Paper Number
ICIS2026-3076
Paper Type
Completed
Abstract
Digital content platforms must recommend new content before knowing its quality or audience fit, while early feedback is selectively shaped by prior allocation rules. Although exploration may support learning, direct evidence is limited regarding whether early random exploration improves the subsequent organic recommendation performance of new content by generating more informative feedback. Using a randomized traffic allocation experiment with over 65,000 new videos on a major Chinese short-video platform, we show that early random traffic increases subsequent views and like ratios and reduces skip ratios. To identify the mechanism, we construct a video-level measure of feedback informativeness from the dispersion and depth of engagement states and instrument it using pre-treatment heterogeneity in assigned viewers’ historical engagement propensities. Videos that generate more informative feedback achieve larger gains later, especially in views and skips. Effective cold-start recommendation, therefore, requires diagnostic early exposure, not simply more traffic.
Recommended Citation
Li, Yilin; Zhou, Ningzhe; Wang, Shuxiao; Yu, Chuan; and Wang, Chong (Alex), "How Does Early Random Exploration Shape Content Allocation on Platforms? Evidence from a Large-scale Field Experiment" (2026). ICIS 2026 Proceedings. 11.
https://aisel.aisnet.org/icis2026/social/social/11
How Does Early Random Exploration Shape Content Allocation on Platforms? Evidence from a Large-scale Field Experiment
Digital content platforms must recommend new content before knowing its quality or audience fit, while early feedback is selectively shaped by prior allocation rules. Although exploration may support learning, direct evidence is limited regarding whether early random exploration improves the subsequent organic recommendation performance of new content by generating more informative feedback. Using a randomized traffic allocation experiment with over 65,000 new videos on a major Chinese short-video platform, we show that early random traffic increases subsequent views and like ratios and reduces skip ratios. To identify the mechanism, we construct a video-level measure of feedback informativeness from the dispersion and depth of engagement states and instrument it using pre-treatment heterogeneity in assigned viewers’ historical engagement propensities. Videos that generate more informative feedback achieve larger gains later, especially in views and skips. Effective cold-start recommendation, therefore, requires diagnostic early exposure, not simply more traffic.
Comments
24_Social_media