Synthetic Market Research / Buyer guide
Synthetic Research Companies for US Buyers: A Comparison by Research Task
Compare providers by research task, published evidence and workflow
The right synthetic research company for a small survey subgroup may be a poor fit for a B2B positioning project. This guide compares providers offering population simulation, generated research responses or sample augmentation, with recommendations tied to the work a buyer needs done.
This guide is an editorial shortlist of potential fits for the tasks below. We have not established an overall winner or a ranking of predictive accuracy. The assessment uses public product documentation and published evaluations, without hands-on testing across providers. Companies appear alphabetically, and their capabilities overlap.
The guide is written for US research buyers, including those considering international suppliers. Inclusion does not establish US headquarters or validated coverage of every US audience. Every figure in this guide was traced to its originating document. The evidence profile shows who produced each published result, what it was benchmarked against, how it guarded against leakage and what kind of agreement it measures. The methodology sets out the selection, verification protocol, threats to validity and limitations.
Disclosure: SMR, the author and contributors report no relevant commercial relationships with the companies covered. No company paid for inclusion or reviewed this article before publication.
Compare the shortlist
Match the research brief to the advertised workflow and the scope of the available evidence.
| Company and offering | Task to consider it for | Basis for the match | Main evidence limitation |
|---|---|---|---|
| Aaru | Broad population simulation | Documented workflow for defining decisions, audiences and questions.[1] | Its principal published survey evaluation concerns aggregate answer distributions.[2] |
| Artificial Societies | Stakeholder reactions and narrative scenarios | Networked personas, group discussions and scenario methods.[20][21] | The cited comparison uses general-purpose language models, not these commercial providers.[22] |
| Electric Twin | Repeated questions to a reusable audience | Customer or other real-world data, browser access and audience licensing.[15][16] | Published agreement scores do not establish accuracy for every new audience or question.[17] |
| Evidenza | B2B positioning and go-to-market research | Professional-audience attributes and a managed research-and-strategy service.[4] | The 88% headline is a self-defined “similarity score” with no published formula; dentsu’s 0.87 is a partner’s early correlation.[5][6] |
| Fairgen — Boost | Augmenting small subgroups in an existing study | Models trained on the study’s own data.[10] | The external evaluation shows that improvement varies by subgroup.[11] |
| NIQ — BASES AI Screener | Early consumer packaged goods (CPG) idea screening | Feedback on ideas and unpriced concepts using behavioral data.[12] | Customer time and cost savings are operational outcomes, not predictive-accuracy measures.[14] |
| Qualtrics — Edge Audiences | Synthetic responses alongside enterprise survey work | Individual-level synthetic data and access to human panels.[7][9] | The cited US experiment covers eight items on one topic, uses demographically matched samples and does not describe a hold-out design.[8] |
| Simile | Behavioral simulation with estimates of likely error | Human-data grounding and a published confidence-model approach.[18][19] | Predicted error remains a model estimate that needs validation on the intended task.[19] |
What separates these companies
Start with the deliverable and the information available. A team with a completed human survey faces a different problem from one exploring an unfamiliar audience. Likewise, a managed strategy engagement and a workspace for recurring questions place different demands on the buyer’s staff.
The evidence also answers different questions. Agreement with human response percentages, preservation of relationships between answers, prediction of observed purchases and faster concept development are separate outcomes. A strong result on one does not establish the others. A simulated scenario can help generate and test hypotheses; claims about what an intervention will cause require an appropriate experimental or causal design.
