Building reliable personas through mixed methods — qualitative approach confirmed by statistical segmentation
How to go beyond intuitive personas with a rigorous mixed-methods approach: 10 interviews, questionnaire with 102 clients, free-text coding and K-means clustering to reveal that role within the company — not sector or purchase frequency — is the only true differentiating factor.
- Role
- UX Research Lead
- Duration
- 1 month
- Tools
- Qualtrics, R Studio, Figma
Context
At RAJA, product decisions relied on a fragmented view of users — a few informal typologies built on internal assumptions. Teams believed they knew their customers: large companies, SMEs, various sectors. But these categories had never been validated against actual user behaviour.
The objective was to build personas grounded in reality: not in intuition, but in a mixed-methods process — qualitative to explore, quantitative to confirm. The challenge was to go from rich qualitative data to statistically robust segments, directly usable by design and product teams.
Methodology
Qualitative exploration
10 semi-structured interviews of 60 minutes conducted in France with clients from RAJA's key sectors, with varied company sizes reflecting the actual client base. Sessions conducted with two interviewers and double coding to ensure reliability of the thematic analysis. This phase identified the nature of purchasing tasks as the key dimension to explore quantitatively.
Questionnaire design
Design of a structured questionnaire based on dimensions identified in the qualitative phase. 28 variables measured covering purchasing role, digital tool usage, objectives, and purchase process. Combination of closed questions (scales, behavioural items) and open questions to capture the diversity of real-world usage. Deployed to 102 clients via RAJA's internal panel and client base.
Free-text coding
Systematic content analysis of open responses collected from the 102 respondents. Categorisation and recoding into binary variables (True/False) to make the data usable by the clustering algorithm. This bridging step between qualitative richness and quantitative rigour was the most time-intensive of the process.
K-means segmentation
Application of K-means on binary variables derived from the coding step — a pragmatic choice given the nature of the data, which prioritised actionability. Optimal number of clusters determined using the elbow method, converging at k=3. The analysis revealed that the only truly discriminating dimension is the client's role within their company: neither industry sector nor order frequency conditioned their relationship to purchasing on RAJA.fr.
Optimal number of clusters — Elbow method
The elbow at k=3 confirmed the optimal number of segments
Persona definition by parangon
For each of the 3 clusters, selection of the parangon respondent — the most representative case of the centroid — to ground each persona in a real individual rather than a fictional average profile. Three personas identified. Persona cards produced in Figma (A5 format, descriptive text and characteristic tags) and shared with product and design teams.
Key finding
Industry sector and order frequency don't differentiate RAJA's buyers.
The only discriminating variable: the client's role within their company. A counter-intuitive finding that initially caused disappointment — and ultimately reshaped the teams' mental model of their users.
The 3 personas
Three segments were identified, each anchored in a real parangon respondent. The detailed persona cards (A5 format, Figma) are confidential. The offline profile was excluded from the digital design scope.
The Offline Orderer
Off-siteThe Professional Buyer
The Entrepreneur at the Helm
Content anonymised — detailed cards subject to NDA
Key learnings
- The expected segmentation — by industry sector or order frequency — did not emerge from the data. The only truly discriminating variable was the client's role within their company. A counter-intuitive result that initially caused disappointment before being accepted.
- One of the 3 clusters revealed an 'offline orderer' profile that does not use the website at all — a segment invisible in analytics but genuinely present in the client base.
- Recoding free-text responses into binary variables was the pivotal step in the process: it was what made statistical analysis possible on initially qualitative data.
- Choosing the parangon over an average profile significantly helped teams take ownership of the personas — a real individual is more memorable than a list of attributes.
Impact
3 B2B personas built through mixed methods and grounded in real individuals (parangon method). 2 of the 3 personas were retained as priority targets for design phases — the offline profile being excluded from the digital scope. The personas are still actively used by teams, and have opened the door to a next iteration incorporating finer psychological dimensions.
- 10
- interviews
- 28
- variables
- 102
- respondents
- 3
- personas (2 retained)
Let's work together
Interested in this case study? Let's discuss what user research can bring to your product.