What it’s for
User simulation is used in three main ways: to run research people already do, faster; to test changes before committing to a pilot; and to give software the ability to ask simulated people before it acts.
Research, faster and cheaper
The same questions a market research, polling or user testing study would ask, answered in minutes or hours rather than the weeks a panel takes.
- Concept and pricing tests
- Polling and policy research
- Usability and onboarding tests
Less than 1/10th the cost of a survey of humans
Aaru
…get real answers in seconds, not weeks
Electric Twin
Feedback in Minutes not Months
Ask Rally
6 months → 6 hours
Evidenza
Testing changes before a pilot
Service, benefit and policy changes can be tried on simulated customers before an organisation commits to a costly pilot or launch, including combinations too numerous to pilot.
- Benefit and service design
- Reminder and outreach cadences
- Policy changes
…to see which combinations increase intent to refill or adopt clinical services–before running costly pilots
Simile
Feedback for AI agents
An agent that writes copy, designs a page or plans a launch can ask simulated people before a person reviews its work. That needs an API, or an MCP server (the protocol AI agents use to call outside tools), and results in minutes.
- Agents testing their own drafts
- Checks inside product and marketing pipelines
From cheapest to most ambitious
-
1
Lower the cost of research you already do
Running existing studies on simulated people instead of panels makes them faster and cheaper. The gain is capped by today’s research budget, and the risk is a wrong answer where a real study would have caught it.
-
2
Research decisions that never had any
Most decisions are too small, too fast or too numerous for a panel: a headline, a pull request, a pricing variant. Simulation can put a question to simulated people for each of them.
Would simply have been impossible with traditional market research
Artificial Societies, a customer quoted on its site…for the decisions that would otherwise be made on instinct because a research study takes weeks
Semilattice…testing what hasn’t happened yet
Simile -
3
Search the space of options
If simulation is fast enough and can be called from code, software can generate many versions of a product, message or policy, test each with simulated people, and keep the ones that do best. That depends on simulated answers staying accurate for options no real person has seen, which no published study has yet tested at that scale.
What can be simulated
Most products predict survey answers, either for an audience as a whole or for each person in it. Some also simulate interviews, group discussions, how opinions spread, or how people use software.
Survey answers, at the audience level
The simulated sample reproduces the share of people choosing each answer. Nobody has to be matched one to one, so this is the most common and the easiest to validate. Shares are what sizing, pricing and polling decisions use.
Research: Argyle et al., 2023; Santurkar et al., 2023Survey answers, at the individual level
Each simulated person answers as one individual, either a twin of a real person or a simulated individual with consistent traits, and is scored against real people’s own answers. Results can be segmented and re-cut. Twins in published studies vary less than the people they copy.
Research: Park et al. (Simile), 2024; Peng et al., 2025; Bisbee et al., 2024Question types
Each type needs its own measure. Single choice and scales are the most studied. None of the benchmarks reviewed here evaluates multi-select, and open-ended answers are judged by comparing themes and wording with real ones.
Research: Park et al. (Simile), 2024 (v1); Łajewska et al., 2026; Hämäläinen et al., 2023- Single choice
- Aaru Artificial Societies Semilattice Simile Ask Rally
- Scales (Likert)
- Artificial Societies Electric Twin Semilattice Ask Rally
- Multi-select
- Aaru Artificial Societies
- Open-ended
- Aaru Artificial Societies Electric Twin Listen Labs Simile Synthetic Users Ask Rally
Interviews
A simulated person answers a semi-structured interview. In studies the themes resemble real interviews, while structure, tone and depth differ.
Research: Amirova et al., 2024; Kapania et al., 2024Focus groups
Several simulated people discuss a topic, with or without a moderator. One published study compares simulated sessions with real ones.
Research: Zhang et al., 2024Word of mouth and spread through networks
Simulated people post, reply and share, and opinions spread through the network between them. Research simulations reach a million agents.
Research: Park et al., 2022; Yang et al., 2024; Gao et al., 2023Using software and websites
Simulated people with personas try to complete tasks on a live site or prototype. Against 31,865 real shopping sessions, prompted models matched the next human action 12% of the time, and a fine-tuned model 17%.
Research: Lu et al., 2025 (UXAgent); Lu et al., 2025Experiments and economic games
Simulated people take part in classic experiments and games, and their treatment effects are compared with human results.
Research: Horton et al., 2023; Aher et al., 2023; Park et al. (Simile), 2024A provider is listed if its own pages describe the behaviour. Dashed names are tools built on prompting general-purpose models.