Synthetic respondents are AI-generated survey participants, usually built to resemble a population or segment. AI personas are conversational customer archetypes, which may be prompted or grounded in research. Digital twins, in the individual-level sense used here, model specific real people from their own records. Customer Models, Meaningful's term, are research-grounded segments that give directional, source-linked reactions to new questions.
Vendors use the labels differently. Ask what evidence goes in, whom an answer represents and whether it is observed or generated. A convincing quotation or neatly formatted percentage cannot settle those questions.
Meaningful makes Customer Models, so we have an interest in this comparison. Apply the same standard to us: inspect the evidence and limits before using output with a client.
The short answer: how do the four approaches compare?
The four approaches differ mainly in their evidence source and unit of analysis; none turns a newly generated answer into a measured response from a human sample. A measured survey result may be displayed alongside a model's answer, but it remains a result from the original study, with its original base.
| Approach | Built from | What it represents | New output | Safe first use |
|---|---|---|---|---|
| Synthetic respondents | Population, panel, survey or other training data; method varies | Simulated survey takers, not necessarily matched to actual individuals | Generated answers, sometimes aggregated as synthetic percentages; not new human measurement | Questionnaire pilots and hypotheses to validate with respondents |
| AI personas | A short prompt, or a research-backed profile if the vendor provides one | A customer type or archetype | Directional, simulated conversation | Brainstorming possible objections and research questions |
| Digital twins, individual-level | Records from specific real people | One model per person, sometimes updated as evidence changes | Predicted or simulated answers, not observed new answers | Exploring gaps in an existing panel, subject to validation and consent |
| Meaningful Customer Models | Research selected from your workspace | Research-grounded customer segments, reviewed by a person | Directional reactions with links to supporting analyses; earlier measured figures kept separate | Comparing likely segment objections before human testing |
Evidence can range from a prompt to your own selected studies. Evidence-grounded does not mean measured: even a model built from interviews has not asked anyone about your new stimulus. These are working definitions, not universal vendor terminology.
What are synthetic respondents?
Synthetic respondents are simulated survey participants that answer questions as if they belonged to a specified population or segment. Some systems use statistical methods, some use large language models, and some combine generated answers with real data. The percentage of simulated respondents choosing an option is a property of that system's output, not a new estimate obtained by asking those people.
You can rehearse a questionnaire or stress-test answer lists before fieldwork. Generated responses might expose ambiguous wording or missing options, provided the pilot is not sold as the study.
What happens when you benchmark synthetic results against people? In STRAT7's July 2026 UK snacking test, two providers' boosted synthetic samples were, on average, 2.0 and 2.8 percentage points from a real holdout sample on brand-awareness toplines. That sounds small until the decision depends on movement rather than a static headline. One provider got the direction of year-on-year awareness change right for just 19% of 16 brands in the comparison.
The test also found 68% internally contradictory price ordering for one provider's purely synthetic respondents. A subgroup result missed the human sample by 17 points despite agreement overall. These are results from particular methods, not universal error rates. Pew Research Center's September 2026 experiment also found synthetic polls poorly reproduced human results, especially for some subgroups; it used individual-level twins, showing that categories overlap in practice.
Use synthetic respondents to find questions worth asking. For trends, segmentation, pricing or a client-facing population figure, collect real responses. If blending generated and human data, demand identification, weighting and subgroup validation.
What is an AI persona?
An AI persona is a conversational representation of a customer type, often asked to explain motivations or react to an idea in that type's voice. At one end, someone writes “you are a busy parent buying snacks” and starts chatting. At the other, a team builds a profile from interviews and can show which observations informed it. Both may be called an AI persona, but the second has a stronger evidential basis.
Suppose you show the line “30% less sugar. Same crunch kids love.” to a prompted “budget-conscious parent”. It may say that the claim raises questions about price and taste. That gives your team a useful hypothesis for an interview guide. It does not establish that parents raised either objection, or that the objections are common among UK snack buyers.
Fluency causes trouble when a vivid quote has no respondent behind it. Put the output under “possible objections to test”, not “what consumers told us”. For a research-grounded profile, inspect the source studies and contrary evidence.
As with general-purpose chatbots versus purpose-built research platforms, persuasive conversation is not an auditable research workflow. The task decides which you need.
What is a digital twin in market research?
