research

Best AI for Synthetic Research & Audience Testing

Compare the best synthetic research, synthetic audience and AI persona tools in 2026 — Simile, Aaru, Artificial Societies, Synthetic Users, GWI Spark, Fairgen, Listen Labs and more — what each does, funding, pricing, and the accuracy evidence on whether synthetic respondents actually work.

Updated June 2026

Quick answer: “Synthetic research” covers five different things, and conflating them is the costliest mistake buyers make. For genuinely synthetic audiences (AI respondents, no humans), the most credible players are Simile — the Stanford spin-out behind the foundational “generative agents” research, which raised $100M in February 2026 — and Aaru, which leads large-scale audience and behaviour simulation. For synthetic data grounded in real first-party panels, GWI Spark and Qualtrics Edge are the safest enterprise picks, and for boosting thin survey segments, Fairgen is the most statistically defensible. For fast, self-serve concept and persona exploration, Synthetic Users and Ditto lead. The caveat that matters more than any tool: the best independent study (Stanford, 2024) shows AI agents can match a person’s own survey answers about 85% as well as that person does two weeks later — but only when built from rich real interviews, and peer-reviewed work also shows synthetic respondents collapse variance and can flip the sign of key relationships. The consensus across ESOMAR, GreenBook, Bain and the major agencies is the same: use synthetic research to explore and augment, then validate with real humans before you decide.

One distinction underpins everything below. Many tools marketed as “synthetic” — including the best-funded ones, like Listen Labs and Outset — actually interview real humans with an AI moderator. That is not synthetic data at all, and it is arguably the most reliable AI-research category. This guide separates the two cleanly. For traditional AI assistants used in research workflows (literature review, analysis, summarisation), see our best AI for research guide; this page is about simulating, augmenting or scaling the audience itself.

The current state of synthetic research: June 2026

Synthetic research has moved from a fringe experiment to one of the most-hyped — and most-contested — categories in market research, in barely 18 months.

The pull is obvious. Survey response rates keep falling, panel fatigue is rising, and a real study costs weeks and thousands of dollars. Synthetic methods promise directional answers in hours at a fraction of the cost. In Qualtrics’ 2025 trends survey, 71% of researchers agreed that within three years synthetic responses will make up more than half of data collection (Qualtrics) — a vendor-sponsored sentiment figure, but a telling one. Synthetic data jumped to a top-three industry trend in a single cycle of the GreenBook GRIT report.

The money has followed. Simile raised $100M in February 2026 led by Index Ventures, with Bain Capital Ventures, Fei-Fei Li and Andrej Karpathy backing the Stanford team that wrote the foundational research (Bloomberg). Aaru raised a $50M+ Series A in December 2025 with Accenture investing (TechCrunch). On the real-human side, Listen Labs reached a reported $500M valuation in 2026 (Forbes). And the incumbents have moved: YouGov acquired synthetic pioneer Yabble in August 2024 (Research Live), and Qualtrics, GWI and Kantar have all shipped synthetic products.

But the validity question is unresolved, and it is the whole story. Two pieces of evidence frame the debate:

• The optimistic anchor: Stanford’s “Generative Agent Simulations of 1,000 People” (Park et al., 2024) built agents from two-hour interviews with 1,052 Americans and found they reproduced each person’s own General Social Survey answers about 85% as accurately as that person did on a two-week re-test, with less bias than demographic-only methods (Stanford HAI). The catch researchers stress: that fidelity depended on rich real-human interview data per person — the opposite of “skip the humans.”
• The warning: Bisbee et al., “Synthetic Replacements for Human Survey Data? The Perils of Large Language Models” (Political Analysis, 2024), found that while synthetic means looked plausible, the variance was drastically too small (overconfidence that breaks statistical power), subgroups were distorted, and several regression coefficients came out wrong-signed — meaning substantive conclusions would reverse (Cambridge).

In short: synthetic research can reproduce averages but routinely misrepresents the spread, the subgroups and the relationships — which is where most research value lives. That is why every credible body lands on “augment, don’t replace.”

The five kinds of “synthetic” research

The single most useful thing to get right is which category a tool belongs to. They differ enormously in how trustworthy they are.

