Affinity answers what goes with what. Which products are bought together, which services are taken together, which combinations recur.
It is the most specialised engine here and the one with the narrowest correct use.
What it does
It looks across transactions for combinations that occur together more than chance would produce, and reports the ones that stand out.
The output is pairs and groups: this and that, this often followed by that.
What you set
The record type and window, and a scope narrowing which transactions count.
The window matters more here than anywhere else in Analytics, and it is covered below.
It is not a recommendation engine
Worth separating, because the words sound alike.
This tells you what has already happened together, across your history. It does not predict what a particular customer will want next, and it does not personalise.
That makes it a merchandising and bundling tool rather than a targeting one, and it is genuinely good at the first job.
What counts as a transaction
The unit matters and it is worth checking before reading anything.
An affinity looks at what appears together within one record: one order, one invoice, one booking. If your business records a customer's monthly purchases as separate transactions, things that genuinely go together will never appear in the same basket and the analysis will find nothing.
That is the most common reason this engine returns a disappointing result, and it is a modelling question rather than a chart problem.
Where it is genuinely useful
Bundling. Two things bought together by a meaningful share of customers are a bundle you have not made yet.
Placement. What to suggest alongside something, based on what people actually do rather than on what seems related.
Noticing a pattern nobody named. Occasionally a combination recurs for a reason nobody in the business has articulated, and that is worth a conversation.
Reading a pair honestly
The question to ask of every result is: would I have predicted this?
Yes, and it is strong. Useful confirmation, and now you can size it.
No, and it is strong. The interesting case, and the one to verify before acting.
No, and it is weak. Noise, however neat the chart looks.
The trap is treating the second and third the same because both were surprising.
Where it misleads
Four failures, and all four are common.
Popular things pair with everything. Your bestseller appears alongside every other product, because it appears in most baskets. That is not affinity, it is popularity, and it is the single most common misreading.
Small numbers look strong. A combination occurring three times can produce an impressive-looking relationship. Check how many transactions are behind a pair before acting on it.
Seasonal pairs are not permanent. Two things bought together in December may have nothing in common in March.
It cannot see intent. Two items in one transaction may be for two different people, two different purposes, or one of them may have been an accident.
Scope it before you read it
An affinity over everything is almost always uninteresting, because the strongest pairs will be whatever you sell most of.
Narrowing it is what makes it useful: one category, one customer segment, one channel. A pattern that holds within a segment and not across the business is frequently the actionable one, because it describes a group you can actually address.
Why the window dominates here
More than in any other engine, the answer moves with the window, because a basket analysis is counting co-occurrence and co-occurrence is seasonal.
The practical test: run it over two adjacent windows of the same length. Pairs that appear in both are worth attention; pairs that appear in one are the season, the promotion, or chance.
Choosing the window
Long enough to have volume, short enough to describe now.
A quarter suits most businesses.
A year flattens seasonality into an average that describes no actual month.
A month in a low-volume business produces relationships from a handful of transactions.
If the answer changes completely when you change the window, you do not have a finding, you have noise.
Saving one
It saves as a first-class widget and can go on a board.
Most affinities should not. This is an investigative surface: you ask, you learn something, you act on it, and the finding stops being news. A board carrying a basket analysis from eight months ago is presenting a conclusion as if it were current.
The exception is a business where the assortment genuinely changes weekly, where it is a standing operational question.
Reading the output
Two numbers travel with every pair and they answer different questions.
How often the combination occurs tells you whether it is worth acting on commercially. A perfect relationship across four transactions is not a bundle.
How much more often than chance tells you whether it is a relationship at all rather than two popular things meeting.
A pair needs both. High frequency with no lift is popularity; high lift with no frequency is coincidence.
Before acting on a finding
Check the volume behind it. Not the strength, the count.
Check it holds in a different window. A real pattern survives being asked again.
Check it is not just popularity. If one side of the pair is your bestseller, be sceptical.
Then test it small. A bundle is cheap to trial and expensive to roll out on the strength of a chart.
Acting on one properly
A finding here is a hypothesis about behaviour, and behaviour is cheap to test.
Bundle two things and watch for a month. If the pairing was real, the bundle sells; if it was an artefact, you have lost a month and learned something.
Change one thing at a time. Two bundles launched together tell you nothing about either.
Re-run the affinity afterwards. A successful bundle changes the baskets, which changes the analysis. That is not a problem, it is the loop working.
When to use something else
A pivot for straightforward counts of what sells with which segment.
A correlation for two measures over time rather than items within transactions.
A report for the actual transactions, which is frequently what somebody wants once they have seen a surprising pair.
Did this answer your question?
No, ask a person