The deck
Built around your objective, ending on who to target.
01 / Your question
It starts with your question.
Agents turn your business question into the criteria every solution will be judged on.
The question threads through the whole run: it sets the criteria, it is the context for every persona, and the deck opens on it word for word.
Before anything is clustered
- Ask for it. “Run a segmentation on our customer base.” The platform recognises the intent and selects Segmentation.
- Map the variables — segmentation, profiling and up to 15 optional additional variables, suggested from column names and SPSS metadata.
- Marketing-Truth review (checkpoint 1). Any segmentation variable where 60% or more already agree or disagree is flagged, and you keep or drop it — a typology built on a variable everyone agrees with is dead on arrival.
- Methods and weighting (checkpoints 2 and 3). Review the ranked method × dimension-reduction recommendations and override if you like; supply population targets for RIM weighting.
02 / Every candidate
Every algorithm, every size.
Six clustering algorithms each try 2 to 10 segments, in parallel — dozens of candidate solutions, not one guess.
- K-Means
- Hierarchical
- Latent Class Analysis
- K-Prototypes
- VarSelLCM
- KAMILA
Each algorithm is paired with a dimension reduction — none, PCA or an Unsupervised Random Forest — and only methods that suit your data are offered: PCA is not paired with K-Prototypes, VarSelLCM or LCA, which need the variable types intact, and LCA needs 300 or more respondents.
How the sweep runs
A typical run produces 30–60 candidate solutions, instead of eyeballing a single silhouette plot. Behind it, an orchestrator types the variables, decides which methods suit the data, scores and prunes the solutions, builds the cross-tab workbooks and hands the chosen one to the report service.
03 / Ranked against it
Judged by what they say about your question.
Every candidate gets a cross-tab and a persona report. The top three come back ranked against your question, with the reasons.
Solutions are scored on silhouette, Calinski-Harabasz and Davies-Bouldin, fused into one 0–5 score; any where one segment holds 70% or more of the sample is dropped.
The rest are ranked on the criteria from your question, plus how distinct and how well covered the segments are. Without a usable question, cluster quality alone decides the order.
04 / The people
Choose one. Meet the people.
You make the pick. Every segment gets a persona, and a difference is only claimed when it is real.
An over- or under-index is called out only when it passes 130 / 70 and is statistically significant — no agency-deck inflation. In this example: a six-segment KAMILA solution, silhouette 0.31, built on 12 attitudinal items.
What is tested
Pairwise z-tests on every cell, ANOVA and pairwise Welch’s t-tests on numeric profiles, and automatic top, middle and bottom-box rows for Likert scales — in an Excel workbook with a hyperlinked table of contents. The persona narratives are written by an LLM, held to those tests.
05 / The answer
Get the answer, in a deck.
The PowerPoint opens on your objective and ends on who to target, who to deprioritise, and why.
Research objective and sample · Building the segmentation · Persona review · Targeting strategy · Strategic recommendations — with a direct answer to your question in the executive summary.