Segmentation — who are our customers really, and how should we organise around them?

“Who are our customers really, and how should we organise our marketing, product and CX work around them?”

Say a beverage brand fields a 60-question study with 1,800 respondents. They feed in twelve attitudinal items — need-states, occasion preferences, brand-attitude statements — as segmentation variables, and twenty more — demographics, usage frequency, brand awareness, NPS — as profiling variables. The orchestrator screens the twelve for items where everyone already agrees, runs several clustering methods in parallel across k = 2..10, prunes the dominant-cluster solutions, and returns the top three. The user picks one — six segments, balanced sizes, silhouette 0.31, a KAMILA solution — and the chat hands back a multi-section markdown report with named personas, a significance-tested Excel workbook, and an executive summary slotted ahead of section one.

Segmentation is the most multi-stage tool on the platform. The orchestrator types the variables, decides which clustering methods suit the data, runs them across a sweep of cluster counts in parallel, scores and prunes the resulting solutions, builds significance-tested cross-tab workbooks, and hands the chosen solution to a deep-research report microservice.

• Seven methods, auto-selected. Six clustering algorithms — K-Means, Hierarchical, LCA, K-Prototypes, VarSelLCM and KAMILA — plus URF importance filtering. The orchestrator scores each method against the data profile and pairs it with a dimension-reduction option: none, PCA or URF.
• Marketing-Truth pre-flight.Any segmentation variable where ≥ 60% of respondents already agree (or already disagree) is flagged before clustering — a typology built on a variable everyone agrees with is dead on arrival.
• A full k-sweep, not a guess. Every k from 2 to 10 runs against every chosen method in parallel — a typical run produces 30–60 candidate solutions, instead of eyeballing a single silhouette plot.
• Composite scoring and pruning.Each solution is scored on silhouette, Calinski-Harabasz and Davies-Bouldin, fused into a 0–5 score. Any solution where one segment holds ≥ 70% of the sample is discarded; the top three are returned for you to compare.
• Significance-tested cross-tabs.Pairwise z-tests on every cell, ANOVA and pairwise Welch’s t-tests on numeric profiles, and automatic top/middle/bottom-box rows for Likert — in a “Corporate Blue” Excel workbook with a hyperlinked table of contents.
• Evidence-bound persona narratives. The report is LLM-generated, but an over- or under-index is only called out when the index passes 130 / 70 and the cell is statistically significant — no agency-deck inflation.

The flow is multi-step with three explicit human-in-the-loop checkpoints — a Marketing-Truth review, a method-selection review, and weighting setup — before clustering runs, then a solution-selection step before the report is produced. You stay in control at every decision the analysis makes.

The deliverables are a multi-section deep-research report (Overview, Sample, Segment Building, Category Deep-Dives, Profile-Target, Market Insights, Next Steps and Appendix, with an executive summary stitched in at the top), a per-solution cross-tab workbook for every surviving solution, a comprehensive cross-tab workbook for the chosen solution, and persona narratives embedded one per segment. The report and the workbook are the two artefacts to call out — they are what traditionally takes a research agency four to six weeks.