Pranjal Aggarwal’s GEO Paper: From KDD 2024 Findings to 2026 Execution

Introduction: The Paper That Named a Discipline, and What Practitioners Still Miss

On November 16, 2023, Pranjal Aggarwal and five co-authors uploaded a preprint to arXiv under the identifier arXiv:2311.09735. In doing so, they formally coined the term “Generative Engine Optimization,” a moment that quietly restructured how the entire marketing industry would come to think about visibility in AI-driven search. The paper was later published at KDD 2024, the 30th ACM SIGKDD Conference on Knowledge Discovery and Data Mining held in Barcelona from August 25 to 29, 2024, appearing on pages 5 through 16 under DOI 10.1145/3637528.3671900. That peer-reviewed pedigree gives it an authority that virtually no practitioner GEO content can match.

Most content about GEO either summarizes the Aggarwal paper superficially or ignores it entirely. This piece treats the paper’s empirical data as a decision framework, mapping what the reported lift numbers actually mean for operational content strategy in 2026.

The stakes are not academic. EMARKETER projects that roughly 31.3% of the US population will use generative AI search in 2026, and Gartner predicted that traditional search volume would drop 25% by this year. The Aggarwal et al. findings are the closest thing practitioners have to a controlled proof of what works.

What follows: who wrote the paper and why that matters, what the benchmark actually tested, a strategy-by-strategy translation of the lift numbers, domain-specific implications from the paper’s Table 3, and what the accumulating 2026 evidence base adds to the original findings.

Who Wrote the GEO Paper and Why the Provenance Matters

The paper is commonly called the “Princeton paper,” but that label is imprecise. Lead author Pranjal Aggarwal was affiliated with IIT Delhi. The full collaboration spans Princeton University (Vishvak Murahari, Karthik Narasimhan, and Ameet Deshpande), IIT Delhi (Aggarwal), Georgia Tech (Tanmay Rajpurohit), and the Allen Institute for AI (Ashwin Kalyan). It is more accurately described as a Princeton and IIT Delhi collaboration with contributions from two additional research institutions.

Institutional provenance matters for practitioners because the multi-institutional, peer-reviewed origin gives the paper a credibility floor that blog-based GEO frameworks cannot reach. When a marketing leader needs to justify a GEO investment to a CMO or a board, this is the academic anchor to cite. It is not one agency’s opinion; it is a benchmark study presented at one of the most respected data mining conferences in the world.

The revision history reinforces that rigor. The paper moved through three versions: v1 on November 16, 2023, v2 on May 28, 2024, and v3 on June 28, 2024. This reflects ongoing refinement in response to community feedback before and after the KDD 2024 presentation, a sign of methodological seriousness rather than instability.

Context matters too. The paper was written when Perplexity.ai was already live but before Google AI Overviews dominated mainstream awareness. That timing makes it a genuinely forward-looking document, one that anticipated the 2026 landscape rather than merely reacting to it.

What the Paper Actually Built: GEO-Bench and the Three-Stakeholder Problem

At the center of the research sits GEO-bench, a large-scale benchmark of 10,000 diverse user queries spanning nine domains and sourced from nine underlying datasets. GEO-bench was designed to evaluate content visibility strategies inside generative engine responses, and it is the empirical foundation for every finding discussed in this article.

To measure visibility, the authors introduced two core metrics. The first is Position-Adjusted Word Count (PAWC), an objective measure that weights the volume of cited text by its citation position within the generated response. The second is Subjective Impression, a qualitative LLM-as-a-judge score assessing how prominently the engine credits the source. Together, these capture both how much of a source is used and how visibly it is attributed.

The paper also introduced a framing that most competitor content ignores: the three-stakeholder problem. Traditional search had two clear winners, users and search engines, while content creators formed a third party who received traffic in exchange for being indexed. Generative engines disrupt this arrangement by synthesizing answers directly, often without sending traffic back to source websites. That disruption reframes GEO as a strategic necessity rather than an optional enhancement. If the engine answers the question and the user never clicks, being cited becomes the new visibility currency.

Two caveats deserve emphasis. First, the paper used Perplexity.ai as its primary generative engine test bed, not ChatGPT or Google AI Overviews, which has implications for how broadly the findings generalize across the 2026 AI search optimization strategy landscape. Second, the framework is deliberately black-box: content creators can define and optimize custom visibility metrics without any access to the engine’s internal workings. That design is democratizing because it makes the methodology applicable across platforms rather than tied to a single vendor.

