What are synthetic respondents in market research? Definition, how they work, and limitations
Synthetic respondents are AI-generated participants that simulate human responses in market research. Learn what they are, how they work, where they're useful, their limitations, and how they compare to real respondents.
Synthetic respondents are AI-generated virtual participants that simulate how real humans would respond in surveys, interviews, and market research studies. They are created by training large language models on real-world data (demographics, behaviors, past survey responses, public opinion data) to produce personas that can be queried for feedback at scale and speed. This guide explains what synthetic respondents are, how they work, where they are useful, where they fail, and how they compare to real human research participants.
Frequently asked questions
What are synthetic respondents?
Synthetic respondents are AI-generated profiles that mimic the responses real humans would give in market research. They are not real people. Instead, they are constructed by large language models (LLMs) trained on public datasets, past survey responses, demographic data, and behavioral signals. When queried, they produce responses that statistically resemble what real humans with similar attributes might say. Synthetic respondents are used in market research, concept testing, persona development, and survey simulation, primarily as a way to scale or accelerate research that would otherwise require recruiting real participants.
How do synthetic respondents work?
Synthetic respondents work in three steps. First, an AI model (typically a large language model) is trained or prompted with data about a target audience: demographics, behaviors, attitudes, past survey responses, and contextual information. Second, the model generates a persona profile representing a specific audience segment (“a 35-year-old marketing manager at a mid-size SaaS company who has used HubSpot for 3 years”). Third, the persona is queried with research questions, and the model generates a response in character. Multiple personas can be queried in parallel to simulate a survey sample of hundreds or thousands of “respondents.”
Are synthetic respondents accurate?
Synthetic respondents are accurate for some research questions and inaccurate for others. For quantitative trends and behavioral patterns that are well-represented in the training data, synthetic respondents typically match real respondents at 85 to 95% accuracy. For qualitative depth, novel concepts, emotional nuance, and lived experiences, accuracy drops to 60 to 80%. They are most accurate when the research question involves predictable patterns and least accurate when the research requires understanding new contexts, unfamiliar products, or unique personal stories.
Can synthetic respondents replace real participants?
No. Synthetic respondents augment but do not replace real participants. They are useful for early-stage hypothesis generation, hard-to-reach audiences, large-scale quantitative simulation, and research at scale that would be cost-prohibitive with real participants. They fail at understanding novel concepts, capturing emotional nuance, surfacing unexpected insights, and validating high-stakes decisions. The dominant view in the research community is that synthetic respondents are a complement to real research, not a replacement.
What are synthetic respondents used for?
Synthetic respondents are used for five primary purposes: hypothesis generation before real research, concept testing at scale, audience segmentation and persona development, survey pre-testing (running a draft survey through synthetic respondents to identify problems before sending to humans), and modeling hard-to-reach audiences (executives, regulated populations, niche specialists). They are most valuable when speed and scale matter more than depth and validity.
What are the limitations of synthetic respondents?
Synthetic respondents have five major limitations. First, they are backward-looking: they reflect patterns in their training data and struggle with anything new. Second, they lack lived experience and emotional depth. Third, they tend toward sycophancy (overly agreeable responses that confirm what the researcher seems to want). Fourth, they inherit biases from their training data. Fifth, they cannot simulate fatigue, distraction, or attention drop-off, which are real factors in human survey responses. For high-stakes research, these limitations are disqualifying.
How synthetic respondents work
Synthetic respondents are generated through a four-stage process. Understanding each stage helps you evaluate when synthetic respondents are appropriate and when they are not.
Stage 1: Training data curation
The AI model that generates synthetic respondents needs training data that represents the target audience. This typically includes:
- Demographic data (age, gender, location, income, occupation, education)
- Behavioral data (purchase history, product usage patterns, app interactions)
- Attitudinal data (past survey responses, public opinion data, social media sentiment)
- Contextual data (industry information, brand familiarity, life stage)
The quality and representativeness of training data directly determines synthetic respondent quality. A model trained primarily on US tech consumers will produce poor synthetic respondents for European blue-collar workers.
Stage 2: Persona generation
The model uses the training data to construct persona profiles representing target segments. A persona is a structured description that the LLM uses as context when generating responses. For example:
“You are Maria, 42, a marketing director at a 200-person B2B SaaS company in Austin, Texas. You have 15 years of marketing experience, manage a team of 8, and have used Salesforce, HubSpot, and Marketo in previous roles. You are evaluating a new marketing automation tool for your team. You are budget-conscious but value reliability over price.”
