Research method
Synthetic Interviews & Personas
- Explore
Interviewing or chatting with AI-generated personas built from an audience description, and sometimes from your own research and customer data. The output reads like interview transcripts, but no real customer is involved, so every finding is a hypothesis to test with people.
When to reach for it
Pilot an interview guide or survey before real fieldwork, and catch confusing or leading questions.
Orient quickly in an unfamiliar audience or category before planning real research.
Generate a long list of possible objections and motivations to probe with real customers.
Stress-test draft copy for unclear claims and jargon ahead of a real copy test.
Often paired with Customer Interviews.
Pros
Results come back in minutes, with no recruiting, scheduling or incentives.
Cheap enough to run many times, so a guide or concept can be iterated before real people see it.
Personas grounded in your own interviews and data perform better than demographic prompts.
Cons
Answers bunch around the most common response and drop the unusual ones. NN/g calls this mode collapse, and it means synthetic interviews miss the long tail that real research is often looking for.
Synthetic personas tend to be agreeable and positive, so they flatter concepts that real users would question.
They are weakest for niche, non-Western or minority audiences, for emotional and habit-driven behaviour, and for anything genuinely new.
No real customer is studied, so this is not user research and cannot validate a decision.
Practical considerations: sample sizes, tools, logistics
Practical considerations
Pair it
Always follow with Customer Interviews, a Customer Survey or Copy Testing with real people, aimed at confirming or killing what the synthetic run suggested.
Customer language
Never take customer wording from synthetic personas for copy, keywords or labels. Use their output to shape questions, then take the words from real customers.
Label it
Reports should say which findings came from people and which came from models.
Model drift
Outputs shift as the underlying model updates, so re-run rather than reuse old results.
Never in isolation
Recipes that use Synthetic Interviews & Personas
A method earns its value when its signal corroborates others. Here is where this one does its work.
Last updated
.svg)
