Management is preparing a funding round and must answer two questions: what is its customer base worth, counting the customers it has not yet acquired; and, to grow it, is it better to lower the acquisition cost, raise the margin or keep customers longer?

The page's data are synthetic. The cumulative number of customers who have bought at least once follows an S-shaped curve with a ceiling of 2,000,000, whose acquisition peaks at quarter 13.33; each quarter's acquisitions are observed with noise of about 10%.

What is the present value of the future margins of all customers, current and future, net of the acquisition cost of the future ones ^{[C15, C21]}? The acquisition cost of current customers is already spent: it no longer counts ^[C30]. The difficulty is twofold. The number of future customers is not observed: the acquisition curve must be extended. And retention, the input that weighs most, is rarely published ^[C31]: the authors obtain it from analysts, from an expert or from an industry average ^{[C32, C33, C34, C35]}.

Three common shortcuts:

• Counting current customers only. Multiplying the number of customers by the value of one customer leaves out those the firm will acquire, and what acquiring them will cost. The authors value the current and future base instead ^[C15].
• The "average lifetime". With 80% annual retention, a customer stays five years on average: five full years of margin are then discounted. The authors show that this method overstates value ^[C18]: for an annual margin of $100, 80% retention and a 12% discount rate, it gives $360 instead of $250, about 44% too much ^{[C19, C20]}. Much of this gap is a matter of calendar: the $250 counts the first margin only after a year of survival, whereas the shortcut counts it as certain. On the same calendar, the correct value is $312.50 and the excess falls to 15.4%: that is the cost, on its own, of treating every customer as staying exactly the average lifetime.
• Extending an acquisition curve too early. Fitting an S-shaped curve requires having observed its inflection point, the moment acquisition peaks ^[C41]; before it, the parameters become unstable ^[C42].

The customer count alone is not enough either. In 1999 and 2000, the number of customers or visitors was used to value internet firms ^[C11]; yet a customer is worth only the margin they leave, how long they stay and what they cost.

The value of a customer is the discounted sum of their future margins, as long as they stay ^[C02]. The value of the base is the sum of two blocks ^[C15]:

1. current customers: each is worth their margins to come; their acquisition cost is already paid ^[C30];
2. future customers: an S-shaped curve extends acquisition; it is similar in spirit to the Bass diffusion model but more convenient to compute ^[C24], and its estimates are very comparable to those of the Bass model according to Bass, Jain and Krishnan, cited by the authors ^{[C25, C84]}. Each acquired customer brings their future margins minus their acquisition cost ^[C21], and the value of each cohort is discounted to today.

Retention matters twice: it sets how long customers stay, and it is used to rebuild acquisition. A firm publishes only its active customers, acquired minus departed: with 100,000 customers then 130,000 and 80% retention, it acquired 50,000 customers, not 30,000 ^[C38]. The authors therefore model the cumulative number of customers who have bought at least once, not the net increase ^[C88]. Assuming another retention rate also means assuming another acquisition path.

The authors use only public information, annual reports and stock-exchange filings, except for retention ^[C12]; their method therefore also serves investors, analysts and acquirers who lack internal data ^[C13]. For each quarter, you need:

• the number of active customers;
• the margin per customer, including direct costs that the firm does not always deduct, such as fulfilment ^[C26]; the authors take the average of the last four quarters ^[C27];
• the acquisition cost: the quarter's marketing cost divided by the number of customers acquired ^[C28]. Part of this cost serves to retain customers, and it cannot be separated ^[C29];
• an annual retention rate, declared and justified ^[C31], and an annual discount rate: the authors use 12%, the middle of the 8% to 16% range of finance textbooks ^{[C36, C37]}.

You also need enough history to have passed the acquisition peak ^[C41].

This section is for practitioners. Decision-makers can skip it and resume at "The declared calculation": the "In plain words" sentence at its end is enough.

