Channel CAC Allocator

Ask a marketing team what it costs to win a customer and you will usually get one number. Ask how that number was made and the room gets quieter. The cost of acquiring a customer is simple to define and surprisingly easy to compute wrongly, and the most useful version of it is often not the one on the dashboard.

The textbook definition is plain. Customer acquisition cost is the cost of persuading a customer to buy, and the simple method divides the marketing cost of acquiring new customers by the number of customers acquired in a defined period. A fuller method adds the wages, software, outside services and overhead that sit behind the campaigns. Both are ratios of a total cost to a total count. That small detail turns out to matter more than any other.

Blended, paid and the customers nobody can place

Two versions of the ratio are quoted most often, and they answer different questions. Blended CAC puts everything on both sides: every channel’s spend plus the shared cost that is not tied to any channel, over every new customer, including the ones who arrived with no traceable source. Paid CAC narrows both sides to the channels you pay for and the customers credited to them. Neither is the same as cost per action, an advertising measure and pricing model for a specified action such as a sale, a click or a sign-up, which is sometimes misread as cost per acquisition.

Neither is the true one. Blended CAC tells you what growth cost the business as a whole, so it is the number that belongs next to cash planning. Paid CAC tells you what the paid programme costs per customer it is credited with, so it is the number a media budget is judged on. The gap between them is mostly the organic share: customers who came from word of mouth, search results nobody bought, or a referral the tracking missed.

That gap depends on attribution, which is the assignment of credit for a conversion to the touchpoints that preceded it. Single-touch models give all the credit to one event, such as the last click; fractional models split it by rule. Studies comparing rules-based attribution with controlled experiments have repeatedly found the two diverge, because attribution is built on observed correlation rather than on deliberately changing anything. Every channel figure below inherits whatever bias your attribution carries. No arithmetic removes it.

A worked month

Here is one month with two paid channels. Search ads spent 6,000 and are credited with 60 new customers. Social ads spent 3,000 and are credited with 20. A further 20 customers arrived with no attributed source, and 2,000 went on shared acquisition cost such as tools and part of a salary.

Channel CAC is easy: 6,000 over 60 is 100 for search, and 3,000 over 20 is 150 for social. Paid CAC is the paid spend over the paid customers, 9,000 over 80, which is 112.5. Blended CAC is all spend over all customers: 11,000 over 100, which is 110. Blended is lower than paid here because the 20 unattributed customers cost nothing directly, and they outweigh the shared cost.

Now the common mistake. Average the two channel CACs and you get 125. Put 110 vs 125 side by side and the error is plain: the mean treats a channel that brought 20 customers as if it brought as many as one that brought 60. Ratios must be built as totals over totals. A cost per customer is only meaningful when every customer counts once and every unit of spend counts once, which is exactly what the plain average breaks.

The fix is boring and absolute: add up all the spend, add up all the customers, and divide once, at the very end.

Fully loaded channel CAC spreads the shared cost across channels. Split by spend share, search carries two thirds of the 2,000, so its loaded cost is about 7,333.33 over 60, or 122.22, and social’s is about 3,666.67 over 20, or 183.33. Split by customer share instead and the numbers move again. Neither split is a fact about the world; it is a policy, and the page you use should say which one it applied.

To reproduce this, type the two channel rows, one unattributed row and one shared-cost row for a single period into the Channel CAC Allocator. You should see blended CAC 110, paid CAC 112.5 and channel CACs of 100 and 150.

What a customer is worth, and the ratio everyone quotes

Cost alone says little until it is set against value. Customer lifetime value is an estimate of the profit a customer brings over the whole relationship, often the present value of projected cash flows, and it is described as an upper limit on what it makes sense to spend acquiring a customer. Plenty of shortcuts exist. One multiplies monthly revenue by a fixed number of months, with no margin, no churn and no discounting, which overstates value and should be labelled as the rough count it is.

Suppose each customer is worth 360. Search then returns 3.6 times its CAC, social 2.4 times, and the blended figure is 36,000 over 11,000, about 3.27. A widely repeated guide treats 3 to 1 as a healthy level and anything under 1 to 1 as a business paying more for customers than they are worth. Treat it as a rule of thumb. The right floor depends on margin, on how fast the money comes back and on what else the cash could do.

Average cost is history; marginal cost is the decision

Here is the question that usually matters: if you spend a little more, what do the next customers cost? Economists call that marginal cost, the change in total cost that comes with a change in output. It is a different quantity from average cost, and in acquisition the two drift apart for a familiar reason.

Diminishing returns describes the fall in extra output as one input is increased while everything else is held equal. Advertising behaves this way: the most responsive audiences are reached first, and each further unit of spend finds people who are slightly harder to persuade. Marketing mix models account for this directly, along with the carry-over of past advertising, which is why they are used to compare spending plans.

You can see it in your own records without a model. Say search spent 4,000 last month and won 48 customers. The step from 4,000 to 6,000 added 2,000 of spend and 12 customers, an incremental CAC of 166.67. Search’s average of 100 looks excellent, yet the last customers cost two thirds more than the average. At a value of 360 and a 3 to 1 floor, the most you should pay for one more customer is 120. Search passes on the average and fails at the margin.

Steps between two months are noisy. A dip in customers after a rise in spend is not a negative cost; it is noise, a holiday, a competitor or an attribution change, and it should be flagged rather than printed. A smoother view fits a curve through every month at once. A power law, customers equal to k times spend raised to a power b, is a common shape for this, and taking logarithms turns it into a straight line.

That line can be fitted by least squares, which picks the parameters that minimise the sum of squared residuals. The coefficient of determination then reports how much of the variation the line explains, on the log scale where it was fitted. When b is below 1 the curve bends over: average CAC rises with spend, and the marginal CAC is the average divided by b. With b at 0.6, a channel whose average is 100 has a marginal cost near 166.67.

Once each channel has a curve, the budget question has a clean answer. A unit of spend should go wherever it buys the most customers, so the best split is reached when the next customer costs the same in every channel that is free to move. This is the condition the method of Lagrange multipliers gives for maximising a total under a fixed budget, with the multiplier playing the part of the common marginal rate.

In practice each channel is held inside the range of spending it has actually seen. The common marginal cost is found by bisection, which halves an interval known to contain the answer until it is as narrow as needed. A channel whose curve shows no bend, with b of 1 or more, has no observed limit, so it is held at its highest observed spend rather than extrapolated. Any range beyond what you have seen deserves a visible label.

What the curve cannot tell you

Fitting your own history has real limits. A curve describes the months you had, under the prices, creative and competition you had. Spend that never varied cannot reveal its curve, and spend that always rose with the season will credit the channel with demand it did not create. The split is what the curves imply about your past, not a forecast of next month.

Use the numbers as a careful argument rather than an oracle. Quote blended and paid CAC together, with their definitions. Keep your raw table in a plain CSV so anyone can recompute it. Report the average and the margin side by side. Then test the biggest moves with a deliberate change and a holdout, because only an experiment can settle what an observational fit can merely suggest.