AI-enabled method enhances marketing research

Instead of relying on a few handcrafted descriptions, Cornell researchers were able to use nearly 120,000 real wine tasting notes to capture complex customer reactions.

This improvement on traditional methods was enabled by BRIDGE, Behavioral Research Through Interpretable, Dimensionality-reduced Generative AI Embeddings, a new method designed by researchers from the Cornell SC Johnson College of Business and Singapore Management University. Published in the Journal of Marketing Research on Sept. 6, BRIDGE uses generative AI to enable high quantities of real-world product descriptions to be used in research settings.

BRIDGE addresses the stimulus sampling problem, which is the tendency for selected, handcrafted experimental product descriptions to poorly represent real-world product descriptions. Using BRIDGE, researchers can use real-world textual product descriptions controlling for differences in style, length and content, allowing them to deploy realistic stimuli in experiments while avoiding confounding variables, which are unseen factors that distort results.

“In my field of marketing, specifically consumer behavior, we really care about interpretability,” said Sachin Gupta, the Henrietta Johnson Louis Professor of Marketing at the Samuel Curtis Johnson Graduate School of Management in the SC Johnson College. “We start with a theory, and we want to end with a story of what’s happening in the data. So simply finding something statistical is not enough. You want to be able to interpret what you’re finding, and I think that’s something BRIDGE does quite well.”

In BRIDGE, a large language model finds real-world product descriptions from sites like Amazon, then converts them into embeddings, numbers that correspond with the meaning in the description. When these embeddings are input into a statistical model, BRIDGE identifies and tests which product features influence consumers’ choices.

Findings from the wine experiment reflected well-known patterns in behavioral science, such as the tendency for initial product offerings to influence later consumer decisions, and successfully controlled for confounding variables.

“What excites me most about this methodology is addressing the challenge of making research more realistic,” Gupta said. “I’ve always been interested in the practical impact of what we do in academia, and when research can be more accurate, it will have a greater impact on all our lives.”

Co-authors were Anirban Mukherjee of Avyayam Holdings, previously a visiting scholar at the SC Johnson College, and Hannah H. Chang, associate professor of marketing at the Lee Kong Chian School of Business within Singapore Management University.