Retail assortment planning AI: what this means
Most retail AI investment goes into replenishment and pricing, because both produce fast measurable wins. Assortment gets less attention and is worth more.
Replenishment optimises the flow of items you have already decided to carry. If the assortment is wrong, excellent replenishment ships the wrong product efficiently.
Why it is avoided
Assortment decisions are slow, political and hard to measure. A range review happens twice a year. The counterfactual — what the other range would have sold — is unobservable. Buyers have relationships with suppliers.
None of this is a modelling problem, which is precisely why software teams avoid it.
Where prediction genuinely helps
Not by choosing the range. By making the trade-off legible: which lines are incremental and which cannibalise, what a delisting does to basket composition, which local variations are real and which are noise.
The buyer still decides. But they decide with the substitution effects visible, which is not how the decision is made today.
What a range review actually decides
Not which products are good. Which products earn their space, given what else is on the shelf and what the shopper would have bought instead.
That second clause is the whole difficulty. A line selling 400 units a week looks strong until you learn that 300 of those shoppers would have bought the adjacent SKU at a higher margin. Delisting it improves the category; the sales report says otherwise.
Every buyer knows this. Almost no system tells them which of their lines are in that position.
Substitution is measurable and rarely measured
The data exists. When a product goes out of stock, some shoppers buy something else and some leave — and the basket data records which. A stockout is an unplanned experiment, and most retailers run thousands of them a year without reading the results.
Estimating substitution from stockout events is unglamorous and produces a matrix a buyer can argue with: this line cannibalises that one at 60%, this other line is genuinely incremental.
That matrix changes range reviews more than any forecast does.
Local variation: real or noise
Central buying assumes one range with regional exceptions. Store managers believe their store is different. Both are partly right, and the argument is usually settled by seniority rather than evidence.
The test is whether a store’s deviation persists across seasons after controlling for space and footfall. A line that outperforms in eleven stores for three consecutive quarters is a pattern. One that outperformed last quarter is a coin flip that landed.
Making this distinction visible removes a recurring argument from the process, which is worth something on its own.
Why the buyer still decides
Supplier relationships carry terms that are not in any sales report. A line that loses on margin may be carrying a rebate that pays for the category. A supplier being delisted may be the only source for something else.
A system that recommends a range without those facts will be overridden, correctly. A system that shows substitution, incrementality and local evidence, and leaves the decision to the person who knows the terms, is used.
The measurable outcome
Category margin per square metre, compared across reviews. Slow, but it is the number the decision exists to move — and unlike replenishment metrics, it cannot be improved by shipping the wrong product more efficiently.
Where to start if the range has never been examined this way
Take one category, pull twelve months of stockout events, and estimate substitution from the basket data around them. It is a fortnight of work and it produces a matrix the buyer for that category can immediately confirm or dispute.
Their reaction is the signal. If it matches what they already believed, the method is validated cheaply. If it contradicts them, the argument is now about evidence rather than instinct — and that is a better argument to be having.