Ambiguity in Multi-modal Digital Ads

Information systems and operations management

Speaker :  Shunyuan ZHANG (Harvard Business School)

Room Bernard Ramanantsoa

"We explore the effect of digital ambiguous ads on consumers’ behavior throughout the purchase funnel, considering a multi-modal perspective of the display ad’s visual banner and its textual caption. We collaborate with a display ad platform, analyzing consumers’ click-through rates (CTRs) and conversions (CRs) for tens of thousands of cross-category digital ads. To operationalize ambiguity, we develop two custom deep learning-based ambiguity prediction models, each for one data modal. We find that beyond a rich set of ad characteristics (e.g., photographic attributes, language features, and image-text coherence), ambiguous ads garnered higher click-through rates. However, these ads resulted in lower conversion rates and efficiency. Next, to verify the causal links suggested in the field data, we conduct a pre-registered randomized field experiment, where we manipulated the amount of ambiguity of in a campaign. In particular, we create four versions of ads for a hearing-aid product with very similar images and texts, but different levels of ambiguity. Our analysis further reveals a negative correlation between ad ambiguity and the end-to-end conversion (conversions/impressions). Overall, our findings suggest that advertisers and scholars are well-advised to assess images and texts together rather than individually, and use ambiguity with care."