How Sunzu simulates human behaviour.

We identify customer groups supported by the supplied research and documents. Interviews, surveys, support themes and other customer evidence help establish their circumstances, motivations and constraints.

Research by Park and colleagues found that agents grounded in interviews or surveys predicted responses more accurately than agents based on demographics alone.¹⁰ This supports the approach of grounding simulations in customer evidence. It does not independently validate Sunzu’s implementation.

Findings extracted for the audience model retain source links so reviewers can check them and update the model when evidence changes. Relevant public reviews, forums, news and published studies provide additional context. Findings are checked against customer evidence and assigned confidence ratings.

Sunzu’s behavioural corpus contains 1M+ profiles, 34M+ attributes and 57M+ data points, according to our internal research. It supplements the context available from customer evidence. These figures describe the underlying evidence base, not recruited participants or the sample size of a benchmark. The underlying proprietary data is not publicly disclosed.

Broader evidence can suggest what to investigate when customer data is sparse. Its relevance depends on the audience and question. Source links and confidence ratings support review; they do not, by themselves, establish that a simulated answer is correct.

The audience model has four components. These describe its contents, rather than the 20+ layers of the AI system:

Coherence and alignment assess consistency with the audience evidence. Predictive validity asks how closely findings match what people report or do.

Back-testing informs model refinement. For your own evaluation, keep the human findings used to judge results separate from the material used to build or adjust the audience.

Coverage of reference findings. The percentages describe themes covered, insights found or defects detected. The human reference is set to 100% to express relative coverage. That baseline does not imply that the reference research captured every possible finding.

Coverage and correctness. Recovering a reference finding and avoiding an unsupported finding are different measures. Coverage alone does not state how many additional findings were incorrect, how serious a missed defect was or whether a proposed decision improved a business outcome.

Sample equivalence. The figures shown are Sunzu’s reported estimates of result equivalence within these comparisons. They are not counts of recruited participants, a statistical margin of error or a guarantee that another study will produce the same result.

Research time. Minutes, days and months describe the compared workflows. Assess total elapsed time for your own process, including preparation, setup and review. The two-week interval belongs to the depth-interview retest reference, rather than the duration of a Sunzu study.

Proprietary research. The underlying datasets and detailed evaluation protocols are proprietary and are not published here. The results are reported by Sunzu from its own research. The external publications cited in this paper provide context for the approach; they are not independent audits of these results.

Application to your audience. A benchmark is a starting point for evaluation. Compare findings with your own human evidence, review misses and unsupported additions, and assess whether the simulation helps your team reach a useful decision.

Work directly with people when a decision depends on lived experience, accessibility, cultural context or behaviour that the available evidence does not capture. Unknowns need to remain visible when the evidence cannot support a conclusion.

Some studies show promising results on specific tasks. Maier and colleagues reported purchase-intent predictions close to human test-retest reliability for personal-care product surveys.¹¹ That is evidence for their method and dataset, not a guarantee for every audience or research question.

Research on designers’ use of persona-based chatbots also examines how these tools fit into design practice.⁶ We use audience models alongside human research, with their sources and limits visible to the team.