Strategic Retrieval Exposure in RAG: Chunk Portfolios, Context Allocation, and Citation Measurement

Retrieval-augmented generation (RAG) systems operate under a finite context budget. Each query admits only a constrained working context, so retrieval policies allocate scarce context capacity across competing chunk portfolios, sources, and citation candidates. We study publisher-side adaptation of chunk portfolios under retrieval-mediated search. This adaptation differs from search engine optimization in its optimization unit, from generative engine optimization in its operating layer, and from retrieval-augmented generation poisoning in its objective: the target is retrieval exposure and citation allocation rather than page traffic, answer synthesis, or system compromise. We formalize strategic retrieval exposure using exposure@k, context share, source concentration (Herfindahl-Hirschman Index and Gini coefficient), citation allocation share, and the citation-retrieval gap. In a benchmark spanning five documentation categories, 30 sources, 622 chunks, 180 held-out query instances, and sparse, dense, late-interaction, hybrid, and reranked retrieval systems, strategic chunk portfolios substantially alter context allocation. In the strongest vector-database setting, a combined portfolio increases hybrid context share from 0 to 0.792 and the Herfindahl-Hirschman Index from 0.358 to 0.707. Across the benchmark, the same portfolio increases mean context share from 0.101 to 0.542 and mean exposure@8 from 0.580 to 0.880, while source concentration rises from 0.226 to 0.536. On the evaluated citation subset, citation allocation differs from retrieval allocation. Mean citation precision is 0.950, mean citation recall is 0.417, mean citation allocation share is 0.171, and mean retrieval-context share is 0.263. These results show that strategic retrieval exposure is a benchmarkable context-allocation problem with measurable strategic incentives and concentration amplification, with effect magnitudes varying substantially across retrieval architectures and query settings.