US Synthetic Respondent Platforms for Consumer Research 2026

Last updated: 8 October 2026

Quick Answer: The synthetic respondent platforms worth knowing for US consumer research in 2026 are Aaru, Evidenza, Fairgen, Perplica, QualiSynth, Simile and Synthetic Users. They sell three different products: prompted AI personas, AI twins grounded in real data, and synthetic boosts to a human panel. Across 285 published comparisons, synthetic and human samples agreed only 24.9% of the time.

A synthetic respondent is a large language model answering a survey or an interview as if it were a person with a stated profile. Its pitch is a panel that never drops out and scales in minutes. Its evidence is that it captures broad patterns, flatters well-known brands, compresses the diversity of real opinion, and in most published comparisons diverges from the humans it stands in for.

That does not make the category useless. It makes the buying question specific: which of three products is on offer, what real data grounds it, and which decisions it is fit for. The answer to the last one is narrower than most vendor pages suggest and wider than most skeptics allow.

Key takeaways

• Three products share one label: prompted personas, twins conditioned on real data, and synthetic rows added to a real survey. Grounding data decides which one you are buying.
• Published evidence agrees on a shape: synthetic samples reproduce direction and rank order reasonably well, overstate magnitude and positivity, and lose the spread of real opinion.
• Vendor-published accuracy figures on these sites range from 85% to 92% parity claims to error metrics on thousands of questions. Ask which humans, which questions and which metric.
• The 2025 ICC/ESOMAR Code requires researchers and clients to disclose when synthetic data played a significant role.
• Use synthetic reads to screen and pretest. Put launch, pricing and claims decisions on real respondents.

How This Guide Evaluates AI Synthetic Respondent Platforms

Each platform is described from its own website as of 8 October 2026 on four checks: the approach it sells, the real data it is grounded in, the validation it publishes, and who runs the study. Three products sit under the label:

• Pure AI respondents: a model prompted with a demographic or personality profile and nothing else. Strengths: cheapest and fastest. Limitations: no real data behind the answers.
• AI twins: a model conditioned on real data from the people it imitates, from a survey base to a brand's own interviews. Strengths: grounded in first-party evidence. Limitations: only as good, and as current, as that data.
• Augmented panels: a real survey with synthetic rows added to segments too small to read. Best for: thin cells in an existing quant study.

Standards now cover the category. The ICC/ESOMAR International Code, revised in 2025, defines synthetic data and synthetic personas and requires researchers and clients to disclose whether AI or synthetic data played a significant role in sampling, analysis or interpretation. The Insights Association's 2026 guidance on synthetic data adds that synthetic data should not be presented as equivalent to observed human responses without appropriate validation.

What Does the Research Say About Accuracy?

Three independent bodies of evidence published in 2025 and 2026 agree on the shape, and none of them is a vendor.

1. Literature review. The Nuremberg Institute for Market Decisions' guidelines for silicon samples, by Marko Sarstedt, Susanne Adler, Lea Rau and Bernd Schmitt, reviewed 285 silicon-to-human comparisons: 24.9% yielded similar results, 65.3% diverged and 9.8% aligned only partially. The authors recommend pretesting stimuli and survey items, generating large-scale data for quantitative work, and synthetic personas for exploratory qualitative insight.
2. Brand funnel experiment. NIM's digital twin study, led by Carolin Kaiser, found synthetic participants matched real choices about 79% of the time and reproduced the overall pattern, but overestimated selection of well-known brands, deviated by an average of 1.2 points on a seven-point scale and showed significantly less variation than real people.
3. Social science experiments. A Nature study of 70 US survey experiments found GPT-4 predictions strongly correlated with real treatment effects, yet the predicted effect sizes were roughly twice as large as the actual ones. Combining the model's forecasts with human forecasts was more accurate than either source alone.

Synthetic Respondent Platforms Compared at a Glance

The eight rows below cover the main approaches on sale in 2026, with Alchemic included as the adjacent real-respondent layer rather than as a synthetic vendor. Rows are alphabetical by platform name, and "not stated" means the vendor's site does not document it.

To set a synthetic read against real buyers, see how a WhatsApp-native interview study is fielded alongside it. Others in the market include Artificial Societies, GWI's synthetic audiences built on two million interviews a year across 53 markets, synthetic respondents inside Qualtrics, and digital twin features from Dovetail and Outset.

