Acquisition cohorts by channel

Rows. One row per acquisition channel, or one row per channel and first-order month if you have enough customers to split that finely. Kettle & Co's constructed table pools the whole of Q3 2026 by channel, because six channels times three months makes eighteen rows, several of them tiny.

Columns. Months since the customer's own first order, not calendar months. A customer who first ordered on 3 July and one who first ordered on 28 September both sit in month 1 at their own month 1. That alignment is the point: it lets you compare customers at the same age.

Cells. The cumulative share of the row's customers who have placed a second order by that month. Cumulative, so the numbers only ever rise along a row; a flat stretch is where the leak is. Always print the base (the "new" column) next to the row, because a percentage without its base is half a number.

Reading it. Read across a row for the shape of one channel's return: does it keep climbing, or stall? Read down a column for a fair comparison between channels at the same age. Never compare paid search at month 6 with paid social at month 2: older cohorts have simply had longer to come back.

"The channel that brought them" sounds obvious until someone asks which one. There are two common answers, and they put different customers in different rows.

First touch: the first channel the customer is recorded interacting with, perhaps a creator video weeks before they bought. It needs journey data that survives across devices, consent choices and cookie lifetimes, and it changes whenever tracking does. Channel of the first order: the channel recorded on the first order itself (the referrer, campaign tag or code on that order). It lives in the order table, so it is stable and anyone can rebuild it, but it gives the row to whichever channel closed the sale.

Neither is right. Pick one, write it down in one sentence at the top of the table, and keep it. A cohort definition that changes mid-year rewrites history: last quarter's paid social customers become this quarter's organic search customers and every trend line breaks in a way nobody can see. Which credit model is fairer is its own field, taught in Marketing Attribution ↗. For a cohort table you need consistency far more than you need the perfect model.

The choice of definition is the marketing lead's to agree, because it decides which channel her team is credited with; the data team writes it down, explains what each option does, and holds the line once it is set. Data informs the bet; the business places it.

This page is acquisition cohorts by channel: which channel's new customers come back. Two neighbouring jobs are taught elsewhere, by design. The store funnel and checkout, from product page to payment, belong to E-commerce Metrics ↗ and User Journey Analytics ↗. The craft of the cohort heatmap itself, colour scales, retention against repeat, survival curves, belongs to Customer Retention's cohort table and heatmap session ↗. Use that session for the drawing; use this one for the channel question.

In Kettle & Co's constructed Q3, paid search and paid social both cost 30.0 pounds per new customer. On a first-order dashboard they are twins. On the cohort table they are not: by month 6, 38.0 percent of paid search's customers have ordered again, against 22.0 percent of paid social's. For every hundred new customers, that is sixteen more second orders from search, and in tea, second orders are where most of a customer's value comes from.

So "which channel is cheaper?" is the wrong question. The useful one is "which channel's new customers are worth more?", and the cohort table is the first half of the answer. The second half, putting a cost and a margin on it and finding the month each channel pays back, is Angles session 5.

A data person should also say plainly what the table does not show. It shows that paid social's customers come back less often. It does not show why: the audience the platform delivered, the offer that brought them, or the product they first bought could each be the cause. Naming the leak is the data team's job; deciding what to change in the campaign is the marketing lead's, because she lives with the result.

Where the climb happens. Every Kettle & Co curve does about a third of its six-month climb in month 1: paid search reaches 12.9 of its 38.0, paid social 7.5 of its 22.0. The first month after the first order is where a cohort is won or lost, which is why the welcome sequence and the first replenishment reminder matter so much more than anything sent in month 5.

Where the leak is. A flattening curve is customers who have stopped coming back. Paid social adds only 1.3 points between month 5 and month 6 (20.7 to 22.0); organic search still adds 2.3 (38.7 to 41.0). The leak is not one dramatic drop, it is a steady shortfall at every month.

The bottom row. Affiliate has the cheapest paid CAC in the constructed table (15.0) and the lowest repeat (14.0 percent by month 6). That is the shape you would expect if many of those "affiliates" are coupon sites catching people who were already buying, the ask Kettle & Co's BD lead brought to the triage desk in Partnership session 2. The cohort table raises the question; it does not settle it.

Six months is not a lifetime. These curves are still rising at month 6. A six-month repeat rate is a fair comparison between channels at the same age, not a forecast of where each one ends. Turning curves into lifetime value is a modelling job, taught in Customer Retention's CLV session ↗.

PR brought only 120 new customers in the constructed quarter, and creators 250. In a cohort of 120, one customer is 1 in 120, so a single extra second order moves PR's rate by close to a point. Three customers more or fewer and PR swaps places with a neighbour in the ranking. Creators' base is about twice as big, so each customer moves it by about half as much.

The professional habit: print the base on every row, flag any row under a threshold you agree in advance (300 customers is the flag in tonight's worksheet, a working choice, not a standard), and say "too small to rank" out loud rather than letting a small row top a league table. Pooling several months for small channels, as Kettle & Co's table does with the quarter, is the usual fix, at the cost of seeing month-to-month change.

A Black Friday cohort is not a normal month's cohort. Customers who first bought on the year's deepest discount include many who were shopping for the discount, and gift buyers who were never going to drink the tea themselves. Their repeat curve will usually look worse than a normal month's, and that says more about the month than about the channel.

Compare like with like: a November cohort against last November's, not against October's. Never let a promotional month be the baseline a campaign is judged against. And when one channel's customers are concentrated in the sale month (affiliates and coupon sites often are), say so next to its row, because its rate is partly a calendar effect. Tonight's constructed Q3 has no Black Friday in it, which is one reason it is a clean quarter to learn on.

Kettle & Co's constructed blended 6-month repeat rate across all 2,670 Q3 new customers is 29.0 percent. Now double paid social's new customers, from 800 to 1,600, and keep every channel's own repeat rate exactly where it was. The blended rate across the new total of 3,470 falls to 27.4 percent. No channel got worse. The blend is a weighted average, and more weight went on the 22.0 percent channel.

This is the trap behind many alarming board slides: "repeat rate is falling" when the truth is "we bought more customers from a channel whose customers repeat less". Both statements can be true at once, and they call for different decisions. The first sounds like a retention problem for the CRM team; the second is a channel mix choice for the marketing lead. The habit: never quote a blended rate without the by-channel rows underneath it, and when the blend moves, check whether the rows moved or the mix did.

The same arithmetic runs the other way. A blended rate can rise while every channel gets slightly worse, if the mix moves toward a high-repeat channel. The rows are the evidence; the blend is a summary of the rows and the mix together.