Can Consumer Research Rely on Synthetic Respondents to Predict What People Will Buy?

Synthetic respondents can support exploration, pretesting, and early-stage research design, but they should not be treated as reliable stand-ins for real consumers when the question is what people will buy. Purchase decisions raise a higher evidentiary bar because a plausible survey answer is not the same as an observed action.

For consumer insights leaders, the practical question is not whether synthetic respondents are simply useful or unreliable. It is where they are sufficiently reliable to inform a decision, where human validation remains necessary, and when behavioral evidence can help test modeled expectations against what consumers actually do.Synthetic Respondents in Consumer Research: Key Takeaways

• Synthetic respondents can support consumer research tasks such as exploration, pretesting, and early hypothesis exploration, but their reliability varies by task and context.
• Reproducing a plausible survey response is not the same as accurately predicting what a consumer will buy.
• Purchase-related decisions warrant stronger validation, particularly when findings will inform significant commercial investment.
• Human research, experiments, and observed behavioral evidence can help test whether synthetic findings hold for the audiences and outcomes that matter.

Measure Data gives research teams access to underlying permissioned behavioral files they can analyze alongside surveys, experiments, or synthetic research. Clients decide how to query, combine, and interpret that evidence for their own research questions. This gives businesses another way to test modeled expectations against observed consumer activity without conflating the data source with the analysis.

What are Synthetic Respondents in Consumer Research?

Many current synthetic respondent approaches use large language models to generate answers intended to resemble those of human participants. Depending on the method, the models generating them may be given demographic characteristics, prior survey data, behavioral information, or other context. They simulate responses; they do not observe what an individual consumer actually does.

The limitation appears when similarity in responses is treated as proof of likely behavior. A synthetic respondent may produce a credible answer to a question about purchase intent, brand consideration, or product preference. An actual purchase can be influenced by conditions a synthetic response may not fully represent: what a consumer searches for, which alternatives they encounter, changing prices and promotions, product availability, retailer choice, and the sequence of interactions before the decision.

Where can Synthetic Respondents add Value to Consumer Research?

One of the clearest current cases for synthetic respondents is in upstream consumer research tasks such as pretesting, piloting, and early-stage exploration. A 2024 review in Psychology & Marketing examined research comparing synthetic and human samples and found that results varied considerably across domains (Sarstedt et al., 2024). The authors identified qualitative pretesting and pilot studies as particularly promising uses, where synthetic samples can support early research design choices.

In practice, synthetic respondents can help explore potential hypotheses, stress-test questionnaires, identify issues for human investigation, or narrow a broader set of concepts before live research. In these cases, the output shapes the next research step rather than serving as the final answer.

That distinction matters. Imperfections may be manageable when synthetic outputs help decide what to investigate next. When they are used to estimate demand, compare investments, or predict purchase behavior, errors carry more direct commercial consequences. The level of validation should therefore rise with the consequence of the decision.

How Reliable are Synthetic Respondents in Consumer Research?

There is no single accuracy figure that describes synthetic respondents across all consumer research. Results depend on the model, prompting method, population being represented, question type, category, and level at which accuracy is assessed.

A 2026 study from the Nuremberg Institute for Market Decisions illustrates the problem. Researchers compared human and synthetic responses in a soft-drink brand study using a quota-matched U.S. sample of 461 regular soft-drink buyers as the human benchmark. The synthetic respondents captured broad patterns, including stronger selection of familiar brands, but they also overstated positive attitudes and selection of well-known brands and produced less variation than the human participants.”

That distinction between the average pattern and the detail is important. A 2026 working paper on local validity in synthetic preference measurement (Leng, 2026) argues that strong study-level agreement can conceal failures in specific survey items, brands, attributes, or audience segments. Because this is a working paper, its conclusions should be treated as emerging evidence rather than settled consensus. The relevant question is whether the method works for the specific question, brand, attribute, or audience segment that matters, not merely on average.

For consumer research, aggregate similarity should not be treated as sufficient on its own. A method can reproduce a broad market pattern and still misrepresent the segment, attribute, or choice that matters most to a business.

Can Synthetic Respondents Predict What Consumers Will Buy?

Synthetic respondents can reproduce some purchase preferences, but the current evidence cited here does not support treating them as consistently reliable predictors of what consumers will actually buy. A 2026 working paper (Zhong, 2026) tested four frontier LLMs across 86,400 predictions against 110,326 transactions on a live-event ticketing platform. The models recovered some preferences but showed large segment-level misses and cautioned against replacing observed customers when decisions depend on realized behavior. As a working paper, it is emerging rather than settled evidence.

