Retrieval Governs Synthesis: A Forensic Audit of Generative Engine Optimization (GEO) Benchmarks, Replication Failures, and Provenance in RAG Pipelines

The integration of Large Language Models (LLMs) into web search via Retrieval-Augmented Generation (RAG) has prompted claims that traditional Search Engine Optimization (SEO) is obsolete, replaced by a novel discipline termed Generative Engine Optimization (GEO). In this study, we perform a line-by-line forensic audit of the GEO ecosystem: examining the foundational KDD '24 GEO benchmark (N = 9 methods; DOI: 10.1145/3637528.3671900), 401 downstream citing works culminating in multi-model replication studies at NeurIPS 2025 (C-SEO Bench; arXiv:2506.11097) and KDD 2026 (SAGEO Arena; arXiv:2602.12187), live-web core update telemetry, an open-access agency framework (WJARR-2026-0152; DOI: 10.30574/wjarr.2026.29.1.0152), and applying a strict Pure GEO Isolation Filter across seven flagship industry guides and a commercial 28-signal protocol (N = 56 distinct optimization signals). We demonstrate four primary findings: By the foundational paper's own architecture (G_qr -> SE -> S), which restates standard two-stage RAG (Lewis et al., 2020) gated by classical PageRank search retrieval (Brin & Page, 1998; Gyongyi et al., 2004) due to quadratic Transformer self-attention scaling O(n^2 * d) (Vaswani et al., 2017) and long-context positional degradation (Liu et al., 2024), generative visibility is 100% gated by Stage-1 search retrieval; documents ranked outside the top-K organic search cutoff receive zero visibility regardless of text modifications. The widely cited "30%-40% visibility boost" in KDD '24 was an artifact of a frozen five-document sandbox (including manual static file uploads to Perplexity.ai bypassing web search) and automated gpt-3.5-turbo prompts (src/geo_functions.py) that explicitly permitted fabricated data ("Addition of fake data is expected"; "You may invent these sources") and caused the model to hallucinate statistics and non-existent research bodies ("The International Chocolate Consumption Research Group"), which were subsequently rewarded by gpt-3.5-turbo evaluator self-bias. In realistic multi-actor, live-retrieval, and core-update replications (NeurIPS 2025, KDD 2026, February 2026 algorithmic telemetry, and October 2026 within-site controls), KDD '24 body-text rewrites yielded 0/10 positive gains on non-GPT evaluators (Claude-3.5-Haiku), actively degraded Stage-1 retrieval rankings across 171,003 documents, converged to zero in competitive games, and were outperformed by traditional SEO ranking by 770% (7.7x), with Stage-1 retrieval outweighing page-level formatting by 11.4x (32% vs. 2.8%) and organic search demotions causing immediate 22.5%-27.8% collapses in AI citations. Passing commercial GEO frameworks through our isolation filter yields a 0% survival rate of unique non-SEO optimization techniques; all legitimate mechanisms reduce to established Technical SEO, Entity SEO, Answer Engine Optimization (AEO), and Digital PR, whereas non-SEO interventions constitute adversarial spam or uncalibrated citation analytics. Finally, we document how high-entropy practitioner lexicon (Inference Optimisation and The Fortress of Facts) exposes automated citation laundering ("DOI washing") in open-access literature. Canonical Web Edition: https://www.hobo-web.co.uk/retrieval-governs-synthesis-a-forensic-audit-of-generative-engine-optimization-geo-benchmarks-replication-failures-and-provenance-in-rag-pipelines/Primary Forensic Dataset (95 Pages): https://www.hobo-web.co.uk/what-is-generative-engine-optimization-geo-a-geo-first-framework-debunked/