This is a category page, not a vendor comparison. Synthetic respondents ship at most large research vendors now — the useful question is no longer whether to allow them, but which decisions they can carry.
Synthetic audiences, and what the vendors selling them say about the limits
Synthetic respondents stopped being a fringe idea some time ago: Qualtrics, YouGov, Kantar, Ipsos, Zappi, Market Logic, Stravito and several AI-native platforms all ship some version of a simulated consumer, with blue-chip buyers behind them. This page is not an argument that they are fake. It is the more useful thing — a collection of what those vendors publish about where synthetic works and where they themselves say not to use it, and where Merciv sits, which is on real licensed consumer signal with a source you can open.
Shipping almost everywhere
Qualtrics Edge Audiences, YouGov Parallax AI twins, Kantar synthetic sample boosting, Ipsos Digital Twins and PersonaBots, Zappi Amplify AI, Market Logic Synthetic Panels, Stravito AI Personas, Panoplai Digital Twins.
The vendors publish the limits
Qualtrics rules synthetic out for go/no-go launches and regulatory work. Zappi calls it directional, not definitive. Kantar says it cannot create authenticity from nothing. Read those before you read anyone's marketing.
Merciv's position
Merciv reasons over real licensed consumer signal — social, reviews, search — plus your syndicated and internal research, and any persona it builds stays traceable to something a real person said or did.
Merciv vs. Synthetic audiences, capability by capability.
The honest version of this comparison is not synthetic-versus-real. It is which decisions a model of what people said can carry, and which ones need evidence of what people actually did — with a source attached.
| Capability | Merciv | Synthetic audiences | Why it matters |
|---|---|---|---|
| Who actually ships synthetic respondents | No synthetic respondent product. Personas exist, but each one is grounded in your real licensed consumer data and traceable back to it. | Most of the large research category: Qualtrics Edge Audiences, YouGov Parallax (twins modelled on individual real panelists), Kantar synthetic sample boosting, Ipsos Digital Twins and PersonaBots, Zappi Amplify AI, Market Logic Synthetic Panels, Stravito AI Personas, Panoplai Digital Twins. | Worth stating plainly because it cuts against the usual pitch: declining to ship synthetic respondents is not a differentiator. A number of vendors ship none either, and at least one made the public anti-synthetic argument before we did. Treat this as a scoping question about your decision, not a loyalty test. |
| Where the vendors themselves say not to use it | The same evidence base carries early exploration and a high-stakes call, because it is real signal either way — so there is no handover point where the method has to change. | Qualtrics: less suited to go/no-go launches, major pricing commitments and regulatory submissions where precision is non-negotiable, to detailed behavioural recall and unaided awareness, and to deeply nuanced cultural or emotional research. Zappi: treat synthetic as directional, not definitive, and high-stakes decisions such as major innovation bets, launch validation or significant brand investments still require human validation. Kantar: synthetic augments rather than replaces real data and cannot create authenticity from nothing. Market Logic: complement rather than replace primary research, output described as directional. | This is the most valuable material either side of the argument has published, and it belongs in your evaluation rather than in a vendor fight. Every one of those boundaries is the vendor's own, and every one of them lands in the same place: synthetic is for cheap early screening, and the decisions an insights leader is actually paid to make sit outside it. |
| How accuracy gets expressed | A confidence signal sits on the individual finding, so a reader can see which specific claims are strong and which are thin before acting on them. | Aggregate agreement against a human sample. Zappi reports 84% of Amplify AI predictions landing within one point of human results. Panoplai reports 91% quantitative alignment and states that accuracy varies by use case rather than as a single global figure. Ipsos describes twins as capable of responses statistically indistinguishable from real participants across a range of scenarios. | These are real numbers and mostly honestly framed — Panoplai's refusal to publish one global accuracy figure is better practice than a single headline. But note what an aggregate agreement figure is: it tells you how the method performed across a set of questions, not which answer in front of you is one of the misses. Zappi's 84% is also, read the other way, 16% landing more than a point out. That is the gap a per-finding confidence score is meant to close. |
| What the model is a model of | Observed signal: what consumers wrote in a review, searched for, or said in public, plus the syndicated and internal research you already own. | Survey and interview responses. A synthetic respondent is trained on what people told a researcher — Qualtrics on anonymized validated survey responses, YouGov on data real panel members share about themselves, Zappi on its own validated human respondent methodology. | Not an argument that declared data is weak — a well-designed survey absolutely can surface a gap between what people say and what they do, and panels answer questions observed signal cannot. The structural point is narrower: a simulated respondent is a model built from declared answers, so it inherits whatever that layer could not see, and it cannot be the independent check on it. |
