In the first three posts in this series, we examined the diagnostic picture of customer research today — the three pillars under pressure — and the first two methodological shifts that follow from it: from sample to trace, and from asking to reading. Both of those shifts work with data that customers already produce. This post is about a different move. The third shift is about anticipating customer response without asking real customers at all.

Synthetic audiences — populations of large-language-model-based simulated respondents, calibrated against target customer segments — have moved over the past three years from a curiosity in academic papers to a method that is being used in production by serious organizations. This post takes the method seriously: what it does, what evidence supports it, where its limits are, and how it fits into a customer-research practice that does not abandon the older methods.

What a synthetic audience actually is

The vocabulary in this space is unsettled, and it is worth being precise.

A synthetic audience is not a chatbot pretending to be a customer. It is a structured arrangement in which a large language model is conditioned — through prompting, fine-tuning, or both — to produce responses that approximate what defined groups of real customers would say. The conditioning typically draws on demographic, attitudinal, and behavioral characteristics of the target population, often calibrated against historical data from that population, so that the responses the system produces resemble in distribution what the real population would produce if asked.

The method rests on a specific claim about what large language models contain. Models trained on the entirety of human-written text encode an enormous amount of information about how different kinds of people talk, what they value, what they worry about, and how they respond to different framings. They are not customers, but they have absorbed the linguistic and attitudinal patterns of customers from training data at unprecedented scale. The question is whether that absorbed knowledge can be turned into responses useful for customer research.

The short answer, supported by a growing body of academic research, is: for certain classes of questions, yes — and the classes of questions are larger than initial skepticism suggested.

What the evidence shows

The methodological literature on synthetic audiences is recent but no longer thin. Work in Political Analysis has shown that large language models, when conditioned on demographic and attitudinal characteristics, can produce response distributions that correlate strongly with those of real respondents from comparable populations. Working papers from Harvard Business School and NBER have shown similar results in market-research and economic-decision contexts: synthetic respondents reproduce many of the patterns that real respondents produce, including preference orderings, price sensitivities, and reactions to product concepts.

The correlations are not perfect. They are typically high enough to be useful for directional decision-making and screening, but not high enough to replace real fieldwork for high-stakes commitments. This is an important distinction we will return to.

The evidence also shows that the quality of synthetic responses depends substantially on how the audience is constructed. Naive prompting — "respond as a 35-year-old woman in Munich" — produces results that look plausible but correlate weakly with real respondents from that demographic. Carefully constructed audiences, drawing on richer characterizations and validated against benchmark studies, produce results that are substantially more reliable. The method, in other words, is not magic; it requires the same kind of methodological discipline that classical research methods require.

Where synthetic audiences are useful

Three classes of question are particularly well-suited to synthetic-audience methods.

The first is rapid concept screening. When an organization has more product concepts, positioning ideas, or campaign variants than it can realistically field-test, synthetic audiences provide a way to triage. A round of synthetic testing can identify the variants that are clearly weak, the variants that are clearly strong, and the variants where the signal is ambiguous and real fieldwork is warranted. This is a different role than full validation: it is allocation of fieldwork budget toward the questions that genuinely require it.

The second is questions about hypothetical futures. As noted in our previous post, reading what customers have already said cannot tell you how they will respond to something they have not yet encountered. Surveys can address this, but they are slow and expensive. Synthetic audiences offer a third option: they can produce directional responses to novel concepts at a fraction of the cost and time of real fieldwork, with the understanding that the results require validation before being treated as ground truth.

The third is multivariate exploration. When the question involves the interaction of many design choices — different price points crossed with different feature sets crossed with different messaging — the combinatorial explosion makes exhaustive field testing impossible. Synthetic audiences can sweep the design space, identify the regions worth investigating further, and produce a manageable set of candidates for real fieldwork.

In each of these cases, the synthetic audience is not replacing real research; it is restructuring the question of where real research budget should be spent.

Where synthetic audiences are not useful

It is equally important to be clear about where the method does not work, because this is where the most damaging overclaiming tends to happen.

Synthetic audiences are weak for questions about specific real-world products and brands for which the model has limited training-data exposure. A category that exists primarily in B2B contexts, a niche local brand, a recently launched product — these are not well represented in the linguistic patterns the model absorbed, and synthetic responses about them tend to be either generic or fabricated.

They are weak for emotional and reactive measurement, where the value of the response comes from the immediacy and authenticity of a real human reaction rather than from a reasoned response. Eye-tracking, implicit-association measurement, gut-reaction testing — none of this can be simulated by a language model, because the response being measured is not linguistic in origin.

They are weak for predicting behavior of real customers under real constraints. A synthetic respondent can be asked about willingness to pay, but its answer is unconstrained by an actual budget, by the inertia of an existing subscription, by the friction of an actual checkout flow. For questions where the constraint matters as much as the preference, simulation is a poor substitute for observation.

And they are weak as the sole basis for high-stakes commitments. No serious practitioner of the method recommends launching a product, repositioning a brand, or making a major pricing change on synthetic evidence alone. The role of synthetic audiences in a disciplined research practice is to inform, to triage, to screen — not to certify.

How this fits with the older methods

The synthetic audience is, on the analysis above, not a replacement for the survey but a new layer in a stack of methods.

The bottom of the stack is trace data: what real customers actually do, observed at scale and continuously. This is the most credible evidence about behavior, but it is silent on motivation and on hypothetical futures.

Above that is reading: what real customers say unprompted, analyzed at scale by large language models. This addresses motivation and perception for the questions customers have already had occasion to express themselves about.

Above that is synthetic simulation: directional responses from large-language-model audiences for questions about hypothetical futures, novel concepts, and multivariate design spaces. This is the layer that addresses what neither behavior nor existing expression can reveal.

And at the top is targeted real fieldwork: surveys, interviews, and experiments with actual customers, applied to the questions where the lower layers have done as much as they can and a genuine remaining gap requires direct asking.

This is a substantially different stack than the survey-centric architecture of classical customer research. It uses real fieldwork where it is most valuable rather than where it was once the only option. It allows decisions to be informed by evidence at a pace and scale that older architectures could not support. And it is honest about the limits of each layer, in a way that the discourse on AI in market research has often not been.

The methodological discipline this requires

Three principles seem to us essential for organizations beginning to use synthetic audiences responsibly.

First, validate against ground truth where you can. When an organization runs both a synthetic and a real study on the same question, the comparison both calibrates the synthetic method for future use and produces an institutional sense of where it can be trusted. Without this validation loop, synthetic results either get overtrusted (by enthusiasts) or rejected outright (by skeptics), with neither response well-founded.

Second, be explicit about the role of the synthetic study. Treating it as screening produces different methodological choices than treating it as validation. The role should be declared before the study is run, not inferred afterwards from whatever the results turned out to be.

Third, maintain genuine fieldwork capability. The risk of the synthetic method is not that it produces bad answers but that organizations let their real fieldwork muscles atrophy because synthetic answers are cheaper. The customer-research function that thrives in the next decade is the one that uses synthetic methods as a force multiplier on real research, not as a replacement for it.

In the next post in this series, we will turn to the fourth shift — from insights report to insights infrastructure — and examine what happens to the customer-research function when continuous instrumentation replaces the quarterly artifact.