Synthetic Research Companies for European Buyers: A Comparison by Research Task
The right synthetic research provider for a European buyer depends on the research task, not on a headline accuracy score. Boosting a 50-person product test in Italy, screening campaign copy in Germany and simulating a hard-to-reach UK stakeholder group are different jobs, and the evidence behind each offering covers some of them and not others.
This guide is an editorial shortlist of possible fits. It does not rank providers, and it does not name a best provider or an accuracy winner. It relies on public documentation and published evaluations, and we did not test any platform hands-on.
How this list was compiled
- Not a ranking. Companies appear in alphabetical order. No scores or positions are assigned.
- Selection rules. Nine offerings were chosen under four written rules, set out in the methodology. The list is a sample, not a market census. Other European providers we identified are named there.
- No payment. No company paid to be included, to be described favourably or to appear in any position. No company supplied content or reviewed the article.
- Commercial disclosure. The publisher runs a directory of synthetic research vendors and sells paid visibility packages to vendors. None of the nine companies profiled here was a paying customer when this article was published.
- Sources and date. Every figure comes from a public source listed in the references, as reviewed on 18 November 2026. Figures reported by a company about its own product are labelled as such.
- Corrections. Any company, profiled or not, can send corrections or further evidence. Changes will be dated in the article.
The guide is written for buyers in Europe. Every offering was developed by a company founded or headquartered in Europe or the UK, but inclusion does not confirm validated coverage of any specific country or language. European buyers also face questions that US guides tend to skip: where data is processed, which model providers sit underneath and whether the evidence was gathered on a European population. We report what each provider states on those points and say where it states nothing.
Compare the shortlist
Nine offerings cover seven distinct research workflows. Match the task first, then check whether the evidence was gathered on a population like yours.
| Company and offering | Task to consider it for | Basis for the match | Main evidence limitation |
|---|---|---|---|
| Artificial Societies (London) | Simulated populations and survey experiments, including hard-to-reach audiences | Personas built from real-world observations and linked in networks; client data can be added; an 11-test validity benchmark across 20 human sources | The company ran its own product through its own benchmark, and public sources mean contamination cannot be ruled out |
| Electric Twin (London) | Repeated questions to a reusable audience built from your own data | Audiences built from client surveys, interviews and focus groups; hold-out testing on 11,000 UK and US personas; an LSE-led comparison with seven sampling services | Evidence covers the UK and US only; the external study was funded by Electric Twin and co-led by its co-founder |
| Ipsos Data Boosting (Paris) | Augmenting a small human product test | Synthetic respondents generated from a human seed of 50 to 100 per product; 260 validations against all-human samples | Co-authored by Ipsos and a client; few per-validation statistics; markets and categories not broken down |
| Lakmoos (Brno) | Turning past survey data into an AI panel; agency white-label | A private model built on the client's own data; a one-month paid pilot | The "over 98% similarity" claim is not defined on the site; the public research repository reports no results on its page |
| Minds (Berlin) | Reusable audiences for qualitative and quantitative studies, including Sinus-Milieu personas for Germany | Audiences built from a brief, files or links; SINUS Milieu-Minds tested against a 2,005-person German SINUS survey | One country, one questionnaire, not preregistered; a fixed answer mix was slightly more accurate on exact shares |
| neuroflash Digital Twins (Hamburg) | Pre-testing marketing copy, ads and posts in German or English | Twins calibrated on more than 1,000,000 survey profiles collected since 2017; processing in Germany | The 85 to 98% accuracy claim comes without a published validation method that we could find |
| quantilope Category Twins (founded in Hamburg) | Interviewing twins built on your own brand tracker | Twins built on the brand's own tracking data, refreshed with every wave | No accuracy or validation figures published |
| Statista SynthiePop (Hamburg) | Sizing, profiling and locating German target groups | 83.6 million virtual individuals built from official statistics and Statista data, with no generative AI | Germany only; no validation results or benchmark figures published |
| YouGov e-twins with OriginalVoices (London) | Twins of known panellists, checked against those same people | Twins of 1,008 ID-verified panellists; the Parallax product pairs simulation with fast human validation | Headline accuracy relies partly on panellists rating their own twin; results come from YouGov's own white paper |
What separates these companies
The first difference is what each offering starts from. Electric Twin, Lakmoos and quantilope build from the client's own surveys or tracker. YouGov, neuroflash and Statista start from data the provider already owns: a verified panel, a survey archive or official statistics. Minds can run on SINUS-Institut's milieu model, Artificial Societies builds personas from public observations, and Ipsos generates synthetic respondents from the human sample collected inside the same study. A buyer with a rich first-party dataset faces a different choice from a buyer starting cold.
The second difference is the deliverable. Some return answer distributions for a survey, some return a go or no-go product decision against an action standard, some return a conversation with a persona and one returns population counts down to a 100-metre grid. These outputs are not interchangeable, and an accuracy figure for one says nothing about the others. Two offerings also sit outside the usual language-model approach: Statista states that SynthiePop uses no generative AI, and Ipsos describes a deep-learning generative model without saying whether it is a language model.
