Internal Methodology Reference
How to design and annotate interview guides that produce data ready for segmentation analysis.
The interview guide is not just a question list. It is the data collection plan that determines which segmentation variables will exist in the final dataset.
No amount of analysis can recover data the guide did not collect. If the guide does not ask about evaluation triggers, you cannot build an Evaluation Trigger dimension. If it does not ask about buying committee composition, you cannot build a Decision Complexity dimension.
The binding constraint: Every clustering variable must come from something you explicitly asked about. Finalize the interview guide only after mapping every question to its segmentation role.
Before writing any question, identify which downstream audiences will use the segmentation findings. Each audience needs a different type of variable.
Needs pain points in the participant's own words, evaluation criteria, and what messaging resonates or triggers rejection.
Questions to include:
Needs observable signals to classify a prospect before a full interview, and differentiated scripts for each segment type.
Questions to include:
Needs feature requirements, table-stakes disqualifiers, and what would accelerate or block the purchase decision.
Questions to include:
If the guide only covers one audience, the segmentation can only make recommendations to one audience. This is the most common avoidable failure in client studies and is detectable from the interview guide before a single interview is conducted.
All interview questions fall into one of two temporal layers. This determines whether a variable can be a clustering input or must be treated as profiling.
Questions about the past evaluation process — what drove the decision, what criteria were applied, who was involved, why alternatives were rejected.
Defining variable candidates
Questions about current satisfaction, gaps, and needs — describing what the participant experiences with the tool they chose.
Profiling variables by default
The temporal contamination risk: A participant who chose Ease-of-Use Tool A may now report ease of use as highly important — because the tool shaped their expectations, not because ease of use predicted the choice. Using Module B data as defining variables introduces reverse causality: the outcome (tool choice) partly causes the variables you are using to predict it. The resulting clusters group people by downstream effects rather than upstream drivers.
The boundary question: "What criteria did you use when evaluating tools?" is Module A (asking about a past decision). "What do you look for in a tool today?" is Module B (describing current preferences, shaped by the current tool). These questions look similar but have opposite segmentation roles. Always anchor the question in time explicitly — "Back then..." for Module A, "Currently..." for Module B.
Annotate every question before finalizing the guide. Annotations force explicit segmentation role assignment and feed directly into the coding pipeline.
Place the annotation immediately after the question number, before the question text:
| Type | When to use | How it encodes | Example |
|---|---|---|---|
thematic | Open-ended qualitative question | Coded to named themes; becomes composite binary dimension | "What motivated you to evaluate new tools?" |
discrete category | Fixed set of non-ordered values | Coded to a category label | "What industry is your company in?" |
ordinal category | Ordered values (size, tenure, recency) | Coded to an ordered label; encoded as integer | "How many people work at your company?" |
binary | Yes/no or forced choice | Coded as 1/0 | "Did you evaluate more than one tool?" |
rank order | Participant ranks a list of options | Binary flags per rank position, or top-N presence | "Rank these 5 features by importance" |
| Role | What it means | Where it goes |
|---|---|---|
| defining | Goes into cluster analysis. Determines which cluster each participant falls into. | segmentation-ready.csv (defining columns) |
| outcome | Validates that clusters predict something useful. Never enters clustering. Must come from screener. | segmentation-profile.csv (outcome column) |
| profiling | Describes what each segment looks like after clustering. No causal role in segment formation. | segmentation-profile.csv (profiling columns) |
Most of your variables are profiling. Only a small, curated set — typically 6-12 for a 70-100 participant study — should be defining. The N/10 rule sets the hard cap: maximum defining dimensions = sample_size / 10. At N=80 that is 8 defining dimensions.
The screener and interview serve different roles. Mixing them up contaminates the outcome variable and weakens the segmentation.
Structured fields captured before the interview. Consistent format, not subject to conversational bias or interviewer effect.
Qualitative depth the screener cannot capture: narrative, emotion, language, criteria, committee dynamics.
The outcome variable must come from the screener. If you infer current tool adoption from what someone says during the interview, the outcome variable is contaminated. The interview topic primes them. Their description of their current tool is influenced by the conversation context. You cannot verify the answer independently afterward. The screener field is the ground truth — the interview version is a recall confirmation only.
Every annotation you write in the interview guide becomes machine-readable context in study_config.json, which feeds Phase 6 of the coding pipeline.
Create study_config.json in the Interview Project folder alongside transcripts.json. The Dimension Architect reads it during Phase 6 to classify themes into segmentation dimensions.