The companies, compared
Aaru: broad population simulation
Aaru’s published workflow begins with the decision, target audiences and questions, then produces simulated results and reports. That makes it a candidate when a brief spans several kinds of audience or market question.[1]
Its September 2026 evaluation reports 2,993 questions across 186 study series and nine industries, all drawn from studies fielded in 2026. Each question was simulated with a knowledge cutoff set before its fieldwork date, and Aaru states that neither the questions nor their published answers were used to develop or tune the system. Observed mean absolute error (MAE) is 3.53 percentage points (median 2.73).[2]
The headline 2.78-point sample-adjusted MAE is a different statistic. For each question it subtracts the error a perfect simulation would still show, because the published survey result is itself a sample. Aaru computes that expected sampling error with a normal approximation, assumes a design effect of 1 and imputes a sample of 1,441.5 for the 345 questions without a printed base. The adjustment describes the average; it is not a bound on any single question. The publication evaluates aggregate answer distributions, not relationships between individual responses, and describes the tested system as a deliberately constrained configuration.[2]
EY separately compared Aaru’s simulation with its global wealth-research survey of about 3,600 affluent investors in more than 30 markets. It reports a median Spearman correlation of 0.90 across 53 single-select questions, a root-mean-square error of 7.1 percentage points and notable divergences, including on inheritance planning. This is evidence from a commercial collaborator, and correlation measures rank alignment rather than percentage correctness.[3]
Buying implication: include Aaru in a broad simulation shortlist. The sources support that consideration; they do not establish an overall lead over Simile, Artificial Societies or the other providers.
Artificial Societies: stakeholder and narrative scenarios
Artificial Societies describes a system that builds personas from real-world observations and connects them in networks. Its methods include interactive focus groups, follow-up interviews and “scenario cascades” (sequences of scenarios); advertised applications include strategic communications, investor relations and market expansion.[20][21]
Its October 2026 benchmark tests population and subgroup answer distributions, the stability of individual respondents’ answers, question-wording effects and relationships among answers. It uses seven human datasets, including the 2024 General Social Survey. It reports an average total variation distance (TVD) of 0.08, against 0.31 for Qwen3.8, and compares with nine off-the-shelf language models. It does not compare the other commercial research systems in this guide. The page does not say whether the human datasets were held out from the system’s development, and a survey benchmark alone does not validate claims about how narratives spread.[22]
Buying implication: consider it where interactions between stakeholders are part of the brief. Evaluate the network assumptions and scenario outputs against relevant observed outcomes before treating them as forecasts of actual reactions.
Electric Twin: repeated research with reusable audiences
Electric Twin describes synthetic audiences built from customer data or other real-world data. Teams can query an audience through a browser interface. Its FAQ describes a license based on audiences and seats, with unlimited queries; the agreed contract determines the actual scope and cost.[15][16]
The company’s methodology page describes held-out testing on about 11,000 personas spanning UK and US populations, with evaluation questions never used in persona creation. It reports 95.5% on 1-MAE and 92% on its Normalised Distribution Accuracy Measure (NDAM), against a human ceiling of about 94% NDAM when the same people answer the same question twice. These are separate measures of distribution agreement. The headline “96%” is not the probability that an individual response or a new question is correct. The page does not publish the NDAM formula or name the human data sources.[17]
The company also refers to external academic validation. We did not find that work documented in the cited methodology, which is written by company staff, so we treat the published evidence as vendor-reported.[16][17]
Buying implication: consider Electric Twin when recurring access to a modeled audience suits the team’s workflow. Confirm which data supports that audience, how it is refreshed and how the supplier detects questions beyond its validated coverage. Customer-owned data is one possible input, not the only route described by the company.
Evidenza: B2B positioning and go-to-market research
Evidenza’s stated audience attributes include job role, company size and industry. Its FAQ says a target-audience description is sufficient to begin, with supporting research optional. The managed offer includes research design, analysis and a strategic presentation.[4]
This is a relevant starting point for a B2B team seeking a positioning or messaging recommendation. It does not establish that every professional role, industry or buying committee is equally well represented.
Evidenza’s homepage advertises 88% accuracy in more than 100 validations, and defines the figure as a “similarity score”: the average match between synthetic and traditional research. No formula, sample sizes or per-study results are published alongside it, so it cannot be compared with other providers’ metrics. Dentsu’s partnership announcement quotes early results showing 0.87 correlation with traditional research, without stating what was correlated. It is a commercial partner’s statement, not a separately established 87% correctness rate.[5][6]
Buying implication: consider Evidenza for a managed B2B research engagement. Ask for validation matching the buyer roles and decision being studied, particularly before translating a positioning result into a revenue or demand forecast.