The third difference is the kind of evidence behind each claim. We found vendor-run benchmarks, one external study funded by the vendor, one paper co-authored with a client, one partner-reviewed case, a white paper scored partly by the panellists themselves and several offerings with no published figures at all. We found no independent replication of any of the nine on a fresh European population. That does not make the products weak; it means the buyer's own test carries more weight than any number in this article.
The companies, compared
Each chart in this section shows one company's own reported results in that company's own metric. The charts are not comparable with each other, and no chart places one profiled company against another.
Artificial Societies: simulated populations and survey experiments
Artificial Societies, a London company founded by behavioural and data scientists from Oxford and Cambridge, builds networks of personas from real-world behavioural observations. Clients can add first-party data to calibrate a bespoke population. The workflow covers survey design, statistical outputs, interviews and simulated focus groups, with answers inside 24 hours.
In October 2026 the company published a benchmark of eleven validity tests across twenty human sources, including Pew surveys of France, Germany and the UK, European Social Survey data for Great Britain and a Finnish survey experiment. On seven of those sources, the company reports an average total variation distance of 0.08 for its product against 0.31 for the best of nine general language models. It reports subgroup distance of 0.063 against 0.222, and a question-wording error of 1.4 points against 6.2. A separate January 2026 evaluation against 1,000 UC Berkeley surveys reported 86% distribution accuracy against a 91% human-replication ceiling.
Source: Artificial Societies, 1 October 2026. Company-run benchmark on public sources; the best general model differs by source.
The limits are stated by the authors themselves. The project is company-funded, every author is affiliated with Artificial Societies and the sources are public, so training-data contamination cannot be ruled out. The product page reports results for only some of the eleven tests and gives no human ceiling for them. The benchmark does not test behavioural prediction directly.
Consider it when the audience is hard to recruit and the question is attitudinal. Ask for product results on the European sources in the benchmark, not only the averaged figure, and for a test on your own category.
Electric Twin: repeated research with reusable audiences
Electric Twin, based in London, builds a synthetic audience from a client's own surveys, interviews and focus-group findings, then lets teams query it repeatedly. The audience persists, so a team can return to it for each new question.
Its February 2026 methodology splits each dataset into persona-creation data and evaluation questions, which are never used to build personas. Across 11,000 UK and US personas it reports 96% on 1-MAE and 92% on NDAM, against a human ceiling of about 94% NDAM when the same group answers the same question twice. A July 2026 study led from the LSE compared Electric Twin with seven human sampling services across 7,755 UK participants. The synthetic audience fell within the range of the human samples on most questions, while voting estimates differed by 10 to 20 points between the human platforms themselves. The study was preregistered.
Source: Electric Twin methodology post, 11 February 2026. Company-reported, UK and US data. The post names four of eight traditional methods.
The methodology page names no independent auditor, uses private evaluation data and covers only the UK and US. Electric Twin funded the LSE study, and its two lead researchers are the company's co-founder and its Chief Science Adviser. The human survey data was collected independently by Stack Data Strategy. The match held on most questions, not all.
Its published evidence is on UK and US populations. For continental markets, ask for hold-out results on that country before signing, and confirm where data is processed: neither the methodology page nor the FAQ says. The FAQ states GDPR compliance and ISO 27001:2022 certification.
Ipsos Data Boosting: augmenting a small human product test
Ipsos, headquartered in Paris, offers Data Boosting inside product testing. A small human sample is collected, and a generative model trained on that human seed adds synthetic respondents. The typical design is 50 humans plus 150 synthetic respondents per product, rising to 100 humans for riskier decisions.
An August 2026 paper co-authored with Barilla reports 260 validations comparing all-human samples with boosted samples. The boosted sample reached the same action-standard decision as an independent all-human holdout in 92% of cases, with estimated cost savings of 20% to 60%. In the Barilla case, an Italian Ragù recipe test used 100 humans plus 300 synthetic respondents per cell; both samples rejected the prototypes, and Barilla did not proceed.
Source: Coquelle et al., Research World, 20 August 2026. Co-authored by Ipsos and Barilla. The share of go decisions in the sample is not reported.
In the 8% of divergent cases the authors found three failure patterns: a skewed human seed carried forward, weak differences amplified into significant ones in flat product sets, and overfitting when too many variables were generated from a small seed. Augmentation cannot recover population features absent from the seed. The paper gives few per-validation statistics and does not break down markets or categories.
Treat it as a cost lever inside a human study, not a replacement for one. Agree the seed size and the action standard before fieldwork, and keep 100 humans per product for high-risk calls, as the authors recommend.
Lakmoos: an AI panel built on your past survey data
Lakmoos, based in Brno, builds a private model on a client's own data and offers AI respondents, a custom AI panel, conversational querying of past surveys and augmentation of new studies. It describes its system as neuro-symbolic but gives no architectural detail. Agencies can white-label it.
