Products and baskets

Revenue grew almost every year, so an absolute stacked chart of categories is five bands all getting taller and a reader who learns nothing. Divide each year by its total. Now the chart answers a different question - not "did we grow" but "did we shift" - and the shifts are the finding: drinkware and home spiked in 2020, bags rose steadily, stationery has faded since 2023. Use complete years only; 2026 is eight months and its share would be a blend of the wrong seasons.

One level down, fifteen lines are too many for a stacked area and exactly right for a slope chart: each line's share of revenue in 2021 on the left, in 2025 on the right, one segment between. Colour only the lines that lost more than a fifth of their share. The eye reads the amber slopes in a second and ignores the rest, which is what a slope chart is for. Two full years four years apart avoids both the 2020 spike and the partial 2026.

Rank every SKU sold by total profit, accumulate, and plot cumulative profit share against cumulative SKU share. The curve bends hard: a small share of SKUs carries 80 percent of profit, and the long tail is hundreds of SKUs that each earn almost nothing. A Pareto curve is a catalogue decision waiting to be made - which SKUs to feature, which to leave to merchants, which to retire - and it is also the reason category totals mislead: a category can look healthy on the strength of three SKUs.

Read the owner table twice. Ninety-eight platform SKUs make just over half the profit: they are the volume. The 1,023 merchant SKUs make just under half at a margin 8.3 points higher: they are the rate. Neither is the "better" half; the platform runs on both.

Customisation is Printwell's thesis: merchants make their own versions of the platform's lines and the platform earns more on them. The data is allowed to say whether the thesis pays. Margin by owner per complete year on the left; share of revenue on merchant SKUs per server on the right, key servers amber. The margin gap holds in every year, so it is a property of the goods, not of one good year.

For the three biggest lines, plot every SKU's list price, platform on the left and merchant on the right. Merchant SKUs sit one rung above platform SKUs on every line: the customised version of a mug costs more than the standard mug, and that rung is the upsell path a store can walk a customer up. Where a line has no merchant SKUs, there is no rung and nothing to upsell to, which is a merchant-success conversation rather than an analytics one. Read the ladder alongside the owner-margin chart: the higher price and the higher margin are the same fact seen from two sides.

Self-join the order-line table on txn_id, keep each unordered pair once with a string comparison, count distinct orders per pair, and divide. That is the whole algorithm at pair level, and doing it by hand once means you will never again read a lift table without knowing exactly what the denominators are. min_support comes first, before ranking: a pair that appears in twelve orders can have a lift of 9 by luck, and it will sit at the top of any table sorted by lift. The 0.002 floor means a pair must appear in at least a fifth of a percent of orders before its lift is allowed to speak.

Read the table with the pair the data planted in mind: the stationery trio co-occurs above chance as two pairs, notebook + sticker pack and sticker pack + pen set, while notebook + pen set sits below 1. Lift is pairwise; a three-item bundle is not the sum of its pairs, and the honest bundle offer here is "sticker pack with either", not a trio.

Lift is the analyst's number. The attach rate is the merchandiser's: given a mug is in the basket, how often is a t-shirt? Draw it next to the base rate of t-shirts in any basket and the gap is the bundle nobody is selling on purpose - customers are already assembling it more often than chance, without a prompt, a discount or a "goes well with" tile. Six pairs, two bars each.

Count distinct product lines per order, then the distribution per year. Most Printwell baskets are a single line, every year. That is the plain statement of the cross-sell headroom: the pair lifts above are computed on the minority of orders with two or more lines, and the majority never got the chance to attach anything. A bundle programme's ceiling is the one-line share, and its progress metric is that share going down. Draw the chart before promising anything about baskets.

Breadth is the exponential of the Shannon entropy of a merchant's revenue shares across lines: the effective number of lines. A merchant selling fifteen lines equally scores 15; one selling almost only mugs scores just above 1. Counting lines with any sale would give both a 15 and tell the success team nothing. The where(shares > 0, 1) guard keeps log(0) out of the sum without dropping rows.

The top five by revenue are all enterprise merchants with breadth between 9 and 12: they already sell across the catalogue and the conversation with them is about depth, not breadth. The file reports/a3_merchant_preference.csv carries all 140, and the ones that matter to a success team are further down: high revenue on one line, breadth under 3, and a pair-lift table that says what their customers buy elsewhere.

A lift of 1.5 is not a business case. Sizing one honestly takes three numbers you already have and one you must decide. The attach gap - attach rate minus base rate - is the share of first-item orders that already carry the second above chance, and the most a prompt can plausibly move is some fraction of the orders that do not. Multiply the orders with the first item by the gap you expect to close, then by the margin of the second item, and you have an annual profit figure for the bundle before any discount. Then subtract the discount on every order that would have bought both anyway; that is the cost of bundling, and it is usually most of the benefit if the discount is generous.

• Size it on the notebook + sticker pack gap of 5.0 points and the 1,692 orders that already hold both, not on the lift, which has no units.
• Decide the target before launch: attach rate on the first item, measured monthly, against a control group of stores that did not get the bundle tile.
• Watch two things after launch: does the attach rate rise, and does the one-line share of baskets in Part 2 fall? If the first moves and the second does not, the bundle is cannibalising two-line baskets rather than creating them.
• Re-run lib.pair_lift on post-launch orders only. A bundle that works shows up as a higher lift on the pair; one that only discounts shows the same lift at a lower margin.