Tiger Tracks · Eye of the Tiger · Measurement & Privacy · October 2026

Qualtrics surveyed more than 3,000 market researchers in 14 countries in 2024 and found that 71% believed synthetic responses would make up the majority of market research data collection within three years, while 87% of those who had used synthetic responses reported high satisfaction with the results [1]. A year later, 45% of researchers who had adopted synthetic data described it as their most reliable data source [3]. Adoption is running ahead of proof.

Synthetic audiences are artificially generated profiles that simulate how real customers think, answer and behave. They promise speed, scale and privacy protection, but they raise hard questions about accuracy, bias and how much weight a simulated answer should carry in a media or product decision. This article sets out what synthetic audiences can and cannot do, and how to keep real customers at the center of measurement.

1. Privacy Shifts Changed the Data Problem, Not the Goal

For two decades, digital marketing relied on deterministic data such as cookies, device IDs and direct customer records. Privacy regulation, starting with GDPR in Europe and California's CCPA in the United States, narrowed how that data could be collected and used. Twenty US states had comprehensive consumer privacy laws on the books by early 2026, with Indiana, Kentucky and Rhode Island taking effect on January 1, 2026 [4]. In June 2026, Vermont became the 23rd state to enact one, enforceable from January 1, 2028 [5].

The browser picture has moved in an unexpected direction. Google kept third-party cookies in Chrome and, in October 2025, retired most of its Privacy Sandbox technologies, including Topics, Protected Audience and the Attribution Reporting API, citing low adoption [6]. The signal loss marketers feared from Chrome did not arrive as planned, but consent rules, state laws and fragmented first-party data still leave real gaps. Synthetic audiences are one response to those gaps, not a replacement for the customer data marketers are still allowed to collect.

2. Synthetic Audiences and Real Customers Do Different Jobs

Machine learning models, most often large language models conditioned on demographic and behavioral profiles, generate synthetic respondents that do not correspond to actual individuals but statistically represent segments of a customer base. Real customer data, by contrast, records what people actually did and said.

Real Customers: The Ground Truth

First-party data from CRM systems, website interactions, purchase history and direct feedback remains the standard for precision marketing. It supports granular targeting, personalization and measurement of campaign return. Its constraints are cost, time, consent and regulation.

Synthetic Audiences: Statistical Proxies

Synthetic profiles fill gaps where real data is sparse, slow or restricted. They are probabilistic rather than deterministic, so their outputs describe likely patterns, not observed behavior. That distinction should shape every decision built on them.

3. Adoption Has Outpaced Validation

Research teams have embraced synthetic data quickly, and teams that adopt it report more demand for their work. Qualtrics' 2026 Market Research Trends report found that 72% of teams using synthetic responses, agentic AI or purpose-built tools said their organizations were using research significantly more than before, and traditional teams were nearly twice as likely to see flat or declining demand (37% versus 20%) [3]. Gartner forecast as early as 2022 that synthetic data would completely overshadow real data in AI models by 2030 [7].

Satisfaction is not the same as accuracy. A researcher can be satisfied with a fast answer that is wrong for the segment that matters most. That is why independent validation matters.

4. The Evidence Shows Promise and Clear Limits

Academic work established the core idea. Researchers at Brigham Young University showed in 2023 that GPT-3, conditioned on socio-demographic backstories from real survey respondents, could produce "silicon samples" whose response patterns tracked those of human subgroups, a property they called algorithmic fidelity [8].

Commercial tests have pushed further. In 2025, EY and synthetic research firm Aaru recreated a six-month wealth survey of 3,600 affluent investors in a single day; across 53 matched questions, Aaru reported a median Spearman correlation of 0.90 with the original results and an average gap of 7.1 percentage points per question [9].

Independent testing is less encouraging. In a September 2025 study, Verasight built personas for 1,500 nationally representative respondents and asked several LLM configurations to answer as those people. Even the best-informed configuration missed immigration attitudes by 11.3 percentage points, adding more information or newer models did not reliably improve accuracy, and the authors concluded that LLM-based imputation "fails to deliver reliable subgroup estimates" [2].

5. Measurement Needs a Validation Loop

Synthetic audiences inherit the biases of their training data and modeling choices. If the source data underrepresents a demographic, the synthetic output will too, and the Verasight results show that subgroup accuracy is where synthetic samples are weakest [2]. Marketers should validate synthetic outputs against controlled real customer segments before acting on them, and repeat that check as models and markets change.

