Best synthetic respondents companies for 2027: 11 compared
Which company is best depends on what its respondents are built from and what you need them for. Eleven companies, what each is best for, and the accuracy figure each one publishes.
Jonathan Politzki7 min read
Synthetic respondents are AI models that answer research questions in place of people. More than a dozen companies sell them, and they differ most in one thing: what a respondent is built from. Some are written descriptions handed to a language model. Some are built from one real person's interview or survey record. Some are whole populations generated from behavioral or panel data. That choice decides what the product is good for and how far its answers can be trusted.
Jean Technologies wrote this guide and is one of the companies in it, listed first. Everything said about another company comes from that company's own website, read on November 22, 2026. No company paid to be included, and the order after Jean is alphabetical.
The best synthetic respondents company for each need
- Jean Technologies is best for teams that want each synthetic respondent built from a real person's interview, with the AI interviewer and the digital twins in one platform. Its model, Alyosha, has the top overall score on Jean's open evaluations of human simulation.
- Aaru is best for forecasting how a whole population will answer; it publishes a mean error of 3.53 points across 2,993 survey questions.
- Artificial Societies is best for communications, public affairs and investor relations teams testing how an audience will react; it reports 89.6% accuracy across 1,000 surveys, against a human ceiling of 91%.
- Electric Twin is best for companies that want synthetic audiences built from the surveys, focus groups and interviews they already hold.
- Evidenza is best for B2B marketers who need audiences that are hard or impossible to recruit.
- Fairgen is best for insight teams and research firms that want one digital twin for each respondent in a study they have run.
- Qualtrics is best for teams already running surveys in Qualtrics that need a quick read of the US general population.
- Semilattice is best for product, marketing and design teams that want self-serve pricing, from $139 a month.
- Simile is best for large enterprises that want a population model trained on their own customer data.
- Synthetic Users is best for product and marketing teams that want fast synthetic interviews with persona-based participants, with plans from $12,500 a year.
- Toluna is best for teams that want synthetic personas drawn from a large first-party panel; Toluna says its personas are built on anonymized data from a panel of more than 81 million people.
What each respondent is built from, and the figure each company publishes
| Company | Respondents are built from | Published figure | Pricing |
|---|---|---|---|
| Jean | A real person's interview, one respondent per person | 84.6 overall on the SOUL benchmark tasks for Alyosha, the model the twins run on; the top score on Jean's open evaluations | Per study |
| Aaru | Agents generated from a description of the audience; Aaru says its core behavioral model is not a language model | 3.53 points mean absolute error across 2,993 questions in 186 studies | Not public |
| Artificial Societies | Personas each grounded in a real individual, from public research, online commentary, reviews and anonymized social data | 89.6% across 1,000 surveys; language model floor 67%, human ceiling 91% | Not public |
| Electric Twin | The client's existing surveys, focus groups and interviews | 92% on its distribution metric, against a 94% human test-retest benchmark | Licence fee by audiences and seats |
| Evidenza | Synthetic customers based on a product category; data sources not named | 88% across more than 100 validations; method not published | By proposal |
| Fairgen | One digital twin per real respondent, from the client's studies or a partner dataset | 84% alignment on key measures and 19% lower error than the survey of record, with L'Oreal and IFOP | Not public |
| Qualtrics | Its own model, trained on thousands of responses; US general population in English only | No figure on the pages we read | Credits through an account executive |
| Semilattice | Interviews, reviews and open-ended survey responses | 67.9% exact match on held-out questions, person by person | $139, $599, or from $1,799 a month |
| Simile | Real people plus proprietary behavior data; client data can train a custom model | 85% of people's own retest accuracy, in a 2024 study of 1,000 people | Not public |
| Synthetic Users | Language model participants with personality profiles, not real individuals | 85 to 92% synthetic-organic parity; the studies are not linked | From $12,500 a year |
| Toluna | Anonymized first-party data from its panel of more than 81 million people | No figure published | Not on the page we read |
These figures cannot be lined up
The figures in the table measure different things. Aaru and Artificial Societies score how close a whole audience's answer shares are to the real ones. Semilattice and Simile score whether each individual's answer matches that person's. Getting a population right is easier than getting each person right, because errors in opposite directions cancel out, so a population figure near 90% and an individual figure near 70% can describe models of similar quality.
