Beyond the Blue Link: Empirical Evaluation of Generative Engine Optimization in Stochastic Retrieval Systems

Abstract

This paper studies how authored hypertext structure affects machine-mediated attribution in generative retrieval systems. We formalize this as Answer Engine Optimization (AEO) and introduce Generative Share of Voice (gSoV), a probabilistic visibility metric for stochastic retrieval. Our core result comes from controlled retrieval injection on open-weights models: structured HTML yields 2.6 × higher citation rates and 4.0 × higher extraction fidelity than equivalent unstructured content, with measurably lower attention entropy. We formalize this advantage using the Dexter Reference Model, showing that SAA Answer Units function as self-describing components whose anchor structure reduces synthesis rejection. Controlled source substitution shows that community corpus placement increases citation probability for subjective queries (ΔgSoV = 18.2 pp, p < 0.001), independent of content quality. A 12-month field study and cross-domain validation across 12 entities (Cohen’s d = 1.42) provide convergent support. The contribution for Hypertext is methodological: we treat semantic markup, node structure, and community-linked traces as first-class determinants of machine-mediated reuse, connecting generative engine optimization to classical concerns about composites, transclusion, and reader agency. We release our evaluation protocol and code.1