| Tracing a specific claim | Each finding carries a citation you can open from inside the exported artifact, back to the real post, review or document behind it. | Varies, and some vendors take it seriously — Panoplai publishes data lineage, labels synthetic versus human-derived output, and flags low-confidence results for human review. What no synthetic product can offer is a link to the individual real person behind a simulated answer, because there isn't one. | This is the mechanical difference, and it is worth being precise rather than sweeping. The question to ask any vendor, us included: when a stakeholder challenges one specific number, what can you open? |
| Best fit | Teams whose recommendations get challenged, and who need a standing read on what consumers are doing outside any study they commissioned. | High-volume, low-risk, early-stage work: screening a long list of concepts or creative assets cheaply before committing budget, and reaching small or hard-to-reach groups where real sample is thin or expensive. | These are complementary, and most large insights functions will end up doing both. The failure mode is not using synthetic — it is letting a screening tool carry a decision its own vendor says it should not. |
Where each tool wins
No tool is the best at everything. Picking the right one means knowing where it pulls ahead — and where it doesn’t.
Where Merciv wins
- Real observed consumer signal — social, reviews, search — connected to your syndicated and internal research in one place.
- A citation on each claim that opens from inside the artifact, plus a confidence signal on the individual finding.
- Surfaces the questions nobody thought to ask, rather than answering the ones already written into a discussion guide.
- No handover point: the same evidence base carries early exploration and the high-stakes call.
- Permission-aware retrieval under SOC 2 Type II, with zero training on your data by policy.
Where synthetic audiences win
- Cost and volume for early screening — a long concept or creative list narrowed before anyone spends real fieldwork budget.
- Reach into small, niche or hard-to-recruit groups where real sample is thin, slow or expensive.
- Iteration speed: a read in minutes lets a team test more ideas than a fieldwork calendar would ever allow.
- Several vendors publish genuine validation work — Panoplai's research-on-research programme and Ipsos's Stanford collaboration are real methodological investment, not marketing.
- Some are notably candid about their own limits, which is more than much of this category manages.
- Twins modelled on individual real panelists (YouGov Parallax) are a meaningfully better design than personas assembled from broad stereotypes.
Five questions worth asking any synthetic vendor
These are the questions that separate a well-built synthetic product from a plausible one, and every serious vendor will have answers. Ask us the equivalents.
- Where is your own published list of decisions this should not be used for? The good vendors have one — Qualtrics, Zappi and Market Logic all publish theirs.
- How is accuracy measured, against what held-out data, and is it reported per use case or as one global number?
- How does it perform on questions requiring lived experience, habit frequency, or predicting future behaviour — the areas where simulation is weakest?
- What was the underlying panel, how were those people recruited, and what did they consent to?
- When one specific answer is challenged, what can you show me?
Screening tool, not a decision instrument
The reading that holds across every vendor's own documentation: synthetic is good at narrowing and bad at deciding. That is a genuinely useful thing to be good at, and it is worth buying for. The risk is scope creep — a tool bought for screening quietly becoming the evidence behind a launch.
- Use synthetic to narrow a long list before spending fieldwork budget.
- Use real respondents where the vendor's own guidance says to: high-stakes commitments, regulated work, deep emotional or cultural nuance.
- Use observed signal for what neither can do — telling you what changed in the market this week, when nobody had written the question yet.
‘We don't use synthetic consumers’ is not a differentiator
We are including this because it is true and because a prepared buyer will already know it. Plenty of vendors ship no synthetic respondents at all, and the public anti-synthetic argument in this industry was made by other people before it was made by us. Anyone pitching abstinence as a unique position should be asked what else they have.
- The claim that survives is mechanical, not moral: a citation on each finding, a confidence score on that finding, over signal from outside your own archive.
- Where a synthetic vendor publishes their error bars and their exclusions, take them seriously — that is better practice than an unquantified promise of trustworthiness.
- Judge us the same way. Ask what Merciv cannot answer, and whether it says so.
