GEO Generative Engine Optimization Paper: 2026 Reversal Data

Introduction: The GEO Paper Everyone Cites, Half Reads

Search for the original Generative Engine Optimization paper, and the same summary appears almost everywhere: add citations, add statistics, add quotations, and a page gains up to 40% more visibility in AI-generated answers. That summary is accurate, but it leaves out the paper’s most consequential finding.

The 2024 study from Princeton, Georgia Tech, the Allen Institute for AI, and IIT Delhi produced the famous 40% headline. The same study also documented a reversal effect: the optimization tactics that lifted low-ranked pages could reduce the visibility of pages already sitting near position one, by roughly 20% to 30% in the paper’s rank-level breakdown.

This article covers five things:

1. It identifies exactly which paper “the GEO paper” refers to.
2. It isolates the reversal from the paper’s own data.
3. It explains why explainers online quote conflicting figures such as 28%, 41.5%, and 132.4% for the same tactic.
4. It tests the original findings against 2026 replication research.
5. It connects the lab results to live 2026 field data.

The timing matters. Data from AthenaHQ, BrightEdge, and Ahrefs published in 2026 shows that traditional rank and AI citation are separating in the real world, which makes the paper’s buried finding more relevant in 2026 than when it was published.

What “The GEO Paper” Actually Is (And Which One You Mean)

The canonical paper is “GEO: Generative Engine Optimization,” by Pranjal Aggarwal, Vishvak Murahari, Tanmay Rajpurohit, Ashwin Kalyan, Karthik Narasimhan, and Ameet Deshpande. The authors were affiliated with Princeton University, the Allen Institute for AI, Georgia Tech, and IIT Delhi.

Publication timeline:

• November 16, 2023: The term “Generative Engine Optimization” entered the literature with the preprint arXiv:2311.09735.
• August 25–29, 2024: The paper was formally published at KDD 2024, the 30th ACM SIGKDD Conference on Knowledge Discovery and Data Mining in Barcelona, under DOI 10.1145/3637528.3671900.

The paper is often confused with later work. arXiv:2509.08919 (“Generative Engine Optimization: How to Dominate AI Search”) and arXiv:2510.11438 (AutoGEO: Automatically Learning Generative Engine Preferences) are separate papers. Neither is the source of the “40%” figure. The same applies to E-GEO (arXiv:2511.20867), which extends the framework to e-commerce. These papers belong to a research lineage that the original started, and none of them should be cited in its place.

The paper’s footprint also shows how much more it is read than formally cited. Princeton’s research repository records 45 Scopus citations. ACM’s metrics page lists 60 total citations and 20,654 downloads. The paper has also reached the general public: Wikipedia credits it with coining the term, which is now used alongside “answer engine optimization” (AEO) and “AI optimization” (AIO).

Inside the Benchmark: GEO-bench and the Nine Tactics

To test its ideas, the team built GEO-bench. The benchmark contains 10,000 queries drawn from nine sources across 25 domains, paired with relevant web sources. Those queries were run through AI search systems so the researchers could measure how content changes shifted visibility inside generated answers.

The paper tested nine content modification strategies:

• Cite Sources
• Quotation Addition
• Statistics Addition
• Fluency Optimization
• Technical Terms
• Easy-to-Understand
• Authoritative (a more persuasive, confident tone)
• Unique Words
• Keyword Stuffing

Results were measured with two core metrics:

• Position-Adjusted Word Count (PAWC): How much of the AI-generated answer draws on a given source, weighted by where that source appears in the response. It measures share of the answer.
• Subjective Impression: A model-judged score of how relevant, influential, and prominent a source feels to a reader. It measures perceived quality rather than raw word share.

Nearly every number in the GEO debate comes from one of these two metrics. Keeping them separate resolves most of the confusion covered later in this article.

The Headline Stat, Verified Against the Raw Tables

The paper’s own framing is that GEO can boost visibility by up to 40%. Its top performers, Cite Sources, Quotation Addition, and Statistics Addition, achieved relative improvements of 30-40% on PAWC and 15-30% on Subjective Impression.

Table 1 gives the position-adjusted PAWC values, in rank order:

1. Quotation Addition: 27.2
2. Statistics Addition: 25.2
3. Fluency Optimization: 24.7
4. Cite Sources: 24.6
5. Technical Terms: 22.7
6. Easy-to-Understand: 22.0
7. Authoritative: 21.3
8. Unique Words: 20.5
9. Baseline (unmodified): 19.3
10. Keyword Stuffing: 17.7

Few summaries mention the last line. Keyword Stuffing is the only tactic that performed worse than doing nothing. One of the most familiar habits from traditional SEO actively hurt visibility in generative engines.

