Synthetic Users in 2026: A Primer on Synthetic Respondents, Digital Twins and Synthetic Data

Introduction

Synthetic users are AI-generated respondents that answer research questions as a described person would. They come in three forms: generic personas built from a model’s training data, digital twins built from real people’s own answers, and synthetic data that tops up a real sample.

Independent 2026 studies find them close on headline results but unreliable on subgroups, pricing and change over time. Digital twins checked against real respondents are the most defensible kind.

Last updated November 2026.

What Are Synthetic Users?

Synthetic users, also called synthetic respondents, are simulated research participants generated by a large language model (LLM), prompted or trained to answer questions the way a defined person or group would. A team asks them survey questions, interview questions or reactions to a concept, and gets answers back in minutes instead of the days a real study takes to field.

The category grew quickly in 2026. Qualtrics now sells synthetic panels inside its survey platform. Dovetail offers Digital Twins built on a team’s own research data. Simile, a company founded by the Stanford researchers behind one of the field’s best-known twin studies, announced a funding round led by Index Ventures in February 2026.

The vocabulary hasn’t kept up. Vendors use "synthetic users," "synthetic respondents," "AI personas," "silicon samples" and "digital twins" for overlapping products. The same term can mean a demographic prompt at one company and a model of a single interviewed person at another. That matters because the evidence differs by kind, and a finding about one kind doesn’t transfer to the others.

The ICC/ESOMAR International Code, the main self-regulatory code for market research, draws the line this guide follows. It defines a synthetic persona as a digital representation of a person generated to mimic the behavior and characteristics of real people or groups. It defines synthetic data as information generated to replicate the characteristics of real-world data. The next section splits those two definitions into the three kinds buyers actually meet.

Why Synthetic Users Took Off in 2026

Three pressures pushed synthetic users from research curiosity to budget line in 2026. Research teams are being asked to turn studies around in days, often with fewer people. Language models got good enough to produce fluent, plausible answers on almost any topic. And recruiting real respondents got harder as bots and AI-written answers crept into online panels and open-text fields.

Fluency is the trap. A synthetic answer reads like a real one, so its errors don’t announce themselves. Research leaders at a dinner in San Diego in October 2026 described a broader AI risk that applies here: output that looks passable to a non-expert even when it’s wrong.

Synthetic Users Are Not Design Personas

A synthetic user isn’t the same thing as a design persona, the one-page profile of a target customer that product teams have used for decades. A design persona is a summary written by people from research. A synthetic user is a model that generates new answers on demand. The difference matters because a design persona can’t be wrong about a question nobody asked it, and a synthetic user can.

The Three Kinds of Synthetic Users

The three kinds differ in what they’re built from, and that single difference predicts most of what they can and can’t do.

Synthetic Personas

A synthetic persona is a model answering in character from a short description, such as "a 34-year-old nurse in Ohio who shops for groceries online." Academic papers often call a set of them a silicon sample.

Self-serve synthetic tools often work this way because a persona needs no data from you. The trade-off is that the persona’s knowledge is whatever the model absorbed in training, so it reflects the model’s idea of a nurse in Ohio, not any nurse.

Digital Twins

A digital twin is a model conditioned on a specific real person’s own data: their answers to earlier surveys, an interview transcript, or their product and purchase behavior. In the best-known study, Park and colleagues built twins of 1,052 Americans from two-hour interviews. In market research, twins are used to explore how a customer segment might react before fielding a real survey.

Some products apply the term at group level. Dovetail’s Digital Twins answer in the voice of one account, segment, persona or behavior group, drawing on research a team has centralized in Dovetail plus any connected sources.

Synthetic Data

Synthetic data, in the survey sense, means modeled responses added to a real sample to enlarge it or fill thin cells. Qualtrics and PureSpectrum describe Edge Audiences as a panel offering that combines traditional and synthetic panels.

Blending changes the error profile. In Strat7’s 2026 test, purely synthetic respondents broke the logical order of price questions 68% of the time, and blending with real data cut that to 32.8%.