• t: time, in quarters; t = 1 for the first observed quarter, t = 0 for the stock of customers before it; T: the valuation quarter, here 16.
• N(t) = α / (1 + exp(−β − γt)): the cumulative number of customers who have bought at least once ^[C24]. α is the ceiling of this curve, the total number of customers the firm will ever acquire; β places it in time; γ sets its slope.
• n(t) = αγ exp(−β − γt) / (1 + exp(−β − γt))²: the number of customers acquired per quarter at date t, that is, the slope of N(t); it peaks at t = −β / γ.
• A_t: active customers at the end of quarter t. r: annual retention; r_q = r^(1/4): quarterly retention, the rate that, repeated over four quarters, gives back r. i: annual discount rate; i_q = (1 + i)^(1/4) − 1: the quarterly rate that, compounded four times, gives back i. This move to quarterly rates is a convention of the page; the authors do not state it.
• Rebuilt acquisitions: n_t = A_t − r_q × A_(t−1); cumulative N_t = N_(t−1) + n_t, with N_0 = A_0, that is, no departures before the first quarter ^{[C38, C88]}. α, β and γ are estimated by nonlinear least squares on N_0, …, N_T ^[C39].
• m: the margin per active customer per quarter; c: the acquisition cost per acquired customer.
• The horizon is infinite: retention and discounting make distant margins weigh much less ^[C17].
• Value of one customer from the next quarter on: LV = m × r_q / (1 + i_q − r_q), the sum of m × r_q^k / (1 + i_q)^k for k ≥ 1: each future margin, weighted by the chance that the customer is still there, then discounted. This is the formula of the authors' example: 100 × 0.8 / (1.12 − 0.8) = 250 ^[C19]. Their equation (2), written from t = 0, would give 350; the page follows their example, and the convention that reproduces their values.
• Value of the base at quarter T: V = A_T × LV + ∫ n(s) × (LV − c) × (1 + i_q)^−(s − T) ds, the integral running over the acquisition date s, from T to infinity. The first term covers current customers, the second future customers ^{[C15, C21, C30]}. The authors do not write the current-customer term in this form: it is a reconstruction, validated by the reproduction of their values. They then apply a 38% tax to approximate firm value ^[C23]; the page reports pre-tax values, except to reproduce their case.
• Lifetime shortcut: each customer is worth the annual margin 4m discounted over 1 / (1 − r) full years, the first margin being counted at the end of the first year ^{[C18, C20]}.
• Elasticity: relative change in V when an input improves by 1% of its value, the curve being held fixed — retention multiplied by 1.01, margin multiplied by 1.01, acquisition cost and discount rate multiplied by 0.99. In this model, the margin elasticity always equals 1 plus the acquisition-cost elasticity.

In plain words: the value of the base is that of today's customers, plus that of tomorrow's customers minus what acquiring them will cost; retention sets both how long they stay and how many customers the firm has acquired.