These figures are also averages across many queries and many starting ranks. An average can hide very different behavior at different positions, and in this paper, it does.

The Reversal Almost No One Reports

The paper’s central finding appears in its rank-dependent breakdown. When results are split by a page’s original search position, the same tactics move in opposite directions.

• Quotation Addition: roughly +100% at rank five, roughly -23% at rank one.
• Statistics Addition: roughly +98% at rank five, roughly -21% at rank one.

Low-ranked pages gained a great deal. Top-ranked pages lost ground.

The mechanism is straightforward. Generative engines already trust and surface top-ranked sources heavily. Adding quotes, statistics, or citations to those pages can dilute the authority signal they already carry or disrupt how the engine synthesizes them. A page at rank five starts out underrepresented, so the same additions help it compete for inclusion in the answer in the first place.

The paper’s deployed test on Perplexity.ai shows the same pattern. Using 200 test queries, with each page’s text uploaded as a file, Quotation Addition scored 29.1 against a 24.1 baseline on PAWC, a relative improvement of about 22%. Cite Sources and Statistics Addition showed gains of up to 9% and 37% across the two metrics. These gains were concentrated at lower starting ranks.

For anyone applying these findings in 2026, the implication is direct. Running these tactics on a page that already ranks first can reduce its visibility by up to roughly 30%. GEO tactics are not safe to apply to every page.

Why Every Article Quotes a Different Percentage for the Same Tactic

Readers comparing explainers often find Cite Sources reported as +28%, +30-40%, +41.5%, or even +132.4%, with no explanation of the difference. The variation is not random. It comes from which part of the paper each writer drew on:

• 28-40% figures generally come from the paper’s headline PAWC range.
• 41.5%-style figures usually come from Subjective Impression deltas or from tests that combined several tactics.
• Outliers like 132.4% come from rank-specific or best-case subsets of the rank-dependent breakdown, where low-ranked pages showed very large gains. They do not reflect the overall average.

A GEO percentage means little unless it states three things:

1. Which metric was used: PAWC or Subjective Impression.
2. Which starting rank the test began from.
3. Whether it was a single-tactic or combined-tactic test.

Most competing coverage omits all three. Any number reported without them, including the well-known “40%,” should be read with caution.

Stress-Testing the 2023/2024 Findings Against 2026 Replication Research

Two years after KDD 2024, researchers tested whether the nine tactics still affect citations once relevance and retrieval position are controlled for. The results are sobering.

The Vishwakarma et al. factorial study (2026) ran 252,000 trials across six LLMs and eighteen factors. It found that relevance and position are the primary determinants of the first citation. The authors’ conclusion is that a perfectly written page that never reaches the engine’s top retrieved results cannot be cited at all.

A September 2026 replication (Bajemon & Rochet’s work on scoring GEO without querying the engine) reported:

• A content score computed without querying any engine predicted citation order with a Spearman correlation of just 0.114.
• The three most commonly recommended GEO tactics moved citations on none of the ten engines tested.
• Volume-controlled statistics edits produced no pooled gain (-0.276 percentage points, 95% CI -0.970 to 0.425). The authors’ summary was that relevance, not volume, earns the citation.
• Between two engines answering the same question, mean overlap of cited URLs was 0.0079, and zero of the top five links were shared.

Structural research points the same way. SAGEO Arena, a February 2026 preprint, found that optimizing body text alone can hurt retrieval and reranking. Enriching structural fields such as schema markup, meta descriptions, and heading hierarchy improved retrieval by roughly 22%. A related study found that 43% of topically relevant pages earn zero AI citations, most often because no passage on the page works as a self-contained answer.

A 2026 critical survey of 45 studies concludes that the original paper’s lasting value is conceptual. It named the phenomenon, modeled generative engines as retrieval plus synthesis, and proposed metrics for citations spread throughout a response. The survey treats those contributions, rather than the specific tactic percentages, as the foundation of the field. It also notes that terminology, metrics, and evidence standards across the field remain inconsistent.

From the Lab to the Real World: 2026 Field Data Confirms the Pattern

If rank determines whether GEO tactics help or hurt in controlled tests, real AI search data should show a winner-take-most pattern. The 2026 field data does.