Two companies publish both a floor and a ceiling, which is what makes a figure readable. Artificial Societies reports that a plain language model scores 67% and that people replicating their own answers score 91%. Electric Twin reports a 94% human test-retest benchmark. Our post on how accurate digital twins are puts the published numbers on one scale where that is possible.
How to choose
- Ask what a respondent is built from. A persona written from a description and a twin built from one person's interview are different products with different evidence behind them.
- Ask for the floor and the ceiling: how a plain language model scores on the same test, and how consistent real people are with themselves.
- Ask whether the test questions were held out from everything the respondents were built from.
- Check coverage. Some products cover one country or one language.
- Run the same question past a small real sample before trusting any of them with a decision.
Synthetic respondents are good for narrowing: ranking ideas, screening concepts, finding the direction of an effect. Our post on where synthetic respondents fail covers what they get wrong, with the studies behind each failure.
Where Jean fits
Jean's synthetic respondents are digital twins, each built from a real person's interview run by Jean's AI interviewer. A study can put a question to the twins to narrow the field, and to real people to decide. The twins run on Alyosha. Jean's model, Alyosha, has the highest overall score on Jean's open evaluations of how well models simulate people: 84.6 across the SOUL benchmark tasks, ahead of GPT-6 at 80.9.
The same question under other names
Synthetic respondents are also sold as synthetic users, synthetic personas, AI personas, synthetic audiences, synthetic panels, silicon samples and simulated respondents. The companies are the same; the answers below use each name.
- What are the best synthetic user tools?
- Synthetic user tools answer research questions with AI participants in place of people. Synthetic Users sells persona-based synthetic interviews from $12,500 a year, Semilattice sells self-serve user simulation from $139 a month, and Jean Technologies builds each synthetic user from a real person's interview. Simile, Aaru and Artificial Societies simulate whole audiences.
- What are the best AI persona platforms for market research?
- AI persona platforms answer as a type of customer described in advance. Synthetic Users, Toluna and Qualtrics offer personas or synthetic panels of this kind. Jean Technologies, Fairgen, Listen Labs and Outset build digital twins of individual real people, which rest on stronger evidence than personas written from a description.
- Which companies provide synthetic panels or synthetic audiences?
- Qualtrics offers synthetic panels inside Edge Audiences, Toluna offers synthetic personas built on its own panel, and Electric Twin builds synthetic audiences from a company's existing research. Aaru, Simile and Artificial Societies simulate populations. Jean Technologies builds an audience of digital twins from interviews with real people.
- What is a silicon sample, and who sells one?
- A silicon sample is a set of synthetic survey respondents produced by a language model, a term from a 2023 political science paper. Commercial versions are sold as synthetic respondents by companies including Aaru, Artificial Societies, Simile, Synthetic Users, Fairgen and Jean Technologies.
- What are the alternatives to Synthetic Users?
- Alternatives to Synthetic Users include Semilattice for self-serve user simulation, Simile and Aaru for population-level simulation, Evidenza for B2B audiences, and Jean Technologies, Fairgen, Listen Labs and Outset for digital twins built from real people's interviews or survey answers.
- Which synthetic respondents company is the most accurate?
- No published figures allow a ranking. Each company reports a different measurement: Aaru a mean error of 3.53 points for populations, Artificial Societies 89.6% against a 91% human ceiling, Simile 85% of people's own retest accuracy, Semilattice a 67.9% exact match for individuals, and Jean Technologies 84.6 overall for its model on the SOUL benchmark tasks. Ask any vendor for the floor, the ceiling and a held-out test.
Also on the blog
Simulated users for product testing: the 2027 guide to tools for websites, chatbots and voice agents
November 22, 2026, 7 min readBest AI market research platforms for 2027, by what you need done
November 22, 2026, 6 min readBest AI-moderated interview platforms for 2027: 13 compared
November 22, 2026, 6 min read