Attribution also changes. Synthetic profiles cannot be tracked through a deterministic journey, so their value shows up in pre-launch decisions: which concept to test, which message to lead with, which segment to prioritize. Campaign impact still has to be measured on real customers through incrementality tests, cohort analysis and media mix modeling.

6. Synthetic Data Should Inform Decisions, Not Make Them

Use Synthetic Audiences as Supplements, Not Replacements

Synthetic audiences work best early in a decision, when speed matters more than precision. Real customer data should make the final call on budget, targeting and creative.

Validate Before You Scale

Run a real-sample benchmark on a subset of questions each time a synthetic model is used for a new category or audience. Pay particular attention to subgroups, where independent testing shows the largest errors [2].

Invest in Data Governance and Model Transparency

Document what data trained the model, how personas were constructed and how outputs were checked. Governance should address bias and fairness alongside accuracy.

Build Cross-Functional Collaboration

Marketing, data science, legal and privacy teams need shared rules on where synthetic data can and cannot inform decisions.

Monitor Regulatory Change

State privacy laws continue to be enacted, take effect and be amended through 2026 [4][5]. Synthetic approaches reduce exposure to personal data, but the real data used to build them remains subject to those rules.

7. The Effects Cascade Across the Marketing Ecosystem

The first-order effect is faster, cheaper research: questions that once took months of fieldwork can be simulated in a day [9]. The second-order effect is a shift in where errors hide. When synthetic answers feed concept selection and audience planning, a model's blind spots on specific subgroups can quietly shape which ideas reach market [2]. The third-order effect is a premium on real customer evidence. As synthetic data becomes common, the brands with clean first-party data and disciplined incrementality testing will hold the benchmark everyone else needs to check their simulations against.

Conclusion

Synthetic audiences are a genuine advance in speed and scale, and research teams are already using them at volume. They are also simulations, and the evidence shows they are least reliable exactly where marketers need precision: specific segments and close calls. The brands that benefit will use synthetic data to ask better questions faster, then let real customers answer them. That balance, with human judgment deciding what to trust and AI accelerating the work around it, is the Human-Led, AI-Augmented advantage.

References

1. Qualtrics. (October 22, 2024). AI to Drive Massive Changes to Market Research in 2025, Qualtrics Report Says. https://qualtrics.com/news/ai-to-drive-massive-changes-to-market-research-in-2025-qualtrics-report-says
2. Verasight, Morris, G. E., Leff, B., and Enns, P. K. (September 26, 2025). Synthetic Sampling Report II. https://www.verasight.io/reports/synthetic-sampling-2
3. Research Live. (November 19, 2025). Most researchers using AI regularly, finds Qualtrics study. https://www.research-live.com/article/news/most-researchers-using-ai-regularly-finds-qualtrics-study/id/5144776
4. MultiState, Rieper, M. (February 4, 2026). All of the comprehensive privacy laws that take effect in 2026. https://www.multistate.us/insider/2026/2/4/all-of-the-comprehensive-privacy-laws-that-take-effect-in-2026
5. National Law Review, Fife, W. (June 17, 2026). Vermont Makes 23: States with Comprehensive Data Privacy Laws. https://natlawreview.com/article/vermont-makes-23-states-comprehensive-data-privacy-laws
6. Google, Chavez, A. (October 17, 2025). Update on plans for Privacy Sandbox technologies. https://privacysandbox.google.com/blog/update-on-plans-for-privacy-sandbox-technologies
7. Gartner. (June 22, 2022). Is Synthetic Data the Future of AI? https://www.gartner.com/en/newsroom/press-releases/2022-06-22-is-synthetic-data-the-future-of-ai
8. Argyle, L. P., Busby, E. C., Fulda, N., Gubler, J. R., Rytting, C., and Wingate, D. (February 21, 2023). Out of One, Many: Using Language Models to Simulate Human Samples. Political Analysis, 31(3). https://doi.org/10.1017/pan.2023.2
9. Aaru. (October 7, 2025). EY Wealth Research case study. https://aaru.com/case-studies/ey-wealth-research

Published by Tiger Tracks. Eye of the Tiger Intelligence Series.