The website claims an "over 98% similarity score" across 20 client benchmark studies in 2025 and more than 15 validation projects. A March 2026 research repository by Lakmoos-affiliated authors compares plain GPT-4-Turbo output, GPT-4-Turbo enriched with Czech Statistical Office distributions and 500 Lakmoos respondents. The benchmarks are Czech Statistical Office and Median research agency distributions plus individual-level data from Ipsos Instant Research.
The site does not define "similarity" or describe its method. The repository page lists tests and effect sizes but reports no findings and links no paper. The site states GDPR compliance but not where data is hosted.
Its public benchmarks are Czech, which is relevant for Central European audiences. Ask for the full paper and the similarity definition before committing to the paid pilot.
Minds: reusable audiences for qualitative and quantitative studies
Minds, a Berlin company founded in 2024, builds reusable audiences from a brief, files or links and reuses them across surveys, interviews and concept tests. In July 2026 it launched SINUS Milieu-Minds with SINUS-Institut: personas for each Sinus-Milieu, two per milieu by default, extendable by age, life situation or other facets. A third-party listing states the application is hosted in Frankfurt, the database in Stockholm and the model providers in the US, at €0.12 per synthetic response.
In an October 2026 case, 200 Milieu-Minds, 20 per milieu, answered 24 concern questions. Results were compared with a June 2026 SINUS survey of 2,005 people, and the survey results were withheld from the Minds at runtime. Per milieu, answer shares came within 8.3 points on average, or 91.7% closeness. Two real survey waves were 96.3% close to each other, and general AI models given only a milieu name reached about 72%.
Source: Minds Team, 22 October 2026. Distribution closeness is 100 minus mean absolute difference, against the June 2026 SINUS survey. Fixed answer mix calculated as 100 minus its reported 7.5-point deviation.
Minds also reports results for its Mosaic forecasts. As of 12 November 2026 it states 93.5% distribution overlap with published survey answers across 95 questions from 23 surveys, where overlap is one minus total variation distance. On 28 questions from 8 surveys where the forecast was locked before fieldwork, it states 97.3%. These figures are company-reported, and we did not review the underlying survey list.
A fixed answer mix taken from an earlier SINUS wave deviated by 7.5 points, slightly more accurate than the Minds on exact shares. The personas came from existing production groups, the comparison was not preregistered and the AI-model controls were added after the main result was known. It covers one country, one questionnaire and attitude questions only.
Its most detailed published case covers German segment work built on Sinus-Milieus. For other countries that case offers no evidence, and buyers with strict data-transfer rules should ask which model providers process prompts.
neuroflash Digital Twins: pre-testing marketing content
neuroflash, founded in Hamburg in 2018, sells Digital Twins for testing posts, ads, calls to action, images and landing pages before launch, with API and MCP access. The twins are based on more than 1,000,000 real survey profiles collected since 2017, and named customers include REWE Group, Migros, T-Mobile, Swisscom and StepStone. The company states that all data is processed in Germany, with no US servers.
The product page claims 85% to 98% accuracy, against about 55% for generic language-model personas. A company article on evaluation metrics cites a different figure, 85 to 95 percent parity with human panels, and refers to validation against more than 80 academic studies without listing them. The studies it does cite tested other systems, not neuroflash's product.
In the pages we reviewed, we found no document that defines the 85% to 98% figure: the metric, the benchmark, the hold-out design and the populations are all unstated.
It fits frequent, low-stakes creative pre-tests where speed and German data residency matter. Ask for the validation method behind the accuracy range, and check a sample of its verdicts against your own live A/B results.
quantilope Category Twins: twins built on your brand tracker
quantilope was founded in Hamburg in 2014, though trade press now describes it as US-based. Category Twins, launched in March 2026, are synthetic consumers built on a brand's own tracking data. They refresh with every tracking wave and use Ehrenberg-Bass Institute measures such as Mental Availability and Category Entry Points, with segments such as heavy category buyers. Full global release followed on 21 October 2026, after beta testing with 30 clients in 10 countries.
quantilope has published no accuracy or validation figures for Category Twins. The release cites beta testing, not measured agreement with human answers.
The twins can only reflect what the tracker measures, and quantilope positions them for early-stage, directional feedback and not as a replacement for primary research.
It is relevant only to brands already running quantilope tracking. Ask how the twins score on a tracker wave held back from their construction.
Statista SynthiePop: a statistical population of Germany
Statista SynthiePop, from Hamburg-based Statista, is a synthetic population of 83.6 million virtual individuals representing Germany. Users combine more than 40 consumer attributes and locate target groups down to a 100-metre grid through a conversational interface.
Statista states that SynthiePop combines official statistics with proprietary Statista data, uses no AI-generated data and is built by experts using statistical modelling. It uses no personal data. A published case on the Germany Ticket projected a survey of 5,000 commuters, mobility studies and Federal Employment Agency data onto more than 40 million commuters.