1. Code check on the published case. Both programs recompute the customer value of the authors' five firms, built on public data ^[C04], as of 31 March 2002, from their inputs: customers, quarterly margin, acquisition cost and retention from their Table 1 ^{[C60, C65, C66, C67, C68]}, curve parameters from their Table 2 ^{[C61, C81, C82]}, 12% discounting and 38% tax ^{[C23, C36]}. The authors do not state how they move from annual to quarterly rates; the convention used, the fourth root of retention and of 1.12, with t = 1 for the first quarter of data, was chosen because it reproduces their values. Independent evidence supports it: with it, the calendar dates of the acquisition peak in their Table 2 are recovered for all five firms, and none with t = 0 ^[C83]. The programs stop with an error if it no longer reproduces four firms out of five. They also recompute Table 5 ^{[C85, C86, C87]}.
2. Synthetic retailer, drawn by a pseudo-random number generator written out in full (seed 20261006; x ← 1103515245 · x + 12345 mod 2³¹), so that Python and R draw the same numbers: sixteen quarters of active customers, each quarter's acquisitions being multiplied by random noise with a standard deviation of 0.10 on the log scale. The design favours the method: the analyst knows the true retention, and nobody left before the first quarter.
3. Estimation as an outside analyst would do it: acquisitions rebuilt with the assumed retention, curve fitted by least squares (Nelder-Mead simplex, an algorithm that searches for the best fit step by step, restarted until stable). Two bounds are set: the ceiling α does not exceed a thousand times the customers already counted, the slope γ does not exceed exp(3) per quarter, about 20.
4. Value of current and future customers, the integral computed by Simpson's rule with a step of 0.05 quarter over 400 quarters. Reference: the same value computed with the true curve and the observed active customers.
5. Three shortcuts: current customers only; average lifetime; curve fitted on the first eight quarters, before the peak.
6. Elasticities, defined as in the authors' Table 4 for retention, margin and acquisition cost ^[C63], and a grid of retention at 70%, 80% and 90% and discounting at 8%, 10% and 12%, modelled on their Table 5, holding the curve fixed and then re-estimating it.
7. Uncertainty: parametric bootstrap, which simulates 1,000 retailers from the estimated curve and the estimated noise, then re-estimates each; the noise is estimated counting the three parameters of the curve; interval from the 25th to the 975th sorted value, at a nominal 95%.
8. Repetition on 100 independent retailers drawn from the true curve, each with a bootstrap of 200 draws, to measure the actual coverage of the interval.
9. Invariants: seven identities, including the authors' numerical example (250 and 360) ^{[C19, C20]}, their example of rebuilt acquisitions ^[C38] and the integral of acquisitions equal to the remaining market; the programs stop if one fails.

The published case, recomputed. Customer value after tax, in billions of dollars, as of 31 March 2002:

For the first four firms, the elasticities to margin and acquisition cost are also recovered, to the published rounding; as one always equals 1 plus the other, this adds one check, not two. Three gaps remain, with different causes.

• The discount-rate column. A rate multiplied by 0.99 gives +0.28% for Amazon, where the authors publish 0.46%. Their column is recovered, to within 0.01 point, for the same four firms if the rate moves from 12% to 11.80%: +0.47% against 0.46, +1.17% against 1.17, +0.64% against 0.63, +1.14% against 1.14. It is as if the authors called "1%" a cut of more than 1% in the rate, without saying so.
• Capital One. Its whole row is recovered if the value of its future acquisitions is 1.25 times larger: 11.00; +5.12%; +0.32%; +1.32%. The cause most likely lies in its curve or its time origin — a typo in Table 2 or a different data window — and cannot be identified without the series.
• Table 5 and the values at 100% retention. Recomputed with the curve held fixed, these values depart from the published ones, and 36 cells out of 39 of Table 5 depart in the direction a curve re-estimated with each retention rate would: the published values are lower than ours when the assumed retention exceeds the firm's own, higher otherwise. Outside these 39 cells, at each firm's own retention, 4 cells out of 6 are recovered; the other two, eBay's at two discount rates, escape this explanation, since retention does not change there. At 100% retention, the page finds 3.72 for Amazon against "about 3" ^[C78], 6.05 for eBay against 5.3 ^[C79] and 3.95 for E∗Trade against 3.89 ^[C47]: all three published values are lower. For Capital One at 90%, 13.35 against 14.1 ^[C64]. If the authors re-estimated their curve in Table 5, they did not do so in Table 4, which is recovered with the curve fixed: on the synthetic retailer, re-estimating the curve brings the effect of retention multiplied by 1.01 down from +3.75% to +3.18%, a gap that would show in Table 4. This explanation therefore implies a change of procedure between the two tables; their quarterly series are not published, and we cannot check it.

The synthetic retailer. With 80% annual retention, quarterly retention is 0.9457; with 10% discounting, the quarterly rate is 2.411%; an active customer is then worth €144.81 of future margins. The rebuilt acquisitions give 1,382,156 customers who have bought at least once, where the true number is 1,379,949. The estimated curve has a ceiling of 2,099,582 customers, where the true ceiling is 2,000,000, and peaks at quarter 13.78 instead of 13.33.

