Resource
User simulation providers compared
Six providers were checked, alongside tools built on prompting general-purpose models. Five train models on data about people, and one uses general-purpose language models. Every published accuracy figure comes from the provider itself or its founders. They use different metrics, some for the audience as a whole and some for each person.
Yes
–No
?Not described
·Not mentioned on its own pages
Scroll sideways to see every provider
| Attribute | Aaru | Artificial Societies | Electric Twin | Listen Labs | Semilattice | Simile | Tools built on prompting |
|---|---|---|---|---|---|---|---|
| Note | Salesforce agreed to acquire Listen Labs in September 2026 | Such as Ask Rally, Evidenza, Minds and Synthetic Users | |||||
| Getting started | |||||||
| Price | Not publishedA pilot first, scoped with Aaru’s team | Not publishedA demo first | Not publishedA walkthrough first. A licence priced by audiences and seats | Not publishedA demo first for Listen Twins. Sign-up with a work email covers its interview research only | $139 / monthSign up on the site. Simulations on public user models start in November 2026; until then, Custom models only, through a call | Not publishedA request form first | From $20 / month (Ask Rally); Synthetic Users from $12,500 / year; Evidenza by proposal |
| Use it in | Web app | Web app | Web app | Web app | Web appMCPAPI | Web app | VariesMinds has an MCP server and an API |
| How it’s built | |||||||
| Prompted | No | No | YesPersonas in general-purpose language models, constrained by survey data | No | No | No | YesPersonas written into a prompt |
| Grounded | No | YesEach persona grounded in a real individual’s public data | ?Not describedNot described whether the survey data is per person | ?Not describedOne model per interviewee; method not given | No | YesEach agent grounded in the participant’s interview | No |
| Trained | YesIts own population model | YesRothko, trained further on decisions | No | YesEach twin trained on the client’s outcomes; method not given | YesA fine-tuned model per cluster of similar people | YesA fine-tune of Qwen | No |
| Detail | A population model, not a language model, built from census, transaction and behavioural dataAaru, 2026 | Its own model, Rothko. Personas grounded in real people’s public online data, connected in a networkArtificial Societies, 2026; Artificial Societies, Rothko | Personas constrained by survey dataElectric Twin, white paper | Listen Twins, one model per real interviewee, built from AI-led interviewsListen Labs | Anacreon, a fine-tuned model per cluster of similar peopleDahiya (Semilattice), 2026 | A fine-tune of Qwen, grounded in interviews and its own studiesSimile; Park et al. (Simile), 2024 | Persona promptsArgyle et al., 2023; Synthetic Users; Minds |
| Accuracy | |||||||
| Published figure | 3.53 points Mean absolute error, audience level Average over answer shares on 2,993 questions Aaru, 2026 | 86% Distribution overlap, audience level 1,000 surveys, against a 91% human ceiling Artificial Societies | 4 points Mean absolute error, audience level Best case, published as 96% (1 − mean absolute error). 92% on its stricter NDAM measure, against a 94% human ceiling Electric Twin | Not published Individual level Tested on held-out interview answers Listen Labs | 0.775 Ordinal alignment, individual level Median cluster, one customer audience; 67.9% exact match Dahiya (Semilattice), 2026 | 86% Ratio to human retest, individual level Interview and survey agents; 83% from interviews alone. 85% in the paper’s first version Park et al. (Simile), 2024 | Method not published Undefined percentage Figures such as “88% similarity” without a stated method Evidenza |
| Scores each simulated person against the real one | NoScored on the percentage choosing each answer | NoScored on the percentage choosing each answer | NoScored on the percentage choosing each answer | YesHeld-out interview answers | YesEach person’s newest records held out | YesAgainst each person’s own retest | ?Not described |
| Tests beyond survey questions | YesBrand rankings against card-purchase data | YesQuestion-wording experiments and coherence | YesFocus groups and image and video tests | ?Not described | NoSurvey items only; journeys not yet validated | YesPersonality tests, economic games and experiments | ?Not described |
| Reports how confident each result is | YesPredicts its own error for each question | ?Not described | ?Not described | YesA confidence score, and abstains without signal | ?Not describedAbstention described in the paper, not tested | YesA trained model predicts each result’s error | ?Not described |
| Simulates a new audience without new data | YesSimulates a described audience from census and licensed data | ?Not described | ?Not describedBuilt per audience, from customer or existing survey data | NoNeeds an interview with each person | NoNeeds a corpus for each audience | YesReturns answers for a described population | YesAny persona can be written into a prompt |
| Built from | |||||||
| Surveys and panels | Yes | Yes | Yes | Not mentioned | Not mentioned | Yes | Not mentioned |
| Interviews | Not mentioned | Yes | Not mentioned | Yes | Yes | Yes | Not mentioned |
| Public text | Yes | Yes | Not mentioned | Not mentioned | Yes | Not mentioned | Not mentioned |
| Census and official statistics | Yes | Yes | Not mentioned | Not mentioned | Not mentioned | Not mentioned | YesSynthetic Users |
| Behavioural records | Yes | Yes | Not mentioned | Not mentioned | Not mentioned | Yes | Not mentioned |
| Experiment responses | Not mentioned | Not mentioned | Not mentioned | Not mentioned | Not mentioned | Yes | Not mentioned |
| A customer’s own data | Yes | Yes | Yes | Yes | Yes | Yes | YesAsk Rally, Synthetic Users |
| Only what a language model learnt in training | Not mentioned | Not mentioned | Not mentioned | Not mentioned | Not mentioned | Not mentioned | YesAsk Rally, Evidenza |
| What it simulates | |||||||
| Surveys | Yes | Yes | Yes | Yes | Yes | Yes | YesAsk Rally, Evidenza |
| Question type: single choice | Yes | Yes | Not mentioned | Not mentioned | Yes | Yes | YesAsk Rally |
| Question type: scales (likert) | Not mentioned | Yes | Yes | Not mentioned | Yes | Not mentioned | YesAsk Rally |
| Question type: multi-select | Yes | Yes | Not mentioned | Not mentioned | Not mentioned | Not mentioned | Not mentioned |
| Question type: open-ended | Yes | Yes | Yes | Yes | Not mentioned | Yes | YesSynthetic Users, Ask Rally |
| Interviews | Not mentioned | Yes | Yes | Yes | Not mentioned | Yes | YesSynthetic Users, Evidenza |
| Focus groups | Not mentioned | Yes | Yes | Not mentioned | Not mentioned | Not mentioned | YesAsk Rally |
| Word of mouth and spread through networks | Not mentioned | Yes | Not mentioned | Not mentioned | Not mentioned | Not mentioned | Not mentioned |
| Using software and websites | Not mentioned | Not mentioned | Not mentioned | Not mentioned | Yesaccuracy not yet validated | Not mentioned | Not mentioned |
| Experiments and economic games | Not mentioned | Yes | Not mentioned | Not mentioned | Not mentioned | Yes | Not mentioned |
From each provider’s own pages and papers, October 2026. “Not described” means the provider does not say either way. An MCP server lets an AI agent such as Claude run simulations as part of its own work. Accuracy figures are self-reported and use different metrics, so they can’t be compared directly; see What a percentage can mean.