One Deep Interview Beats 100 Surveys (Mom Test)

One hour of interview is 100x more accurate than a survey.

One team tanked with a survey, another doubled revenue from interviews

In 2023 a SaaS team in Seoul was building a marketing automation tool. To get the question "what feature do we most need" answered, they made a Google Forms survey. They emailed 100 people, 23 responded. "We need a reports feature" 48%, "API integration needed" 35%, "We could accept higher pricing" 41%. They collected the data.

The team lead decided to build the reports feature first. 3 weeks of hard coding. Shipped it. Nobody used it.

Another team took a different route: "let's just talk to 5 of our customers for an hour each." They bought them Americanos at a café and watched how they actually worked. At some point they overheard something: "Setup is too complicated. I just did the basic settings for the first 3 days." That one sentence told them onboarding mattered more than features. In 3 hours they made a guide video. Revenue doubled.

Surveys give you wide, shallow data; interviews give you narrow, deep data. MVP validation needs deep.

Mom Test, the question even your mom can't lie about

There's a book worth recommending. The Mom Test (Rob Fitzpatrick, 2013). The title is funny on its own. It means "questions even your mom can't lie about."

Your mom loves you, so she says nice things. "Oh, that idea is great!" But whether the idea will fly in the market — she doesn't know. The core of this book is: avoid leading questions, and focus on behaviors and facts.

In Korean culture, saying nice things to be polite is very natural. Especially when a founder asks "would you want our service to have this feature?" the answer is always "yes, sounds good." That isn't fact.

Bad questions

• "Would our service be helpful?"
• "Do you think this feature is needed?"
• "Does our app look easy to use?"

All of these elicit "yes."

Good questions

• "How did you do this task last week?"
• "Who did you work with, and what frustrated you most?"
• "What tool are you using now, and what annoys you every day?"

These ask about specific behavior and emotion.

Sketched example: a team built "AI diary for daily emotion logging." A survey said 73% "would find it helpful." But in interviews, when asked "Are you logging your emotions these days?" the common answer was "No, I'm too busy." Drilling into "why," the real problem wasn't "emotion logging" but "fatigue management." They pivoted the feature, and actual users started showing up.

Good interview questions ask about the "past"

UX research pros use this frame: "Past → Present → Future" ordering.

• Past: "How did you solve this before?" (factual data)
• Present: "How are you doing it now?" (current pain)
• Future: "Ideally, how would you want to do it?" (desire)

Important: future questions go last. If you ask "ideally how would you want to?" first, the respondent will shape the answer to fit your product. That's not truth — that's "trying to be a good guest."

People are way more accurate when described by their current behavior than by their desires. "How many hours of SNS do you want to use a day?" gives worse data than "How many hours did you use SNS yesterday?"

30-minute interview question template

[Opening — 3 min]
"Thanks for the time. I'd like to learn about your day-to-day.
We'll talk about our service later."

[Background — 5 min]
- What kind of work do you do?
- How did you work last week, Monday through Friday?

[Problem discovery — 10 min]
- What part of that took the most time?
- What have you tried to solve it?
- What tool are you using, and what annoys you daily?

[Deepening — 10 min]
- When did this problem last cost you something?
  (Money, time, or emotional energy)

[Close — 2 min]
- Do you know anyone else in a similar situation?

Three things matter in this structure.

1. Specificity: if "usually" or "generally" comes up, drag them to a specific case — "What about last Monday?"
2. Don't bring up tech: your solution shouldn't appear. "Our app can automate this!" is forbidden. Just dig into the problem.
3. Tolerate silence: silence is awkward in Korean culture but you need to give the respondent thinking time after a question. Hold 3 seconds of silence.

1-hour interview, 5 steps in practice

Step 1. Pick interviewees (1 week in advance)

Find 5 actual target users. If none in your circle, post in communities like Disquiet or Ohou and ask, "I'd love to interview you for an hour — I'll buy the Americano." Not friends — people with the actual problem.

Step 2. Pick the venue

Cafés are recommended. In Korea, a neighborhood café beats Starbucks or Ediya. Less formal, more relaxed conversation atmosphere.

Step 3. Recording & notes

"Mind if I record what we say? I won't write down your name." Most agree. Phone voice memo is enough.

Step 4. Conduct the interview (60 min)

Use the template, but don't break the flow. If they say "by the way, my 3 teammates always miss something," dig into that. It might be more important.

Step 5. Wrap up

"Thank you so much for sharing. If you know anyone in a similar situation, I'd appreciate an intro." The first interview opens the door to the second.

AI transcript analysis: finding patterns

Now the annoying part. Turning the recordings into text and finding patterns.

Step 1. Audio → text (OpenAI Whisper, free)

Available free on Google Colab. Even simpler with Claude Code. A 30-minute audio file becomes text in 2 minutes.

Step 2. Text → analysis (Claude)

Drop the transcript into Claude with a prompt like:

Below is an interview transcript with a marketer.

Find the "real problem" this person was expressing.
(Not what they directly said — what was implied.)

Analysis criteria:
(1) Recurring keywords
(2) Emotional intensity
(3) Actual behavior
(4) Parts not solved by current tools

Step 3. Find patterns (after 5 interviews)

After interviewing and analyzing 5 people, you see: "wait, this problem shows up in all 5?" That's signal. You've moved from "coincidence of 5" to "pattern of a market."

One team's workflow: 5 café interview recordings analyzed by Claude revealed "users churn at 90% during onboarding, not at the UI." They invested in a guide video instead of feature development. Retention went from 35% to 62% in a month.

Post-interview checklist

• Did I get the audio file?
• Did I hear at least 1 specific pain?
• Did I get 3+ concrete examples, not generic "would've answered the same on a survey" answers?
• Did I record based on "what they did last week" rather than "what they'd want in the future"?
• Did I get a referral to the next interviewee?

Wrapping up

A 1-hour interview boosts survey accuracy by 5–100x. But there's one limit. "Knowing they want it" and "actually paying for it" are different.

Will someone who said "yes, that feature would be great" really pay? Next post: the last gate that interviews alone can't validate. Fake Door Test, sell before you build.

References

1. Rob Fitzpatrick, The Mom Test, 2013.
2. How to talk to customers for PMF validation: The Mom Test summary — OTHERHAND VENTURES
3. Korean startup community user-interview recruiting cases on Ohou, Disquiet (2024–2026)
4. OpenAI Whisper — Free speech recognition API
5. How to use NotebookLM with Claude Code

Previous: Use AI as a Mirror, Not an Answer Machine
Next: Sell Before You Build: 10,000 KRW Validates Intent from 100 People (Fake Door Test)

Minchul Kim, CEO of Freeive, Fail School