"Go find some customer insight" is one of those instructions where the goal is obvious but the actual procedure is almost never spelled out.
Building on the basics of insight marketing, this article lays out a working process for actually discovering insight — from n=5 interview design to AI-assisted open-text analysis — at a level you can take into the field on Monday.
The full method set
Insight discovery isn't one technique; it's a combination:
| Method | Strength | Effort |
|---|---|---|
| User interviews (n=5–10) | Depth, context, emotion | Medium |
| Structured open-text analysis | Scale, pattern detection | Medium |
| Behavioral observation (ethnography) | Things people don't say | Large |
| Outlier analysis on data | Hypothesis triggers | Small |
| Extreme-user research | Hidden motivations | Medium |
| Competitive interview comparison | Positioning understanding | Medium |
This article focuses on the two most generally useful methods: interviews and open-text analysis.
Method 1: n=5 user interviews
"Interviews don't count unless you do them at scale" is a common misconception. For insight discovery, n=5 is enough. Large-sample surveys are the wrong tool for finding insight in the first place.
Why n=5 works
A well-established UX research finding: roughly 85% of the major themes surface in the first 5 interviews. Going from 5 to 10 to 20 gives diminishing returns on new insight.
That said, who those 5 are matters a lot:
- 5 people from the same segment skews everything
- Build for diversity (demographics, usage intensity, satisfaction)
- Include one "domain expert" and one "casual user" by design
Standard interview shape (60–90 min)
0–10 min: Icebreaker, background
"Tell me about your recent work on X"
10–30 min: Concretize the usage moment
"When did you last use it? What was the situation?"
"Walk me through it as a story"
30–50 min: Excavate the "why"
"Why did you pick this one?"
"What were the alternatives?"
"What would you do if you couldn't use it anymore?"
50–70 min: Throw hypotheses at them
"Some people describe this as X — does that match?"
→ Watch their reaction to validate
70–90 min: Open space
"Anything you want to ask me?"
"Anything you didn't get to say?"
The deepest disclosures tend to land in that last open space.
Things you must not do
Don't 1: Leading questions
✗ "This is convenient, right?" (fishing for YES)
○ "How did this feel to use?"
Don't 2: Stacking "why"
✗ "Why? Why? Why?"
○ "Can you give me a specific situation?" "What else?"
Three "whys" is a ceiling. More than that turns into interrogation and shuts the respondent down.
Don't 3: Imposing your hypothesis
✗ "So you're saying X, right?" (pushing your interpretation)
○ "Just to make sure I heard right — you're describing X?" (checking)
Recording discipline
- Always record (with consent)
- Keep notes light; make eye contact
- Process the recording within 30 minutes of the session while it's fresh
- Quote verbatim for striking lines; don't paraphrase
Method 2: Structured open-text analysis
Open-text survey responses are packed with insight seeds. But "reading them" isn't analysis.
Step 1: Categorize
Take 100+ open-text responses and:
- Cluster by theme
- Use AI (ChatGPT/Claude) for first-pass classification
- Collapse to 5–10 categories
Example categories:
- Pricing complaints
- Friction with feature A
- Support quality
- Competitor comparison
- Usage scenarios
- Expected-but-missing features
Step 2: Frequency + sentiment
- Frequency per category
- Positive / negative / neutral distribution
- Pattern differences across segments
Step 3: Hunt for "weird phrases"
The most valuable lines are the ones that don't fit a category — the unexpected, the strange:
"I just kind of keep using it"
"I don't really want anyone to see I use it"
"It feels a little embarrassing"
"I didn't really choose it consciously"
Lines like these are usually the entry point to unsurfaced insight.
Step 4: Form hypotheses
Convert weird phrases into testable hypotheses:
Phrase: "I just kind of keep using it"
Hypothesis: This product retains not through explicit value
but through lack of a switching trigger.
→ "Making it harder to leave" is more important than
"giving more reasons to use"
Step 5: Validate with n=5 interviews
Take the hypothesis into interviews and watch reactions. When you hear "yes — exactly that," you've found insight.
Method 3: Extreme-user research
The median customer gives median answers. Outliers are where hidden motivations live.
Extreme users worth studying
- Super heavy users — using the product in ways no one else does
- About-to-churn users — why leaving, why they stayed this long
- Switchers from competitors — what tipped the decision
- Trial users who didn't convert — what was missing
- Long-tenured users — what keeps them locked in for years
Targeting these segments with surveys + interviews reaches insight that averaged-out feedback can never expose.
Method 4: Data outliers as hypothesis triggers
Anomalies in quantitative data are insight entry points.
What counts as an anomaly
- A metric far above expectation (NPS well above industry average)
- A metric far below expectation (one segment churning at unusual rates)
- Unexpected (lack of) correlation
- Seasonal weirdness
Throw "why?" at each, then go deeper with interviews or surveys.
The hard part — four common traps
Trap 1: Validating your own hypothesis
The most common failure: steering the interview toward the conclusion you wanted.
Counters:
- Always include questions that could disprove your hypothesis
- Deliberately interview people who hold opposing views
- Have a third party review the interview record
Trap 2: Calling an n=1 finding "insight"
"One person said it" is not insight. It's a hypothesis. Generalizing from a single case is the classic mistake.
Counters:
- Require insight candidates to be observed across multiple respondents
- Validate with quantitative surveys at scale
- Distinguish "hypothesis" from "confirmed" explicitly
Trap 3: Taking what they said at face value
"It's too expensive" rarely means price is the actual issue. It often means the real complaint is being expressed as price because price is the easiest thing to articulate.
Counters:
- Probe: "If it were half price, would you definitely keep using it?"
- Cross-check stated reasons against behavioral data (actual retention)
Trap 4: Discovering insight, then doing nothing with it
If insight doesn't change a decision, its value is zero.
Counters:
- Share insight in exec staff meetings
- Bake it into marketing copy, product priorities, sales talk tracks
- Measure whether insight-aligned changes moved metrics in the next round
Quick health check
A self-check for your discovery process:
- Have you interviewed at least 5 people?
- Is there diversity in attributes / usage / satisfaction?
- Have you categorized the open text?
- Did you hunt for "weird phrases"?
- Did you talk to extreme users?
- Did you go looking for opposing views?
- Did you confirm scale quantitatively?
- Is the insight scheduled to inform a real decision?
Where Repoan fits
Repoan supports what happens around the insight discovery work:
- Pre-interview screening surveys — to pick the right n=5 and seed early hypotheses
- AI-generated questions — for open-text designs that surface insight
- AI theme extraction on open text — to find "weird phrases" inside 100+ responses
- Per-segment dashboards — extreme-user identification baked in
- Continuous surveys — to track insight hypotheses over time
Summary
To find customer insight:
- n=5 interviews are the most efficient unit
- Mine open text for "weird phrases"
- Pay attention to extreme users
- Use data anomalies as hypothesis triggers
- Always validate scale quantitatively
- Avoid the "validate-your-own-hypothesis" and "n=1 means insight" traps
Insight discovery is observation, dialogue, and interpretation — deeply human work. AI can carry the heavy mechanical pieces (open-text structuring, theme extraction), but the interpretive core is human. That division of labor is how you mine real insight in the AI era.