Segment and value: who deserves the next dollar

What goes on top. The narrowest numerator is media spend: what the ad platform invoiced. A fully loaded numerator adds agency and creator fees, tools, and the people who run the channel. Neither is wrong; they answer different questions. Media-only CAC tells the marketing lead what the auction charges her. Fully loaded CAC tells the finance partner what a customer really costs the business. Organic search shows why the choice matters: it has no media bill at all, so in the constructed table its "spend" is the 4,000 of content cost that produced the pages people found. Leave that out and organic CAC is zero, which flatters it into meaninglessness; put it in and organic CAC is 8.0.

What counts underneath. A "new customer" should mean a first order ever, deduplicated: the same person reached by paid social on Monday and a creator on Wednesday is one new customer, not two. If each channel counts its own "new" from its own reporting, the six denominators add up to more people than actually arrived, and every CAC looks cheaper than it is. The company's own customer table, not the platforms, is where "first order ever" is decided.

So a CAC on a slide should never travel alone. Write it as "30.0 (media only, first-ever orders, deduplicated)" and the next person who compares it with a number from finance will know whether they are comparing like with like.

Blended CAC is easy to compute, stable from quarter to quarter, and it fits on one slide. In the constructed Q3 table it is 64,000 over 2,670 new customers, 24.0. It is a fair number for one question: what did acquiring customers cost the business overall? It is the wrong number for every channel question, because it mixes a channel that costs 8.0 with one that costs 36.0. Worse, it moves when the mix moves even if no channel changes: shift money towards the expensive channels and blended CAC rises with no channel getting worse, the same mix effect Angles session 4 found in blended repeat. Report blended for the overall question, by channel for every allocation question, and say which one you are showing.

The constructed model is deliberately simple. Every new customer's first order earns 14 of gross margin. Each customer who orders again within six months adds 48 more over that period, spread evenly. So a channel's six-month margin per customer is 14 plus its repeat rate times 48: paid search, with 38 percent repeat, reaches 32.24; paid social, with 22 percent, reaches 24.56.

Organic search pays back at month 0: its CAC of 8.0 is covered by the first order alone. Affiliate pays back at month 1, CAC 15.0 against a first-order margin of 14. Paid search and PR pay back at month 6, just inside the window. Paid social and creators do not pay back within six months at all: paid social is still 5.44 short of its CAC at month 6, creators 7.12 short.

Notice what this does to the part 1 picture. Paid search and paid social had the same CAC, 30.0. On payback they are on opposite sides of the line, and the whole difference is the repeat rate Angles session 4 drew as a cohort curve. Cost tells you half the story; margin over time tells you the other half.

"Not within 6" is not "never". With a longer window, more repeat orders could close the gap. Whether to judge channels on six months, twelve or a lifetime is a finance and marketing choice about how long the business can wait for its money; the data team's job is to show the answer at each window, not to pick one quietly.

The six-month margin per customer above is a crude customer value: one window, one average repeat rate, one margin. A proper customer lifetime value (CLV) models how many more orders each customer is likely to place and when, and it has its own methods, assumptions and failure modes. Computing it is taught in Customer Retention, the CLV session ↗.

For allocation work, treat CLV the way you treat a supplier's price: an input with a source and a date. Write down which model produced it, over what horizon, on margin or on revenue, and from which cohort. Then the payback table can use it, and anyone who disagrees with the CLV knows exactly which input to argue with, rather than arguing with the whole recommendation.

Affiliate looks excellent on payback: month 1, the fastest of the paid channels. But its six-month repeat is 14 percent, the lowest of any channel, so its six-month margin per customer, 20.72, is the lowest too. It pays back quickly because it is cheap, not because the customers are valuable.

Angles session 3 drew the mechanism that produces exactly this pairing: coupon sites intercepting buyers who were already at checkout, collecting the last click and the commission for an order that was going to happen anyway. A cheap CAC with a weak repeat rate is the signature to check for. If many of affiliate's "new" customers were already on their way, its true CAC is higher than 15.0, because the denominator counts people the channel did not bring. The split of affiliate orders into new and returning customers, and what each did next, is the evidence; what to do with the coupon partners is the BD lead's call.

Average CAC is total spend over total new customers: for paid social in the constructed Q3, 24,000 over 800, 30.0. Marginal CAC is the extra spend over the extra customers it brings. In the constructed response curve, the next 6,000 on paid social buys 120 new customers: 6,000 over 120, 50.0. The next customer costs two thirds more than the average one.

Put that next to Part 2. At an average CAC of 30.0, paid social was already 5.44 short of paying back by month 6. If the extra 120 customers behaved like the current ones, which is itself an assumption, each would bring about 24.56 of six-month margin against a cost of 50.0. The marketing lead's plan to grow paid social is not obviously wrong: it may be buying reach for the slow brand clock of Angles session 1, or the longer window may close the gap. But "it has the same CAC as paid search" is no longer a reason.

Where does a response curve come from? In practice, from the company's own history of spend changes, from deliberate spend tests, or from a marketing mix model. Each is an estimate with uncertainty, and a curve fitted to a narrow range of past spend says little about spend far outside it. Mix models and spend tests at depth are taught in Marketing Attribution ↗.

It is tempting, with a table this clear, to write "move 6,000 from paid social to organic search content" at the bottom and send it. Resist it. The marketing lead knows things the table does not: a product launch that needs reach next quarter, a creator contract already signed, a brand campaign whose payoff will not show in six months. She will also live with the outcome, in her targets and in her next budget round. Whoever lives with the outcome owns the decision. The data team's deliverable is the evidence, laid out as options with their numbers and their uncertainty, by a named date. The choice stays with her.

Channel is only one way to cut "who deserves the next dollar". The same arithmetic, CAC with its base, margin-based payback, average against marginal, applies to any segment that costs something to acquire and earns something afterwards. Two that matter at Kettle & Co are worth working through on your own data:

Business accounts against consumers. Office-tea accounts, the BD lead's business customers, are acquired through trials, samples and sales time rather than ads, and they reorder on a different rhythm. Their CAC numerator is mostly people's time, which is easy to leave out; their repeat pattern looks nothing like a consumer's. Mixing them into consumer CAC distorts both. The BD lead's ask about office-tea trials not converting is ask 24 on the Partnership session 2 triage desk.

Subscribers against one-off buyers. A subscriber's repeat orders are close to scheduled; a one-off buyer's are not. A channel that brings a higher share of subscribers will look better on six-month margin for a structural reason, and the honest comparison splits the channel's customers by what they signed up for.

No constructed numbers are given for these cuts on purpose. Build them from your own customer table and see whether the ranking of your channels survives the split.