How to Tell Which Kind a Vendor Sells

Vendor marketing rarely says which kind it sells, but four questions settle it:

Does the product need any of our data to work? If not, it’s a synthetic persona or a population model.
Is each simulated respondent tied to one real person’s records? If so, it’s an individual twin.
Does it answer as a segment, drawing only on research we upload? If so, it’s a segment twin.
Does it add rows to a sample we fielded? If so, it’s synthetic data.

The answers predict the failure modes in the rest of this guide. A product that needs none of your data can only tell you what the model already believes about people like your customers. The next section explains how each kind is built.

How Synthetic Users Are Built

Four construction methods cover the products reviewed here, and the evidence for each is different.

Prompted Personas

A prompted persona is the cheapest method: describe the person in the prompt and ask the question. It’s also the weakest at reproducing a population. A 2026 Prolific study found that simulating individuals from demographic personas roughly tripled distributional error compared with simply asking the model for the overall distribution of answers.

Fine-Tuned Population Models

A fine-tuned population model is trained on large volumes of real survey responses so it predicts how a population will answer. Qualtrics describes its synthetic panels as powered by a first-party model trained on thousands of responses from respondents with varied demographic backgrounds. This method typically performs best on the kinds of questions that dominate its training data and worst on the kinds that don’t.

Individual Twins

An individual twin conditions the model on one person’s own history. Park and colleagues found that twins built from interviews reproduced participants’ General Social Survey answers at 83% of the consistency those same people showed when retaking the survey two weeks later. Twins built from demographics alone reached 74%. That benchmark is important: it measures the twin against how consistent the person is with themselves, not against perfection.

Retrieval-Grounded Segment Twins

A retrieval-grounded twin doesn’t retrain a model at all. It answers by searching a fixed set of your own research, such as interview transcripts, support tickets and survey responses, and writing a reply in that segment’s voice. Dovetail’s documentation describes twins that answer from attached sources, which fits this method. The strength is traceability, since an answer can point back to the quote it came from.

The limit is coverage. If nobody in the source data discussed a topic, the twin has nothing real to draw on, and Dovetail’s documentation says the twin should say so rather than guess.

Elicitation Methods

How the answer is extracted matters as much as who is simulated. Asking a model for a number on a rating scale tends to produce answers that bunch together.

Asking for a text answer and mapping it to the scale, a method called semantic similarity rating, produced more realistic spread in the 2026 NIM study, though not enough. Prolific found that asking for a probability distribution per persona cut error by 35% to 48%.

Why Construction Predicts Failure

Each construction method fails where its inputs run out. A prompted persona fails on anything the model’s training data got wrong about a group. A population model fails on question types its training responses didn’t cover. An individual twin fails on topics the person was never asked about.

A retrieval-grounded twin fails on anything missing from the uploaded research.

How Accurate Are Synthetic Users? What the 2026 Studies Show

Synthetic users are close on averages and unreliable on differences. The five studies below land on some version of that sentence.

| Study | Design | Where it held up | Where it failed | Caveat | |:------------------------:|:-----------------------------------------------------------------------:|:--------------------------------------------------------------------------:|:---------------------------------------------------------------------------------------------------------------:|:-------------------------------------------------------:| | Strat7, July 2026 | 3,000 real respondents compared with two synthetic vendors | Headline results 2 to 3 points from real on average | Willingness to pay about 16% high, price logic broken 68% of the time, change tracked correctly 47% of the time | Industry paper, not peer reviewed | | Prolific, August 2026 | 996 U.S. participants simulated with increasing persona detail | Probability-distribution prompting cut error by 35% to 48% | Demographic personas roughly tripled distributional error, and no method matched a second human sample | Preprint, not peer reviewed | | NIM, Kaiser et al., 2026 | Synthetic consumers compared with real participants in a brand funnel | Brand choice matched about 79% of the time | Attitudes 1.2 points off on a seven-point scale, skewed positive, with too little variance | Tested demographic personas, not twins | | Bisbee et al., 2024 | Synthetic personas compared with the American National Election Studies | Averages close to the human baseline | 48% of regression coefficients differed significantly, and 32% of those flipped sign | Earlier model generation | | Park et al., 2024 | Twins of 1,052 people built from two-hour interviews | 83% of people’s own test-retest consistency, 86% with survey answers added | Demographics-only agents reached 74% | Authors later founded Simile, a commercial twin company |

