Transparency & Methodology

How we built the engine.
And our methodology.

We know "AI personas" sounds like magic. Here is the exact methodology, the models, the data sources, and the validation approach — no marketing fluff.

Architecture

Persona Architecture

Our personas are not chatbot prompts with random attributes slapped on. They are built from real demographic distributions and behavioral embeddings.

1M+ Real Demographic Base

  • Synthesized from US Census ACS distributions (age, income, education, geography)
  • Occupation and industry matched to BLS labor statistics
  • Behavioral traits inferred from Pew Research & Simmons survey distributions

1024-Dim Semantic Embeddings

  • Every persona is embedded using high-dimensional vectors for semantic similarity search
  • K-means clustering finds opposed pairs for social simulation debates
  • Similarity matching ensures demographic queries return representative samples

Multi-Model Inference

  • Primary: frontier reasoning model for persona responses
  • Analysis: large reasoning model for thematic codebook generation
  • Fast: open-weight model for JSON repair & structured extraction
  • Temperature tuning (0.1–0.7) controls response variance per simulation

Data Privacy Guarantee

  • Your product ideas are never used to train models
  • Inference calls are ephemeral — no prompt/response retention
  • Responses stored only in your Supabase instance, isolated per team

Validation

Methodology

We do not claim perfection. We claim directional accuracy at 1/10th the cost and time.

~90%
Correlation to Real Panels

In backtests against 12 published consumer panels (n=200–1,500 each), TestSynthia aggregate preference rankings correlated at r≈0.89–0.92 with real survey rank-order correlations. Not identical. Directionally aligned.

K-Means
Segment Separation

Persona embeddings cluster into demographic segments that mirror real-world variance. Silhouette scores >0.6 confirm distinct persona profiles, not statistical averaging into a bland median.

15 min
vs. 3–6 Weeks

Time-to-insight is deterministic. The trade-off is precision vs. speed. Use this to kill bad ideas early. Use real panels to confirm winners.

What the 90% number actually means

It does not mean "90% of individual responses are word-for-word identical to real humans." It means: if you run a simulated survey and a real survey on the same concept with the same target demographic, the rank order of preferred features, the direction of sentiment, and the magnitude of intent will align with Pearson r≈0.9. It is a wind tunnel, not a crystal ball.

Trust

How We Prevent Hallucinations

The biggest fear in survey responses is "is this just telling me what I want to hear?" Here are the guardrails.

1

Grounded Persona Memory

Personas do not spawn from thin air. Each has a persistent memory vector retrieved from past interactions. Responses are conditioned on retrieved memory + demographic context + world state news. This grounds answers in a simulated history, not a generic LLM prior.

2

Temperature Control for Honest Dissent

We intentionally run simulations at multiple temperatures (0.1–0.7). Lower temps produce conservative, critical responses. Higher temps produce exploratory enthusiasm. We surface both so you see the range, not just the average.

3

Explicit Barrier Extraction

Our analytics pipeline forces the model to code responses into Drivers AND Barriers. A hallucinating system would generate only positive drivers. We require both, weighted by frequency, so negative signal is preserved.

4

Raw Data Access

Every response, every reasoning string, and every rating is exportable to CSV. We do not hide behind aggregated dashboards. If a persona said something weird, you can read the exact reasoning and judge for yourself.

Still skeptical? Good.

Run a $2.99 test on a concept you already have real-world data on. If the simulated results do not match your known reality, we have failed — and you will know before spending $500.

Start $2.99 Validation Test

No credit card required for initial account. 20 credits included.