Fairgen Boost: augmenting an existing human study
Fairgen says Boost generates “predictive respondents” from patterns in each study’s own data, so that segments too small to report become readable. Its models are trained only on that study’s data, and it supports questionnaire features such as skip logic, piping and multiple selections.[10] This makes the existing dataset central to the purchase.
Dig Insights evaluated the method on a 26,372-respondent telecom dataset and a 2,258-respondent quick-service restaurant (QSR) dataset. It split each dataset once: the training set went to Fairgen, and the test set, never exposed to the algorithm, served as ground truth. Dig states that it had no commercial ties to Fairgen and that Fairgen allowed publication regardless of outcome. The report also describes the work as a partnership, and the copy we cite is hosted by Fairgen.[11]
In the telecom table, distribution error falls for three subgroups and rises slightly for two (18–24-year-olds and newcomers), whose effective sample size also falls. Correlation error falls for all five. In the QSR table, which starts from subgroups of 10 to 35 respondents, distribution and correlation error fall for all five.[11]
Our reading: the pattern fits the bias–variance trade-off. Borrowing information from the rest of the sample helps most where a subgroup is tiny and noisy. It can add bias where a subgroup genuinely differs from everyone else, and that bias can outweigh the reduced noise once a subgroup is large enough to stand on its own. The report uses a single split, without repeated draws or significance tests, and does not state the size of the benchmark subgroups in the test set. Where those benchmark subgroups are small, part of the measured error belongs to the benchmark rather than to either estimate.
Buying implication: consider Boost when the task is to improve estimates within an existing study. Generated records are not additional independently recruited people. Validate uncertainty using a method appropriate to augmentation, rather than calculating a conventional margin of error from the combined record count.
Evidence check: results varied by subgroup
View the underlying chart data
| Subgroup | Human n | Original error | With boost |
|---|---|---|---|
| Saskatchewan | 21 | 6.9 | 4.3 |
| Nova Scotia | 40 | 4.9 | 3.0 |
| NoTweets | 237 | 2.0 | 1.5 |
| Age 18–24 | 157 | 2.4 | 2.6 |
| Newcomers | 198 | 2.4 | 2.5 |
NIQ BASES AI Screener: early CPG idea screening
NIQ positions BASES AI Screener around early ideas and unpriced concepts, using purchasing-behavior data and language models to produce feedback, drivers, barriers and improvement suggestions.[12] Its April 2025 launch announcement listed the United States among several launch markets and described a DIY workspace.[13] NIQ’s 2026 materials describe availability in select markets and categories, with expansion ongoing.[14]
NIQ’s April 2026 Reckitt case reports up to a 65% reduction in research timelines and 50% lower research costs. The release headline calls this “65% faster research”, which is a different quantity; we use the more specific body wording. These are vendor-published customer outcomes. They describe that implementation, rather than a guaranteed saving or a comparison of prediction accuracy.[14]
Buying implication: consider the product when prioritizing early CPG ideas. Confirm the required US category and concept format are covered. A screening score does not, by itself, establish sales volume, price acceptance or a final launch decision.
Qualtrics: synthetic responses within a survey workflow
Qualtrics describes its synthetic research approach in terms of simulated individual-level data, and says it currently offers synthetic panels for the US general population in English.[7] Its Edge Audiences offering includes both traditional and synthetic panels, making it relevant to teams already working with structured questionnaires.[9]
Its January 2026 publication replicates a published US study of attitudes to Google Search, using eight attitudinal test items. Of 620 human respondents, 102 were removed by quality checks. The remaining 518 were matched one-to-one with synthetic records on age, gender and product-use frequency; 584 synthetic records were produced after 16 were dropped for hallucinations. Qualtrics reports an average deviation from human item means of 0.07 standard deviations, against 0.87 for GPT-5 and 0.88 for Gemini.[8]
That result measures agreement in item means, after matching, on one topic. The publication does not describe a hold-out design, and Qualtrics notes weaker performance on context-specific past behaviors.[8]
Under the August 2025 agreement, PureSpectrum selected Qualtrics as its exclusive synthetic-panel provider, and PureSpectrum became Qualtrics’ exclusive quantitative panel provider. We cover that announced capability once, while recognizing that procurement channels and services may differ.[9]
Buying implication: consider Qualtrics where survey workflow is a priority. Evaluate the relevant synthetic product on the actual question types and audience; agreement after matching is not a guarantee for an unmatched new study.
