Definition
Cluster randomization allocates an experiment at the level of a group rather than each individual shopper. It can help when treatment applied to one person also changes another person's experience, such as a shared account workflow or store-wide merchandising process. The assignment unit and analysis must reflect that shared exposure.
Why It Matters
- Different users of one buying account may share quotes, recommendations, and purchase decisions.
- A store team cannot always operate two incompatible service procedures at the same time.
- A Commerce Intelligence OS should evaluate interventions using an assignment design that matches how the commerce workflow actually operates.
How It Works
- Map likely spillovers before choosing the cluster: account, store, region, or another operational group. Explain why interference should be limited across groups.
- Define eligibility, the commercial outcome, assignment ratio, and test duration. Check the number and size distribution of available clusters before planning power.
- Randomize groups and retain membership history. Log cross-group exposure and concurrent interventions that could compromise the intended comparison.
- Use an analysis that accounts for the grouping, such as a suitable cluster-level analysis or cluster-aware uncertainty estimate. Do not count every shopper as an independent randomized unit; report remaining spillover and design limitations.
Ecommerce Example
Context: Illustrative example: a B2B retailer tests a quote assistant across 40 business accounts, each with several buyers.
Recommended move: Assign the entire account to one variant so colleagues sharing a quote receive consistent treatment. Analyze outcomes with account grouping represented.
Why it matters: Forty assigned accounts do not become thousands of independent assignments just because many sessions occur. These numbers illustrate design, not a powered experiment or proven lift.
iKawn Framework
Map
The iKawn framework records relationships through which an intervention can spread.
Assign
Keep experiment membership attached to the operational group.
Measure
Connect orders and mature return outcomes to assignment.
Interpret
Judge commercial effects with uncertainty appropriate to the design.
Concise Summary
Choose the randomization unit from the workflow. Group assignment can reduce contamination, but analysis and power must account for the limited number of independent groups.