Behavioral clustering of ordering customers
Behavioral clustering that turns each customer's journey into a vector and groups customers who behave alike, with every cluster named from the sessions at its center. Shown with every customer identifier removed.
- Kind
- project
- Role
- led
- Shown
- anonymized
Field innovationTechnical leadership
Each cluster is a named group of customers who behave alike: direct-to-checkout buyers, browse-heavy shoppers who convert once, habitual order trackers. Every cluster carries its size, conversion rate, average order value, and sessions per member, and the members are plotted in a three-dimensional behavioral space.
How the clusters are built. The pipeline runs directly in BigQuery, on the behavioral data Fullstory already holds for a customer. A semantic layer maps technical fields to business meaning, so URLs become funnel stages and custom events become named moments. Each customer’s path becomes a weighted journey, where the steps that matter most repeat more often. A sentence-embedding model (all-MiniLM-L6-v2) turns each journey into a vector, UMAP reduces the space, and HDBSCAN finds the clusters without a preset count. All of it is open source, and it uses the named elements and pages from existing Fullstory configuration, so any Fullstory customer can do the same.
How the clusters are named. Claude, working through Fullstory MCP agentic session review, walks the centroid sessions of each cluster, the ones closest to the middle. It writes a semantic name from the behavior it observes and how alike those sessions are.
About the figures. The sizes and values shown are real, but they come from a sample of partial data capture and can’t be tied to any business. They illustrate the method and carry no analytic weight. The brand and its product names are removed, and the hosted data is a fixed-seed sample of 250 points per cluster, rounded to three decimals.
Choose a cluster to read how its members behave.