Whose Content, Whose Benefit? A Creator Value Retention Audit Framework for AI Wenyiwen Search Summaries on Xiaohongshu

Abstract

Generative search can present an answer before a user reaches the creators whose posts supplied the underlying knowledge. This paper introduces the Creator Value Retention Audit (CVRA), a reproducible computational framework for examining interface-level retention without claiming unobserved economic harm. Friction redistribution serves as an interpretive mechanism, not a measured outcome: the experiment records source interface layers and visible creator pathways but does not measure user effort. The framework combines rendered attribution salience, 1-indexed source interface-layer depth, creator affordances, relational cues, identity representation, source diversity, and lexical coverage in a transparent Creator Value Retention Index (CVRI). We executed a controlled browser experiment with 40 matched scenarios and 80 synthetic local HTML stimuli at a fixed 1280 by 900 viewport. The original-note-first condition achieved a mean CVRI of 97.92, compared with 57.07 for the AI-summary-first condition. Both conditions retained visible attribution, while summary-first stimuli placed the source at layer two instead of layer one, retained 25% of creator affordances, retained 40% of relational cues, and compressed visible text to 36.38% of source length. Equal-weight and component-ablation analyses preserved the gap direction. The experiment demonstrates internal metric sensitivity to prespecified synthetic differences, not real-world Xiaohongshu traffic, income, attention, creator loss, or perceived friction.