Synthetic Users vs digital twins

What Actually Scales

‍

We get asked this a lot: what’s the difference between Synthetic Users and digital twins, and why didn’t we build a “twin” of every user?

‍

Short answer: digital twins don’t scale for user research. They try to model individuals one-by-one. Real-world audiences are high-dimensional and constantly changing. Enumerating every combination of traits, contexts, and behaviors is a losing game.

‍

Think of it like this: we don’t judge intelligence by how many answers are memorized; we judge it by how well it generalizes to new questions. The same principle applies to understanding customers.

‍

‍

Why digital twins break down

‍

1. Combinatorial explosion.
   Ten common attributes (location, income band, family status, retailer preference, brand loyalty, price sensitivity, skin concerns, dietary needs, device, channel) with just five levels each already yields 510=9,765,6255^{10} = 9,765,625510=9,765,625 combinations. Add context (season, promotions, life events) and it blows up further. You end up “collecting twins,” not understanding people.
   ‍
2. Drift.
   People change. Twins need constant re-training per individual or segment. That’s cost and latency, and it quietly decays accuracy.
   ‍
3. Data requirements and privacy.
   Faithful individual replicas often need sensitive, granular data. That’s heavy governance with questionable ROI, especially when you mostly need representative behavior, not a perfect clone.
   ‍
4. Brittleness.
   Instance-based models overfit to the data you have about that person. Move the scenario even slightly and they fail to generalize.
   
   ‍

‍

What we do instead

‍

We built Synthetic Users to model the space of likely behaviors rather than individuals. Technically, this means factorizing people into stable psychological traits, situational context, and domain knowledge. Then sampling and composing those factors to generate responses that match real-world distributions.

‍

‍

Our architecture (at a glance)

‍

1. LLM ensemble + routing.
   We use multiple large language models with different training pedigrees and biases. A lightweight router chooses which model(s) to use for a given prompt (e.g., “clinical interview with an oncologist in Berlin”). We also vary generation order and aggregation to reduce single-model bias and improve reliability.
   ‍
2. Personality & preference layer.
   We represent stable human variation with a Big Five (OCEAN)–style factor model, enriched by behavioral signals. Big Five isn’t “the only” personality model, but it is one of the most validated, cross-culturally replicated frameworks in psychometrics. We use it because it’s measurable, predictive at the group level, and composes well with context.
   Important nuance: traits influence tendencies; they don’t deterministically fix choices. That’s why we combine traits with situational variables.
   ‍
3. Your data, layered in via RAG
   Past interviews, surveys, CRM notes, and insight reports are mapped to the same factor space. This lets Synthetic Users reflect your audience while still generalizing to new questions and scenarios.

‍

Link to our architecture explained here.

‍

How this scales in practice

‍

• Coverage without enumeration.
  You don’t need a twin for “a parent in rural Ohio who shops weekly at Walmart, prefers fragrance-free, and has a toddler with eczema.” You sample a parent profile with relevant traits and constraints, add retail and dermatology context, and generate behaviors consistent with both.
  ‍
• Scenario agility.
  Change the dynamic parameters (seasonal allergy flare-ups, supply chain delays, a new price anchor) and you recompute behavior.
  ‍
• Bias control by shuffling between models.
  Ensemble methods plus explicit trait/context sampling reduce hidden skews you’d get from a single model or a single “perfect twin.”
  ‍
• Freshness.
  Because behavior is generated from factors + context, updates to either flow through immediately.

‍

‍

Try it (fast)

‍

The point isn’t to read a manifesto, it’s to compare outcomes.

1. Pick a recent study or interview guide.
   ‍
2. Run it with Synthetic Users (takes minutes).
   ‍
3. Compare against your human results on: directionality, rank-ordering of options, edge-case discovery, and time-to-insight.
   ‍

This is how we measure Synthetic Organic Parity.

‍

The proof is in the delta: faster iteration, better coverage, lower cost, and results that hold up when reality changes.

‍

Feel free to book a demo with us here.

‍

‍