Two bad extremes
One range everywhere is simple to buy and operate, and guarantees every store carries products its customers do not want while missing ones they do.
A bespoke range per store captures local demand and is unmanageable - buying, allocation, planograms and replenishment all multiply, and the data supporting each decision gets thin.
The workable answer sits between: a small number of store groups, each with a range, plus limited local flexibility on a defined share of space.
Cluster on behaviour, not geography
Stores are usually grouped by region or size because those are the obvious attributes. Neither necessarily predicts what sells.
- Category mix - what share of sales each category represents
- Price architecture - where the store sells in the range
- Basket composition and average basket size
- Demand shape over the week and year
- Local demographics and competition, where you have that data
Clustering on these frequently produces groups that cut across regions - a city-centre store may behave like another city-centre store two hundred miles away far more than like its neighbour in a retail park.
How many clusters
More clusters capture more local variation and cost more to operate. The right number is a business constraint rather than a statistical one.
| Clusters | Benefit | Cost |
|---|---|---|
| 1 | Simplest operation | Poor local fit everywhere |
| 3-5 | Most of the available gain | Manageable for buying and allocation |
| 10+ | Diminishing additional gain | Buying and planogram complexity rises sharply |
| Per store | Best theoretical fit | Unworkable in practice |
Somewhere in the middle is almost always right. Ask the buying and space teams how many ranges they can genuinely maintain, and work within that.
The problem of what you never stocked
Sales data tells you what sold where it was available. It cannot tell you what would have sold in a store that never carried it - and that is exactly the range decision you are trying to make.
This is a real limitation. The partial answer is to infer from similar stores that did carry it, and to run deliberate trials placing products in stores that have not stocked them. Without some trial activity, the range gradually calcifies around what it already is.
Space, not just presence
Ranging decides what is carried; space decides how much. A product with the same range status in two stores can perform very differently depending on facings and position.
Modelling sales without accounting for space attributes it incorrectly - a product given twice the space will sell more and look like stronger demand. If space data exists, use it; if not, be cautious about conclusions drawn from raw sales comparisons.
Sales data shows what sold where you stocked it. It is silent about everywhere you did not.