Strat7: Headlines Hold, Tracking Fails

Strat7 compared a nationally representative sample of 3,000 real respondents with surveys run by two synthetic data companies, repeating its 2025 experiment with one synthetic vendor replaced and a tracking test added. Synthetic headline results landed 2 to 3 percentage points from the real ones on average.

But synthetic willingness-to-pay prices ran about 16% above what real people said, and synthetic data tracked the real change between 2025 and 2026 correctly 47% of the time, which Strat7 called essentially a coin toss.

Strat7’s group AI lead, Hasdeep Sethi, told Research Live that simulating full surveys is probably the wrong use of the technology. He suggested capping synthetic respondents at 10% to 20% of a sample and named digital twins as a better application.

Prolific: Richer Personas Aren’t Enough

Prolific’s researchers asked 996 U.S. participants about political opinions and consumer topics, then tried to reproduce the results with simulated respondents given no context, demographic personas, or demographics plus full interview transcripts. Demographic personas narrowed the spread of answers and roughly tripled distributional error. Interview transcripts cut that collapse by about half but still didn’t beat the zero-context baseline. No method matched the accuracy of a second human sample.

The paper also found that the worst method at reproducing the full distribution was among the best at predicting the most popular answer. A vendor that reports one accuracy metric can pick the one that flatters its method.

NIM: Too Positive, Too Uniform

Kaiser and colleagues at the Nuremberg Institute for Market Decisions (NIM) asked synthetic consumers to move through a realistic marketing funnel and compared them with real participants. Synthetic brand choices matched real ones about 79% of the time but overstated the pull of well-known brands.

Attitude ratings ran 1.2 points off on a seven-point scale, skewed positive, with significantly less variation than real answers. The authors called the better elicitation method an improvement, not a replacement for real data.

Bisbee: Averages Look Right, Relationships Don’t

Bisbee and colleagues, writing in Political Analysis in 2024, compared synthetic personas with the American National Election Studies. Average scores matched closely. But 48% of regression coefficients estimated from synthetic responses differed significantly from the human ones, and among those, the sign of the effect flipped 32% of the time. The same prompt run months later didn’t reproduce the original results, which the authors linked to model updates.

Park: The Strongest Case for Twins

Park and colleagues interviewed 1,052 Americans for two hours each and built a twin of every participant. The interview-based twins reproduced General Social Survey answers at 83% of participants’ own two-week test-retest consistency, against 74% for twins given only demographics. Interview-based twins also reduced accuracy gaps across racial and ideological groups. Several of the authors, including lead author Joon Sung Park, went on to found Simile.

How to Read a Vendor Accuracy Claim

Vendor accuracy figures are hard to compare because they measure different things. Five checks help:

Ask what the benchmark is. Park’s 83% is a share of people’s own test-retest consistency, not raw accuracy, so it reads higher than a raw match rate would.
Ask whether the claim covers averages or distributions. Prolific showed a method can predict the most popular answer well while badly misreading the spread.
Ask whether subgroups and change over time were tested. Strat7’s headlines held while its tracking failed.
Ask who ran the test. A vendor-run validation on studies the vendor chose is weaker evidence than an independent comparison.
Ask whether the work is peer reviewed. Several of the most-cited 2026 results are preprints or industry papers.

Qualtrics publishes no quantitative accuracy figures on its synthetic panels help page. A missing figure is itself an answer, and a vendor that can’t answer these five checks hasn’t shown you accuracy yet.

