The GEO industry has a marketing problem disguised as a research question. Vendors sell optimization methods with confident case studies; meanwhile, the peer-reviewed literature contains a paper — published at NeurIPS, one of machine learning's most selective venues — whose headline conclusion is that "most current C-SEO methods are largely ineffective."
Both things are true. This piece walks through the evidence bench by bench, because the disagreement is not random: it tracks identifiable differences in methodology, and once you see them, you know which findings to trust under which conditions. If you buy or sell GEO services, this is the due-diligence layer.
The Founding Claim: +40% Visibility
The field's founding evidence is the original GEO paper (Aggarwal et al., KDD 2024; arXiv:2311.09735). Princeton and IIT Delhi researchers proposed the GEO concept, built GEO-bench (10,000 real queries across domains), and tested nine optimization methods — adding citations, quotations, statistics, fluency optimization, easier language, authoritative tone, technical terms, unique words, and keyword stuffing.
The results that launched an industry: citations, quotations, and statistics boosted visibility by 30–40%; keyword stuffing did essentially nothing; effects varied by domain.
The caveat that most follow-up coverage omits: the experiments ran in single-actor conditions on GPT-3.5-class models — one document being optimized, older generation models. Whether the gains survive when everyone optimizes at once, on modern engines, was an open question. The next benchmark answered it.
The Reversal: "Largely Ineffective"
C-SEO Bench (Puerto et al., NeurIPS 2025; arXiv:2506.11097) is the study the industry quotes least. Its design fixed the founding study's blind spots: two tasks (question answering and product recommendation) across six domains, 1,921 queries, 16,360 document records, four modern models (GPT-4o mini, Claude 3.5 Haiku, o3, o4-mini) — and crucially, multi-actor conditions where many documents adopt the same optimization simultaneously, as they would in the real world.
The findings, in escalating order of discomfort:
Testing nine C-SEO methods (seven inherited from the GEO paper, two new), only 3 of 54 significance tests survived multiple-comparison correction
In 19 of 24 model-domain combinations, one popular tactic — rewriting content to add statistics — significantly lowered the document's citation rank
The strongest lever in the entire study was not a GEO tactic at all: simply placing the source document first in the retrieval context (a proxy for classic search dominance) improved citation rank 7.6× more than the best content-rewriting method
When many actors adopt the same tactics, gains collapse — the authors call the condition "congested, zero-sum": everyone optimizing shifts citations around without creating new visibility
Read carefully, C-SEO Bench does not say "GEO is fake." It says something sharper: content rewriting, as a category, mostly does not move modern engines — and the battle is decided upstream, at retrieval.
The Retrieval-Layer Verdict: Structure Beats Rewriting
The 2026 generation of benchmarks converged on exactly that conclusion, from three independent directions:
SAGEO Arena (arXiv:2602.12187) built the first full-pipeline benchmark — real retrieval, real generation, 170,000 documents — and found existing optimization methods "remain largely impractical under realistic conditions." The degradation concentrates in the retrieval and re-ranking stages, precisely where content rewriting never reaches. Notably, it found that structural signals like schema markup — which earlier benchmarks ignored — matter, because they operate at the stage where the battle is actually decided.
GEO-SFE (2026) reached the same verdict with different vocabulary: document structure, not content phrasing, largely determines whether a page gets cited — "Structural Feature Engineering."
FeatGEO (2026) re-tested the original nine GEO tactics on modern AI engines and found them largely ineffective — then redirected the field toward feature-level optimization of the retrieval pipeline itself.
RAID G-SEO (2025) added the final twist: optimization methods that model search intent — without even looking at the user's surface query — outperformed the entire original GEO tactic set.
And the observational layer agrees: GEO-16 (arXiv:2509.10762; 1,702 citations audited across three engines) found the strongest citation-correlated page attributes are metadata & freshness, semantic HTML, and structured data — infrastructure, not prose — with pages scoring G ≥ 0.70 and hitting at least 12 of 16 quality pillars earning significantly higher citation rates.
