[4Yin-B] Poster 7

[4Yin-B-39] Relative quantification of brand personality using large language models

〇Yutaka Kuroki^1,2, Tomonori Manabe^4,3, Kei Nakagawa¹ (1. Osaka Metropolitan University Graduate School of Business, 2. Sansan, inc., 3. University of Tsukuba, 4. Sompo Risk Management Inc.)

Keywords:

Brand personality,Large language models,Bradley-Terry model

Brand personality measurement often relies on surveys, which can be costly and difficult to scale across many brands and categories.We propose a survey-free approach that uses a large language model (LLM) as a proxy rater in within-category paired comparisons.To mitigate order/position effects, we query each brand pair in both orders and aggregate responses.For each of Aaker’s five personality dimensions, we fit a Bradley-Terry model to obtain softmax-normalized latent strengths.We evaluate external validity on three categories (Financial services, Cars, and Beverages) by assessing association with reference indicators derived from a consumer-survey dataset introduced in prior work.We find that the inferred scores often align with the external indicators, although validity remains strongly dimension- and category-dependent.When association is weak, we discuss a plausible explanation: differences between publicly available knowledge reflected in LLM outputs and the consumer perceptions summarized by the reference indicators.We also report simple diagnostics of internal consistency to help judge when inferred rankings are stable and when additional evidence may be needed.