What the Studies Didn’t Test

The published evidence has gaps that matter for business buyers. None of the five studies tested twins of a company’s own customers built from that company’s own survey data. Most tested consumer or political questions in the U.S. or Europe, not niche business-to-business roles.

And Qualtrics limits its synthetic panels to English, so evidence outside English is thin. Treat any result in those areas as untested until your own hold-out check says otherwise.

What the Pattern Means

The pattern across five studies is consistent. Synthetic respondents reproduce the center of a distribution reasonably well and miss its spread, its subgroups and its movement.

Most business decisions depend on those three things: which segment differs, which concept beats which, and whether a number moved. Twins built from real individual data close part of the gap, which is why the next section starts with where synthetic users earn their place.

Where Synthetic Users Work

Synthetic users are a useful research instrument in five situations, and teams that dismiss them outright give up real speed in each one.

Pre-Testing a Questionnaire

A synthetic run can expose confusing wording, double-barreled questions and broken logic paths before a single real respondent sees the survey. A bad question costs a synthetic run minutes. The same question sent to, for example, 1,000 real respondents costs a field period.

Generating Hypotheses Early

When a product has no users yet, synthetic personas can generate objections, use cases and vocabulary to test. Research leaders at the same San Diego dinner broadly accepted this zero-to-one use, as long as real data validated the output afterward. Synthetic personas are a starting point for foundational research, not a substitute for it.

Narrowing Message and Stimulus Variants

Synthetic personas are a cheap filter for a long list of options. A team with, for example, 20 draft headlines can use a synthetic run to find the ones that are confusing or off-tone and cut the list to five before a real message test. The synthetic run decides what gets tested, never which variant wins.

Estimating Headlines in Familiar Categories

In its national consumer sample, Strat7 found synthetic headline results averaged 2 to 3 points from real ones. For a rough first read on a well-known brand’s overall standing, that’s often close enough to decide whether a real study is worth running.

Boosting Hard-to-Reach Groups

Strat7 flagged boosting hard-to-reach groups as potentially promising, with caution. A small number of modeled cases, clearly labeled and capped, can fill a cell that real recruiting can’t reach in time.

Moving Faster and Spending Less

Synthetic answers arrive in minutes and cost a fraction of incentives and panel fees. Qualtrics and PureSpectrum say synthetic responses can cut fielding costs by up to 50%, a vendor claim published without methodology. Speed is a real advantage for exploratory work, and it’s the reason the category exists.

Rehearsing Stakeholder Questions

Twins of a known audience can rehearse a conversation before it happens. Simile’s chief executive told Bloomberg, as SiliconANGLE reported, that its model correctly forecast eight of 10 analyst questions on a simulated earnings call, a company claim rather than an independent result.

The same idea applies to a research readout: ask a twin built from stakeholder interviews which objections a finding will meet, then prepare the evidence for each one.

In every one of these uses, a wrong synthetic answer costs little and gets caught before anyone acts on it. Apply that test to any new use case.

Where Synthetic Users Break, and the Better Instrument for Each

Synthetic users fail in predictable places. Each failure below names the instrument that answers the question better.

Comparing Subgroups

Strat7 found large gaps between real and synthetic answers in subgroup analysis, and Bisbee found relationships between variables were unreliable. Even Qualtrics' own guidance plans about 350 responses per subgroup analyzed separately. The better instrument is a real survey with a sample sized for every subgroup you’ll report.

Tracking Change Over Time

Synthetic data tracked real year-over-year change correctly 47% of the time in Strat7’s test, and model updates can shift results on their own. The better instrument is a real tracker for experience measurement, with a frozen questionnaire fielded to the same population each wave.

Setting a Price

Synthetic willingness to pay ran about 16% high, and purely synthetic respondents broke price-question logic 68% of the time. The better instrument is a pricing study such as Van Westendorp or Gabor-Granger fielded to real target buyers.