In this guide, a digital twin is a model of a particular real person's attitudes or behaviour, built from that person's own records. It may use past survey answers, interviews or interactions to simulate an answer the person has not actually given. The engineering term “digital twin” is sometimes applied much more broadly in customer research, so always ask whether the counterpart is an individual, account or segment.
There is evidence for the potential of individual grounding. In Joon Sung Park and colleagues' study, researchers created agents from interviews and surveys with 1,052 people. On held-out General Social Survey items, interview-only agents performed at 83% of the participants' own two-week test-retest consistency, compared with 74% for demographics-only agents. Those figures are relative to human self-consistency, not claims of 83% and 74% absolute accuracy on future customer decisions. The study suggests richer individual evidence helps for its tested tasks.
Junsol Kim and Byungkyu Lee's related AI-augmented survey work found stronger performance filling historical gaps than predicting entirely unasked opinions. It is not an individual-twin test; imputed answers must remain labelled as inference.
There is evidence against assuming that more personal data makes a twin a reliable stand-in for a person in a new study. Tianyi Peng and colleagues' Science Advances paper, led by researchers at Columbia, tested twins built from each person's answers to more than 500 previous questions. Across 19 preregistered studies and 164 outcomes, twin answers were only modestly more accurate than a base language model's and correlated weakly with human answers, with an average correlation coefficient of 0.20. The authors documented insufficient individuation, stereotyping, representation bias, ideological bias and hyper-rationality.
These studies test different tasks. Park's held-out survey result does not promise accuracy on a new concept; Peng's finding does not negate every backfilling task. A twin may help you decide what to ask next. It is not the person you would ask.
Are AI personas the same as digital twins?
No: under the working definitions here, an AI persona describes a type of customer, while an individual-level digital twin simulates one identified person using that person's records. An evidence-grounded persona can still be a segment, and a twin's generated response is still not an observed response.
Market Logic calls research-backed segments “synthetic personas”. Dovetail allows a “customer digital twin” to represent a customer, account, segment or persona. Neither label reveals the unit or source.
Ask a vendor: “This answer stands for ______, and it is based on ______.” Then ask what a percentage means and which observations can be inspected. Compare methods, not labels.
What is a customer model in market research?
A customer model, as Meaningful uses the term, is a set of segment-level representations built from research you select, then used to explore directional responses to new questions. It is not an industry-standard label that settles how every vendor builds a model. Meaningful's Customer Models introduction describes the product and its boundaries in more detail.
You select research from your workspace. The system drafts four groups; you can edit, add or merge them. Reasoning links to supporting analyses, source types stay distinct, and answers use a saved model version. It does not guarantee that every raw passage is reread per query or that an analysis captures every minority view.
Take the fictional snack line “30% less sugar. Same crunch kids love.” One group might question taste, another might care more about price and a third might wonder whether the claim fits a school lunchbox. In the product's illustrative UK snack-shopper example, those are hypothetical, directional reactions, not quotations from shoppers or a fresh quant study. The useful outcome is a sharper stimulus and a better list of follow-up questions.
The model can sit beside earlier survey figures, if those figures came from an actual study and retain their base. It cannot create a new “percentage of shoppers who would buy” simply because a group sounded favourable. That separation matters when you are combining social listening with surveys and interviews: posts, respondent quotes, segment summaries and survey bases each represent different units. A customer model is neither a twin of an individual nor a survey emulator, and Meaningful does not present its directional reactions as purchase, share or conversion predictions.
Measured or directional: what does the output actually mean?
A measured result describes responses actually collected under a stated method; a directional output is a generated interpretation or simulation that helps decide what to investigate. Both can be useful. They answer different client questions, and a chart does not change which kind of evidence it contains.
| What you see | What it means | What it does not mean |
|---|---|---|
| “42% chose option A” in a documented survey of 500 respondents | 210 observed selections out of that study's stated base, subject to its sampling and design limits | That the percentage applies to a new market or a changed question without further work |
| “42% chose option A” among 500 generated synthetic answers | 210 selections made by the simulation under its settings | That 210 people answered, or that 42% of a target population agrees |
| “This customer group may object to price” in a source-linked model | A directional reading of selected research, to check against evidence and test | A count of unique respondents or a predicted share of buyers |
The unit matters: ten excerpts might come from one interview or ten. A segment is a grouping, not a respondent. A twin's new answer remains simulated even when linked to one participant; 500 synthetic “respondents” means 500 model outputs, not 500 newly observed humans.
