What market basket analysis is
Market basket analysis finds items that appear together in the same transaction more often than chance would explain. It works on the receipt, not the shopper: no identity, no loyalty card, no history. Just the contents of one basket, repeated across every basket a store rings.
The output is a set of rules of the form "shoppers who buy A also buy B", each carrying three numbers that decide whether the rule is worth anything.
Support, confidence and lift
These three get confused constantly, and using the wrong one produces confident bad merchandising.
Support is how often the combination appears at all, as a share of every basket. It tells you whether the rule is big enough to matter.
Confidence is how often B appears given A appeared. This is the same calculation as an attach rate.
Lift is confidence divided by B's overall rate. It tells you whether A actually predicts B, or whether B is just popular. Lift above 1 means a real association; lift at or below 1 means the pairing is coincidence dressed up as insight.
Sunrise Market, an illustrative single-site operator, over 3,892 weekly baskets:
| Pair | Support | Confidence | B's base rate | Lift |
|---|---|---|---|---|
| Coffee and sandwich | 7.0% | 22% | 9% | 2.4 |
| Coffee and energy drink | 1.0% | 3% | 10% | 0.3 |
| Candy and packaged beverage | 11.0% | 35% | 42% | 0.8 |
| Sandwich and chips | 3.0% | 34% | 15% | 2.3 |
The second and third rows are the useful ones, because they are the rules a confidence-only reading would get wrong. Coffee and energy drink shows 3% confidence, which looks weak and is: at a 0.3 lift these are substitutes, and a shopper buying one is markedly less likely to buy the other. Meanwhile the row that looks strongest by raw co-occurrence, candy and packaged beverage at 11% support, carries a 0.8 lift because packaged beverages appear in 42% of baskets anyway. Pairing on support alone would put those two together and gain nothing.
Rows one and four are genuine. Sandwich and chips at 34% confidence and 2.3 lift is a real affinity worth merchandising against.
Where basket analysis misleads
Popular items pair with everything. The highest-support rules in any store involve its best sellers, which is why lift exists. Ranking rules by support produces a list of the store's top items paired with each other.
Correlation is not a merchandising instruction. Two items co-occurring because they serve one occasion will not necessarily co-occur more if you move them next to each other. Sometimes proximity helps; sometimes the shopper was always going to buy both and you have given up a display slot for nothing. The rule identifies a candidate, and a test settles it.
Baskets are not shoppers. One person's four visits are four baskets. Any conclusion phrased as "our customers prefer" is an overreach from data that only knows transactions.
The daypart is inside the average. A pair that shows 2.4 lift across the week may be 4.1 in the morning and 0.9 in the afternoon. Basket rules computed on a weekly file describe an average occasion that does not exist.
Why convenience baskets are different
Small baskets change the statistics. A grocery basket with 24 items generates 276 possible pairs; a convenience basket with two items generates one. That sparsity means convenience basket analysis works on fewer, stronger signals, and the rules that emerge tend to be occasion-driven rather than category-driven. Coffee and a breakfast sandwich is not an affinity between two products, it is one purchase decision that happens to ring as two lines.
The practical consequence is that convenience basket rules should be read as occasion evidence. When a rule breaks, an occasion is moving, and that is a more actionable finding than a change in any single item's velocity. It is also how share of wallet is estimated in a channel with no panel to lean on.
The short version
- Market basket analysis finds items co-occurring in one transaction, using support, confidence and lift.
- Lift is the one that matters: it separates a real association from an item that is simply popular. Lift at or below 1 means no association.
- Rules are candidates for a test, not instructions, and baskets are transactions rather than people.
- Convenience baskets are small and occasion-driven, so rules there describe occasions rather than product affinities.