Testing a New Brand or Category

Strat7 flagged new entrants and low-penetration brands as weak spots, and NIM found synthetic consumers overstated the pull of well-known brands. A model can only echo what it has read. The better instrument is a concept test with real respondents from the target market.

Measuring Past Behavior, Recall and Awareness

Qualtrics' own help page says its synthetic panels are poorly suited to past behavior, detailed recall and brand awareness. A model doesn’t remember what a person bought last month. The better instrument is the behavioral data itself, since real recall is also imperfect, or a real survey where no records exist.

Claiming Statistical Significance

Synthetic responses vary less than human ones, so adding synthetic rows to a sample shrinks the error bars without adding information. Bisbee and colleagues showed the undersized variance would make power analyses optimistic. The better instrument is a real sample sized for the effect you need to detect, with synthetic rows excluded from any significance test.

Asking About Something That Just Happened

A model can’t know about an outage last Tuesday or a pricing page that changed this morning. Strat7 found situational questions were a weak spot. The better instrument is an in-product survey that reaches real users at the moment the experience happened.

Representing People the Model Rarely Sees

Language models learn from text, and some groups write far less online than others. Synthetic answers for those groups lean on stereotypes absorbed in training, which is the opposite of what research on an under-served segment needs. Bisbee and colleagues found synthetic answers varied far less than human ones on questions about racial and religious groups.

The better instrument is targeted recruiting of real respondents from that group, even if the sample is small.

Reaching Niche Business Roles

The same problem hits business-to-business research. A model has read plenty about consumers and very little about how a procurement lead at a mid-sized hospital network weighs a software renewal. The better instrument is a real sample of that role, recruited through a business panel with job-level targeting or drawn from your own customer list.

Confirming What Leadership Already Believes

A synthetic study can be prompted, rerun and reframed until it agrees with the plan, and nobody outside the team can tell. Research leaders at the San Diego dinner named the risk of telling leadership what it wants to hear as a central worry.

The better instrument is a real study with the decision metric and the threshold written down before fielding, so the result can’t be tuned after the fact.

Governance, Disclosure and the Questions to Ask a Vendor

The ICC/ESOMAR Code adds disclosure duties for synthetic research, and twin source data is personal data.

What the ICC/ESOMAR Code Requires

The 2025 ICC/ESOMAR International Code sets disclosure at three levels. Article 4 requires researchers to notify a data subject at the start of the research when a synthetic persona is used for data collection.

Article 7 requires telling the client when synthetic data or synthetic personas are used, including the extent of human oversight. Article 9 requires disclosing on publication whether synthetic data played a significant role in sampling, deployment, analysis or interpretation.

The Vendor Question Checklist

The answers separate a vendor with documented limits from one without them. Ask every synthetic research vendor:

What data trained the model, and when was that data collected?
Does any of our data train a model that other customers use?
How long is our data retained, and can we delete it on request?
Which underlying model runs the product, and is the version pinned between studies?
What validation evidence is published, against which human benchmark, and was it independent?
Are synthetic responses labeled at the row level in exports and dashboards?
Which question types and topics does the vendor document as weak?
Can we run the same instrument on real respondents to check the result?

Question 2 is often the hardest for a vendor to answer. A twin built from your own consented respondents keeps the answer simple, because you know what it learned from. Question 4 matters because Bisbee’s reruns drifted after a model update. Qualtrics answers question 6 well: it tags synthetic responses with their own response type for filtering.

Labeling Synthetic Results Inside Your Company

Internal readouts need the same honesty as published ones. Label every synthetic result as synthetic on the chart itself, not in a footnote, and keep synthetic rows in a separate column or filter from real ones. Record the model, version and date beside the result. A readout that mixes synthetic and real answers without labels can’t later say which findings came from people.

Treating Twin Source Data as Personal Data

A twin built from identifiable people’s answers is built from personal data, so the privacy rules that covered the original survey still apply. Involve the privacy team before building one, confirm the original consent covers modeling, and make sure a deletion request removes the person from the twin’s source data as well as from the survey.