Leader and laggard gaps. AthenaHQ’s State of AI Search 2026, published September 29, analyzed millions of responses across eight LLMs. The average brand appears in only 16.3% of AI answers about its category, while category leaders reach 56.5%, about 3.5 times the average. A brand’s own domain goes uncited in 84% of AI answers about its own category.

Measurable concentration. Google AI Mode shows the highest citation concentration, with a Gini coefficient of 0.782. Perplexity is less concentrated. ChatGPT activates web search only for certain queries, so 57.8% of its runs produce no citations at all.

Rank and citation are separating. BrightEdge reports that 83% of Google AI Overview citations come from outside the top 10 organic results. Ahrefs found that only about 38% come from positions 1–10. Traditional rank still influences visibility, but it no longer guarantees a citation.

Instability across and within engines. A late-October 2026 Axios/Muck Rack study ran the same 1,020 prompts through ChatGPT, Claude, Gemini, and Perplexity and found that each engine favored different sources:

• ChatGPT: 34.6% government and nonprofit sources
• Gemini: 34.4% company-owned content
• Perplexity: 27.8% journalism
• Claude: 23.5% third-party corporate content

When ChatGPT was asked the same questions twice on the same day, only 41% of cited domains overlapped. Journalism’s overall share of citations also fell from about 27% in May to 22.2%, below third-party corporate content at 25.1%. A strategy of optimizing a page once and leaving it does not hold up under these conditions.

What This Means for GEO Strategy in 2026: Rank-Aware, Not Tactic-Obsessed

The original paper’s reversal effect, the 2026 replications, and the field data all lead to the same conclusion. A page’s existing position and relevance determine whether the classic GEO tactics help, do nothing, or cause harm.

In practice, a rank-aware approach looks like this:

• Diagnose before editing. Establish whether a page is already a top performer in traditional search and AI citations, or an underperformer.
• Use restraint on top performers. For pages that already lead, structural refinement (cleaner headings, schema, self-contained answer passages) is safer than adding quotes or statistics.
• Apply additive tactics to underperformers. Lower-ranked pages are where citations, quotations, and statistics are most likely to help.
• Invest in retrieval first. A page that never reaches an engine’s retrieved set cannot be cited, so relevance, structure, and internal linking come before copy changes.
• Measure repeatedly. Given the instability across and within engines, a single snapshot is not a reliable measure.

This aligns with KOZEC’s Search Compliance Optimization (SCO) philosophy. SCO focuses on following search engines’ recommended best practices, useful content, clear pages, smart internal links, and consistent publishing, rather than chasing algorithmic shortcuts or copy tricks applied the same way to every page. KOZEC’s approach combines structured content ecosystems, schema and structured data optimization, and ongoing performance tracking. It is designed for the rank-aware conditions described above, not the uniform checklist that 2026 research shows can backfire.

Conclusion: The Paper’s Most Important Finding Was Never the 40%

The “40% boost” is real but incomplete. The paper’s own rank-level data shows the same tactics can cost top-ranked pages up to about 30% of their visibility.

The 2026 replication research, from Bajemon & Rochet’s September study to Vishwakarma et al.’s 252,000-trial experiment, supports the directional point that relevance and position dominate. It also undercuts the universal magnitude claims that many explainers still repeat. The field data from AthenaHQ, BrightEdge, Ahrefs, and Axios/Muck Rack shows that position and relevance, not individual content tactics, now govern AI search visibility in practice.

In 2026, GEO depends on a page’s existing standing. Content strategies that ignore that standing risk undoing the visibility they already have.

Build Rank-Aware GEO Content With KOZEC

Applying GEO tactics without regard to rank or relevance carries real risk. KOZEC’s agentic platform and SCO framework are built around structured, interlinked content that accounts for each page’s position, rather than isolated copy changes.

Relevant capabilities include:

• Competitive and topic analysis to find genuine content gaps and relevance opportunities
• Structured data optimization, the kind of structural enrichment that 2026 research links to better retrieval
• Internal linking into topical ecosystems rather than standalone pages
• Ongoing performance tracking that helps separate pages already performing well from those that need support

KOZEC reports results from client businesses, including +386% AI Overview citation growth and +621% keyword visibility increase, as examples of this structural approach. These figures are self-reported and not independently verified.

Teams ready to build a GEO strategy based on 2026 research can book a demo at kozec.ai/schedule-a-demo/ or compare KOZEC’s plans, starting with the Foundation tier at $600 per month with no long-term contract.