Which product to offer next: a market basket analysis

Recommendation. We recommend making the six offers that held, each to the customers who have bought its main product and not yet the offer, and holding back the other two, the offers to customers who bought filter beans and ground coffee, until a test shows they add sales. Leave a random half of each offer’s customers without it, so the sales the offers add can be measured. The ranking reads only how strongly two products go together in past orders and leaves out what an offer is worth. The recommendation holds for the products and prices of the year of orders; a new range or a price change calls for mining the rules again.

Support, confidence and lift

The analysis reads each order as a basket: which products are in it, whatever the quantities. For two products it counts the orders holding the first, the orders holding both and all orders, and three measures come from those counts. Here they are for the pour-over dripper and paper filters, over the 109,172 orders of October 2025 to June 2026:

• Support is the share of all orders that hold both: 4,476 of 109,172, 4.1%.
• Confidence is the share of the orders holding the first product that also hold the second: 4,476 of the 8,403 dripper orders hold filters, 53.3%.
• Lift is the confidence over the second product’s share of all orders. Filters are in 13,721 orders, 12.6%, so the lift is 53.3% ÷ 12.6% = 4.2: an order with a dripper holds filters 4.2 times as often as the average order.

Confidence alone favours whatever sells most. Ground coffee is the shop’s best seller, in 33.3% of all orders, so it turns up beside everything: in 33.2% to 39.8% of the orders holding any other product. A travel mug order holds ground coffee 37.2% of the time, more often than an espresso machine order holds descaling tablets, 27.7%. The lifts tell the two apart: 1.1 for the travel mug and ground coffee, 11.3 for the machine and the tablets. A lift of 1 means the first product says nothing about the second, and above 1 the pair shares an order more often than chance.

The rules

The analysis keeps a rule when its pair passes two filters: the pair is in at least 0.5% of the orders, 546 of 109,172, and its lift is above 1. The first keeps out pairs seen in a handful of orders, whose lift is mostly chance. The 20 products make 190 pairs; 109 are in at least 546 orders, and 102 of those have a lift above 1.

Most pairs pass the second filter. An order holding any one product holds more products than the average order, between 2.8 and 4.0 against 2.1, so it has more chances to hold any other. A lift just above 1 says little, and the pairs worth acting on sit at the top of the list.

The lift is the same both ways. The confidence differs whenever the two products are in different numbers of orders, because it divides by the first product’s orders: 53.3% of dripper orders hold filters, and 32.6% of filter orders hold a dripper.

The next offer for each main product

The shop’s eight main products are those that brought in the most revenue over the nine months, 76.4% of it between them. For each, the next offer is its kept partner with the highest lift.

We recommend offering a lasting product, such as a milk jug, only to customers who do not already own one, and a refill, such as filters or beans, each time.

Did the offers hold on months the analysis never saw?

The rules were mined on October 2025 to June 2026. The test reads the three months after, July to September 2026. For each offer it takes the customers who ordered in the nine months, ordered again in the three, and had not bought the offer by the end of the nine, and asks whether those who had bought the main product went on to buy the offer more often than all of them. It counts only customers who ordered again, 14,233 of the 38,173 who ordered in the nine months: a customer who stopped ordering could not buy an offer, and counting them would let a group that simply orders more often pass for one that wants the offer more.

Six of the eight held: the customers who had bought the main product bought its offer more often than all customers by more than chance explains, at two standard deviations. The other two, the offers to customers who bought filter beans and ground coffee, did not: 7.8% against 8.3%, and 10.7% against 10.1%, within what chance explains. They are the offers for the two products in the most orders, which turn up beside everything and so say little about what else their buyers want.

What the test does not show

The shop made no offers in the year, so these customers bought the offer unprompted. The gap shows who is likely to buy, not that an offer adds a sale. Some of the 25.6% of dripper buyers who bought filters would have bought them anyway, and an offer to them, at a discount, gives margin away.

A test is how you would know: offer paper filters to a random half of the customers who buy a dripper without them, not to the other half, and compare what the two halves buy. Seeing the offer raise the 25.6% by 5 points takes 1,264 customers in each half, at 95% confidence and in 8 tests of 10. The shop takes about 436 dripper orders without filters a month, so the two halves fill in about 6 months. Tested together, a random half of each offer’s customers against the other, the six offers reach about 5,000 customers a month, and the test answers the question the shop acts on: whether the offers, together, add sales. How large a test must be is what our price-test calculator works out; an offer test is the same arithmetic, two groups and one share each.

Not covered

• Rules with two products on the first side, such as a dripper and a kettle together. Three products share fewer orders than any two of them, so fewer such rules pass the support filter.
• What each offer is worth, its margin after any discount. A fuller next-best offer weighs the chance of taking each offer by its value, as set out on customer lifetime value.
• Where and when an offer is made: in the order confirmation, on the product page, or by email later.
• Products with no orders yet, which have no rules until they sell.

Checks and sign-off

Sign-off. Putting offers live is a decision you act on, so it carries the analytics director’s sign-off. Mining the rules again on new orders is a routine refresh: it carries its checks and no new signature.

This analysis is an example on an invented shop and is never signed, so its stamp stays at “Sign-off pending”.

What it reads. Order lines with a customer, a date and a product; a product list with ranges and prices; and, where you have them, past offers and who took them. Where the data goes is on data handling.

How it is built. AI agents rebuild your order history, mine the rules, run the test on the latest months and draft the analysis. Number and claim checks run as scripts, not AI. Before the offers go live, a review panel of three reviewer agents and a judge reads the work, and the analytics director signs it. How the checks work is on how it works.

What arrives. An analysis like this one: a memo of up to two pages, with its tables as an appendix. Then the offer list for each customer, as a file or a recurring report, and the method note listing the checks that ran. The analysis and its offer list are one sprint, from €12k, fixed before work starts; mining the rules again on new orders runs on an async retainer, from €8k a month.

Not covered. Sending the offers and running the test, which stay with you, and installing anything on your systems. The analysis says which offer to make and how to measure what it adds.