A Practical Guide to Synthetic Data in Market Research: Part 1

AI claims are loud and timelines are shrinking. Among the many solutions we’re seeing pushed is synthetic data, often positioned as a replacement for human research. The argument appears simple: if AI can model respondents, why invest the time and cost in recruitment and fieldwork? Over the past few months, we set out to test that assumption.

Welcome to Part 1 of A Practical Guide to Synthetic Data in Market Research where we’ll discuss when synthetic data works and how to build the right foundation. The goal of this guide is simple: to help researchers approach synthetic data the way we’ve learned to approach it ourselves – not as a last-minute replacement for human research, but as a new tool in the research toolkit. And like any tool, there are situations where it works well and others where it doesn’t.

Through a series of pilots across multiple vendors, we explored what synthetic data can realistically deliver in a research context. We tested different modeling approaches, audiences, and questionnaire structures. We benchmarked outputs against real human datasets. And we stress-tested the use cases that matter to research, product, marketing, and innovation teams. Our conclusion is straightforward: Human-Guided Synthetic Research can create some advantages, but only if you build the right foundations to use it.

Synthetic data isn’t magic, it’s modeling

Before diving into how to build the right foundation, it’s important to recognize that synthetic data isn’t just a single technique. It’s a collection of modeling approaches that generate simulated observations based on patterns in real datasets.

It also isn’t new; in fact, many of the techniques behind synthetic data predate the current wave of AI tools. Early versions were developed in public health and census research to analyze sensitive datasets while protecting the privacy of individual respondents.

For market research, we see three main approaches. We’ve listed them below but a word of caution: there is no standard taxonomy (yet) for synthetic data as it’s constantly evolving. For clarity, we’ve defined how we see it:

• Sample Augmentation – Fills in the gaps in your data. AI generates additional responses for underrepresented groups, using the same questions and structure as your original survey
• Digital Twins – Creates an AI version of each respondent. These “twins” reflect how individuals think, feel and behave based on their actual responses. They’re used to explore the why behind the data, simulating rich, qualitative feedback at scale.
• Persona Bots – Turns your segments into conversations. Built from real data patterns, these AI personas let you “talk” to a specific audience and explore how they might think, decide and act. It’s a fast, flexible way to bring your segments to life and uncover deeper perspective.

Because these approaches operate differently, their strengths and limitations vary as well. However, across the pilots we ran, several consistent patterns emerged about where synthetic data can, and cannot, support research.

The foundation of good synthetic data: Finding the right use case

One of the most common questions we hear from researchers, innovation teams and marketing leaders is: Where does synthetic data truly create value?

Across our pilots, we found that synthetic data’s effectiveness and impact depend on choosing the approach based on the research objective rather than treating it as a universal solution. Finding that fit is the first step to building good synthetic data.

There are several use cases where you might want to use synthetic data; here are some examples:

• Privacy-sensitive research environments: Analyze patterns without exposing identifiable participant information, particularly in regulated sectors such as healthcare, finance or employee research
• Rapid concept testing and early exploration: Quickly pressure-test ideas, messaging or product concepts before committing to full fieldwork
• Simulation and modeling scenarios: Support forecasting or behavioral simulations when real-world datasets are incomplete or restricted
• Hard-to-reach or niche audiences: Extend insights for small or specialized populations where recruiting large samples is costly or impractical
• Iterative testing without respondent fatigue: Conduct early questionnaire testing or concept refinement without repeatedly returning to real participants
• Balancing representation in limited samples: Address gaps in small or uneven datasets where key segments are underrepresented

Many of these use cases can offer cost and time efficiencies as they offer larger datasets without the operational overhead of recruiting equivalent human samples. However, that is often a side benefit, rather than being the sole focus. This is because there remains a need for a strong foundation which is time consuming and often effort intensive.

If your goal is to maximize cost and time efficiencies, you should target a study where you are completing the same thing repeatedly, for example, a programmatic brand tracker. A caveat to this is when the targets are an extremely expensive, hard-to-reach population.

That is in part why, of all these use cases, we explored solutions around niche or hard-to-reach audiences. It is also one of the most consistent requests we’re hearing from research end-users.

That said, no matter how you apply synthetic data, it is not the same as recruiting additional respondents. In the example of niche audiences, increasing a dataset from 100 to 300 with 200 additional synthetic observations, does not reduce sampling error in the same way as collecting responses from 200 more people. The goal is not simply a larger “n” but greater confidence in the patterns that matter.

To build your foundation right, it’s an equal balance of data, design and people

Let’s assume you’ve picked a promising use case. The next step isn’t jumping straight into synthetic modeling; it’s making sure you’ve set up the right foundation to make it work.

