User Simulation Primer

A primer on
user simulation

User simulation puts questions about a feature, a price or a message to simulated people instead of real ones. Answers come back in minutes or hours rather than weeks, so ideas that would never have justified a study can be tested too. It goes by at least seven names, and the evidence on it looks contradictory: providers publish accuracy figures as high as 96%, while a mega-study of 19 pre-registered studies found digital twins only slightly better than a model with no persona. Both can be true, because they measure different things on different kinds of model. This primer is for product and marketing teams deciding whether to use it.

In this primer
  1. 00 Introduction The names, how it works, and how accurate it is
  2. 01 Data What the models are built from
  3. 02 Modelling How the models are built
  4. 03 Accuracy How close the answers are
  5. 04 Uses What it’s for
  6. 05 History From life tables to language models
  7. 06 Providers How the providers compare

Written by Semilattice, one of the six providers covered. Providers are listed alphabetically, and every claim about research or a provider links to its source.

The names

Why there are so many names

The names come from different fields, and most of them describe the same kind of product. Few say how the product is built, and how it is built decides how accurate it is and which published critiques apply to it.

Named for what you get

These name the output. A product under any of them can be prompted, grounded in a person’s data, trained, or a combination.

Synthetic respondents
Market research
Simulated people who answer a survey.
Synthetic users
UX and product research
Simulated users of a product, answering questions or using software. Synthetic Users is also the name of one prompting tool.
Synthetic data
Statistics and machine learning
The answers these models produce, used in place of data collected from people. The same term is also used for artificial copies of datasets made to protect people’s privacy.

Named for how it is built

These describe a method, and research on that method applies to products that use it.

AI personas
UX and marketing
A description of a person written into the prompt of a general-purpose model.
Silicon samples
Political science
A language model given demographic backstories to stand in for a survey sample.
Argyle et al., 2023
Generative agents
Computer science
Language-model agents with a memory that act as a character or, more recently, a specific real person.
Park et al., 2023; Park et al. (Simile), 2024

Named for who is simulated

Says that each simulated person is a copy of one specific real person.

Digital twins
Marketing science and engineering
A model of one specific real person, built from their own answers.
Toubia et al., 2025

The Modelling chapter explains the three ways a model is built.

How it works

How a simulation is built and checked

Most products follow these four steps, and results take minutes or hours. They differ in the data at step 1 and the method at step 2, and those decide how close the answers come at step 4.

  1. 1People

    Data about real people, such as survey answers, interviews, reviews or purchases. Tools built on prompting skip this step and rely on what a language model learnt in training.

  2. 2Model

    A model is prompted with a description of a person, given one real person’s data, or trained on many people’s records.

  3. Example
    3Ask

    You put a new question to the simulated people. Each answers as one person, and the answers are counted like a survey.

  4. Example
    4Check

    To measure accuracy, the simulated people are asked questions real people have already answered, with the real answers held back, and the two are compared.

Accuracy

How accurate is it?

Individual level

Each simulated person’s answer is compared with the real person’s own answer. Two of the six providers report this, as do the research papers on twins and trained models since 2024.

Ordinal alignment
Questions with ordered answers, such as scales
0.72 to 0.78
Out of 1, in research papers. Random answers score about 0.6, and the same people asked again score 0.82.
Toubia et al., 2025; Peng et al., 2025; Dahiya (Semilattice), 2026
Exact match
Any multiple-choice question
66% to 69%
Share of answers exactly right. On the General Social Survey chance is 27%, and the same people asked again matched 81%.
Park et al. (Simile), 2024 (v1); Park et al. (Simile), 2024; Dahiya (Semilattice), 2026

Audience level

The share of simulated people choosing each answer is compared with the real share. Three of the six providers report this.

Mean absolute error
Answer shares
3.5 to 4 points
The average gap between simulated and real shares, from one provider’s average and another’s best case. A model can match shares while giving most individuals the wrong answer.
Aaru, 2026; Electric Twin

Tools built on prompting

Products that write a persona into the prompt of a general-purpose model.

Method usually not published
“88% similarity”
Figures like this are published without the formula behind them, so they cannot be checked or compared.
Evidenza

Models are prompted, grounded in a person’s data, or trained, and each widely cited critique tested one of these. Whether a critique applies to a product depends on how the product is built.

The Accuracy chapter lists each critique under the way of building a model it tested.

Chapters

What each chapter finds

  1. 01 Data What the models are built from Data can record what people say, who they are, or what they do. Each provider uses a different combination.
  2. 02 Modelling How the models are built A model is prompted, grounded in one person’s data, or trained on people’s records, and each simulated person is either a twin of a real person or a simulated individual.
  3. 03 Accuracy How close the answers are Research papers report 0.72 to 0.78 on ordinal alignment, the most common individual-level score, against 0.82 for the same people answering again. Each widely cited critique tested one of the three ways of building a model.
  4. 04 Uses What it’s for Running existing research faster, testing changes before a pilot, and letting software ask simulated people before it acts. Most products predict survey answers; some simulate interviews, discussions or software use.
  5. 05 History From life tables to language models Statisticians have predicted populations since 1662. After 2020, language models made it possible to try predicting each individual.
  6. 06 Providers How the providers compare Six providers were checked. Five train models on data about people, and one uses general-purpose language models. Every published accuracy figure comes from the provider or its founders, and they use different metrics at the audience or the individual level.