The gap, +4.4%, comes entirely from future customers: the ceiling of the curve is overestimated. Future customers account for 30.7% of the estimated value.

The shortcuts.

Counting current customers only amounts to removing future customers: the gap is, by construction, their share of the reference; it would be reversed if acquiring a customer cost more than the customer brings in. The average lifetime credits each customer with €181.96 instead of €144.81, 25.7% too much, and the error carries over to future customers. This gap mixes two effects. Compared with an annual value that, like the shortcut, counts the first margin at the end of the first year, €160.00, the shortcut is 13.7% too high; that is the effect of the average lifetime itself. The rest comes from the first year: the annual value of €160.00 counts that year's margin as secured, whereas the quarterly value lets customers leave from the first quarter on; paying margins quarterly, which works the other way, does not offset this gap — an annual value that applies retention from the first year is only €128.00. Fitted on the first eight quarters, before the peak, the curve has a ceiling of 1,677,096 customers; the value at quarter 8 falls to €161.43M against a €188.30M reference, −14.3% — on this draw.

Which lever? Effect of improving each input by 1% of its value, the curve held fixed:

Retention clearly beats margin, as in the authors' results ^{[C07, C08]}. The order between margin and acquisition proves nothing: the margin elasticity here equals exactly 1 plus the acquisition-cost elasticity, an identity of the model. With the same definitions, the effect of retention is 10.4 times that of the discount rate; the authors' ratio of "almost five" ^{[C09, C55]} compares 1% of retention with a larger cut in the rate.

Retention, the assumption that weighs most. Value of the base in €M; on the left, the curve held fixed; on the right, acquisitions rebuilt and the curve re-estimated with the assumed retention:

The two columns answer two questions. The fixed curve measures a lever: if the shop keeps its customers longer from now on, its past acquisitions do not change, and the value moves from €223.73M to €376.39M at 90% retention. The re-estimated curve measures an uncertainty: if past retention was in fact 90% rather than 80%, the shop also acquired fewer customers than believed, and the base is worth €350.08M. Assuming 70% instead of 80% rebuilds 1,560,456 customers who have bought instead of 1,382,156, and a ceiling of 2,449,181 instead of 2,099,582: more assumed departures mean more assumed acquisitions. Between 70% and 90%, the value more than doubles.

On 100 independent retailers, the full method deviates from the reference by −0.3% on average (standard deviation 3.4%, at most 10.6%), a bias that the simulation error, that is, the imprecision due to the finite number of draws, 0.3%, does not distinguish from zero; current customers only, by −27.8% (standard deviation 0.6%), a stability that follows from the construction of the design. Fitted before the peak, the curve gives a median gap of −8.8%, and only 30 draws out of 100 stay within 20% of the reference. In 6 draws, the ceiling runs to its bound of a thousand times the customers counted, and the gap there measures nothing but that bound; the other draws range from −53.7% to +2,616.9%.

• Constant retention ^[C16]. The authors adopt it for convenience; it implies that a mature firm, which hardly acquires anymore, would end up losing all its customers ^[C57]. Do your long-standing customers leave less than new ones?
• Known retention. It is almost never published ^[C31]; the authors take it from an analyst's report for Ameritrade, extend it to E∗Trade by similarity ^[C32], ask an expert for Capital One ^[C33] and take the average of US firms for eBay ^[C34]. Where does your retention figure come from, and on what definition of an active customer?
• The same margin profile for all cohorts ^[C22], and a stable margin: the authors take the average of the last four quarters ^[C27]. Do your latest customers spend like the first ones?
• Acquisition that follows an S-shaped curve, fed by all customers, active or departed ^[C40]. For eBay, a marketplace, network effects between buyers and sellers are not captured ^[C48]. Do your customers bring in other customers?
• A history that goes past the acquisition peak ^{[C41, C42]}. Have your quarterly acquisitions already started to decline?
• Independent levers. The method ignores the links between acquisition, retention and margin ^[C53]: a price promotion can acquire customers who stay only briefly ^[C54]. Do customers acquired through promotions stay as long as the others?
• Customers who make up most of the value. The method sees neither growth options, that is, the value of future activities that are still uncertain ^[C06], nor non-customer assets; it does not replace financial methods, it extends one of them ^[C59]. Does part of your value lie in products or markets you have not yet launched?
• What the data cannot settle: the acquisition ceiling of a young firm ^[C58], and future retention, which history does not show. On what do you base retention for the years ahead? The rest can be checked: the curve on the following quarters, the margin in the accounts.

