How to Interview a Synthetic User

Interviewing a synthetic user works like interviewing a human — open questions, concrete scenarios, follow-ups on surprises — with one inversion: the participant is generated, so preparation shifts from recruiting the right person to grounding the right person. Here's the craft, step by step, including the mistakes that make synthetic interviews worthless.

Step 1: Start from a grounded person, not a prompt

The single biggest quality lever is decided before the first question. A persona typed into a prompt — "Maria, 34, busy professional" — produces the model's stereotype of that phrase. A grounded synthetic user is generated from real population data: a statistically real combination of age, income, occupation, household, and language, with personality traits modeled on top and a daily context to answer from.

The test: could you be surprised by who you're talking to? If you authored every attribute, you can't be. If the person was drawn from a Census-grounded population, you'll find yourself interviewing a 58-year-old first-time borrower who works nights — someone you wouldn't have invented, which is the point.

Step 2: Set the scene, not the script

Give the interview a concrete situation: "You've just gotten a text that your prescription is ready, but the pharmacy closes in 40 minutes and you're at work." Context activates the person's attributes — commute, patience, obligations — and produces answers anchored in a life rather than free-floating opinions.

What you don't do is tell them how to feel. "You're frustrated with your pharmacy" pre-loads the finding; the interview will dutifully return your own assumption with quotes around it.

Step 3: Ask like a researcher, not a fan

Every rule of human interviewing applies, amplified — because LLM-driven participants are more agreeable than people, leading questions are even more poisonous.

• Past behavior over hypotheticals. "Walk me through the last time you disputed a charge" beats "would you use a dispute-resolution feature?"
• Concrete over abstract. Show the actual copy, the actual price, the actual flow — reactions to specifics are worth more than opinions about categories.
• Open, then narrow. Start with "what's annoying about managing prescriptions?" before you ever mention refill reminders.
• Hunt for disconfirmation. Explicitly ask what's bad, what they'd skip, what would make them stop using it. A synthetic interview with no bad news is a broken interview.

Step 4: Probe the way the person answers

The follow-up is where synthetic interviews come alive, because the person stays in character under pressure. When the 71-year-old ESL retiree says the pricing page is "fine," push: which plan would she pick? What does "cancel anytime" mean to her? Push-back is also where distinctness shows — the skeptic gets more skeptical, the rambler digresses, the anxious one asks what happens if she picks wrong. If every probe returns smooth, compliant elaboration, your participant isn't grounded enough to be worth interviewing.

Step 5: Use memory for the second interview

Real research relationships are longitudinal, and synthetic ones can be too. A synthetic user with persistent memory remembers last week's conversation — so you can re-interview the same person after changing your pricing, your onboarding copy, or your product name, and see what moved for them specifically. This is a study design human research can rarely afford: a perfectly stable panel with total recall and no attrition.

Step 6: Scale the interview into a survey

One-on-one interviews find the language — the objections, the misreadings, the words people actually use. To find out how widely each reaction holds, take the sharpest questions from your interviews and run them across the whole audience as a batch survey: the same question, hundreds of synthetic respondents, answers split by age, income, and language. The interviews generate the hypotheses; the cohort distribution tells you which ones describe a segment rather than an individual.

Step 7: Read the results as what they are

A synthetic interview produces hypotheses with excellent provenance — every quote traces to a specific, demographically real person — but hypotheses nonetheless. The honest reading discipline, in three lines: treat surprising findings as questions to put to real users; treat pricing and emotion signals as noise until validated; label synthetic-sourced insights as synthetic wherever they're stored, so a rehearsal never quietly ossifies into evidence. (The full argument for where synthetic research helps and misleads is in synthetic users for UX research.)

The short version

Ground the person in real data, set a concrete scene, ask non-leading questions, probe hard, re-interview with memory, scale to cohorts, and validate what you'll bet on. Interviewing synthetic users well is mostly interviewing well — the technology just removed the six weeks and the recruiting bias between you and the conversation.