Simile: behavioral simulation and predicted error
Simile describes populations grounded in human studies and behavioral datasets, with optional customer data and refreshes as information changes.[18]
Its confidence-model publication trains a model to predict the error of each simulated answer distribution, measured as total variation distance (TVD), and shows that estimate for each question in four bands. The evaluation uses approximately 8,600 questions, described as roughly representative of production queries, with five-fold cross-validation that keeps each question’s text and experimental sample together in the same fold.[19]
The final model separates acceptable from unacceptable simulations (a TVD threshold of 0.16) with an AUROC of 0.736, and its predicted error correlates with observed error at 0.565. Calibration is shown graphically rather than as a summary statistic.[19]
The capability is relevant when a team wants uncertainty to influence which outputs it uses. An estimated error is not a guarantee, a conventional survey margin of error or proof that the system knows every case it will get wrong. Aaru also publishes work on estimating its own errors; the capability is not exclusive to Simile.[2][19]
Buying implication: include Simile when evaluating behavioral simulation and confidence reporting. Test whether low predicted error corresponds to low observed error on fresh cases drawn from the intended US audience.
How strong is the published evidence?
A single accuracy figure says little until you know who produced it, what it was compared with and what it actually measures. The table profiles the principal published result for each offering on those dimensions. It deliberately does not combine them into a score: the differences between rows are the information a buyer needs.
| Offering and source | Conducted by | Human benchmark | Leakage safeguard described | Compared with | What agreement is measured |
|---|---|---|---|---|---|
| Aaru [2] | Vendor | 2,993 published questions from 186 study series fielded in 2026 | Per-question knowledge cutoff before fieldwork; questions not used in development (vendor-stated) | Direct prompting of general-purpose models; survey-to-survey replication | Aggregate answer distributions (MAE, TVD) |
| Aaru with EY [3] | Commercial collaborator | EY global wealth survey, about 3,600 investors in 30+ markets | Not described in EY’s account | No baseline reported | Rank order (median Spearman 0.90); RMSE 7.1 points |
| Artificial Societies [22] | Vendor | Seven datasets, including the 2024 General Social Survey | Not described | Nine general-purpose language models | Population and subgroup distributions, opinion spread, wording effects, relationships |
| Electric Twin [17] | Vendor | Private survey data; about 11,000 UK and US personas | Evaluation questions excluded from persona creation | Traditional survey methods; human test–retest ceiling | Aggregate distributions (1-MAE, NDAM) |
| Evidenza [5][6] | Vendor; commercial partner | “Traditional research”; not specified | Not described | Traditional research only | Undefined “similarity score”; unspecified correlation |
| Fairgen Boost [11] | External firm stating no commercial ties | Held-out test sets from two studies (n = 26,372 and 2,258) | Test set never exposed to the algorithm; single split | The same subgroup without boosting | Subgroup distributions, pairwise correlations, effective sample size |
| NIQ BASES AI Screener [14] | Vendor-published customer case | None reported | Not applicable | The customer’s previous process | Operational time and cost, not predictive accuracy |
| Qualtrics Edge Audiences [8] | Vendor | 518 US adults after quality checks, matched one-to-one | Not described | GPT-5 and Gemini | Item means (Cohen’s d); correlations with product use |
| Simile [19] | Vendor | About 8,600 questions resembling production queries | Five-fold cross-validation grouped by question and sample | Simpler error-prediction methods | Accuracy of predicted error (AUROC, correlation), not the simulation itself |
The gap that matters: no offering in this guide has a published evaluation that is independently run, uses a common task with commercial rivals and is replicated. Only one external evaluation appears (Dig Insights on Fairgen), and it tests augmentation within a study rather than simulation. That is why this guide names no overall winner.
