From Page Form to Context Admission: Stage-Dependent Effects of Content-Preserving Rewrites in Generative Search

Generative AI answers are becoming a main entrance through which users meet brands, and companies routinely rewrite their official websites into FAQ or entity-page formats to be cited. We ask how such formatting acts across the stages of a controlled retrieval-augmented pipeline, using a testbed of 35 Chinese home-appliance and consumer-electronics brands whose pages were rewritten under strict proposition-level content preservation (46,139 frozen facts; fact equivalence model-verified) and substituted through experimenter-assigned full-crossover arms across 87,975 runs (BM25 primary, dense contrast; the designed reranker was not executed in the main batch, deviation D-36, and is executed in a replication). Effects concentrate at the context-admission margin. Among the 15 brands whose treatment verifiably took effect, FAQ pages raise the probability of entering the answer context from 28.6% to 47.2% (log risk ratio +0.501, 95% CI [+0.19,+0.89]; Holm-corrected significant), holding under single-adopter semantics (+0.501, [+0.17,+0.93]); aggregates are stable across background-composition semantics although unit-level entries churn and offset. Conditional on entry, generation-side estimates are near zero to negative, with equivalence to zero not established; significance is anchored to one design-preserving bootstrap pipeline, with a conservative alternative disclosed. Entity-page rewriting instead cuts lexical-retrieval admission (market-level −1.20 [−2.01,−0.65]; single-adopter −1.857 [−3.05,−1.16]). The FAQ advantage is lexical-specific and inverts under dense retrieval (+18.6pp sparse → −16.9pp dense), while the entity-page penalty persists under every retriever tested (sparse/dense/hybrid −20.0/−29.7/−22.8pp, all significant). An LLM evidence filter empties the pack for 23% of queries (50% on brand-evaluation) and yields admission −11.9pp with citation +4.2pp on survivors. Citation markers are additionally contaminated by ritual citations, markers attached to refusal answers, spanning 5–59% across generators (re-annotation: κ 0.72/0.51; DS cells robust, Doubao-F annotator-sensitive), so marker-based visibility metrics require content-level auditing. Effects are stage-dependent on this testbed's configuration; commercial engines' evidence-processing stages are not observed, and validation on production engines remains future work. All analyses are exploratory (the study was not preregistered).