When a Synthetic Platform Is the Better Choice

Screening dozens of message or concept variants down to a shortlist in an afternoon is a job no human panel does at that speed, and Synthetic Users, Evidenza and Aaru are built for it. The same goes for:

• Pretesting a questionnaire or discussion guide for double-barreled questions and vague options, the first use NIM recommends.
• Rehearsing a concept on personas built from your own transcripts before fieldwork, which is Perplica's stated purpose.
• Pressure-testing a qualitative study design several times before funding it, QualiSynth's pitch.
• Reading a thin segment in an existing survey, Fairgen's Boost.

The human study belongs on the shortlist that survives, which is the argument made at length in synthetic users versus real respondents.

1. Alchemic: Best for the Real Respondents Behind a Synthetic Study

Alchemic is adjacent to this category, not part of it: one of its own product pages puts it as "real interviews, not synthetic answers." It sits on a synthetic shortlist as the human layer that grounds and checks a simulation. Its AI-moderated interviews reach people by text natively inside WhatsApp with voice notes, by outbound AI phone call to any working number, or by voice and video on the browser, in 60+ languages including Spanish, Hindi and Arabic. Recruitment is managed fieldwork or bring your own list, across 14 markets including the USA and the UK, and its stated benchmark is about three days from brief to live dashboard for a 200-interview qualitative study.

• Strengths: a knowledge base that persists across waves, giving a twin a first-party base that grows with every study; reach into the subgroups where synthetic error concentrates.
• Limitations: no published pricing; SOC 2 Type II is in progress rather than certified.
• Best for: grounding a twin on real interviews and validating a synthetic read before a decision.

2. Aaru: Best for Simulating Behavior at Population Scale

Aaru describes its product as behavior simulation at the scale of the real world, building simulated populations grounded in real-world behavior and outcomes. Its framing is that people answer when asked but act when simulated, so its pitch is predicting what people do rather than what they say. Listed uses include product innovation and feature testing, pricing, marketing messages, segmentation, scenario planning and competitive response. It publishes a mean total variation distance of 7.62% and a mean absolute error of 3.53% across 2,993 questions, and cites research with EY on investor preferences.

• Strengths: published error metrics at question scale and a behavior-first model.
• Limitations: population-level simulation answers aggregate questions better than it explains an individual's reasons.
• Best for: product, price and message decisions that need a fast directional read.

3. Evidenza: Best for Hard-to-Reach B2B Audiences

Evidenza sells AI market research for what it calls impossible audiences: surveying AI copies of a company's customers, including the hardest-to-reach B2B buyers. Users build a synthetic sample of thousands of customers for a product category, run a customizable survey and real-time interviews with it, then export a go-to-market plan. It claims 88% accuracy in 100+ validations against traditional research and says it is trusted by hundreds of enterprise brands. Listed uses include sizing addressable buyers, clustering by need, ranking category entry points and scoring ads.

• Strengths: speed on audiences that are slow and expensive to field, such as senior B2B decision-makers.
• Limitations: rare audiences are also the thinnest in any model's training data, so validation on a handful of real buyers matters most here.
• Best for: B2B marketing teams testing positioning and messages on senior buyers.

4. Fairgen: Best for Augmenting an Existing Survey

Fairgen calls itself an AI and synthetic data research suite. Boost generates niche data to read segments too small to analyze, Check spots and removes low-quality respondents, and Impute, listed as coming soon, will fill in missing answers. Its Twins product builds premium simulated audiences from a brand's own past studies, quantitative and qualitative, so the twin draws on that brand's own research rather than on general model knowledge alone. Its solutions are pitched at both brand teams and research providers.

• Strengths: works on top of research a team already owns, including data-quality checks.
• Limitations: a boost extends a real sample, so a weak or biased base survey carries its problems into the synthetic rows.
• Best for: insights teams with an existing quant program and thin subgroups.

5. Perplica: Best for Rehearsing on Your Own Research

Perplica builds AI personas from a team's own interview transcripts, briefs and documents, then lets the team interview them, run focus groups and test concepts. Every answer comes with the reasoning behind it and the source material that supports it. The company states in its product and terms that a persona is a rehearsal, not a replacement, for talking to real customers, and contrasts its approach with synthetic users built from a one-paragraph brief and general model knowledge.

• Strengths: auditable answers tied to the team's own evidence.
• Limitations: a persona can only be as current and as broad as the material it was built from.
• Best for: teams with a library of interviews that want to rehearse concepts before new fieldwork.