Protecting the Real Half of the Data

Synthetic research raises a governance question on the real side too. Bots and AI-written answers are increasingly hard to spot in online samples, and a hold-out check is only as good as the real sample behind it. Use attention checks, screen open-text answers for copied or generic text, and prefer sources where you know who the respondent is, such as your own customers.

Synthetic Research Vendors in 2026

The vendor market sorts by kind more cleanly than by feature list. The four below are representative of the main approaches, and this list doesn’t rank them.

Qualtrics Synthetic Panels

Qualtrics sells synthetic panels through Edge Audiences, bought with subscription credits through an account executive. Qualtrics' help page limits them to a U.S. general population panel in English, recommends closed-ended attitudinal questions, and lists past behavior, detailed recall and brand awareness as poor fits. The page asks users to report results transparently as generative AI output and publishes no quantitative accuracy metrics.

Dovetail Digital Twins

Dovetail’s Digital Twins answer in the voice of a customer segment and draw only on research a team has centralized: interviews, transcripts, support tickets and survey responses. Dovetail’s documentation is direct about limits. It says twins don’t model statistics or forecast numbers, that quotes should be checked against the source, and that a twin doesn’t replace talking to customers.

Simile

Simile builds digital twins of individuals and was founded by Joon Sung Park, Michael Bernstein and Percy Liang, authors of the Park twin study. SiliconANGLE reported in February 2026 that Simile trained its model on interviews with hundreds of people plus transaction logs, named CVS Health and Telstra as customers, and that Index Ventures led its funding round.

Electric Twin

Electric Twin is a London-based company that builds simulations of a customer’s target audiences to test decisions, according to Research Live, which reported a funding round led by Atomico in February 2026. Research Live’s report doesn’t describe its construction method, so buyers should put the vendor checklist above to it directly.

Platforms built around AI-moderated interviews increasingly add synthetic features as well. Those remain research platforms with a synthetic capability, and the same checklist applies.

How to Run a Vendor Pilot

A short pilot on a study you’ve already fielded is the fastest fair test of any vendor. Pick a past study with known real results, ideally one with subgroups and a tracked metric, and give the vendor the identical questionnaire without the answers. Then score the output on five checks:

The five checks are headline, ranking, subgroups, spread and change, and each should be scored against a threshold you write down before seeing the output. A vendor that passes the headline and fails the rest is a brainstorming tool, not a measurement tool.

How to Check a Synthetic Result Against Real Respondents

A synthetic result becomes evidence only after a real sample agrees with it. The hold-out check below takes a real field period and gives the synthetic result a measured gap against real respondents.

The Hold-Out Protocol

Freeze the instrument. Use identical questions, wording, answer options and order for the synthetic and real runs.
Name the decision metric before either run, such as the share choosing each concept or the ranking of three messages.
Run the synthetic study and record the model, version and date.
Field the same instrument to real respondents from the population the synthetic sample was meant to represent.
Split the real sample into two random halves and compare them. The gap between the halves is your noise floor.
Compare the synthetic result with the real one on the headline, the ranking, the subgroup differences and the spread of answers.
Decide before fielding what size of gap would change the decision, and act on the real result if the synthetic one crosses it.

Step 5 is easy to skip and the most useful. Prolific found no synthetic method matched a second human sample, so the human-to-human gap tells you the best any synthetic run could do. A synthetic result that falls inside the noise floor is as close as a second fielding on that metric. One that falls outside it isn’t.

A Worked Example

Here is an illustrative case, with invented numbers. A product team uses synthetic personas to compare three onboarding messages, and the synthetic run ranks message A first by a wide margin.

The team fields the same three messages to real respondents. The real halves agree with each other, and both rank message B first, with A second. The synthetic run got the headline interest level about right and got the decision wrong.

