Query-Conditioned Generative Engine Optimization: Per-Document Reflective Search Against a Live Citation Oracle
Year: 2026
Type: article
Abstract: Generative engines expose content through selective citation: a few sources are quoted andthe rest go unread. Dominant optimizers mine global rewrite rules from a corpus and stripquery dependence out of them by design, so the rewriter never sees the question the documentmust answer. That is defensible but not forced: in commercial content operations target queriesare routinely known, as in keyword-level search engine optimization, and instance-level queryconditioning remains largely unexplored. We study that setting and introduce RIO, a reflectiveoptimizer that evolves a rewrite instruction for each query–document pair against the engine’smeasured citation behaviour, under a fidelity-aware objective. Across two benchmarks, citationshare roughly doubles, rising by 0.205 and 0.198 where an equal split of five sources is 0.200.All twelve arms run on one scoring path with one rewriter model. Ablations bound it: most —62 % and 80 % — comes from knowing the query rather than from the optimizer, and at matchedinformation corpus-mined global rules significantly beat us. Against a budget-matched sampler,reflection reaches the same score at a third of the engine calls. Query conditioning also preserveslexical retrievability, which query-blind rewriting degrades. Gains survive a change of enginefamily, attenuating by two thirds, and the strongest engine we test is the hardest to move. Thesetting matters more than the mechanism: query-conditioned optimization is the right tool wherequeries are known and the document is worth the compute; mined global rules remain the rightchoice elsewhere.
Author: Koralewski, Marcin
Language: English
Subfield: Information Systems
Field: Computer Science
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W7203578667