• Curve validation: on real data, hold the last quarters out and compare forecast and observed acquisitions. The estimated peak, 13.78 quarters, must fall within the data; if it falls outside, the ceiling is not identified ^[C41]. The page does not run this validation on the example: the repetition on 100 retailers stands in for it.
• What the data do not verify: the value of future customers, 30.7% of the total, rests entirely on extending the curve beyond quarter 16; that of current customers, on retention assumed constant for ever ^[C16].
• Estimation uncertainty, by parametric bootstrap: the acquisition noise is estimated at 0.0865, where the true value is 0.10; the interval of the ceiling runs from 1,902 to 2,346 thousand customers; that of the value of the base, from €208.10M to €244.16M. These intervals are at a nominal 95%; over 100 retailers, the interval of the value contains the reference 92 times (simulation error 2.7%): slightly less than the stated 95%, though 100 draws cannot establish it; one possible, partial cause is noise underestimated over sixteen quarters (0.0865 against 0.10 in the detailed draw). The +4.4% gap of the detailed draw is within chance variation, the standard deviation being 3.4% over 100 draws. These intervals assume the S-shaped curve is right and retention known; they cover neither the error on retention, nor that on the margin, nor the choice of discount rate.
• The retention assumption weighs far more than estimation uncertainty: from 70% to 90%, at a 10% discount rate, the value goes from €156.33M to €350.08M. It must be presented as scenarios.
• Comparing with market value is not a validation. The authors find customer value close to stock-market value for Capital One, Ameritrade and E∗Trade ^{[C05, C43, C45, C47]}, but far below it for Amazon, 0.82 billion against 5.36 ^[C44], and for eBay, 1.89 against 15.85 ^[C46]. Over the five firms and four quarters, customer value explains only 14% of the variation in stock-market value, an R² of 0.139 ^[C80], the R² being the share of variation a model explains; without Amazon and eBay, that is, twelve observations, it explains 93%, an R² of 0.927 ^[C49]; with three firms of very different sizes, this R² most likely rests on the gap between Capital One and the two brokers. And the Capital One value on which this regression rests is not recovered here. A gap says that the market is wrong or that the model is missing something ^[C06]: it does not say which.
• What the simulation does not prove: the data are drawn from the model, which is right by construction, with the true retention; the measured gaps come only from fitting the curve. The simulation measures the cost of the shortcuts when this model is true; it does not show that the model suits your firm.

Three lessons. First, a customer base is valued with its future customers: without them, the value of the base is understated here by more than a quarter, that is, the share of future customers, who bring in more than they cost to acquire. Second, the average lifetime inflates the value of a customer: it acts as if each stayed five years, whereas departures start in the first period ^[C18]; on the same calendar, the excess is 13.7% here and 15.4% in the authors' example.

Finally, retention is the most powerful lever: raising retention from 80% to 80.8% adds 3.75% to the value, a 1% better margin adds 1.12%. The authors find an elasticity — the percentage of value gained for a 1% improvement — of 3 to 7 for retention in their abstract, of 2.45 to 6.75 in their Table 4, about 1 for margin and 0.02 to 0.3 for acquisition ^{[C07, C08]}, and an effect of retention all the stronger as the discount rate is low ^[C56]. These elasticities measure a lever: they take past acquisitions as known. When past retention itself is uncertain, the re-estimated grid is the one to read.