When to Re-Run the Check

A passed check isn’t permanent. Re-run it when the vendor changes the underlying model, when you move to a new category or audience, and on a regular schedule for any synthetic method you rely on. Bisbee and colleagues found the same prompt produced different results months apart after a model update, so a check from last spring says little about this fall.

Fielding the Real Half With Sprig

Sprig is built for the real-respondent half of this check. Teams can field the frozen instrument to an external research panel, by email to their own customers, or through in-product surveys on a website or web app.

The Design Agent flags unclear questions and broken logic before launch, and a researcher approves the final study. Sprig can’t simulate a population it has no respondents for, and a real field period takes longer than a synthetic run.

How to Build Digital Twins From Your Own Respondents

Digital twins are only as good as the real data underneath them, so a twin program starts with collecting the right first-party responses. Five steps cover it:

Collect consented data with twin use in mind. Tell respondents their answers may be used to model customer segments, which keeps the original consent aligned with how the data is used.
Scope each twin to one segment, account type or behavior. Dovetail’s documentation warns that a twin blended from all customers averages away the differences you’re trying to find.
Give each twin both kinds of data: structured survey answers for what people choose, and open-text answers or interviews for how they explain it.
Document the sources, dates and model behind every twin, so anyone reading an answer can trace it.
Refresh the source data as customers change, and re-run the hold-out check after each refresh.

Step 3 has modest support. Park’s twins did best when interviews and survey answers were combined, at 86% of participants’ own test-retest consistency, against 83% for interviews alone and 82% for survey answers alone.

The bottleneck is step 1. A twin built from a few interviews collected years ago reflects that year’s customers. Teams that run regular surveys with their own customers already hold the raw material, and their twins stay as current as those surveys.

Collecting Twin-Ready Data With Sprig

Sprig collects that raw material from real customers through email surveys and in-product surveys that pair rating questions with open-text follow-ups. The Field Agent adds AI follow-up questions in field.

For any wave used as a hold-out or tracker, keep the instrument identical for every respondent by leaving AI follow-ups out of the measured questions. The same studies then serve as the real-respondent check for any twin built from them.

Which Kind of Synthetic User for Which Question

The router below assigns 12 common research questions to the instrument that answers them best.

The rule behind the table is simple. If the answer will be reported as a number someone acts on, it comes from real respondents. Synthetic personas belong to questions about wording and ideas. Digital twins belong to questions about how a known customer base thinks. A concept and prototype test or market sizing study belongs to real respondents.

A worked pairing shows how the two halves connect. A team builds a digital twin of its enterprise admin segment from, for example, 40 interview transcripts and asks it to react to a new permissions model. The twin surfaces three likely objections.

The team writes one survey question per objection and fields it to real admins, which tells them how many admins hold each objection and which one to fix first.

Frequently Asked Questions

What Is the Difference Between Synthetic Users and Digital Twins?

Synthetic users is the umbrella term for AI-generated research participants. A digital twin is one kind: a model built from a specific real person’s or segment’s own data, such as past survey answers or interview transcripts. A generic synthetic persona is built from a demographic description and the model’s training data instead. Twins are grounded in real individuals, which is why they outperform demographic personas in the published evidence.

Are Synthetic Respondents Accurate?

Synthetic respondents are accurate on averages and unreliable on differences. Strat7’s 2026 study found headline results averaged 2 to 3 points from real ones, but willingness to pay ran about 16% high and synthetic data tracked real change correctly only 47% of the time. Prolific found no method matched a second human sample. Treat synthetic results as directional until a real sample confirms them.

Can Synthetic Respondents Replace a Survey Panel?

Synthetic respondents can’t replace a survey panel for measurement. They miss subgroup differences, change over time and pricing, which are the outputs most panel studies exist to produce. They can replace some exploratory steps that came before a panel study, such as pre-testing wording or generating hypotheses. Strat7’s AI lead suggested capping synthetic respondents at 10% to 20% of a sample.

What Is Synthetic Data in Market Research?

