Can synthetic data replace real people in market research?

Synthetic data promises faster insights, lower costs and a new way to understand audiences. But can it really tell us what people think, feel and do?

Ask a room full of researchers whether synthetic data can replace real respondents, and you're unlikely to get a unanimous answer. As with any discussion on emerging technologies, the debate around synthetic data in market research is shaped by who’s doing the talking. Everyone has an agenda, and that inevitably influences how they view its potential.

We’re no different. The Research Agency (TRA) is a full-service market research agency built around the idea of ‘The art of knowing people’. We believe in the power of speaking to real people, quantitatively and qualitatively, and understanding the full spectrum of perspectives, behaviours and experiences.

But we're also big believers in what technology can make possible. We've helped build world-class ResTech businesses, invested heavily in new techniques and embraced multiple data sources to solve complex research problems.

These aren't competing beliefs. Together, they inform our position: synthetic data has a useful role in market research, but we have not yet seen sufficient evidence that it can replace real respondents. We see the clearest current value in persona-based exploration, helping researchers connect evidence across sources, interrogate patterns and develop richer hypotheses.

Beyond that, the evidence for replacing real respondents is less convincing. Synthetic data may reproduce an average accurately, but averages can conceal the differences, contradictions and emerging behaviours that make research valuable.

Our position isn't that synthetic data is inherently good or bad. It's where it adds value, what evidence supports its use and how well we understand its limitations.

Where synthetic data works

The strongest current use case is personas. Synthetic personas can provide a useful interface for exploring existing knowledge: bringing together disparate research studies; enabling repeated questions; and helping teams examine a topic from different angles.  

Imagine being able to explore years of existing audience research through a single interface, asking new questions, testing different perspectives and identifying connections that might otherwise take hours to find.

Crucially, the persona is a thinking tool, not a respondent. Grounded in existing evidence, it can make patterns easier to identify, challenge assumptions and support faster hypothesis generation. Its value lies in helping researchers make better use of what they already know, rather than producing new quantitative data.

Where it falls short

Beyond personas, synthetic data is being applied in several ways: imputation to fill gaps in existing datasets, audience boosting to generate additional modelled respondents, predictive models that estimate responses to new questions or stimuli, and the generation of entirely synthetic respondents.  

These approaches differ in purpose and sophistication, and some may prove valuable in specific circumstances. But they share an important consideration: their outputs are modelled from existing information rather than independently observed from new respondents.

Synthetic systems can reproduce established relationships and central tendencies, but their ability to capture the full distribution of human attitudes and behaviours requires careful validation.

That's where things get more complicated.

Outliers, weak signals, changing attitudes and smaller but commercially significant groups may be compressed, exaggerated or missed altogether. Getting the average right is not the same as accurately representing the range of opinions or the relationships between attitudes and behaviours.

An average might tell us what most people think. But what about the people who think differently, the behaviours we haven't encountered before or the small shifts that signal something bigger is coming? These are often the findings that lead to the most interesting research questions. ‍

People don’t follow the model

People are contradictory. What we say, believe and do doesn’t always align. We make trade-offs, respond differently depending on context and behave in ways that aren’t consistently rational.  

We might say price is our biggest consideration, then happily pay more for a brand we love. We might express concern about sustainability, then choose convenience when we're short on time. And sometimes, we struggle to explain our own decisions.

As we explore in our article What behavioural data doesn't tell you about people's behaviour, even identical behaviours can be driven by very different motivations.

Traditional research doesn't perfectly predict behaviour either. But speaking to people allows us to encounter those contradictions, investigate them and uncover something unexpected.

Synthetic systems risk smoothing these inconsistencies into coherent, statistically likely answers, potentially removing the very differences we're trying to understand.

What’s the research problem we’re solving?

There's a temptation with any exciting new technology to start looking for ways to use it. But in research, the better question is whether it helps us solve a problem.

Consider five common challenges in quantitative market research and where synthetic data can, or cannot, help.

1. More sample: If the goal is greater statistical confidence and a wider range of opinions, replicating existing respondents doesn't provide the same evidence as collecting independent responses. Increasing the number of real respondents is generally the more defensible solution.

2. Access to niche audiences: Synthetic data may help explore existing knowledge about hard-to-reach groups, but where that knowledge is limited or unrepresentative, modelled responses carry considerable uncertainty. If reliable information about an audience is scarce, there is little basis for assuming a synthetic model can represent it accurately.

3. Speed: Modelled responses can support rapid exploration when speaking to people is impractical. But speed should not be confused with validity. The risk is reproducing established patterns rather than identifying what's changing.

4. Connecting evidence across sources: This is where synthetic personas can be particularly valuable, allowing teams to query, compare and explore a grounded body of knowledge repeatedly.

5. Better response quality: The priority should be working with panel providers to verify participants, design surveys that respect respondents' time and encourage thoughtful answers and considering whether people are being fairly compensated for their participation. These are industry responsibilities that synthetic data cannot resolve.

Ultimately, different research problems call for different approaches. As we explore in Comparing brand health tracking methods, understanding what each method can and cannot tell us is essential to choosing the right one.  

Where TRA stands on synthetic data

TRA supports thoughtful experimentation with synthetic data, but we do not believe the current evidence supports treating synthetic respondents as equivalent to real people in quantitative research.

Today, we see a credible role for well-grounded synthetic personas as interfaces to evidence, tools for interrogation and catalysts for better questions.

Imputation and modelled or fully generated responses may also have applications, but their use should be guided by three considerations: a clearly defined research problem, transparent assumptions and robust evidence that the approach is fit for purpose.

For our teams and clients, this means being clear about what is observed, what is modelled and what remains uncertain. It also means recognising when direct engagement with people is essential to answering the research question.

Technology is already helping us work with research in ways that weren't possible a few years ago. And we're excited to see where synthetic approaches might take us next.

But the standard of evidence matters. Until synthetic approaches can demonstrate that they reliably capture the variability of human attitudes and behaviours, and the relationships between them, we don't see a basis for treating them as a like-for-like replacement for research with real people.

After all, the aim of market research isn't simply to produce more answers. It's to understand people well enough to ask better questions, challenge what we think we know and discover something new.

Explore more of TRA's approach to combining technology with human insight.  

‍