How Source Attribution Visualization Shapes User Attention and Preference: An Eye-Tracking Study of Four AI Chatbot Layouts

초록

As generative AI chatbots become a primary information channel, users increasingly accept answers without verification, and citations can raise trust even when sources are irrelevant or fabricated. How source-attribution visualization shapes the visual preconditions of verification remains unknown: users can notice, read, or compare a source without clicking. This within-subjects eye-tracking study (N = 23; 92 trials) evaluated four attribution visualizations abstracted from commercial AI chatbots and rendered as simulated screens: inline component (sentence-end chips), card list (cards above the answer), side panel (adjacent panel), and raw hyperlink (bare URLs), combining gaze metrics, surveys, and interviews. Repeated-measures ANOVAs revealed strong layout effects on source discoverability and engagement, largely robust to sensitivity checks (the panel's discovery latency was order-sensitive): the card list was discovered almost immediately, with the raw hyperlink last. Yet no self-reported measure differed detectably. The most frequently nominated format, the inline component, attracted about half the dwell time of the stand-alone formats, whose prolonged fixations suggested citation-to-text mapping cost rather than genuine engagement. This attention-preference gap means both must be measured jointly. We contribute a four-layout gaze-based comparison, a reproducible participant-level analysis workflow, and three design principles (pre-click identifiability, sentence-level claim-source mapping, and in situ preview) within a proposed two-stage attribution architecture.

키워드