Big Five traits for synthetic customer simulations

A room of synthetic customers is only as useful as the personality spread inside it. If every simulated persona reasons the same calm, agreeable way, the debate goes flat, and a flat debate misses the exact objections you needed to hear: the skeptic who reads the refund policy, the worrier who fixates on the one thing that could break, the loyalist who will not switch no matter how good your pitch sounds. Personality variance is what pulls those voices into the open. The science that describes that variance is the Big Five.

The Big Five (openness, conscientiousness, extraversion, agreeableness, neuroticism, or OCEAN) grew out of the lexical hypothesis: the traits people care about get baked into language, so you can recover the structure of personality by factor-analyzing the words we use for each other. Decades of work, much of it by McCrae and Costa, kept landing on the same five dimensions across cultures and methods (McCrae & Costa, 1989). Each dimension maps to buying behavior a founder already worries about:

• Openness separates early adopters, who try the new thing because it is new, from habit loyalists who trust what already works and need a reason to move.
• Conscientiousness is the fine-print reader: the customer who checks the terms, plans the purchase, and, once committed, churns less because they honor the decision.
• Extraversion sets how loud a customer is in both directions, the word-of-mouth engine and the voice that makes a boycott carry.
• Agreeableness is the say-do gap in one trait. Agreeable people stay polite in a survey and go quiet in defection: they will not tell you the price stings, they just leave.
• Neuroticism is the source of risk and safety objections: refund anxiety, "what if it breaks", the friction that kills a checkout even when the product is fine.

These are not just intuitions. In a broad review, Ozer and Benet-Martinez cataloged how the Big Five predict consequential outcomes across work, health, and relationships (Ozer & Benet-Martinez, 2006). And when Soto ran a large preregistered replication of that trait-outcome literature, 87 percent of the associations held up (Soto, 2019). Traits predict behavior replicably in humans, which makes them a sound scaffold for persona diversity, though not a guarantee that a synthetic persona forecasts an individual buyer.

Why the Big Five beats MBTI for research

Founders know the sixteen-letter types, but the research community treats them as shaky. A large share of people land in a different type on retest, and Pittenger's review walks through why that makes the instrument a poor basis for prediction (Pittenger, 2005). The deeper problem is the boxes. MBTI forces continuous traits into either/or categories, and even its own dimensions, measured properly, come out as Big Five gradients rather than types. For simulating a market you want the gradient. A customer is 70 percent of the way toward skeptical, not "a skeptic".

Inside a PredictAible simulation

The cast stage writes each persona a role, a stance, and a short backstory, then assigns it a Big Five vector drawn from a deterministic spread that covers the whole trait space on purpose. That spread deliberately includes low-agreeableness skeptics and high-neuroticism worriers instead of clustering around the bland middle. At debate time, the two strongest traits on each persona turn into a short cue for how it argues, not what it knows. In the café run, that is what keeps the cranks cranky and the skeptics skeptical while the rest of the room negotiates. It also pushes back on a failure mode we cover in our piece on whether AI agents can mirror real people: models tuned to be agreeable make simulated rooms read calmer and more consensual than a real market. A deliberate temperament spread can reduce that flattening, though it does not remove model bias. Worth being exact about what this is: the Big Five vector is a behavioral overlay engineered to force disagreement, not a psychometric reading of any real person.

Honest limits

Traits are distributions, not destinies: a high-openness customer usually tries new things, not always. The personality a language model expresses is an approximation of human personality structure, not a person. Researchers have shown you can measure and even shape it in model outputs, yet it stays a simulation of the shape (Serapio-Garcia et al., 2023). And personality steers stated reasoning better than it predicts revealed behavior, so read a simulated room as a way to surface objections early, not as a forecast of the exact churn number. That is the same posture we take toward AI focus groups in general.

FAQ

Do AI agents really have personalities?

Not in the human sense. They express a personality structure you can measure and steer in their output. PredictAible assigns each persona a Big Five profile so that expression is deliberate and spread across the room, not left to the model's default.

Why not MBTI?

It is popular but psychometrically weak: low retest reliability, and categories where the science keeps finding gradients. The Big Five gives you continuous traits with replicated links to real outcomes, which is what you need to model behavior.

How does personality change simulation results?

It changes who objects and how hard. Tilt the room toward skeptics and worriers and the edge cases come out earlier and louder: price resistance, refund anxiety, switching friction. A uniform, agreeable room hides them.

Build a room of synthetic customers with a deliberate Big Five spread at /builder.