The whitepaper reports no validation results or benchmark figures. Statista itself notes that emotion-driven or situational questions can be modelled only to a limited extent, and that biases in input data carry through.
This is a market-sizing and location-planning tool, not a respondent simulator. It complements the persona tools on this list and does not compete with them, and it covers Germany only.
YouGov e-twins: twins of known panellists
YouGov, the London-listed research group, works with OriginalVoices to build e-twins of 1,008 government-ID-verified members of its own panel. Its Parallax product, launched in July 2026, pairs answers from these twins with fast automated surveys of real people, with configurable sample sizes and validation depth. YouGov accesses AI models through enterprise agreements with providers such as OpenAI, Anthropic and Google, or through managed cloud services such as AWS Bedrock.
In a November 2026 white paper, 693 of the real panellists answered the same new questions as their twins. Panellists judged that their twin's open-ended answers described their own views 75% of the time, falling from 81% for high-confidence answers to 67% for low-confidence ones. On fiction books, a topic absent from the training data, twins came within one point of their human 90% of the time, and adding ratings of more than 4,000 entities lifted person-level accuracy from 78% to 90%.
Source: YouGov white paper as reported by Research Live, 10 November 2026. Self-rated by 693 panellists; answers rated as describing their views very or somewhat well.
The 75% figure rests on panellists' own judgement of their twin, and the results come from YouGov's own paper without independent review. Twins exist only for people already on the panel.
It fits buyers who want a simulated sample they can re-contact and verify. Parallax builds a human check into the product, which is the step this guide asks buyers to run with any provider.
How strong is the published evidence?
We found no independent replication of any offering on a fresh European population. Six of nine have published vendor-run or vendor-sponsored results, one states an accuracy range without defining it and two have published no figures. Read each row for who ran the test and what agreement means.
| Offering | Accuracy figures published | Metric defined | Safeguard against leakage described | Party outside the vendor involved | Population behind the published evidence |
|---|---|---|---|---|---|
| Artificial Societies | Described | Described | Described | Not found in sources reviewed | Benchmark design includes France, Germany, UK and Finland; product results reported on seven sources, including the US General Social Survey |
| Electric Twin | Described | Described | Described | Partly | UK and US |
| Ipsos Data Boosting | Described | Described | Described | Partly | Italy case; 260 validations not broken down by market |
| Lakmoos | Described | Not found in sources reviewed | Not found in sources reviewed | Not found in sources reviewed | Czech Republic |
| Minds | Described | Described | Described | Partly | Germany |
| neuroflash Digital Twins | Described | Not found in sources reviewed | Not found in sources reviewed | Not found in sources reviewed | Not stated |
| quantilope Category Twins | Not found in sources reviewed | Not found in sources reviewed | Not found in sources reviewed | Not found in sources reviewed | None published |
| Statista SynthiePop | Not found in sources reviewed | Not found in sources reviewed | Not found in sources reviewed | Not found in sources reviewed | None published |
| YouGov e-twins | Described | Described | Described | Not found in sources reviewed | Not stated in the coverage reviewed |
Described Described in the sources reviewed Partly Partly Not found in sources reviewed Not found in the sources reviewed
Source: public documents listed in the references, reviewed 18 November 2026. This is a record of what is published, not a quality score. Rows are alphabetical and are not totalled.
| Offering and source | Conducted by | Human benchmark | Leakage safeguard described | Compared with | What agreement is measured |
|---|---|---|---|---|---|
| Artificial Societies benchmark | The vendor; company-funded | 20 public sources, including GSS 2024, ANES, Pew Western Europe and ESS Great Britain; product scored on 7 sources with 2,796 synthetic respondents | Target answers withheld; linked records kept together across splits; contamination not ruled out | Nine general language models, two statistical controls, human repeat items | Total variation distance and other distribution, consistency and effect measures |
| Artificial Societies, January 2026 evaluation | The vendor | 1,000 UC Berkeley surveys | Not described on the page | A 91% human-replication ceiling | Response consistency and distribution accuracy |
| Electric Twin methodology | The vendor | Private survey data, 11,000 UK and US personas | Evaluation questions excluded from persona creation | Eight traditional survey methods; human test-retest ceiling | Aggregate answer distributions: 1-MAE and NDAM |
| Electric Twin, LSE-led study | LSE-led, preregistered; funded by Electric Twin and co-led by its co-founder | 7,755 UK participants across seven sampling services | Not described in the coverage we read | Two opt-in panels, Prolific, two aggregators, two social-media river samples | Whether synthetic estimates fall within the range of human platforms |
| Ipsos Data Boosting | The vendor, with client co-authors | Independent all-human holdout samples across 260 validations | Judged against a holdout, not the training seed | All-human samples | Same action-standard decision |