The persona becomes the system prompt that shapes how the model responds to queries.
Stage 3: Query and response generation
Researchers send research questions to the model, which generates responses in character. Questions can be open-ended (“What concerns would you have about adopting this product?”) or structured (Likert scales, multiple choice). The model produces responses based on what its training suggests a person matching the persona would say.
Stage 4: Aggregation and analysis
Responses from multiple synthetic respondents are aggregated and analyzed using the same methods you would apply to real survey data: calculating averages, identifying patterns, segmenting by attributes, comparing groups. The output looks identical to a real survey result, even though no humans were involved.
Types of synthetic respondents
Not all synthetic respondents are created the same way. The four main types differ in how they are constructed and what they are best suited for.
| Type | How it works | Best for | Limitations |
|---|---|---|---|
| LLM-only personas | Generated entirely from a large language model’s training data | Hypothesis generation, exploratory research, persona drafts | Limited to model’s training data; high hallucination risk |
| Data-grounded personas | LLM responses constrained by real-world datasets (demographics, behavior data) | Survey pre-testing, audience modeling | Better accuracy than LLM-only but still limited by data scope |
| Twin-based synthetic respondents | ”Digital twins” of real individuals built from their actual past responses | Longitudinal modeling, individual-level prediction | Requires extensive prior data; privacy concerns |
| Hybrid synthetic-real | Synthetic respondents augment real survey data (e.g., filling in underrepresented segments) | Boost sample size for hard-to-reach segments | Requires careful weighting; risk of distorting findings |
Where synthetic respondents work well
Synthetic respondents add value in specific research scenarios where their limitations are manageable.
1. Early-stage hypothesis generation
When you are exploring a new problem space and need to generate hypotheses fast, synthetic respondents can help you frame the research question, identify potential audience segments, and surface initial themes. The output is a starting point, not a conclusion.
2. Survey pre-testing and pilot
Before sending a survey to 5,000 real respondents, run it through 100 synthetic respondents first. This catches confusing questions, response option gaps, and survey logic errors at near-zero cost.
3. Concept testing at scale
For early-stage concept testing where you want directional feedback on dozens of variations, synthetic respondents allow rapid iteration. Real participant testing should still validate the final candidates.
4. Modeling hard-to-reach audiences
C-suite executives, regulated populations, niche specialists, and other hard-to-reach audiences are expensive and slow to recruit. Synthetic respondents can model these audiences for early-stage exploration, with the understanding that real validation is required before high-stakes decisions.
5. Large-scale quantitative simulation
For research questions that need 10,000+ responses to be meaningful (statistical power, segmentation analysis, geographic comparisons), synthetic respondents enable scale that would be cost-prohibitive with real respondents.
Where synthetic respondents fail
The cases where synthetic respondents fail are equally important to understand.
1. Novel concepts and emerging behaviors
Synthetic respondents are trained on historical data. They cannot give you reliable feedback on something that did not exist when their training data was collected. New product categories, disruptive features, and emerging behaviors are blind spots.
2. Emotional depth and lived experience
A real participant can tell you about the time their kid had a meltdown in the grocery store and how it changed how they shop. A synthetic respondent will give you a generic answer that pattern-matches to grocery shopping research it has seen. The story is what makes qualitative research valuable, and synthetic respondents cannot produce real stories.
3. Sycophancy and confirmation bias
Large language models tend to produce responses that align with what the prompt seems to want. Ask a synthetic respondent “Would you be excited about this new feature?” and you will likely get an enthusiastic answer regardless of whether real users would care. This is a systematic bias that is difficult to correct.
4. Attention, fatigue, and distraction
Real survey respondents skip questions, satisfice on long surveys, abandon mid-survey, and give inconsistent answers when tired. These behaviors carry information about the survey itself (it is too long, too complex, too boring). Synthetic respondents do not exhibit these patterns, so survey design lessons learned from synthetic testing may not transfer to real respondents.
5. High-stakes decisions
For decisions with significant financial, regulatory, or reputational consequences, the limitations of synthetic respondents become disqualifying. Critics have called synthetic respondents “homeopathy for market research” for their unreliability in high-stakes contexts. Real human validation is required for go/no-go decisions, pricing decisions, regulatory submissions, and brand strategy.