Synthetic data in market research is modeled survey responses generated to replicate the characteristics of real data. It’s typically used to enlarge a real sample or fill thin cells for hard-to-reach groups. Blending matters: in Strat7’s test, purely synthetic respondents broke price-question logic 68% of the time, and blending with real data cut that to 32.8%. Exclude synthetic rows from significance tests.

Do Digital Twins Work Better Than Synthetic Personas?

Digital twins built from real individual data outperform demographic personas in the published evidence. Park and colleagues found interview-based twins reached 83% of people’s own test-retest consistency on the General Social Survey, against 74% for demographic-only agents. But Prolific found interview transcripts still didn’t beat a zero-context baseline in its test, so twins narrow the gap without closing it.

When Should I Use Synthetic Users?

Use synthetic users for work where a wrong answer is cheap: pre-testing survey wording, brainstorming objections and use cases, preparing interview guides, and getting a rough headline read in a familiar category. Don’t use them for pricing, tracking, subgroup comparisons, new brands, recall or anything reported as a number someone will act on. Those questions need real respondents.

Do I Have to Disclose Synthetic Respondents?

Researchers following the 2025 ICC/ESOMAR International Code have to disclose synthetic research at three levels. Article 4 covers notifying data subjects when a synthetic persona collects data. Article 7 covers telling clients when synthetic data or personas are used, including the extent of human oversight. Article 9 covers disclosure on publication. Qualtrics also asks users to report synthetic panel results transparently.

What Should I Ask a Synthetic Research Vendor About My Data?

Ask what data trained the model, whether your data trains a model other customers use, how long your data is retained and whether you can delete it, and which underlying model runs the product. Also ask whether the model version is pinned between studies, since Bisbee and colleagues found results drifted after a model update. A twin built from your own consented respondents keeps these answers simple.

How Do I Validate a Synthetic Research Result?

Validate a synthetic result by fielding the identical instrument to real respondents from the same population. Split the real sample into two halves to measure the human noise floor, then compare the synthetic result on the headline, the ranking, subgroup differences and the spread of answers. Decide in advance what size of gap would change the decision.

Are Digital Twins the Same as Customer Personas?

Digital twins aren’t the same as customer personas. A customer persona is a written profile that summarizes research about a type of customer, and it doesn’t answer new questions. A digital twin is a model built from real people’s data that generates new answers on demand. A persona can be out of date, but it can’t invent an answer. A twin can, which is why twins need a hold-out check.

Will Synthetic Users Get More Accurate?

Synthetic users are improving, and the 2026 studies show where. Strat7 found willingness-to-pay answers bunched less around round numbers than in its 2025 test, and Prolific found better prompting cut error by 35% to 48%. But no method in Prolific’s study matched a second human sample, and tracking change stayed close to a coin toss in Strat7’s test. Expect better exploration tools before a replacement for measurement.

How Many Real Respondents Do I Need to Check a Synthetic Result?

The real sample needs to be large enough to measure the decision metric with the precision the decision requires, including every subgroup you’ll compare. Qualtrics' own guidance for its synthetic panels plans about 350 responses per separately analyzed subgroup for a margin of roughly five points at 95% confidence. The same arithmetic applies to the real hold-out.

How Is Synthetic Research Priced?

Synthetic research pricing varies by vendor and is often sold inside a broader subscription. Qualtrics sells synthetic panels through Edge Audiences with subscription credits bought through an account executive, and each synthetic response consumes both interactions and credits. Qualtrics and PureSpectrum said in August 2025 that synthetic responses can cut fielding costs by up to 50%, a vendor claim published without methodology. Ask each vendor for its pricing model directly.

How Does Sprig Fit With Synthetic Research?

Sprig is an enterprise survey platform for collecting responses from real people through research panels, email, links and in-product surveys. In a synthetic workflow, Sprig fields the real-respondent check and collects the first-party survey responses a twin can be built from and checked against.

The Synthesize Agent groups open-text responses into themes a researcher can trace back and review, which is useful when comparing real and synthetic answers side by side.