← All work Next caseA price for every item, set by math instead of a meeting →
Customer segments that hold up under pushback
15communities, tested
- Scale
- Survey data from every site
- Methods
- Latent class analysis, HDBSCAN, a stability test
The problem
Survey data is hostile. Mixed question types, "pick three" answers, small samples per site. Standard clustering gives segments that look fine and mean nothing.
What I did
I used methods that fit the data instead of forcing it into k-means. I ran the clustering two ways, with and without outside market data, to measure how much the fill-in data moved the answer, then tested which segments were real.
FIG.06 ยท pick a site, then re-run it with outside datatry it · sample data
What changed
- A customer mix for every site.
- Segments that survive a re-run of the analysis.
- Recommendations operators acted on that month.
What this looks like for you
A segmentation you can defend in front of a skeptic is worth ten that merely look good.
Let's talk data →