Reading the accuracy numbers
Providers report different statistics, and the same percentage can mean very different things. Each metric below answers a narrower question than “how accurate is it?”
| Metric | What it measures | What it does not tell you | Used by |
|---|---|---|---|
| Mean absolute error (MAE) | Average gap, in percentage points, between predicted and observed answer shares | Whether individual answers or relationships between questions are right. It shrinks mechanically as answer options increase. | Aaru, Dig Insights |
| Sample-adjusted MAE | MAE minus the error expected from the human sample’s own sampling noise | Error on any single question; it depends on assumed sample sizes and design effects. | Aaru |
| Total variation distance (TVD) | Share of respondents who would have to change answer for the two distributions to match | Which options are wrong, or whether subgroup patterns hold. | Aaru, Artificial Societies, Simile |
| 1-MAE | MAE expressed as a percentage score (100 minus MAE) | The share of people or questions predicted correctly. Uninformative guesses can score highly (see below). | Electric Twin |
| NDAM | Electric Twin’s distribution-accuracy score, normalized to 0–1 regardless of the number of options | Comparability with other providers; the formula is not published on the cited page. | Electric Twin |
| Spearman or Pearson correlation | Whether predictions rise and fall with the human results | Whether the levels are right: every estimate can be 20 points too high with a perfect correlation. | EY, dentsu, Dig Insights |
| Cohen’s d of mean deviation | Average gap between item means, in standard-deviation units | Whether the spread or shape of answers matches. | Qualtrics |
| AUROC | How well a confidence score separates good simulations from poor ones (0.5 is chance, 1.0 perfect) | How large the errors are. | Simile, Aaru |
| Effective sample size (ESS) | The real sample size that would give the same measured reliability | That new, independent people were surveyed; it is specific to the tested subgroups and benchmark. | Dig Insights |
Why a high 1-MAE can mislead: suppose the true split on a five-option question is 40%, 30%, 15%, 10% and 5%. A model that knows nothing and guesses 20% for every option misses by 20, 10, 5, 10 and 15 points: an MAE of 12 points, or 88% on 1-MAE. Scores in the 90s therefore need a baseline beside them. Normalized measures such as NDAM, and comparisons with simple baselines, exist to address this.
How the workflows compare
Aaru and Simile: compare on the same simulation brief
Both describe broad simulation and error-estimation capabilities. Their separate publications do not establish which will perform better on a buyer’s task. A useful comparison supplies the same audience definition, questions and decision criteria, then examines error, subgroup performance and uncertainty calibration.[1][2][18][19]
Aaru’s publication sets its own error-prediction AUROC of 0.845 beside the 0.736 Simile published. The two figures come from different question sets and evaluation designs, so the comparison is not a head-to-head test.[2][19]
Evidenza and Artificial Societies: positioning or stakeholder scenarios
Evidenza’s professional-audience attributes and managed deliverable make it relevant to B2B positioning. Artificial Societies’ network and scenario methods make it relevant when stakeholder interactions matter. These are reasons to shortlist each; they are not exclusive capabilities or evidence of comparative superiority.[4][21]
Fairgen and Electric Twin: an existing study or continuing questions
Fairgen Boost targets augmentation within an existing study. Electric Twin offers repeated queries to an audience built from supporting data. The distinction concerns the workflow under review: it should not be taken as a description of every product either company offers.[10][16]
Qualtrics and NIQ: survey operations or early product ideas
Qualtrics is relevant to teams working with survey-shaped datasets, while BASES AI Screener is specifically described for early innovation ideas and unpriced concepts. The choice depends on the output required, available category coverage and validation on the proposed task.[7][12]
Compare the total cost of the research job
We did not obtain comparable commercial quotations, so this review makes no cheapest-provider recommendation. Compare data preparation, audience creation, analyst support, usage, refreshes, validation, integration and any human fieldwork needed to reach a usable decision.