Read it as a reliability ladder: augmentation (anchored to real data) and AI-moderated interviews (real humans) sit well above pure synthetic respondents and personas. The job of this guide is to help you pick the right rung for the decision you are making.

The synthetic research landscape (June 2026)

The core “genuinely synthetic” tools — AI audiences, personas and simulations with no humans in the loop. Funding and pricing are current as of June 2026; where a figure is not public we say so. Most accuracy claims here are vendor self-reported and not independently validated — treat them accordingly.

A few honest caveats on this table. Simile is the best-resourced and most academically credible, but it is very new as a product, with little public pricing or independent validation. Aaru’s “$1B headline valuation” is an artefact of a multi-tier deal structure, not a clean post-money figure (TechCrunch). And the self-reported accuracy numbers some vendors quote (Ditto’s “92% overlap with focus groups,” Lakmoos’ “98%+ similarity”) have not been peer-reviewed — the only independent benchmark in the category remains the Stanford study. Below the leaders sits a long tail of newer, lower-cost self-serve entrants (such as testfeed.ai) aimed at founders and small teams; they are cheap and quick but unproven, with no published validation, so treat their output as a first directional signal.

The synthetic leaders up close

Simile — the most credible new entrant

Simile is a Stanford spin-out from the team behind the foundational “generative agents” research (Joon Park, Michael Bernstein, Percy Liang). It builds behavioural “digital twins” that simulate how specific people respond to product, pricing and feature changes, combining qualitative interviews with transaction and behavioural data; early work includes CVS Health. Its $100M February 2026 round led by Index Ventures makes it the deepest-funded, most academically credible bet in the category — but it is very new as a commercial product, with little public pricing or independent validation yet (SiliconANGLE).

Aaru — large-scale audience and behaviour simulation

Aaru generates large populations of AI agents to predict how groups react to products, messaging, policy or events, positioning itself as a “prediction engine” rather than a survey tool, at very low per-simulation cost. It raised a $50M+ Series A led by Redpoint Ventures in December 2025, with Accenture both investing and partnering (Accenture). The forecasting framing invites scrutiny that softer survey use-cases would not, and the headline valuation is a deal-structure artefact — but at scale and price, nothing else matches it.

Synthetic Users — the self-serve qual pioneer

Synthetic Users runs qualitative “interviews” against AI personas built on OCEAN personality profiles, routing across multiple models to reduce single-model bias, at roughly $2–27 per interview with a free trial. It is an early mover and the most accessible entry point for synthetic qual — but pure-LLM personas draw the strongest “sycophantic and flattened” criticism, and its funding and true scale are not publicly disclosed.

Evidenza — built for B2B

Evidenza generates dynamic B2B buyer personas that take qualitative interviews or quantitative surveys, and even “clones” marketing-science thinkers as synthetic CMOs and CFOs. Founded by ex-LinkedIn B2B Institute figures, it has a Dentsu partnership for media planning and a credible marketing-science niche (Dentsu) — though its funding amount is undisclosed and B2B personas are inherently harder to validate.

Ditto — always-on personas in your design tools

Ditto offers 300,000+ census-calibrated personas with self-serve plug-ins for Figma, Canva and Framer, priced at roughly $50,000–75,000/year for unlimited studies — one of the few pure-plays with public pricing. Its “92% overlap with focus groups” figure is self-reported, not independently verified, and the company also publishes a market map of its rivals, so read its commentary with that in mind.

Artificial Societies — networked audience simulation

Artificial Societies builds networks of AI personas to simulate how a post, message or campaign spreads across an audience’s social graph — modelling collective dynamics rather than one respondent at a time. It is a smaller, earlier-stage company than the funding leaders above: backed by Point72 Ventures out of Y Combinator (W25) with about $4.9M raised, it reports 15,000+ users and 100,000+ simulations run (EU-Startups). Its networked approach makes it the most differentiated option for pressure-testing messaging and campaigns, though, like all pure-synthetic tools, its validation is largely anecdotal.

Grounded synthetic: tools anchored in real first-party data

The strongest defence against “made-up data” is to build the synthetic layer on top of real, proprietary human responses. These incumbents lead that approach.