| Lakmoos research repository | Vendor-affiliated authors | Czech Statistical Office and Median distributions; Ipsos Instant Research data | Not described on the page | Plain and statistically enriched GPT-4-Turbo | Significance tests and effect sizes; no results on the page |
| Lakmoos website claim | The vendor | 20 client benchmark studies in 2025 | Not described | Not described | "Similarity score", undefined |
| Minds, SINUS case | The vendor; SINUS-Institut supplied the benchmark and reviewed the case | SINUS survey, June 2026, n = 2,005; earlier wave, n = 1,047 | Survey results withheld at runtime | A second survey wave, two general AI models, a fixed answer mix | Per-milieu and overall answer shares, mean absolute difference |
| Minds, Mosaic forecasts | The vendor | Published survey answers: 95 questions from 23 surveys, as of 12 November 2026 | Forecasts locked before fieldwork for a 28-question subset | Not described on the page we read | Distribution overlap (one minus total variation distance), average miss per answer option, correlation |
| neuroflash Digital Twins | The vendor | Not stated | Not stated | Generic language-model personas | "Accuracy", undefined |
| quantilope Category Twins | No evaluation published | None reported | Not applicable | None reported | None reported |
| Statista SynthiePop | No evaluation published | None reported | Not applicable | None reported | None reported |
| YouGov e-twins | The vendor, with OriginalVoices | 693 of 1,008 panellists answering new questions | New questions; one topic absent from training data | The same real person | Panellists' rating of their twin's open answers; numeric closeness; person-level accuracy |
Reading the accuracy numbers
The percentages on this list measure different things, so a higher number is not automatically a better product. A 92% decision match, a 92% NDAM score and 91.7% distribution closeness answer three separate questions.
| Metric | What it measures | What it does not tell you | Used by |
|---|---|---|---|
| Mean absolute difference (MAE) | Average gap, in percentage points, between synthetic and real answer shares | Whether any individual answer is right; it shrinks as answer options multiply | Minds |
| Distribution closeness | 100 minus MAE, expressed as a score | Anything beyond MAE; it needs a baseline to mean much | Minds |
| 1-MAE | One minus MAE of response proportions | It flatters questions with many options; a three-option question has a 33% floor | Electric Twin |
| NDAM | Electric Twin's distribution score, normalised from 0 to 1 regardless of option count | Comparability with other vendors' scores | Electric Twin |
| Total variation distance (TVD) | Share of respondents who would need to switch answers for two distributions to match; lower is better | Individual-level fit or relationships between answers | Artificial Societies, Minds (Mosaic) |
| Business decision consistency | Share of tests where boosted and all-human samples reach the same go or no-go decision | The size of score errors, or how often the human answer was "go" | Ipsos |
| Correlation (r) | Whether synthetic and human results move in step | The level: a constant offset can still correlate perfectly | Ipsos, sample-size research |
| Self-rated match | Share of open answers a person says describe their own views | Numeric accuracy; raters may be lenient | YouGov |
| Person-level accuracy | How close one twin's answers are to its own human's | Whether the panel represents the wider population | YouGov |
| Undefined "accuracy" or "similarity" | Unknown | Everything a buyer needs: metric, benchmark and population | neuroflash, Lakmoos |
A worked example. Suppose 70 of 100 product prototypes fail their action standard with human respondents. A rule that answers "fail" every time, with no model at all, matches the human decision 70% of the time. A 92% decision-consistency rate is therefore strong or modest depending on how many tests were "go" decisions, and the Ipsos paper does not report that split. The same logic applies to every score above: ask what a naive baseline would have scored on the same questions.
How the workflows compare
Four pairings cover most of the decisions a European buyer will face. In each, the deciding question is what data you bring, not which vendor scores higher.
Artificial Societies and Electric Twin: modelled populations or your own data
Both simulate audiences, but from opposite starting points. Artificial Societies can model an audience with no prior research archive, using real-world observations, and add client data where it exists. Electric Twin builds from the client's own surveys and interviews, then tests every audience on held-out questions. A buyer with years of UK tracking data has the inputs Electric Twin needs. A buyer facing a new, hard-to-reach audience has fewer options, and Artificial Societies' approach is designed for that case. Run both on the same brief only if you can supply a human benchmark neither has seen.
Ipsos and YouGov: boosting a human sample or checking twins against it
Both keep real people in the loop, in different places. Ipsos collects a small human seed and adds synthetic respondents to reach the sample size a product test needs. YouGov simulates first, then validates with fast surveys of real panellists, at a depth the client chooses. Ipsos fits a defined go or no-go product decision with a known action standard. YouGov fits questions where a buyer wants simulation breadth with a measured error check on the same people.