| Buying pattern | Published example | What to establish in a quotation |
|---|---|---|
| Managed project | Evidenza describes pricing customized to the scope and complexity of a managed research-and-strategy engagement.[4] | Audiences, deliverables, revisions, validation and support. |
| Recurring audience access | Electric Twin describes audience-and-seat licensing with unlimited queries.[16] | Audience creation and refreshes, export rights, seat limits and contract term. |
| Self-service screening | NIQ describes a DIY idea-screening workspace.[13] | Eligible categories and markets, usage allowances, support and further testing. |
What should determine the final choice?
Audience coverage. Specify the US population, subgroup, time period and buying context. A global or UK result may inform a review without establishing performance on that particular US audience.
The right metric. A correlation of 0.90 is not 90% correctness. In the hypothetical example below, modeled purchase-intent rates preserve the ranking of all three concepts while overstating every rate by 20 percentage points.
A fresh, blinded evaluation. For consequential decisions, give shortlisted providers the same brief and permitted inputs, withhold the human comparison results, and agree the acceptable error before viewing outputs. Include a simple baseline and enough fresh cases to examine subgroup errors and uncertainty. For augmentation, compare both the original and augmented estimates with a held-out human benchmark. Human samples also have sampling, recruitment and measurement limitations.
Clear reporting. Distinguish generated records from human respondents, document the source data and describe the human benchmark’s recruitment and quality controls. AAPOR’s code, revised in June 2026, states that AI-generated responses are not research participants and must be identified as such in public-opinion research. It also requires disclosure of AI use in data collection and processing, and says the terms “poll” and “survey” should not describe data created through AI.[23]
Use this guide to build a shortlist around the decision and data available. Make the purchase on demonstrated fit for that brief, contractual terms and a validation result the team can inspect.
Research methodology
This section is written so that another analyst, working from the same sources, could repeat the review and reach the same factual findings. Where we exercise editorial judgment, chiefly in matching offerings to tasks, we say so.
1. Review question and design
The review asks: for a US research buyer, which synthetic research offerings are plausible fits for which research tasks, and how much weight can their published evidence bear? It is a structured desk review of public documentation. The unit of analysis is a named offering rather than a company, because evidence about one product or configuration does not automatically transfer to a supplier’s other products. All sources were accessed and checked on 22 October 2026.
2. Selection
The eight offerings form a purposeful shortlist spanning seven distinct workflows: broad population simulation, networked stakeholder simulation, reusable modeled audiences, managed B2B research, augmentation of an existing study, early CPG concept screening and synthetic panels within a survey platform. Each included offering is commercially marketed, relevant to US buyers and has enough public documentation to answer all five assessment questions below. The shortlist is not a census of the market. Omission is not a negative assessment, inclusion is not an endorsement, and US audience coverage must be confirmed for any proposed engagement.
3. Assessment framework
Each offering was reviewed against the same five questions: what task it addresses; what input data it requires; what output and delivery model it describes; what published evaluation supports it; and what limits affect use of that evidence. Task matches are qualitative editorial judgments. We assigned no numerical scores, weights or overall winner, and a task label does not imply exclusive capability.
4. Verification protocol
Every factual and quantitative statement in this guide went through the same steps:
- Trace to origin. Each figure is cited to the document that first reports it, not to a press summary or a third party’s repetition.
- Record the full specification. For each figure we recorded the metric, unit, denominator, population, geography, fieldwork or knowledge-cutoff date, comparator and stated qualifications.
- Prefer the specific over the headline. Where a headline and body text differ, we report the body text. NIQ’s “65% faster research” headline, for example, is reported as the body’s “up to 65% reduction in research timelines”.
- Use technical definitions. Where marketing copy describes a metric differently from the provider’s own technical definition, we use the technical definition. A 1-MAE score is not the share of the time a model “matches what people say”.
