What the technique actually does
Market basket analysis looks across transactions and finds combinations that occur together more than chance would suggest. The output is a set of rules: customers who buy A also tend to buy B.
It is genuinely useful for cross-sell prompts, bundle design, range decisions and store layout. It is also the source of a great deal of confident nonsense, because the raw output rewards popularity rather than insight.
Three numbers, three different questions
| Measure | Question it answers | Trap |
|---|---|---|
| Support | How often does this combination occur at all? | High support favours common items regardless of relationship |
| Confidence | Given A, how often does B appear? | High if B is simply popular, with no real link |
| Lift | Does A make B more likely than baseline? | Can be high on tiny, unstable samples |
Illustrative arithmetic: if 60% of all baskets contain milk, then a rule 'bread implies milk' with 62% confidence looks strong but has a lift of about 1.03 - essentially no relationship. Confidence alone would have promoted it.
Use lift to find real association and support to check the rule occurs often enough to be worth acting on. A rule with excellent lift appearing in a handful of baskets is noise.
Association is not a reason to act
Even a statistically sound rule may be commercially useless. Products bought together because customers were always going to buy both do not need a bundle - discounting them gives away margin on a sale you had.
The commercially interesting rules are the ones where a prompt changes behaviour. That is a different question from co-occurrence, and the only reliable way to answer it is to try the change on part of the traffic and compare.
- Obvious pairs - razor and blades. Already happening; little to gain.
- Seasonal artefacts - two items sold together in December only. Check the rule across time before acting.
- Promotion echoes - the pair was on a shared offer. The rule describes your past marketing, not customer preference.
- Genuinely surprising, repeated across periods - worth a controlled test.
Practical uses that tend to pay
In our experience the reliable wins are less glamorous than the case studies suggest: identifying products that should not be out of stock together, spotting range gaps where a complement is missing, and informing which item to show in a basket prompt rather than choosing it by margin alone.
Store and site layout changes are harder to evaluate and slower to reverse, so they warrant a stronger evidence bar than a recommendation widget that can be switched off in an afternoon.
A rule that says customers buy two popular things together has told you they are popular.