Minds and neuroflash: segment personas or content pre-tests
Both are self-service and both serve German-language buyers. Minds supports full studies with reusable audiences, and its SINUS partnership gives it a recognised German segmentation. neuroflash scores specific marketing assets, such as a post or a landing page, against twins from its own survey archive. Consider Minds for audience exploration and concept work by segment. Consider neuroflash for high-volume creative screening, keeping in mind that we found no definition of its accuracy figure.
quantilope and Statista: tracker-based twins or a statistical population
Neither is a general respondent simulator. quantilope's twins extend a brand's existing tracker into on-demand conversation. Statista's SynthiePop answers how many people of a given profile live where, across Germany, without generative AI. A brand team asking "how would our heavy buyers react?" needs the first. A strategy team sizing a regional market or planning locations needs the second.
Compare the total cost of the research job
The price of synthetic answers is rarely the main cost. The total includes the human data needed to build or check the model, the validation work and the decisions that ride on the result. We obtained no price quotes; the examples below are published terms only and may have changed since the review date.
| Buying pattern | Published example | What to establish in a quotation |
|---|---|---|
| Boost inside a human study | Ipsos Data Boosting, with a human seed of 50 to 100 per product and estimated savings of 20% to 60% against all-human samples | Seed size per product, the action standard, how divergent cases are flagged and whether a holdout is included |
| Paid pilot, then licence | Lakmoos, a one-month pilot at €10,000 for two to five users, up to 10 datasets and 500 survey questions | What carries into production, how client data is used and kept, languages covered and the definition of the accuracy score |
| Self-service, pay per use | Minds, prepaid at €0.12 per synthetic response, shared across a workspace; neuroflash, free to start | Usage limits, export rights, where prompts and data are processed and which model providers are involved |
| Audience licence | Electric Twin, a licence fee set by audiences and seats, with unlimited queries once an audience is built | Audience build and refresh, seat limits, export rights, contract term and where data is processed |
For tracker-based twins such as quantilope's, the dominant cost is the tracking programme itself; we found no separate published price for Category Twins.
What should determine the final choice?
Five factors decide the choice more than any vendor's headline number. The first and fourth matter more for European buyers than for US ones.
Audience coverage. Specify the country, language, subgroup, period and buying context. Europe is not one population, and the published evidence is narrow: Electric Twin's covers the UK and US, Minds' covers one German questionnaire and Lakmoos' benchmarks are Czech. Independent German work shows the gap. A Technical University of Munich study of the German Socio-Economic Panel reached at best 78.8% accuracy on held-out questions for individual respondents, and its authors suggest English-language results may be stronger. A Nuremberg Institute for Market Decisions study found synthetic answers overstated positive attitudes and showed less variation than real consumers, especially for well-known brands. Neither study tested a product on this list.
The right metric. Agree which number matters before the test. A Prolific study found that the method worst at reproducing answer distributions was among the best at predicting the plurality winner, so a single metric can favour the wrong method. Decide whether you need shares, rankings, a go or no-go call or individual answers.
Fresh, blinded evaluation. Give shortlisted providers the same brief, withhold the human results, agree the acceptable error in advance and include a naive baseline. Pew Research Center's September 2026 test of digital twins against its own US panel found synthetic surveys did a poor job of estimating public opinion and were especially error-prone for some subgroups. Pew tested a general AI model, not a product on this list, but the lesson holds: a vendor's average does not show how it performs on your subgroups.
Data location and model providers. Ask where client data and prompts are processed, and by whom. Stated positions differ: neuroflash says all processing happens in Germany; a listing for Minds places its application in Frankfurt and its model providers in the US; YouGov uses enterprise agreements with providers such as OpenAI, Anthropic and Google, or managed cloud services; Lakmoos and Electric Twin state GDPR compliance without naming a hosting location in the pages we reviewed. We did not assess legal compliance for any provider.
Clear reporting. Require that generated records are kept separate from human respondents, that sources and the human benchmark are documented and that results are reported for every task attempted. YouGov has publicly called on the industry to publish validation methods and outcomes and to agree shared standards; buyers can make that a contract term.
Research methodology
1. Review question and design
The review asked which synthetic research offerings developed in Europe suit which buyer tasks, and how much weight their published evidence can carry. It is a desk review of public material, completed on 18 November 2026, using each company's own website and trade-press coverage. Each named offering was assessed separately, even where a company sells several products.
2. Selection
Nine offerings were profiled, chosen to span seven workflows. Four rules were applied in this order:
- The company was founded or is headquartered in Europe or the UK.
- The offering is commercially marketed and generates synthetic respondents, personas, twins or populations for research.
- Public documentation was sufficient to answer the five assessment questions below.
- Where several eligible offerings served the same workflow, we profiled those with the most detailed published evaluation or the clearest difference in method, up to a limit of nine.
Two offerings with no published accuracy figures, quantilope Category Twins and Statista SynthiePop, were kept because they represent workflows a European buyer is likely to meet.