GWI Spark — synthetic audiences on real survey data

GWI builds synthetic personas and focus groups on its own bank of roughly two million annual surveys across 53 markets — “what real people told us, not what a model inferred.” Its Agent Spark runs inside ChatGPT, Claude and Copilot, and the Spark API launched on 30 September 2025 (Research Live). It is enterprise-priced and bounded by GWI’s market coverage, but it has the clearest “grounded in real first-party data” story in the category.

Qualtrics Edge — synthetic responses inside the survey platform

Qualtrics Edge Audiences simulates consumer behaviour using public data plus its 25-year response repository and a fine-tuned model, and pitches a “human plus synthetic” workflow. Qualtrics reports that 87% of synthetic-response users are highly satisfied and claims up to 50% cost reduction — vendor figures, but backed by genuine proprietary data and enterprise trust (TechTarget).

Yabble and Panoplai — established insights, synthetic layer

Yabble (now part of YouGov) offers “Virtual Audiences” from a subscription starting under $800/month, blending LLMs with real behavioural data. Panoplai builds “digital twins” from verified, opted-in respondents and won GreenBook’s 2026 Insight Innovation Industry Impact Award; its 91–97% accuracy claim is self-reported.

Data augmentation: the most defensible synthetic method

If you only adopt one synthetic technique, this is the one with the firmest statistical footing — because it starts with real human data.

Fairgen — boosting thin survey segments

Fairgen’s “Fairboost” learns relationships across a real survey and extrapolates extra responses for under-sampled segments, roughly doubling a subgroup’s effective sample without re-fielding. It raised an $8M seed in May 2024 led by Maverick Ventures Israel (TechCrunch). Crucially, it is narrow by design — it augments a real base survey rather than inventing a standalone audience, which is exactly why statisticians are more comfortable with it. Kantar offers a similar boosting approach inside brand-health tracking.

AI-moderated interviews: real humans, not synthetic (but often confused)

These are the best-funded companies in the broader space — and they research real people. The AI runs the interview and the analysis; the respondent is human. They belong here as a clearly-labelled contrast, because for many decisions they are the more reliable choice.

One comparative study (Glaut) reported AI-moderated interviews produced 129% more words per response than surveys, with far less “gibberish” (Listen Labs). If you need qualitative depth and can accept the cost of real participants, this category — not synthetic respondents — is usually the better answer.

The counter-movement: “proof of human”

Worth knowing, because it sharpens the validity question. Some companies are betting against synthetic respondents.

• Prolific sells access to vetted real participants and launched a “100% Human Guarantee,” arguing that synthetic substitutes erase authentic and minority voices (the “surrogate effect”).
• Roundtable started with a “synthetic humans” tool, then pivoted to “Proof of Human” — detecting bots and AI-generated survey responses, noting that up to ~30% of B2B survey responses can be AI-generated.

When a company abandons synthetic respondents to police them instead, it tells you something about where the unsupervised version of this technology stands today.

Best for: segmented recommendations

The right tool depends entirely on the decision in front of you.

Best for behaviour prediction and frontier credibility

Winner: Simile — the Stanford team that wrote the foundational research, the deepest funding, and a focus on predicting individual decisions. Watch for independent validation as it matures.

Best for large-scale audience and outcome simulation

Winner: Aaru — built for population-scale prediction at very low per-simulation cost, with a real (if scrutinised) forecasting track record.

Best for messaging and virality testing

Winner: Artificial Societies — uniquely simulates a networked audience, so it models how a message spreads, not just isolated reactions.

Best grounded in real first-party data

Winner: GWI Spark (or Qualtrics Edge if you already run Qualtrics) — synthetic layers built on genuine proprietary survey data, the safest enterprise option.

Best for boosting thin survey segments

Winner: Fairgen — the most statistically defensible method, because it augments real data rather than replacing it.

Best for B2B research

Winner: Evidenza — purpose-built B2B personas and marketing-science framing, with agency distribution through Dentsu.

Best for qualitative depth (real humans)

Winner: Listen Labs or Strella — if you need genuine human nuance, use AI-moderated interviews, not synthetic respondents.

