LLMs Reproduce Human Purchase Intent via Semantic Similarity Elicitation of Likert Ratings

TL;DR

This paper introduces a semantic similarity rating (SSR) method that uses LLMs to simulate consumer responses with realistic distributions and qualitative feedback, enabling scalable and interpretable consumer research.

Contribution

The paper presents SSR, a novel approach that elicits textual responses from LLMs and maps them to Likert scales, improving response realism and interpretability in synthetic consumer surveys.

Findings

Achieves 90% of human test-retest reliability

Maintains realistic response distributions (KS > 0.85)

Provides rich qualitative feedback from LLM respondents

Abstract

Consumer research costs companies billions annually yet suffers from panel biases and limited scale. Large language models (LLMs) offer an alternative by simulating synthetic consumers, but produce unrealistic response distributions when asked directly for numerical ratings. We present semantic similarity rating (SSR), a method that elicits textual responses from LLMs and maps these to Likert distributions using embedding similarity to reference statements. Testing on an extensive dataset comprising 57 personal care product surveys conducted by a leading corporation in that market (9,300 human responses), SSR achieves 90% of human test-retest reliability while maintaining realistic response distributions (KS similarity > 0.85). Additionally, these synthetic respondents provide rich qualitative feedback explaining their ratings. This framework enables scalable consumer research simulations while preserving traditional survey metrics and interpretability.