From our pilots, three things consistently determined success: the data you start with, how you design your study and how you align stakeholders and their expectations.

Foundation pillar 1: Start with robust human data

Synthetic data learns from patterns in real data. If those patterns are clear, structured and reasonably current, the outputs can mirror human distributions fairly closely. If the dataset is shallow, the models will start to fill in the blanks, leading to inaccuracies.

However, this isn’t about having more data, it’s about having the right data.

To get good synthetic data, you need both a strong anchor dataset and data that is relevant to the problem you’re trying to solve. Is it for the right audience? Is it recent? You might have large amounts of older data, but in a changing world that doesn’t necessarily translate into something you want to base insights on.

This is also where synthetic data can break down. If the underlying data is biased, incomplete or misaligned, those issues don’t disappear, they are reinforced. Synthetic data doesn’t correct weak inputs, it amplifies them.

So, while strong data can unlock meaningful extensions of your dataset, weak or irrelevant data will limit what synthetic approaches can deliver from the outset. Knowing how to choose data, and what makes data appropriate for synthetic data approaches is key to building out your foundation.

Foundation pillar 2: Designing & structuring research for modeling

The second requirement is carefully structured research design.

Question and survey design repeatedly came up as a challenge in our experience with synthetic data. In our pilots, we found that survey design had a significant impact on synthetic performance. The challenges ranged from inaccuracies to a complete failure to model respondents.

More broadly, we observed that simpler survey structures tend to perform more reliably in synthetic data, while complexity can quickly degrade model performance. This is a trade-off that has no concrete answer and is something that needs to be accounted for when these projects are set up. Do you prioritize accuracy across a few questions or the value of affecting a much larger piece?

Our workaround was to adapt existing surveys to work with synthetic approaches, but this typically required additional time and iteration. It taught us how we could approach similar work again, and what we would likely change if we could redesign the survey from scratch.

These are not reasons to avoid synthetic data but reminders that synthetic approaches are designed and trained in ways that may require researchers to approach survey design differently to get the best results. These learnings only can come through experience and experimentation, another key reason to see this as part of the foundational skills for successful synthetic data.

Foundation pillar 3: Build trust by aligning stakeholder expectations and communicating limitations

Synthetic data still raises questions for many organizations. Even when the methodology is sound, teams may hesitate if they believe the data is ‘made up.’ Transparency about how synthetic data works, and how it interacts with real human data, is essential if insights are going to be accepted and used.

At the same time, it’s important to communicate to stakeholders that synthetic data is better suited to some types of studies, objectives and questions than others. Finding all these nuances can take time and effort.

While exploration will reveal these challenges, it’s best to remember synthetic data is produced by a machine. Studies that rely heavily on context and human judgment, such as purchase intent or ad recall, will struggle as synthetic data cannot fully model such complex concepts.

To make synthetic work, you need to bring stakeholders along and set clear boundaries around what the method can realistically deliver. This means being transparent about how the data is generated, choosing the types of questions you apply it to and positioning it as a supporting approach, not a replacement for human research.

Without that alignment, there is a risk of either mistrust or overconfidence. With it, synthetic data becomes far more effective and usable. Setting up that foundation is key to ensuring that you can have credible conversations to talk through the realities of this ‘new’ tool/technique.

Key takeaways: Building a strong foundation for synthetic data

• Synthetic data is a modeling approach—not a substitute for human behavioral insights.
• Different approaches, including sample augmentation, digital twins and persona bots, solve different research challenges.
• Strong, relevant human data remains the foundation of successful synthetic modeling. Read more here about how synthetic data is only as reliable as the human signals behind it.
• Research design directly influences synthetic data quality and reliability.
• Human-Guided Synthetic Research requires both AI capabilities and human expertise.
• Synthetic data is most effective when positioned as a complement to traditional research methods.

What should market researchers do next?

Synthetic data is here, and it’s not a shortcut. The real value comes to those who take the time to understand it.

The smartest research teams aren’t replacing traditional methods. They’re learning where synthetic data shines, where it doesn’t and how to use it alongside what already works. Like any good research technique, it rewards curiosity, discipline and a willingness to test and learn.

Our advice is simple:

• Start with a focused pilot and a clearly defined use case.
• Benchmark results against real human data.
• Build organizational understanding before scaling.

We’re continuing to explore this space every day. If you’re doing the same, it’s worth comparing notes. In the meantime, we invite you to watch “Synthetic Data Without the Hype,” Escalent’s on demand webinar hosted by Chris Barnes and Dyna Boen where they share practical guidance based on what we’ve learned from training our teams and working with F100 clients.