Purchase intent is stated evidence; a purchase is an observed action. Search, comparison, competitor exposure, price, promotions, availability, and changing priorities can all intervene between the two.

That makes purchase prediction a harder test for any research method. A synthetic model may reproduce how a respondent describes a preference without accurately predicting whether that preference leads someone to buy, delay a purchase, switch brands or abandon the purchase altogether. Research that smooths over this variation can look convincing in aggregate while missing commercially important differences.

The appropriate question is therefore not simply, "Did the synthetic sample look like the human sample?" Researchers also need to ask whether it reproduced the specific behaviors, segments, and conditions relevant to the decision, and whether another source of evidence can test that conclusion.

How Should Researchers Validate Synthetic Respondents Against Human Data?

Consumer research should validate synthetic respondents against independent human, experimental, transactional, or observed behavioral evidence appropriate to the decision being made. Validation should be proportionate to both uncertainty and consequence: the more a decision depends on a synthetic finding, the stronger the case for testing it rather than relying on apparent plausibility.

• Define the decision first. Separate exploratory questions from decisions that directly affect pricing, product development, media, audience strategy, or revenue expectations. The same level of validation is not necessary for every use case.
• Benchmark the method against humans where it matters. Assess performance on the relevant questions, brands, attributes, or segments rather than relying only on a study-level similarity score.
• Use more than one type of evidence. Surveys, qualitative research, experiments, transaction data, and observed behavioral data answer different questions. Agreement can strengthen an interpretation; disagreement can reveal where the research needs more investigation.
• Check variation, not only averages. A convincing mean result can conceal overly uniform synthetic responses or poor performance for smaller but commercially important groups.
• Document the boundaries. Make clear what inputs informed the synthetic respondents, what population they are intended to represent, what has been validated, and what remains an assumption.

This treats synthetic respondents as one research input rather than assuming the same level of reliability in every context. It helps separate useful exploratory signals from evidence strong enough to support a commercial recommendation.

How can Observed Behavioral Data Help Validate Synthetic Respondents?

Observed behavioral data can help research teams test whether synthetic assumptions correspond with activity from real consumers. Measure Data supplies the underlying permissioned behavioral files for that analysis, which clients can compare with surveys, models, experiments, or synthetic research in their own workflows. This adds observed activity to an evidence base that might otherwise rely on stated or modeled responses.

Our multi-platform search research illustrates the point. In the analysis, periods of active ChatGPT engagement coincided with a median 189 Google searches per month, versus 152 during periods of lower ChatGPT engagement. This does not validate synthetic respondents; it shows that real information-seeking behavior can span multiple environments that a single stated response may not capture.

Measure Data is the underlying data source, not the interpretation layer. We deliver permissioned behavioral files for clients to analyze according to their own research questions; Measure Data itself does not identify patterns, generate charts, or produce insights. Measure Predict sits on top of Measure Data as the intelligence layer, allowing users to query that behavioral evidence, identify cross-platform patterns, create relevant charts and surface insights.

Clients can analyze Measure Data alongside synthetic research to test whether patterns suggested by the model also appear in observed activity where coverage is relevant. That analysis may reveal different paths, competitor interactions or audience differences that the modeled response does not represent.

Observed data has limits too: it does not explain why someone acted, and no dataset captures every influence on a purchase. A stronger design can combine stated evidence for attitudes, synthetic methods for exploration, and observed behavioral data for measurable activity, with interpretation remaining the responsibility of the client or the analytical tool they choose to use.

When Should Consumer Research Use Synthetic Respondents?

Synthetic respondents can support selected exploratory and design tasks, but they should not substitute for human participants or observed purchase behavior in higher-consequence decisions. Those decisions warrant independent validation.

If you need underlying behavioral data to test research assumptions against observed activity, talk to our team about Measure Data.

References

Sarstedt, M., Adler, S.J., Rau, L. and Schmitt, B. (2024) ‘Using large language models to generate silicon samples in consumer and marketing research: Challenges, opportunities, and guidelines’, Psychology & Marketing, 41(6), pp. 1254–1270.

Leng, Y. (2026) ‘Where synthetic respondents fail: Diagnosing local validity in LLM-augmented preference measurement’. SSRN working paper.

Zhong, Z. (2026) ‘Where synthetic consumers fail: Evidence from four frontier LLMs and a large-scale ticketing platform’. SSRN working paper.