6. QualiSynth: Best for Pressure-Testing a Qualitative Study

QualiSynth targets independent, boutique and in-house qualitative researchers, pitching dozens of synthetic interviews for a fraction of fieldwork cost. It generates and interviews the segmented synthetic consumers a study needs, so a researcher can refine guides, messages and segments many times before funding the real fieldwork. The platform is self-serve: the researcher sets the brief, audience and interview guide, each synthetic respondent is checked against the segment definition, and every quote in the report traces back to its transcript.

• Strengths: cheap iteration on study design, with traceable output.
• Limitations: it creates respondents rather than analyzing real ones, so its findings are a dress rehearsal for the study, not a substitute for it.
• Best for: qualitative researchers tightening a guide or a segment plan before field.

7. Simile: Best for Enterprise Behavioral Simulation

Simile describes itself as a simulation company committed to training the safest, most accurate behavioral models, showing how and why customers, employees or populations respond to change. Every Simile population starts with real people: its models train on 2.9M responses from 210 social science experiments, its interview-grounded agents were validated against 1,000 real participants, and customer data such as loyalty histories can train custom models. It reports 85% of human self-retest accuracy and tags results with a predicted accuracy level. Named customers include CVS Health, Wealthfront, Banco Itaú and Suntory.

• Strengths: a published grounding base and per-result confidence labels.
• Limitations: built for enterprise programs, which is more model than a single concept screen needs.
• Best for: large organizations modeling customer or employee responses to change.

8. Synthetic Users: Best for Early Problem Exploration

Synthetic Users simulates how an audience thinks, decides and reacts, testing a product, message or policy on synthetic respondents in minutes. It runs a multi-agent architecture in which each participant develops a stable personality profile based on the OCEAN model and keeps context across every question. It reports 85% to 92% parity with organic research across what it calls independent comparison studies, varying by audience type and measured on thematic overlap, depth and qualitative alignment, and it publishes the methodology and the gaps.

• Strengths: speed, interview continuity and an openly stated view that real research stays essential for validation.
• Limitations: parity is a vendor-reported range that varies by audience, so ask which audiences and questions the comparisons covered.
• Best for: product and marketing teams exploring a problem space before spending research budget.

Where Synthetic Respondents Break

Synthetic respondents break on novelty, subgroups, magnitude and anything that depends on lived experience. Each failure is documented, and each maps to a decision a brand should not make on synthetic data alone.

• Novelty. A twin is conditioned on what people have already said. A product with no precedent or a market the training data underrepresents gives the model nothing to imitate.
• Subgroups and spread. NIM's twins lost the variation between respondents, and its guidelines list missing variance and skewed distributions among the known limits. A synthetic read of a first-generation buyer is where the error sits.
• Magnitude. The Nature study's twofold overestimate and NIM's 1.2-point positivity bias mean a synthetic study can rank concepts and still misjudge the size of the win. Purchase intent and go or no-go thresholds need a human number.
• The threat runs both ways. A paper in the Proceedings of the National Academy of Sciences built an autonomous synthetic respondent that passed 99.8% of 6,000 standard attention-check trials. A human panel that cannot tell its people from a model has the same validity problem as a synthetic one.

Where Does a Synthetic Panel Need Real Respondents?

In three places: as the grounding data a twin is conditioned on, as the validation sample its output is checked against, and as the source of any answer the model has no basis to give. The subgroups where synthetic error concentrates are often the same people a browser panel under-reaches, so the real sample has to reach them too, a selection effect examined in sample validity and who a mode leaves out. A practical validation sequence:

1. Run the synthetic read and record which segments and outcomes it ranks highest.
2. Field the same instrument with real respondents in the segments the decision depends on.
3. Compare distributions, not just means, because compressed spread is the most documented failure.
4. Check magnitude separately from rank order, since rank tends to hold while size inflates.
5. Report which findings rest on which source, as both codes now require.

For launch, pricing and claims decisions, a concept test with real shoppers or personas built from real interviews is the instrument the codes describe. The wider tool landscape, synthetic rows included, is mapped in AI market research tools for consumer insights.

Sources and Methodology

• Vendor facts: approach, grounding data, published validation and named customers were read from each vendor's own website on 8 October 2026 and are reported as each vendor states them.
• ICC/ESOMAR International Code (2025) and the Insights Association (2026): definitions, disclosure and validation standards.
• Nuremberg Institute for Market Decisions: the 285-comparison review and the digital twin brand funnel experiment.
• Nature (2026): LLM predictions of 70 US survey experiments.
• Proceedings of the National Academy of Sciences: the attention-check evasion study.