MRSS Talk by Shane Wang, Virginia Tech University (US)

Shane Wang opened this semester’s Marketing Research Seminar Series with his talk “Beyond What Is Said: Using Language Model Agents to Find Presentation Effects in Written Content.” The research shifts attention from what a message says to how it is presented: its stylistic choices, formatting, and structure. The central question is whether the internal representations of language models can help discover features beyond semantic content that influence marketing-relevant outcomes. Features such as layout, organization, spacing, capitalization, and visual emphasis are difficult to measure at scale: their meaning depends on context, they are hard to define in advance, and conventional text-cleaning procedures often remove them altogether. The authors introduce an approach that extracts interpretable features from the hidden layers of language models, links them to outcomes through predictive modeling, and uses a team of AI agents to label what each feature captures. They apply the method to Yelp reviews of Mexican restaurants, using readers’ helpfulness votes as a behavioral measure of review usefulness. The analysis identifies a range of presentation features associated with these votes. Follow-up experiments test two features related to text organization and show that they increase perceived usefulness by making the content easier to read and understand.

What I found particularly interesting was that the method can uncover features researchers might not have thought to look for in the first place. Details such as spacing, capitalization, and text organization are often overlooked by conventional text-analysis tools or lost during preprocessing, even though they can affect how readers process a message. I also appreciated the combination of behavioral evidence from Yelp and experimental validation of selected features. The review data show which features predict readers’ helpfulness votes, while the experiments help establish how particular organizational choices affect perceived usefulness and why.