• Allowed: "If annual retention is 80% and the S-shaped curve is right, the customer base is worth about €224M of margin discounted at 10%, before tax, fixed costs and retention spending, with a nominal 95% interval of €208.10M to €244.16M, whose measured actual coverage is about 92%; between 70% and 90% retention, €156.33M to €350.08M."
• Allowed: "In this simulation, counting current customers only understates the value of the base by a little more than a quarter, the share of future customers."
• Allowed: "If both improvements cost the same, raising retention from 80% to 80.8% would be worth far more here than cutting the acquisition cost by 1%."
• Forbidden: concluding that the firm should invest in retention rather than acquisition. The elasticities do not count what each improvement costs ^{[C50, C51]}; eliminating all departures is not desirable, according to Shaffer and Zhang, cited by the authors ^[C52].
• Forbidden: presenting the value of the base as the value of the firm, or of its shares, without saying what it leaves out — tax, fixed costs, retention spending booked as acquisition ^[C29], debt, growth options, non-customer assets, network effects ^{[C06, C48]}.
• Forbidden: presenting these figures as a market result; the data are synthetic.

• Present the value of the base as retention scenarios, not as one figure: about €224M at 80% retention, €156M to €350M between 70% and 90%. For a funding round, then move from the value of the base to that of the shares: deduct tax, fixed costs, retention spending and debt. The discount rate remains a choice for the finance department.
• Treat acquisition as an investment: an acquired customer is worth their future margins minus their acquisition cost, and the authors argue for counting acquisition spending as investment, not as expense ^[C14].
• Cost a better retention before buying it. The value shows how much raising retention from 80% to 80.8% brings in; an experiment with a control group shows whether a loyalty programme achieves it. Both together make it possible to compare the levers.
• Monitor the acquisition curve every quarter and re-estimate the value; measure retention by cohort rather than assume it.

The decision remains human and involves general management, marketing and finance.

• Use it: to value a growing firm whose earnings are negative or low ^[C10], to compare marketing levers in terms of value, or to estimate a firm's customer base from the outside from its published reports ^[C13].
• Use it with caution when acquisition has not yet peaked: you then need an external estimate of the market potential or a Bayesian approach ^{[C42, C58]}.
• Do not use it for the value of a specific customer: see the page on estimating noncontractual CLV, the one on subscriber value or the one on the value of a seasonal customer.
• Do not use it to measure retention itself: see the page on retention and survival analysis.
• To express the uncertainty of a result, see the page on expressing uncertainty; to judge the curve on held-out quarters, the one on validating a marketing forecast; for repeated draws, the one on Monte Carlo simulation.

• Python 3.14.5, standard library only: msc-p045-reference.py, run with python msc-p045-reference.py
• R 4.6.1, base only: msc-p045-reference.R, run with Rscript msc-p045-reference.R
• Both programs print exactly the same lines; their sha256 fingerprint, after line-ending normalisation, is recorded in the dossier manifest. Declared seed 20261006. No third-party library: the software does not validate the method, it reproduces it.

Active customers by quarter, the assumed retention and its source, the rebuilt acquisitions, the estimated curve with its peak and ceiling, the value of current customers and that of future customers with its interval, the retention and discount grid, the elasticities, what the value leaves out on the way to firm value, and the assumptions to check before deciding.

• Gupta, S., Lehmann, D. R. & Stuart, J. A. (2004). Valuing Customers. Journal of Marketing Research 41(1), 7–18. ↗ — the authors' accepted manuscript, revised February 2003, posted by Columbia Business School, full text verified ^[C01–C88].
• Level of evidence: demonstration on synthetic data drawn from the model itself, one detailed draw, a bootstrap of 1,000 retailers and 100 repeated draws, reproducible in Python and R; code checked against the values published by the authors for four firms out of five, from their published inputs and a calendar convention chosen to reproduce them. This is not a market result.