- Attribute independence. Statements of independence are attributed to the party making them; Dig Insights’ statement of no commercial ties is reported as Dig’s statement, not as audited fact.
- Do not double-count. A partner or client repeating a vendor result counts as one source, not as replication.
- Recompute where possible. Chart values were checked line by line against the source tables, and worked examples were recalculated. Inconsistencies inside sources, such as conflicting dates, are recorded in the references rather than silently resolved.
5. Evidence profile
Instead of collapsing evidence into a grade, each principal published result is profiled on five dimensions in the evidence table: who conducted it; what human benchmark was used; what safeguards against leakage are described; what it was compared with; and what level of agreement it measures. The levels range from item means and answer distributions to rank order, relationships between answers, error prediction and operational outcomes. These dimensions are reported side by side and not weighted, because a composite score would hide exactly the differences a buyer needs to see.
6. Threats to validity we looked for
| Threat | Why it matters | How it appears in this evidence |
|---|---|---|
| Leakage | A model that has seen the benchmark results, or close copies, can appear accurate without predicting anything. | Aaru uses per-question knowledge cutoffs; Electric Twin, Simile and Dig describe held-out designs; the Qualtrics, Artificial Societies, EY and Evidenza accounts do not describe one. |
| Benchmark noise | Human results are samples too. Part of any measured gap belongs to the benchmark. | Aaru adjusts for sampling error; Electric Twin reports a human test–retest ceiling; Dig does not report the size of its benchmark subgroups. |
| Level of aggregation | Matching overall answer shares does not show that individual responses, subgroups or relationships between questions are right. | Most evidence here is aggregate. Dig and Artificial Societies also report relationships between answers. |
| Conditioning | Matching or weighting on known variables can raise agreement on correlated measures. | Qualtrics matched synthetic and human records on age, gender and product use, which is related to the attitudes tested. |
| Weak comparators | Outperforming general-purpose language models does not show an advantage over specialist rivals or conventional research. | Artificial Societies, Qualtrics and Aaru compare with general-purpose models; none compares with another provider in this guide. |
| Selective publication | Vendors choose which evaluations to publish; unfavorable results are rarely visible. | Most results are vendor-run. Dig’s and EY’s accounts include unfavorable rows and divergences, which supports their credibility. |
| Transfer | Accuracy in one market, category or period may not carry over to a new US brief. | EY’s study is global; Dig’s telecom study is Canadian; Electric Twin pools UK and US populations. |
| Metric framing | Correlations and 1-MAE scores presented as percentages invite misreading as “percent correct”. | EY describes a median correlation as “90%”; Evidenza labels a similarity score as accuracy; Electric Twin describes 1-MAE as a match rate. |
7. What it would take to name a “best” provider
A defensible comparative claim would need: a defined candidate universe with inclusion rules; criteria and acceptable error fixed before any results are seen; fresh tasks fielded after every system’s knowledge cutoff; identical briefs and permitted inputs; human benchmarks withheld from providers and large enough that their own sampling error is small relative to the differences sought; at least one simple baseline; results at subgroup and relationship level with uncertainty intervals; and reporting of every task attempted, not only the successful ones. Even then, the conclusion would apply to the tested versions and tasks. None of the evidence reviewed meets this standard, which is why the guide names no winner.
8. Checks not performed
We did not commission platforms, recruit human samples, inspect proprietary training data, rerun published experiments, audit contamination controls, verify contractual US coverage, assess security or privacy compliance, or obtain comparable quotations. Confirming that a source reports a result does not independently reproduce it.
9. Independence, disclosure and corrections
SMR, the author and contributors report no relevant commercial relationships with the companies covered. No company paid for inclusion, and none reviewed or approved this article before publication. Factual corrections will be made on this page with the date of the change, and assessments will be revisited when providers publish new evidence.