Synthetic respondents vs real respondents: side-by-side
| Dimension | Synthetic respondents | Real respondents |
|---|---|---|
| Quantitative accuracy on familiar topics | 85-95% match for behavioral patterns | Gold standard |
| Qualitative depth | 60-80% surface insights only | Superior nuance, stories, emotion |
| Speed | Minutes to hours | Days to weeks |
| Cost | Often 90%+ cheaper | Higher direct cost (incentives, recruitment) |
| Scale | Thousands of “respondents” easily | Limited by recruitment budget |
| Bias profile | Reduced social desirability; introduces model bias and sycophancy | Higher social desirability; introduces fatigue and lying |
| Novel concepts | Poor; cannot reason about unfamiliar contexts | Strong; can react authentically |
| Emotional content | Synthesized, not lived | Real and contextualized |
| Edge cases and outliers | Smoothed out by model averaging | Captured naturally |
| Reproducibility | High (same prompt produces similar outputs) | Lower (real humans vary across sessions) |
| Validity for high-stakes decisions | Low; not suitable alone | High |
| Time savings vs traditional research | 70-90% time reduction | Baseline |
When to use synthetic respondents (and when not to)
Use synthetic respondents when:
- You need fast directional input for early-stage research
- You are testing a survey before sending it to real participants
- You need to model hard-to-reach audiences for hypothesis generation
- The decision is reversible and low-stakes
- You can validate findings with real research before acting
Do not use synthetic respondents when:
- The research will inform a high-stakes decision (pricing, launch, regulatory submission)
- You need emotional depth and lived experience
- The topic involves novel concepts not represented in training data
- The audience is poorly represented in public training data
- The research is intended for academic publication or external sharing without validation
How to use synthetic respondents responsibly
If you decide to use synthetic respondents, follow these practices to manage their limitations.
1. Always disclose synthetic data
Never present synthetic respondent output as real participant data. Internal stakeholders and external audiences should know what they are looking at. Misrepresenting synthetic data as real is an integrity violation.
2. Validate with real research
Use synthetic respondents for hypothesis generation and pre-testing, then validate findings with real participants before acting. The combination is more powerful than either alone.
3. Know your model’s training data
Different synthetic respondent platforms use different training data. Ask your vendor what their model was trained on and how representative it is of your target audience. A model trained on US-only data will fail for global research.
4. Test for sycophancy
Run the same question through synthetic respondents with neutral framing, positive framing, and negative framing. If the model produces dramatically different answers, it is exhibiting sycophancy and the data is unreliable.
5. Triangulate with multiple methods
Synthetic respondents should be one input among many: real research, behavioral data, market data, expert input. Triangulation catches errors that no single method catches alone.
6. Document model versions and prompts
Synthetic respondent outputs depend heavily on the model version and the exact prompts used. Document both so your research is reproducible and so you can understand drift when models update.
The synthetic respondents debate
The market research community is divided on synthetic respondents. Three camps have emerged.
The optimists argue that synthetic respondents democratize research, reduce costs, accelerate iteration, and will continue to improve as models advance. They point to studies showing 85-95% accuracy on quantitative trends and argue that the convenience justifies the limitations for many use cases.
The skeptics argue that synthetic respondents are fundamentally unreliable for the qualitative and emotional dimensions that make research valuable. They cite the sycophancy problem, the backward-looking limitation, and the risk of teams substituting synthetic data for real understanding. Some critics have called synthetic respondents “homeopathy for research” to emphasize the gap between marketing claims and actual reliability.
The pragmatists argue that synthetic respondents are a tool with specific use cases: fast and cheap for hypothesis generation and survey pre-testing, unreliable and dangerous for high-stakes decisions. The right approach is to know which scenarios fit and which do not, and to validate everything important with real research.
The pragmatist view is winning in mature research organizations. Synthetic respondents are entering the toolkit alongside real research, not replacing it.
The future of synthetic respondents
Synthetic respondents will continue to improve as LLMs improve and as more behavioral and survey data becomes available for training. Expect three trends in 2026 and beyond:
1. Hybrid synthetic-real research becomes standard. Most research programs will use synthetic respondents for early-stage work and real participants for validation, rather than choosing one or the other.
2. Vendor differentiation around training data. Synthetic respondent platforms will compete on the quality and breadth of their training data. Platforms with proprietary survey databases or behavioral data will have advantages.
3. Increasing skepticism in regulated industries. Healthcare, financial services, and regulated industries will continue to require real participant data for compliance reasons. Synthetic respondents will be excluded from regulated decision contexts.
For teams evaluating synthetic respondents, the AI in user research guide covers the broader landscape of AI tools in research, and the AI research vs human moderated research comparison provides a deeper analysis of where AI tools work and where human judgment remains essential. The fundamental principle is unchanged: synthetic respondents are a tool, not a replacement for understanding the people you are building for.