Best for fast, self-serve concept and persona testing

Winner: Synthetic Users or Ditto — self-serve synthetic panels that return directional reads on concepts, personas, ads and copy in minutes to hours, at a fraction of panel cost. Whatever you use, treat the output as a first read to confirm with real people, not a verdict.

How accurate is synthetic research, really?

This is the question that decides whether you should trust any of these tools — and the honest answer is “it depends heavily on what you ask of it.”

Where it holds up. On calibrated quantitative trends and on testing names, claims, packaging and price points, synthetic methods often land within a usable range — and the Stanford study’s ~85% fidelity is a real result for agents built from rich individual data (Stanford HAI). Data augmentation that extends a real survey is the most reliable use of all.

Where it breaks. Several failure modes are now well-documented:

• Variance collapse. Synthetic samples are “uncannily precise” — they understate the natural spread of human opinion, producing overconfidence that can break power calculations (Cambridge).
• Wrong-signed relationships. In the same study, some regression coefficients flipped sign versus real data — the kind of error that reverses a business decision.
• Identity flattening. Models build one-dimensional, stereotyped personas per demographic and structurally miss within-group diversity (Wang, Morgenstern & Dickerson, Nature Machine Intelligence, arXiv).
• Sycophancy. Persona prompts supply user context, which tends to increase a model’s tendency to tell you what it thinks you want to hear.
• No grasp of true novelty. Synthetic respondents are backward-looking — good at “does this match known patterns,” poor on genuinely new concepts, and weak on sensitive or thinly-documented topics (NIQ).
• Willingness-to-pay is unreliable. Harvard Business School researchers found LLM willingness-to-pay estimates “sometimes comparable… but often inaccurate and in some cases wrong-signed,” and fine-tuning did not fix new categories or between-segment differences (HBS).

There is also no agreed accuracy standard. “90% accurate” claims rarely come with peer review, and results are sensitive to prompt wording and even model drift over time. Methods researchers recommend include Train Synthetic, Test Real and holdout validation — ask any vendor which they use.

What the industry actually thinks

The striking thing is how much the standards bodies, consultancies and incumbents agree.

• ESOMAR has updated its code to explicitly distinguish a real person from a “synthetic, virtual or digitally created persona,” published buyer guidance, and set a “Minimum Viable Data” threshold — augmentation fails without enough real data behind it (ESOMAR recap).
• GreenBook tracks synthetic as a top trend but publishes pointedly balanced guidance — the widely-shared line is “synthetic sample is not the market; decision-grade data is” (GreenBook).
• Bain calls synthetic personas “force multipliers, not foundations — they should supplement, not supplant, direct engagement with real customers” (Bain).
• NIQ and Kantar frame it as bounded augmentation, trustworthy mainly when the supplier holds real validating data (NIQ).
• Gartner places synthetic data prominently on its AI hype cycle — strategically important and hype-prone in equal measure.

The shared playbook that emerges: use synthetic research early and for exploration, ground it in real data, validate the survivors with real humans (often n≈300–500), avoid it for novel or sensitive topics, and never present exploration as if it were validated fact.

Recent developments (2024–2026)

Simile raises $100M (Feb 2026). The Stanford “generative agents” team — including Joon Park, Michael Bernstein and Percy Liang — raised a $100M round led by Index Ventures to build behavioural digital twins, the largest single bet in the category (SiliconANGLE).

Aaru’s $50M+ Series A with Accenture (Dec 2025). Accenture both invested in and partnered with the prediction-engine startup, signalling enterprise-services interest in large-scale simulation (Accenture).

The AI-moderated interview round (2025–26). Listen Labs ($69M, $500M valuation), Outset ($30M Series B), Strella ($14M Series A) and Genway ($6M, a16z) all raised to scale real-human interviews — the better-validated adjacent category.

YouGov buys Yabble (Aug 2024). A major panel incumbent acquiring a synthetic pioneer marked the moment the establishment took the category seriously (YouGov).

GWI launches the Spark API (Sep 2025). Synthetic audiences grounded in real first-party survey data, distributed inside the major AI assistants.

Pricing: what you’ll actually pay

Pricing in this category is unusually opaque — many tools are demo-and-quote only. What is public, in USD:

As a rule, self-serve persona tools start cheap (tens of dollars per study), grounded-data and augmentation platforms are enterprise-priced, and AI-moderated interviews with real humans cost more again because you are paying real participants. Match the spend to the stakes: cheap synthetic for early exploration, real humans for the decision.

