EXPRESS: Behavioral Research Through Interpretable, Dimensionality-reduced Generative AI Embeddings (BRIDGE): A Method to Incorporate Real-World Stimuli in Consumer Experiments
TL;DR
By incorporating many unaltered product texts into experiments, BRIDGE enhances the realism, generalizability, and practical relevance of consumer research in information-rich environments.
Abstract
Traditional experiments often rely on a few, stylized stimuli, which can limit realism and undermine generalizability beyond the sampled stimuli—known as the stimulus-sampling problem. To address this challenge, this paper introduces BRIDGE, a novel analytical method that enables the use of many unaltered real-world descriptions as experimental stimuli. Leveraging foundational generative AI embeddings, BRIDGE develops (1) structured, low-dimensional, and interpretable representations of focal constructs and (2) statistical controls for non-focal nuisance variations, facilitating causal inference. Extensive Monte Carlo simulations, two coffee certification experiments, and a large-scale choice experiment (plus its validation study) show that BRIDGE recovers true parameters even when textual stimuli contain unobserved nuisance variations, and can effectively account for different sources of confounding. In the choice experiment, 1,000 participants evaluated approximately 50,000 unique product descriptions randomly sampled from a corpus of nearly 120,000. Results show that entirely incidental initial products can shape participants’ subsequent preferences. By incorporating many unaltered product texts into experiments, BRIDGE enhances the realism, generalizability, and practical relevance of consumer research in information-rich environments. A detailed researcher’s guide and Python package bridge are provided.