What user simulation is for
User simulation is used in three ways. The first makes the research teams already do cheaper. The other two are things real panels are too slow or expensive for: testing decisions too small or too many to research with real people, and software that tries many versions of a product and keeps the best.
-
1
Lower the cost of research you already do
Run the market research, polling or user testing you already do on simulated people, and get answers in minutes or hours instead of weeks. The most you can save is what you spend on research today. The risk is a wrong answer that a real study would have caught.
- Concept and pricing tests
- Polling and policy research
- Usability and onboarding tests
-
2
Test decisions that never got user research
Most decisions are too small, too quick or too many to justify research with real people: a headline, a pull request, a pricing variant. With simulation, each one can be put to simulated people. An organisation can also try service, benefit or policy changes before paying for a pilot, including combinations too many to pilot one by one.
- Headlines and copy
- Pricing variants
- Benefit and service design
- Policy changes
-
3
Test many options and keep the best
If simulation is fast and can be called from code, software can write many versions of a product, message or policy, test each on simulated people, and keep the best. An AI agent writing copy or designing a page can do this before a person reviews its work, using an API or an MCP server (MCP is the protocol AI agents use to call outside tools).
- Agents testing their own drafts
- Checks inside product and marketing pipelines
What can be simulated
Most products simulate the research methods teams already use, most often surveys. A few simulate behaviour that no survey or interview records: how opinions spread between people, and how people use software.
Research methods
Surveys
Simulated people answer a questionnaire. Most products check the results for the audience as a whole, which is easiest. Some check each person, so results can be split by segment. In published studies, twins’ answers are less varied than the real people’s. Argyle et al., 2023; Park et al. (Simile), 2024; Peng et al., 2025
Question types in surveys
Each type of question needs its own way of scoring. A score built for scales, such as ordinal alignment, can’t score unordered choices or written answers. Single choice and scales are the most studied. None of the benchmarks here test multi-select. Open-ended answers are judged by comparing their themes and wording with real ones. Park et al. (Simile), 2024 (v1); Łajewska et al., 2026; Hämäläinen et al., 2023
Single choiceScales (Likert)Multi-selectOpen-ended
Interviews
A simulated person is interviewed. In studies, the themes were similar to real interviews, but the structure, tone and depth were different. Amirova et al., 2024; Kapania et al., 2024
Focus groups
Several simulated people discuss a topic, with or without a moderator. One published study compares simulated sessions with real ones. Zhang et al., 2024
Experiments and economic games
Simulated people take part in classic experiments and economic games, and the results are compared with the human results. Horton et al., 2023; Aher et al., 2023; Park et al. (Simile), 2024
Beyond answering questions
Word of mouth and spread through networks
Simulated people post, reply and share, and opinions spread from person to person. Research simulations have run with up to a million simulated people. Park et al., 2022; Yang et al., 2024; Gao et al., 2023
Using 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%. Lu et al., 2025a; Lu et al., 2025b
The providers page lists which providers offer each.