The list is a sample, not a market census. Omission is not a negative judgement, and inclusion is not an endorsement. The companies below met the first two rules and were identified during the review, but were not profiled in this edition. We did not assess their evidence, so nothing here is a finding about them.
| Company | Base | Offering as the company describes it |
|---|---|---|
| Appinio | Hamburg | Exploratory conversations with synthetic personas |
| experial | Cologne | Digital twins of target groups |
| GWI | London | Synthetic audiences grounded in its own survey data |
| Kantar | London | Synthetic Sample Boosting for small or hard-to-reach groups |
| OriginalVoices | London | Digital twins linked to real people; the technology partner in the YouGov entry |
| SYMAR (formerly OpinioAI) | Brno | Self-service synthetic surveys, focus groups and interviews |
This second list is also incomplete. Any European provider can submit documentation through the directory's submission page, and it will be reviewed against the same four rules and five questions at the next update.
3. Assessment framework
Each offering was checked against five questions: what task it addresses, what input data it needs, what output and delivery model it offers, what published evaluation supports it and what limits apply to that evidence. For this European edition we also recorded what each provider states about data location and model providers. Task matches are editorial judgements. No scores or overall winner are assigned.
4. Verification protocol
Every factual or numerical claim followed seven steps:
- Trace each figure to the document that first reported it.
- Record its specification: metric, denominator, population, dates, comparator and caveats.
- Use the body text over a conflicting headline.
- Use the provider's technical definition over marketing wording.
- Attribute independence claims to whoever made them.
- Count a partner or trade outlet repeating a vendor result as one source.
- Recalculate where possible, and record source inconsistencies openly.
Two inconsistencies were recorded. Electric Twin's methodology post reports 96% on 1-MAE, while third-party summaries of the same post cite 95.5%; we use the post's figure. Ipsos' 2025 infographic describes its sample-size database as 184 products and more than 40,000 consumers, while the 2026 paper describes 185 products and more than 36,000 responses; we use the paper's figures.
5. Evidence profile
The evidence table records, for each source, who ran the test, the human benchmark, any leakage safeguard described, the comparators and the agreement measure. Where a source did not state an item, the table says so and does not infer it.
6. Threats to validity we looked for
| Threat | Why it matters | How it appears in this evidence |
|---|---|---|
| Vendor-run evaluation | The party with the commercial interest designs, runs and reports the test | Artificial Societies, Electric Twin's methodology, Ipsos, Lakmoos, Minds and YouGov |
| Sponsored external study | Funding and co-authorship can shape design and framing | The LSE-led study was funded by Electric Twin and co-led by its co-founder |
| Training-data contamination | Public benchmarks may sit inside a model's training data | Artificial Societies states it cannot rule this out |
| Comparators added late | Controls chosen after results are known can flatter the product | Minds added its AI-model controls after the main result |
| Self-rated accuracy | People may judge a plausible answer as their own view | YouGov's 75% figure |
| Undefined metrics | A score without a definition cannot be compared or checked | neuroflash and Lakmoos |
| Biased human seed | Augmentation carries forward any skew in the human sample | Ipsos' divergent cases |
| Single-country evidence | Results on one population may not transfer to another | Minds (Germany), Lakmoos (Czech Republic), Electric Twin (UK and US) |
7. What it would take to name a "best" provider
A defensible claim would need a defined candidate pool with inclusion rules and success criteria fixed before results are seen. It would need fresh tasks fielded after each system's knowledge cutoff, identical briefs and human benchmarks withheld from providers and large enough to be precise in each European market tested. It would need a naive baseline, subgroup results with uncertainty and reporting of every task attempted. Even then, the conclusion would cover only the tested versions, countries and tasks. None of the reviewed evidence meets this standard, so no winner is named.
8. Checks not performed
We did not commission platforms, recruit human samples, inspect training data, rerun experiments, audit contamination controls, verify country or language coverage, assess GDPR or other legal compliance, verify hosting locations or obtain price quotes. Finding that a source reports a result does not reproduce it.
9. Independence, disclosure and corrections
The publisher runs a directory of synthetic research vendors and sells paid visibility packages to vendors. None of the nine companies profiled was a paying customer at publication, and paying customers receive no preference in editorial comparisons. No company paid for inclusion, supplied content or reviewed this article before publication. Companies were not contacted for comment before publication. Any company, profiled or not, can send corrections or further evidence, which will be reviewed against the same five questions. Corrections will be dated in the article when made.
Company and product names are used only to identify the offerings discussed and belong to their owners. This article is general information for research buyers and is not legal, procurement or compliance advice.
10. Figures
Figures 1 to 5 each plot one company's reported results in that company's own metric, taken from the source named under the chart. They are not comparable with each other. Figure 6 records whether each item was described in the sources we reviewed; it is not a score. In Figure 4, closeness for the fixed answer mix was calculated by us as 100 minus the reported 7.5-point deviation. In Figure 2, river sampling is plotted as the reported range of 93 to 94%, and the axis starts at 90%.
References
- Artificial Societies. (2026, October 1). Artificial Societies is more accurate than state-of-the-art LLMs. Company summary of its benchmark results against nine general models.
