<![CDATA[AI browser agents increasingly promise to turn the mouse pointer into a situated interface for selecting page content, asking follow-up questions, scrolling to relevant sections, and opening contextual panels. For graphic design research, this shift matters because the pointer card is both a computational output and a visual communication device: it must make the target, supporting evidence, and next action legible at the moment of use. This paper evaluates evidence-aware web element reranking as the grounding layer for AI pointer cards. The empirical study uses the WebLINX candidate-reranking data, including validation, same-distribution test, and unseen-website test partitions. Each query is paired with candidate DOM/page elements and binary relevance labels. Nine reranking methods were evaluated, ranging from random and DOM-salience controls to BM25 variants, a transparent hybrid formula, a logistic reranker, and a depth-limited decision-tree reranker. Across 5,883 queries and 1,556,655 query–element pairs, the tree reranker achieved the highest same-distribution nDCG@10 (0.272) and Recall@10 (0.473), while the fixed hybrid achieved the highest unseen-website Top-1 accuracy (0.104). Combining lexical and DOM evidence improved ranking, whereas layout features transferred less consistently across websites. Because the first-ranked target was correct in only 10.4% of unseen-website queries, the findings support a confirm-before-act grounding prototype with visible alternatives and correction controls rather than autonomous execution. The reported metrics evaluate target ranking rather than card usability.]]>
Copyrights © 2026