How to choose in June 2026

Synthetic research is real, useful and overhyped at the same time. The way to get value without getting burned is to pick the right category, then the right tool.

• For directional, early-stage exploration on a budget: a self-serve persona tool — Synthetic Users or Ditto.
• For enterprise work that has to stand up: grounded-data platforms (GWI Spark, Qualtrics Edge) or augmentation (Fairgen).
• For behaviour and audience prediction at scale: Simile or Aaru.
• For messaging and campaign testing: Artificial Societies.
• For genuine qualitative depth: AI-moderated interviews with real humans (Listen Labs, Strella) — not synthetic respondents.
• Always: validate anything decision-critical against real people, and treat synthetic output as a fast first read, not the verdict.

The labs and agencies that use this well treat synthetic research as the first layer of a stack, not a replacement for it. Explore synthetically, then confirm with humans — that is the entire skill.

Frequently asked questions

What is synthetic research?

Synthetic research uses AI to simulate human research participants — generating survey answers, qualitative interview responses or whole-audience reactions — so teams can test products, messaging and concepts without (or before) recruiting real people. It spans five sub-types: synthetic respondents, AI personas, audience simulation, data augmentation, and AI-moderated interviews (the last of which actually uses real humans).

What is a synthetic audience?

A synthetic audience is a population of AI-generated personas, modelled on demographic, survey or behavioural data, that responds as if it were a real target market. Tools such as Aaru and Artificial Societies use synthetic audiences to predict how content, ads or products will land before spending on real fieldwork.

Are synthetic respondents accurate?

Sometimes, within limits. The best independent study (Stanford, 2024) found AI agents matched people’s own survey answers about 85% as well as the people did two weeks later — but only when built from rich real interviews. Peer-reviewed work also shows synthetic respondents collapse variance and can produce wrong-signed relationships, so they are reliable for directional exploration but risky as a basis for decisions.

Can AI replace market research or focus groups?

No — and the industry consensus is clear on this. ESOMAR, GreenBook, Bain, NIQ and Kantar all frame synthetic research as a supplement that should be validated against real humans, not a replacement. It is best for early exploration, screening many ideas quickly, and reaching hard-to-recruit segments.

What is the best synthetic audience tool?

It depends on the job. Simile leads on credibility and funding, Aaru on large-scale prediction, Artificial Societies on messaging and virality, GWI Spark and Qualtrics Edge on real-data grounding, Fairgen on statistically sound augmentation, and self-serve tools like Synthetic Users and Ditto on fast, affordable concept testing.

What is the difference between synthetic respondents and AI-moderated interviews?

Synthetic respondents are entirely AI-generated — no humans involved. AI-moderated interviews (such as Listen Labs and Outset) use an AI to conduct and analyse interviews with real people. The latter is real human data and is generally more reliable; it is frequently, and wrongly, marketed under the same “synthetic” label.

Is synthetic data reliable for surveys?

Synthetic augmentation of a real survey — boosting thin segments, as Fairgen does — is the most reliable use, because it is anchored to real responses. Fully synthetic survey data is far less reliable: it tends to be overconfident (too little variance) and can distort subgroup differences.

Which synthetic research tool is best for a startup or small team?

Self-serve tools like Synthetic Users and Ditto are the most accessible, returning directional reads on concepts and personas in minutes to hours at low cost; a long tail of cheaper, newer entrants exists too. Use them to explore early and screen many ideas quickly, then validate anything important with real participants — the cheaper the synthetic read, the more it should be treated as a hypothesis rather than an answer.

Is synthetic research worth it?

Yes, if you use it for the right job. It is genuinely valuable for screening many concepts quickly, exploring early ideas, and reaching segments that are slow or expensive to recruit — provided you validate anything important with real people. It is not worth it as a wholesale replacement for human research, especially for novel products, sensitive topics or precise willingness-to-pay.

For the language models that power these tools, see our best AI models ranking and best AI apps comparison; for AI assistants used in research workflows like literature review and analysis, see best AI for research.