Synthetic personas.Research at scale.

What is a synthetic persona?

A synthetic persona is a modeled profile used to generate a simulated response. A profile might describe a situation, a goal, or a constraint relevant to a product decision. The resulting answer is produced by a model; it is not a recorded opinion from the person the profile resembles.

This distinction belongs in the research brief and the report. AAPOR’s standards explicitly distinguish AI-generated responses from human research participants. Labeling the source of a response helps readers understand what kind of evidence they are looking at.

Reference: AAPOR — Standards and ethics

Where does scale help?

Our view is that the strongest early use is exploring the question space. A team can examine how a draft message might be interpreted across many modeled contexts before choosing what to test with people. The aim is to uncover possible objections, ambiguous wording, or missing use cases—not to manufacture a percentage that looks like a representative population estimate.

More generated profiles do not automatically create more independent evidence. If every answer comes from a similar model and prompt, the apparent diversity may hide a shared source of bias. Treat the number of simulated responses as a description of the run, not as proof of statistical validity.

Design a useful simulation

Start with a decision that is narrow enough to act on. “What do customers want?” is difficult to evaluate. “Which part of this onboarding explanation is unclear?” leads to a more useful next step. Define the audience contexts before generating responses, and keep the wording of the question consistent across them.

Illustrative exercise: a team compares two descriptions of a budgeting app. One emphasizes seeing spending patterns; the other emphasizes planning a weekly allowance. It uses fictional profiles with different budgeting habits to identify questions worth asking in subsequent interviews. The output is an interview guide, not a declaration that one message has won.

• Write the decision and the hypothesis before the prompt.
• Record profile assumptions, model settings, and question wording.
• Keep contradictory responses instead of selecting only convenient ones.
• Turn recurring themes into questions for real participants.

Separate a generated theme from a finding

A useful readout has three columns: what the simulation produced, what might explain it, and what evidence would change the team’s mind. This format makes uncertainty actionable. It also helps prevent a vivid generated quote from being mistaken for an observed customer need.

For example, several fictional responses may mention setup effort. A next step could be a task-based usability session measuring where people need help. If people complete setup easily, the simulated objection may not be relevant. That disagreement is valuable information, not a reason to reinterpret the observed result to fit the simulation.

Make the method visible

AAPOR’s transparency guidance asks researchers using AI-generated responses to explain how those responses were generated. For an internal exploration, a practical starting point is to preserve the prompt, profile definitions, model version where available, and how the output was filtered or summarized.

Atakan AI is developing Persona Research for exploring virtual audiences through simulated surveys. The personas and responses shown on our site are illustrative. We are designing for breadth of exploration while keeping synthetic output clearly distinct from research with real participants.

Reference: AAPOR — Transparency Initiative