Store clusters as a planning unit
A network of thousands of independent stores cannot be planned store by store and should not be planned as one undifferentiated estate. Clustering — grouping stores that behave alike — is the practical unit, and the quality of the clustering determines how much of the channel's targeting value is actually usable.
What should clusters be built from?
Behaviour rather than geography alone. Traffic profile, basket composition, category performance and daypart shape describe how a store actually trades, and stores that share those characteristics respond similarly to the same campaign.
Geography is a convenient grouping and a weak one. Two stores a street apart can serve entirely different trades.
Why not cluster by demographics?
Because the store's trade is a better predictor of category performance than the surrounding area's profile. A store's own transaction pattern is direct evidence; area demographics are an inference about who might shop there.
Where transaction data exists, it should dominate the clustering. Where it does not, area data is the fallback and should be labelled as such.
How many clusters are useful?
Few enough to plan against and many enough to be different. A handful is usually the practical range, because each cluster may need its own creative or schedule and the production cost scales.
Clusters that cannot be acted on differently are a segmentation exercise rather than a planning tool.
What does a cluster enable?
Different store lists, different dayparting, different creative, and different measurement. All four are the channel's targeting proposition, and none of them work without a grouping to apply them to.
It also enables honest reporting: results by cluster show where a campaign worked rather than averaging it away.
How stable are clusters?
Reasonably, and they should be refreshed periodically. A store's trade changes with its neighbourhood, and a clustering built three years ago will have stores in the wrong group.
Refresh annually, and after any structural change in the estate.
What should a buyer ask?
How clusters are built, from what data, how often they are refreshed, and whether activation and reporting can follow them. The last is the practical test: a clustering that cannot be activated against is a description rather than a capability.
How does this relate to a trade area?
A trade area is the geography a store draws from; a cluster is a group of stores that behave alike. They are complementary — trade area informs who might be reached, and cluster describes how the store trades.
Frequently asked questions
Can I define my own clusters?
If the network provides store-level data, yes, and a brand's own clustering can reflect its categories better than a general one. The constraint is whether activation can follow your definition.
What data is needed?
Store-level transaction data over enough time to characterise the store, plus whatever attributes matter to your category. Traffic alone produces clusters by size rather than by behaviour.
How small can a cluster be?
Large enough to measure, which is the binding constraint. A cluster too small for a store-level test cannot have its result read separately.
Should clusters be named descriptively?
Yes, and cautiously. Names that describe the trade are useful; names that describe assumed shopper demographics assert something the data may not support.
Does clustering help with creative?
It is what makes localised creative practical, because it reduces the number of variants from thousands to a handful.
How do I check a cluster is real?
Look at whether the stores in it actually behave alike on the measure you care about. A cluster whose members vary widely on your category is not a cluster for your purposes.
Can clusters change during a flight?
They should not, because changing the grouping mid-flight makes the reporting incoherent. Refresh between flights.
Does this apply to a small buy?
Less so. A buy across a small number of stores can be planned individually, and clustering earns its keep at scale.
What is the first cluster split worth making?
By daypart profile, because it is the split that most changes what creative should run. Channel context is collected by NACS.
### How should clusters be validated?
By checking whether stores within a cluster actually behave alike on the measure that matters to the campaign. A cluster built on general trading characteristics may not be homogeneous for a specific category.
Where it is not, a category-specific clustering is worth building, because the general one will average away exactly the differences the plan was trying to exploit.
Can clusters be shared with the network?
They usually have to be, since activation follows them. Agreeing the cluster definitions with the network before flight is what makes reporting by cluster possible afterwards.
### How should clusters be reported on?
Delivery and effect by cluster, every flight. That is where the channel's targeting value becomes visible, and an aggregate report discards it entirely.
It also builds a picture over time of which clusters respond, which is the input to the next flight's store selection and the compounding benefit of the approach.
How many stores should a cluster contain?
Enough to measure separately, which is the binding constraint. Clusters too small to have their own result are a planning convenience rather than an analytical unit.
Should clusters be shared with the retailer?
In aggregate terms they can be a useful contribution to a category conversation, provided store-identifiable detail is not disclosed.
Considering NRS Digital Media? Start at nrsdigitalmedia.com, read the media deck, or contact the team.