10. Figures
The Fairgen chart reproduces all five telecom subgroup rows from Dig Insights’ Table 1, p. 5; the QSR results from Table 2, p. 6, are described in the text. Values are rounded as published, and no significance test is inferred. The purchase-intent chart is hypothetical and does not depict any company’s performance.
References
Sources reviewed on 22 October 2026. Undated pages are marked “n.d.” Vendor descriptions establish what is offered; vendor and partner performance claims retain the qualifications stated in the article.
1. Aaru. (n.d.). Simulation. Product and engagement description.
2. Aaru. (2026, September 24). Population simulation at the replication floor. Vendor-published evaluation; not a common commercial-provider test.
3. EY. (2025, October 7). How AI simulation accelerates growth in wealth and asset management. Commercial collaborator’s account of a global wealth-research comparison.
4. Evidenza. (n.d.). Frequently asked questions. Product, data-input, managed-service and pricing descriptions.
5. Evidenza. (n.d.). AI market research for impossible audiences. Vendor positioning and performance claim.
6. dentsu. (2025, June 17). Dentsu partners with Evidenza to integrate synthetic audiences into next gen media planning. Commercial partnership announcement; not an independent audit.
7. Qualtrics. (2026, February 19; updated February 20). Synthetic data for market research: Your questions, answered. Product scope, methodology and current US general-population coverage.
8. McLean, D. (2026, January 5). Qualtrics’ fine-tuned AI outperforms general-use LLMs on survey research. Qualtrics-authored US survey experiment; comparators GPT-5 and Gemini.
9. Qualtrics. (2025, August). Qualtrics and PureSpectrum partner to deliver advanced synthetic research capabilities. Provider and distribution relationship. The page header reads 4 August and the dateline 5 August 2025.
10. Fairgen. (n.d.). For research firms and panel providers. Current description of Boost and its study-specific inputs.
11. Dig Insights. (n.d.). When synthetic data works: A validation of augmented sample usage. External evaluation hosted by Fairgen; Dig states it had no commercial ties and could publish regardless of outcome. Chart: Table 1, p. 5; QSR results: Table 2, p. 6.
12. NIQ. (n.d.). BASES AI Screener. Product purpose, inputs and outputs.
13. NIQ. (2025, April 15). NIQ unveils BASES AI Screener: Helping CPG brands thrive in the age of hyper-innovation. Launch announcement identifying US availability and a DIY workflow.
14. NIQ. (2026, April 7). Reckitt accelerates innovation with NIQ AI insights, reporting up to 65% faster research. Vendor-published customer case; the headline’s “65% faster” differs from the body’s “up to 65% reduction in research timelines”. Reported reductions are not a universal service commitment.
15. Electric Twin. (n.d.). Synthetic audiences for enterprise research. Current product and workflow description.
16. Electric Twin. (n.d.). Synthetic audience FAQs: Accuracy, data and cost. Audience inputs, browser workflow and licensing terms.
17. Warner, B., & Bailey, A. (2026, February 11). How accurate are synthetic audiences? Electric Twin’s 96% result explained. Vendor explanation of held-out evaluation and distinct distribution-agreement metrics.
18. Simile. (n.d.). Simulating human behavior. Human grounding, optional customer data and population-simulation positioning.
19. Wesel, A., Chen, S., & Liang, P. (2026, August 25). Building confidence in Simile. Vendor technical evaluation of predicted error on held-out questions. Date from Simile’s blog index; the article page is undated.
20. Artificial Societies. (n.d.). Understand people. Make better decisions. Product positioning and application categories.
21. Artificial Societies. (n.d.). Solutions: AI audience simulation. Networked personas, research methods and stakeholder scenarios.
22. Artificial Societies. (2026, October 1). Artificial Societies is more accurate than state-of-the-art LLMs. Vendor benchmark against nine general-purpose language models, not the other commercial companies in this report. Date from page metadata; no visible byline.
23. American Association for Public Opinion Research. (2026, June). AAPOR Code of Professional Ethics and Practices (revised June 2026). Definitions 2–3 and Section III disclosure items 5 and 12 cover AI-generated responses and disclosure of AI use.