- Chidichimo, E., Jung, M. J., Wallis, F. P. S., & He, J. K. (2026, October 1). Artificial Societies Benchmark: A validation framework for synthetic research. arXiv preprint 2609.30030, company-funded.
- Artificial Societies. (n.d.). Artificial Societies vs Simile: Behavioural simulation platforms compared. Product description and the January 2026 evaluation against UC Berkeley surveys.
- Warner, B., & Bailey, A. (2026, February 11). How accurate are synthetic audiences? Electric Twin's hold-out methodology, 1-MAE and NDAM results.
- Research Live. (2026, July 21). Synthetic surveys comparable to traditional research, finds study. Coverage of the LSE-led comparison of seven sampling services.
- CMOtech UK. (2026, July 20). AI synthetic audience matches human polling in study. Coverage noting the study's preregistration.
- Coquelle, P., Pederzolli, B., Reynolds, N., Ho, C., Rimmer, J., & Greco, C. (2026, August 20). The empirical reality of synthetic data in product testing: Methodological validation and a Barilla case study. Research World. Ipsos Data Boosting, 260 validations.
- Ipsos. (2025, February). The power of product testing with synthetic data. Infographic on the sample-size research behind Data Boosting.
- Polišenský, J., Pouč, P., & Zahradníčková, K. (2026, March 4). Which AI for silicon sampling? A comparative evaluation of generative and neuro-symbolic approaches. Zenodo research repository.
- Lakmoos. (n.d.). Lakmoos AI. Product, pilot terms and accuracy claims.
- Minds Team. (2026, October 22). SINUS Milieu-Minds in a reality test against a SINUS survey. Case study reviewed by SINUS-Institut.
- SINUS-Institut. (2026, July 9). SINUS-Institut und Berliner Start-up Minds launchen KI-Personas zu den Sinus-Milieus. Press release, in German.
- European Purpose. (n.d.). Minds. Third-party listing with hosting locations and pricing.
- neuroflash. (n.d.). Digital twins. Product page with calibration base, customers and accuracy claim.
- neuroflash. (n.d.). neuroflash for performance marketing. States data processing in Germany.
- quantilope. (2026, March). quantilope launches synthetic "Category Twins" for early-stage research. Launch announcement.
- TechIntelPro / PR Newswire. (2026, October 21). quantilope makes Category Twins globally available after co-developing with clients. Full release and beta details.
- Daily Research News Online. (2026, March 17). Synthetic research entry for quantilope. Trade coverage of the launch.
- Statista. (n.d.). Statista SynthiePop. Product page.
- Statista+. (n.d.). Rethink research: Synthetic populations in market research. Whitepaper on method, limits and the Germany Ticket case.
- YouGov. (2026, November). All players need to work together to establish widely agreed standards for validating simulated data. Includes the white paper An evaluation of digital twins based on YouGov panel data.
- Research Live. (2026, November 10). Standards needed for validating digital twins, says YouGov. Coverage of the e-twin results.
- YouGov. (2026, July 8). YouGov Parallax: The world's first research product to combine AI twins with validation from real consumers. Product launch.
- YouGov. (2026, June 9). Our approach to artificial intelligence. Model access and data handling.
- Kinzinger, L., & Hartmann, J. (2026, June 3). Synthetic personalities: How well can LLMs mimic individual respondents using socio-economic microdata? arXiv preprint 2606.04592, Technical University of Munich.
- Kaiser, C., Kaiser, J., Schallner, R., Manewitsch, V., & Rau, L. (2026). Leaving insight to digital twins? Promise, progress and limits of synthetic respondents. NIM Marketing Intelligence Review, 18(1), 48 to 53.
- Pew Research Center. (2026, September 30). Can AI stand in for human survey-takers? Not really. Independent test of digital twins against the American Trends Panel.
- Research Live. (2026, August 24). Synthetic respondents can't replace human measurement, finds Prolific study. Coverage of a Prolific preprint.
- CB Insights. (n.d.). quantilope. Company profile with founding year and Hamburg headquarters.
- Minds. (n.d.). Synthetic research and AI audience simulation. Company home page with Mosaic forecast results, as of 12 November 2026.
- Electric Twin. (n.d.). Synthetic audience FAQs: Accuracy, data and cost. Licensing, compliance statements and study roles.
- neuroflash. (n.d.). Evaluation metrics and statistical reliability for synthetic respondents. Company article.
- GWI. (n.d.). Synthetic audiences. Use-case page.
- Kantar. (n.d.). Artificial intelligence at Kantar. Describes Synthetic Sample Boosting.
- Greenbook. (n.d.). OpinioAI. Company listing with Brno location.
- Dealroom. (n.d.). OriginalVoices. Company profile with London headquarters.
- Startbase. (2025, March 27). 2 million euros pre-seed for experial. Funding announcement.
- Synthetic Market Research Directory. (n.d.). Providers. Directory entry for Appinio.