{
"study_name": "HR Leaders",
"sample_size": 80, // sets N/10 ceiling: max 8 defining dimensions
"outcome_variable": {
"question_id": "Q4",
"field_name": "current_tool",
"description": "Primary recruiting tool — verified from screener"
},
"question_context": [
{
"question_id": "Q1",
"temporal_layer": "firmographic",
"purpose_hint": "defining",
"note": "seniority and function"
},
{
"question_id": "Q7",
"temporal_layer": "module_a", // upstream of tool choice
"purpose_hint": "defining",
"note": "evaluation trigger — primary defining variable"
},
{
"question_id": "Q16",
"temporal_layer": "module_b", // downstream of tool choice
"purpose_hint": "profiling",
"note": "current tool strengths — profiling only"
}
]
}
If study_config.json is absent, Phase 6 runs with no question context and infers temporal layers from question text alone. The Dimension Architect will do its best, but it cannot know which question holds the outcome variable without being told. Always create this file before running the pipeline.
The sample_size field drives the maximum number of defining dimensions. The N/10 rule (at least 10 participants per variable for stable clusters) sets the ceiling.
| Study size | Max defining dimensions |
|---|---|
| 50 participants | 5 |
| 70 participants | 7 |
| 80 participants | 8 |
| 100 participants | 10 |
| 200 participants | 20 |
Six questions from the HR Leaders study showing annotation decisions and the reasoning behind each classification.
Q4 is annotated outcome because it is the ground truth for cluster validation. The answer is already known from the screener. The interview version is a recall check — if screener and interview responses conflict, flag the participant. Never use this field as a clustering input.
Q7 is Module A (asking about a historical decision) and thematic. It will likely form the Evaluation Trigger composite dimension — a primary clustering input. All themes should be coded, not just the first one mentioned. This is one of the highest-value questions in the guide.
This question could be mistaken for Module B because it asks about "problems." But the explicit temporal anchor ("drove you to look for a new tool") places it in the evaluation period. The problems described here caused the switch evaluation — upstream of the choice. defining candidate. If rephrased as "What problems do you have with your current tool?" it would become Module B and must be annotated profiling.
Buying committee complexity is Module A (the decision has been made) and produces a categorical code (sole decider / small group / formal committee). It reliably predicts evaluation process length and sales motion complexity. defining.
Q6 asks about the prior tool — Module A in temporal terms. But the specific prior tool name is competitive intelligence, not a structural segmentation variable. Segments defined partly by "was using Tool X before" cannot be activated — sales and marketing cannot target people based on their prior tool without a full research interview. Annotated profiling for competitive win/loss analysis post-clustering.
Q16 is squarely Module B — the participant is describing their current experience with the tool they chose. These themes are downstream of the tool choice. They describe what the tool is good at, not what drove the selection. profiling only.
Five errors that produce bad segmentation variables — all detectable before fieldwork begins.
| Error | Example | Correct approach | Why it matters |
|---|---|---|---|
| Module B as defining | "What frustrates you about your current tool?" annotated defining | Annotate as profiling; move the corresponding Module A question to defining | Current frustrations are shaped by the tool chosen, not the other way around. Produces reverse-causal clusters. |
| Outcome as defining | Q4 (current tool) annotated defining | Annotate as outcome; ensure screener captures it | Putting the outcome into clustering builds the answer into the question. Clusters become self-fulfilling. |
| Module A / B conflation | "What were your evaluation criteria?" (Module A) and "What do you look for in a tool?" (Module B) both annotated defining | Module A version is defining candidate; Module B version is profiling | These questions look similar but have opposite segmentation roles due to temporal anchoring. |
| Near-constant variable as defining | All participants are mid-market HR leaders; company size annotated defining | Check screener distribution first; if near-constant, annotate profiling | Near-constant variables waste a defining variable slot without adding discriminative power. |
| Rank-order with no encoding plan | "Rank 6 features" annotated defining with no note | Specify encoding in the coding note: binary flag for rank-1 position, or presence in top 3 | The Dimension Architect needs clear encoding guidance for rank-order data. Without it, it will guess. |
The annotation key, decision rules, and pre-finalization checklist on one page.
thematic Open-ended qualitative; codes to themes discrete category Fixed non-ordered values ordinal category Ordered values; encodes as integer binary Yes/no or forced choice rank order Ranked list; specify encoding in note