Synthetic personas

What synthetic personas are

Synthetic personas are AI-generated user profiles built from your research data. Unlike static persona documents, you can interact with them.

Creating personas

Personas are generated from your research findings. The more research you have, the more grounded and useful they become. Each persona includes demographics, goals, pain points, behaviors, representative quotes, and a confidence score showing how well-supported the persona is by real evidence.

Persona views

Click a persona to see its detail view. Switch between interaction modes:

Overview

See the full profile — demographics, goals, pain points, and the evidence that supports this persona. The confidence score shows how well-grounded the persona is in real research data.

Interview

Click the Interview tab to have a conversation with the persona. Type questions and get responses from the persona's perspective. Use this for:

• Testing messaging and value propositions
• Understanding edge cases in the user journey
• Getting reactions to feature ideas before building them
• Exploring scenarios you have not researched yet

Critique

Click the Critique tab, then share a feature, flow, or design direction. The persona evaluates it from their specific perspective, highlighting what works and what does not.

Screen Testing

Click the Screen Testing tab to present a design screen to the persona. Get feedback on whether the screen makes sense, what they would expect, what confuses them, and how they would navigate.

Grounding report

Each interaction shows a grounding report indicating how much of the response is supported by actual research versus extrapolation. Check this to know how much to trust each response.

Confidence levels

Synthetic personas are better than guessing — but they are not the same as talking to real people. Every interaction carries a confidence score so you can judge how much weight to give the feedback.

What the confidence score measures

The confidence score (0–100%) reflects how well a persona's attributes are supported by evidence in your project. It is not a measure of how accurate the persona is in the real world — it measures how grounded it is in the data you have provided.

Each persona attribute is tagged with a grounding source:

• Source: Real data · What it means: Derived from uploaded transcripts, surveys, or user interviews · Weight: Highest
• Source: Research · What it means: Inferred from AI-generated research reports · Weight: High
• Source: Manual · What it means: Created directly by you in the persona editor · Weight: Medium
• Source: Inferred · What it means: Gap-filled by AI when no evidence exists · Weight: Lowest

The overall confidence score is a weighted average across all attributes. More real data and research grounding means higher confidence. More inferred attributes means lower confidence.

Confidence labels

• Score: 60–100% · Label: High · What it means: Most attributes are grounded in research or real data. Responses are likely representative.
• Score: 35–59% · Label: Medium · What it means: A mix of grounded and inferred attributes. Useful for exploration, but verify key assumptions.
• Score: 0–34% · Label: Low · What it means: Most attributes are inferred by AI. Treat responses as hypotheses, not evidence.

Where synthetic personas are valuable

Synthetic personas are useful when you need directional signal faster than you can recruit and interview real users. They work well for:

• Early-stage exploration — testing messaging, feature concepts, or value propositions before investing in full research.
• Edge case discovery — surfacing scenarios, objections, or pain points you may not have considered.
• Design critique — getting a structured perspective on a design before usability testing.
• Interview practice — refining your research questions before talking to real users.
• Broadening perspective — challenging assumptions when a team has limited exposure to certain user segments.

Where synthetic personas fall short

Synthetic personas have fundamental limitations that confidence scores alone cannot capture:

They reflect your data, not reality. A synthetic persona can only be as good as the research it was built from. If your research has blind spots, the persona inherits them. If your sample is biased toward a specific demographic, the persona will reflect that bias — with high confidence.

They cannot surprise you the way real users do. Real user interviews surface unexpected behaviors, workarounds, and emotional reactions that no AI model can predict. Synthetic personas generate plausible responses based on patterns, but they will never say "actually, I use your product completely differently from how you intended."

They lack real context. Real users have jobs, deadlines, coworkers, and bad days. They use your product on a cracked phone screen while commuting. Synthetic personas simulate context, but they do not have lived experience that shapes real behavior.

They are systematically agreeable. AI models tend to be more cooperative and articulate than real users. A synthetic persona asked "would you use this feature?" is more likely to say yes than a real person who would shrug and say "maybe, depends." This means satisfaction scores trend higher than reality, usability concerns may be understated, and negative feedback is often more polished than real frustration.

They cannot validate product-market fit. A synthetic persona saying "I would pay for this" is not evidence of demand. It is a language model generating a plausible response. Product-market fit can only be validated with real users, real money, and real behavior.

Using confidence scores responsibly

Do:

• Use synthetic persona feedback to generate hypotheses, then validate with real research.
• Check the grounding chart before acting on feedback — know how much is evidence versus inference.
• Combine synthetic persona insights with other signals (analytics, support tickets, market data).
• Use higher-confidence personas for prioritization, not for final decisions.
• Upload more research data to improve grounding over time.

Do not:

• Quote synthetic persona responses in stakeholder presentations as if they were real user quotes.
• Skip user interviews because a synthetic persona "already answered that question."
• Treat a high confidence score as proof that the persona is accurate — it only means the data is well-grounded, not that the data is complete.
• Rely on synthetic personas for accessibility, cultural sensitivity, or emotional safety assessments — these require real human perspective.

The evidence hierarchy

When making product decisions, weight your evidence accordingly:

• Evidence type: Real user behavior (analytics, A/B tests) · Reliability: Highest · Example: "40% of users drop off at checkout step 3"
• Evidence type: Direct user research (interviews, usability tests) · Reliability: High · Example: "Users said they could not find the save button"
• Evidence type: Synthetic persona feedback (high confidence) · Reliability: Moderate · Example: "This persona predicted confusion with the navigation"
• Evidence type: Synthetic persona feedback (low confidence) · Reliability: Low · Example: "This persona suggested they might prefer dark mode"
• Evidence type: Team intuition with no data · Reliability: Lowest · Example: "I think users would like this"

Synthetic personas sit in the middle of this hierarchy. They are a step up from gut feeling, but they are not a substitute for real research. Use them to move faster and explore more broadly — then validate what matters with real people.

• For the research that feeds personas, read Research and strategy.