Before a deck uses a new percentage, ask for its human base, question wording, field dates, recruitment and weighting. For an earlier measured figure, show those original details and keep it visually separate from the model's new directional answer. “This may be a concern” is a useful hypothesis; “42% of customers are concerned” would be a very different claim.
Which approach should your agency use first?
Start with the decision you need to support, then choose the tool whose evidence and unit match it. There is no universal first purchase for every agency, and the right way to validate an output depends on whether you need a question, an interpretation or a measurement.
- Piloting a questionnaire? Try synthetic respondents for obvious wording and logic problems, then field a human pilot. If you need prevalence, trend or pricing data, plan a proper human study.
- Brainstorming possible objections? Use a prompted AI persona as stimulus. Mark every objection as a hypothesis and check it against interviews or existing research.
- Exploring missing answers in an existing panel? Consider individual-level twins where you have appropriate data and permissions. Flag imputed answers and validate them against held-out human answers before making a claim.
- Responding to a new message using studies you already own? Use a research-grounded customer model for source-linked, directional segment reactions. Test promising and contradictory interpretations with people before giving a client a number.
Separate those jobs in your shortlist of AI tools for market research agencies. “Can generate a customer answer” is not enough.
What should you ask a vendor before trusting the output?
Ask a vendor to show the route from input evidence to a claim you could defend in a client meeting. These five questions expose the differences more quickly than a polished demo:
- Who or what does one answer represent? A simulated survey taker, a real individual's model, an archetype or a segment?
- What sources went in, and who chose them? Can you remove a study, distinguish source types and see what is missing?
- What is actually traceable? Is the link to a raw respondent record, an analysed excerpt or a summary, and can a reviewer check the original?
- Which outputs are measured, inferred or generated? If there is a percentage, where is the observed sample and base?
- How do you validate and govern it? Ask for a held-out test on your use case, model versioning, consent boundaries and data-handling terms.
See the six questions in the Customer Models introduction for more depth. If a vendor cannot answer them for your dataset, keep its output at the hypothesis stage.
Frequently asked questions
The short distinction is between what a model represents and what its output can prove. The answers below use “digital twin” in the individual-level sense unless they explicitly note a vendor's broader usage.
What is the difference between a synthetic persona and a digital twin?
A synthetic persona usually represents a customer type, while an individual-level digital twin simulates a particular real person from that person's records. Vendors use both names more broadly, so check the evidence source and unit of analysis before relying on either label.
What are synthetic respondents in market research?
Synthetic respondents are AI-generated survey takers that produce simulated answers, often as a population or segment. STRAT7 found close brand-awareness toplines in one test but a serious tracking failure for one provider; that does not make synthetic outputs a substitute for a measured survey.
Are AI personas the same as digital twins?
No. A prompted AI persona is an archetype and may have no study or real person behind it; an individual-level twin draws on a particular person's records. Both can generate convincing answers that the person has not actually given.
What is a customer model in market research?
Meaningful's Customer Models are sets of research-grounded segments built from studies you select and review. They provide directional, source-linked reactions tied to a saved model version, not fresh measurements of how many customers agree.
Can synthetic respondents replace a survey?
Not when you need an observed estimate of what people think or do. Synthetic answers may help pilot questions, but STRAT7's tracking and subgroup misses show why a client-facing measurement needs appropriate human fieldwork and validation.
How accurate are digital twins?
There is no single accuracy rate: it changes with the source data, question and validation task. Park and colleagues found gains on held-out survey items relative to a demographics-only model, while Peng and colleagues found weak individual-level correlation on a wider set of new outcomes.
Can I put a percentage from an AI persona or customer model in a client deck?
Only if the percentage comes from an earlier measured study and is labelled with its original question, sample base and method. A generated persona's or customer model's new leaning is directional, not a percentage of real customers.
Which should an insight agency use first?
Start with the decision: a persona for brainstormed objections, synthetic respondents for a questionnaire pilot, or a source-linked customer model for reactions grounded in research you own. Use individual twins only with suitable person-level data and a validation plan; field with people when the client needs a new number.
None of these methods replaces asking people when the decision requires a measured answer. Their best contribution is to focus your next fieldwork; our guide to introducing AI into an insight agency covers the wider workflow. To see a directional model before uploading any research, request access to Meaningful and try the fictional UK snack-shopper model in the Models tab. It uses illustrative material, not real customer data.