The Synthesis: What Died, What Survived
Laying the six benchmarks side by side, a consistent verdict emerges:
Probably dead as standalone tactics:
Content rewriting for fluency, tone, or keyword variety (the original GEO method set)
Statistics-insertion as a content trick — the single most counterproductive finding in the literature
Any tactic assumed to work identically across engines and domains
Alive and repeatedly validated:
Winning retrieval — classic search strength, placement in trusted corpora, presence on pages systems actually retrieve (the 7.6× finding; consistent with the rank-citation overlap evidence in this series)
Document structure — semantic HTML, schema markup, extractable passages, clean heading hierarchies
Freshness and metadata — GEO-16's strongest correlated pillars
Earned-media presence — the cross-engine study (Chen et al., arXiv:2509.08919) confirms AI search's overwhelming preference for third-party authoritative sources, a finding this series has documented from industry data as well
Entity clarity — the substrate every benchmark implicitly depends on
What This Means for Buyers and Practitioners
The evidence trial resolves into an honest purchasing guide:
Be suspicious of any offer whose mechanism is "we rewrite your content." The literature's most replicated negative finding targets exactly that. Rewriting is, at best, a rounding error on top of structural work.
Prioritize the levers the evidence supports: retrieval access (can the engines fetch and parse you at all), document structure, freshness, entity consistency, and earned mentions. These are the findings that survive methodological scrutiny.
Discount single-actor case studies. The congestion finding means "we improved one client's citations" and "this method works" are different claims — under realistic competition, method-level gains compress sharply.
Ask vendors which benchmark their method would fail. Every method has conditions of validity. A vendor who cannot name theirs has not tested rigorously.
Expect this literature to keep moving. SAGEO Arena appeared in 2026; the benchmark generation it inspired is young. The responsible posture is to re-verify annually — which is also why this series dates every figure it cites.
The Honest Caveats
Benchmarks are synthetic conditions. Real retrieval systems, user personalization, and conversational context differ from test harnesses — in both directions (some methods may work better in production; others worse).
The field's baseline keeps shifting. C-SEO Bench tested 2024–2025 models; SAGEO Arena tested newer pipelines. "Modern" has a half-life of roughly two quarters here.
Absence of evidence is not evidence of absence — but it is evidence of priority. The tactics that died in testing were not proven harmful in every context (statistics-insertion excepted); they were proven unreliable. Unreliable tactics do not deserve line-item budgets.
Publication bias runs both ways. Negative results are rarer everywhere in applied ML; the consensus against content rewriting is nonetheless unusually strong because multiple independent teams reproduced it.
The Bottom Line
Does GEO work? The evidence-supported answer is precise: the answer layer rewards infrastructure — retrieval access, document structure, freshness, entity clarity, and earned presence — and largely ignores prose-level rewriting. The founding study's +40% was real under its conditions; the follow-up literature established that those conditions rarely hold in production, and redirected the field toward the layers where durable advantage actually lives.
That is not a disappointment. It is a clarification — and it happens to favor exactly the unglamorous, compounding work (structure, access, authority) that survives contact with any search technology, this one included.
Sources
Aggarwal et al., "GEO: Generative Engine Optimization" (KDD 2024; arXiv:2311.09735): +30–40% visibility for citations/quotations/statistics; nine methods; GEO-bench (10K queries); single-actor, GPT-3.5-class conditions
Puerto et al., "C-SEO Bench: Does Conversational SEO Work?" (NeurIPS 2025; arXiv:2506.11097): 1,921 queries, 16,360 records, 4 modern models; 3/54 significant tests; statistics-rewriting harmful in 19/24 combinations; context-placement baseline 7.6× strongest method; congested zero-sum dynamics
SAGEO Arena (2026; arXiv:2602.12187): first full-pipeline realistic benchmark (170K documents); methods "largely impractical"; degradation in retrieval/re-ranking; structural signals previously ignored
GEO-SFE (2026): document structure determines citation; FeatGEO (2026): original tactics fail on modern engines; RAID G-SEO (2025-08): intent-modeling beats original tactics
Kumar & Palkhouski, "GEO-16" (2025; arXiv:2509.10762): 1,702 citations, three engines; top pillars — metadata & freshness, semantic HTML, structured data; G ≥ 0.70 + ≥12 pillars operational point
Chen et al., "Generative Engine Optimization: How to Dominate AI Search" (2025; arXiv:2509.08919): earned-media preference across engines
Cross-context from this series: The Great Decoupling, The Mention Economy, The Invisible Supplier (structural mechanics)
Benchmark results reflect the models and conditions tested as of each paper's date; this literature moves quarterly.