Modeling User Audiences — Part 1

The Problem With Most “AI User Simulation”

Most AI user simulation is structurally wrong. They start with a prompt like “Act like a product manager evaluating this feature” or “Pretend you are a consumer deciding whether to buy this product.” This sounds reasonable, but it hides a fundamental problem: it assumes user as a single archetype. In reality, users are collections of people with different backgrounds, experiences, knowledge levels, and decision styles. Two product managers evaluating the same tool might reach completely different conclusions—one prioritizes speed of execution, another prioritizes system stability, another worries about organizational risk.

Yet most simulations reduce this diversity to a single fictional persona. That works for storytelling, but it fails for simulation. If we want AI systems to simulate real audiences, we need something more structured than “pretend you are a user.” At Vectorial, we’ve been building systems that simulate how users react to products, messaging, and ideas, and very quickly we ran into the question: how do you represent an audience realistically enough for simulation? That led us to develop what we call the Tribe Framework.

This article is the first in a short series explaining how we model synthetic audiences. In this part we’ll cover why personas break down in simulation, what we mean by tribes, and the first two layers of audience modeling.

Personas Work for Storytelling, Not Simulation

Personas have been a staple of UX research for years. They typically look something like “Sarah — 34, Product Manager at a SaaS startup.” Personas help teams humanize users and think about their needs, but they come with an important limitation: they collapse entire user segments into a single individual. Real audiences are not one person—they are distributions of people.

Even within a tightly defined group like product managers at growth-stage SaaS companies, there is huge variation in career paths, technical exposure, company cultures, risk tolerance, and decision styles. When we simulate a persona, we are effectively modeling the average user, but the average user rarely exists. For simulation to produce meaningful insight, it must capture diversity within an audience, not erase it. This is where tribes come in.

Introducing Tribes

A tribe represents a structured audience segment rather than a single fictional user. It is not a single persona—it is a structured distribution of users. For example, instead of modeling “a product manager evaluating this analytics tool,” we might model a tribe defined by mid-seniority product managers, working at growth-stage SaaS companies, based in North America, and familiar with modern analytics platforms.

Each tribe represents a coherent segment of users, but not a single identity. The key difference is that tribes preserve internal variation, which allows simulations to produce multiple, divergent responses like a group of users, not a single averaged opinion.

The Vectorial Audience Modeling Stack

At Vectorial we model synthetic audience behavior using a layered approach. At the highest level, we think about responses like this:

Each layer answers a different question: Attributes define who the user is, Exposure defines what the user knows, Traits define how the user thinks, States define situational pressure on the user, and Response is what the user says or does. In this article we focus on the first two layers—Attributes and Exposure—which define who is in the room when we simulate an audience.

Attributes represent the objective characteristics of a user. These are the background factors that shape their environment and context. Because B2B and B2C audiences behave differently, the attributes we model vary slightly.

For business audiences, typical attributes include location, education, seniority, industry, company size or growth stage, and years of experience. One important design choice we made is not treating “function” as an attribute. Instead, the function (for example Product Manager, Designer, or Researcher) typically defines the starting audience, and attributes then refine that audience into meaningful tribes.

Consumer audiences rely on a different set of attributes. Typical examples include age band, income band, geography, life stage or family structure, gender, and education. These attributes help represent population segments realistically, but attributes alone are still not enough. Two people with identical demographics can behave very differently depending on their experience with the product category, which leads to the second layer.

Exposure captures a user’s experience with the ecosystem around a product. Two users might share identical attributes but differ dramatically in exposure. For example, one may have used several competitor products while another may be encountering the category for the first time. Typical exposure dimensions include product familiarity, competitor familiarity, and technical fluency.

Exposure shapes how informed a user’s evaluation is. Highly experienced users may evaluate subtle feature differences, while newcomers often focus on clarity, onboarding, and ease of understanding. This layer is essential for simulating how knowledgeable an audience is, not just who they are.

Why This Structure Matters

Most AI user simulations rely on simple prompts, but prompts struggle to represent diversity inside real audiences. By explicitly modeling attributes and exposure, tribes allow simulations to represent structured variation across users. Instead of generating one averaged answer, the system can simulate multiple audience segments with different backgrounds and different levels of category knowledge.

This shifts simulation from “a single opinion” to “a distribution of perspectives,” which makes simulations far closer to how real markets behave.

What Comes Next

Attributes and exposure define who the audience is and what they know, but identity alone does not explain behavior. Two users with the same background can still make very different decisions because they think differently. Traits help capture factors like how users verify claims, how they weigh cost vs convenience, how they seek information, and how they respond to risk and uncertainty. In other words, they model something far more important than demographics: how people actually think.

In the next article, we’ll introduce the psychographic layer of the model—Traits, the patterns that shape how users evaluate trade-offs, trust information, and make decisions.