Synthetic Consumer Insight Generation with Large Language Models

Stephen L. France and Pia. A. Albinsson submitted a paper titled "Synthetic Consumer Insight Generation with Large Language Models" to arXiv on 7 Jul 2026 (arXiv:2607.05761, file size 985 KB). The paper tests whether large language models can generate synthetic consumer data for projective techniques by comparing LLM outputs with human responses collected in a primary study about perceptions of city tourism destinations.

What did the authors test and how?

The authors tested LLM-generated responses across multiple projective tasks, multiple LLMs, different prompting strategies, and varying temperature settings, and then compared those outputs with human responses from a primary research study on city tourism perceptions. They analyzed both human and LLM responses using linguistic measures, diversity and concentration metrics, topic models, and top-term analyses to identify similarities and differences in content and form.

The experimental scope therefore covered model choice, prompt design, and temperature as tunable variables, while relying on standard analytical tools from computational text analysis. That approach allowed the authors to evaluate not only what topics LLMs produced but also how those topics were expressed and how diversity emerged across responses.

How did LLM outputs compare with human responses?

The study found substantial overlap between human and LLM responses in broad topics and associations, but also important differences in style, linguistic structure, and the way diversity is generated. In other words, LLMs captured many of the same high-level associations that human respondents provided, yet they produced those associations with different linguistic patterns and different mechanisms for diversity.

France and Albinsson applied topic modeling and top-term analyses to map themes, and used diversity and concentration metrics to quantify variation. Those analyses showed that while the set of broad topics often matched between humans and models, the surface-level language and the distributional patterns across responses diverged. The paper therefore separates content overlap from stylistic and distributional differences when assessing whether synthetic responses can substitute for human data in projective techniques.

What recommendations do the authors give?

The paper gives guidance on when and how to use LLMs for generating synthetic consumer data, emphasizing that model and prompt choices shape response quality and that limitations must be recognized. The authors recommend careful selection of prompting strategies, tuning of temperature settings, and model choice to improve the utility of LLM outputs for projective tasks. They also caution against treating LLM-generated data as a drop-in replacement for human-collected responses because of the documented stylistic and diversity differences.

Why it matters

Marketing and consumer-research teams run projective techniques to surface implicit associations, emotions, wants, and needs, but collecting human responses is often costly and slow. If LLMs can reproduce the same broad topics reliably, they offer a way to prototype projective tasks at lower cost and at scale. The study shows that LLMs can already match humans on theme coverage, but the documented differences mean researchers must validate synthetically generated results against real respondents before drawing decisions from them.

What to watch

Look for applied validations that test the paper's recommendations in live market-research settings and for replication across different consumer domains beyond city tourism. Subsequent work that quantifies when stylistic or diversity differences materially change interpretation of projective tasks will determine whether synthetic responses can be used routinely in place of human-collected data.

References: the arXiv submission is listed as arXiv:2607.05761, submitted 7 Jul 2026 by Stephen L. France and Pia. A. Albinsson.