The Nine Strategies: What the Lift Numbers Actually Mean in Operational Terms

The paper tested nine content optimization strategies against an unoptimized control, with Word Stuffing serving as the negative baseline representing traditional SEO orthodoxy. The point of this section is not to summarize those strategies but to translate their measured lift into what a content team should do differently on Monday morning.

The Top Performers: Citation-Oriented Strategies That Beat the Baseline by 30 to 40%

Three strategies dominated the results, and all three share a common logic.

Quotation Addition improved PAWC by 41% and Subjective Impression by 28% over baseline, making it the single best-performing individual strategy. The operational translation is direct: embedding quotes from credible sources, experts, or primary documents hands the generative engine pre-packaged, citable text that requires minimal synthesis effort.

Statistics Addition improved PAWC by 31% and Subjective Impression by 23%. Specific, sourced numerical claims (not vague assertions) function as high-confidence anchors that engines preferentially cite because they are verifiable and compact.

Cite Sources achieved a 30 to 40% relative improvement on PAWC. Explicitly referencing authoritative sources within the body of the content signals that the content is itself a reliable secondary source, a recursive credibility signal that engines reward.

The unifying principle is this: generative engines are retrieval-and-synthesis systems. Content that pre-does the synthesis work by packaging quotable, citable, verifiable units reduces the engine’s cognitive load and increases citation probability.

The Combination Advantage: Why Fluency Plus Statistics Outperforms Any Single Tactic

The paper’s most underreported finding is that Fluency Optimization combined with Statistics Addition outperformed any single GEO strategy by more than 5.5%. GEO is, in other words, a combinatorial optimization problem, not a flat checklist.

The operational translation: a content piece that reads naturally while embedding specific, sourced data points achieves machine-readability and human credibility simultaneously. Most competitor content misses this because presenting GEO tactics as a flat list implies equal weight and independent application. The data suggests a layered approach instead, with fluency as the foundation and citation-oriented strategies as the amplifier.

For content teams, the practical implication is to audit existing content not only for which tactics are present but for which combinations are present. A piece with strong statistics but poor fluency, or fluent prose with no citable anchors, is leaving measurable visibility on the table.

The Strategies That Underperform, and What That Reveals

Keyword Stuffing performed at or below the unoptimized baseline, frequently posting negative performance. This is one of the paper’s most unambiguous results: keyword-density optimization, the default output of most legacy SEO tools, is actively counterproductive in generative engine contexts.

The Easy-to-Understand and Technical Terms strategies showed moderate, domain-dependent performance; neither universally effective nor universally harmful, they are context-sensitive tools rather than default settings. Authoritative Tone produced meaningful gains in Subjective Impression but more modest PAWC improvements, suggesting it influences how prominently an engine credits a source without necessarily increasing the volume of cited text.

These findings validate the logic behind Search Compliance Optimization, the methodology that platforms like KOZEC build around. SCO explicitly rejects algorithmic shortcuts in favor of genuinely useful, well-structured content, and the paper’s controlled benchmark data now provides empirical backing for that stance. For businesses evaluating whether to shift away from legacy approaches, understanding why automated SEO beats traditional agencies becomes a natural next question.

Table 3 Decoded: Domain-Specific Strategy Selection for Industry Verticals

One of the paper’s most consequential findings is that GEO strategy efficacy varies significantly by content domain. A one-size-fits-all GEO checklist is not merely suboptimal; it is empirically incorrect.

For Law and Government and Opinion domains, Statistics Addition delivered the strongest results. For legal and financial services content, that means embedding specific regulatory data, case statistics, and sourced numerical claims, making this a primary content directive for businesses in those verticals. This aligns directly with how KOZEC approaches financial services and insurance firms, where data-dense, citable content is the default standard.

For People and Society, Explanation, and History domains, Quotation Addition performed best. For healthcare and educational content, the directive shifts toward direct expert quotes, attributed client or patient testimonials, and sourced definitional statements. KOZEC’s AI content platform for healthcare marketing applies this logic at scale, prioritizing quotable, attributed content structures over keyword-dense alternatives.

For B2B SaaS content, the directive similarly shifts toward direct expert quotes and sourced definitional statements.

This is why content strategy should differ between, for example, a personal injury law firm (Statistics Addition priority) and a functional medicine practice (Quotation Addition priority). It is the same reason KOZEC applies vertical-specific workflows across the industries it serves rather than a single generic template.

Two important qualifications remain. Because the paper tested on Perplexity.ai rather than Google AI Overviews or ChatGPT, practitioners should treat domain findings as directional signals rather than universal constants while the 2026 evidence base continues to mature. Additionally, the paper found that GEO methods disproportionately benefit websites ranked lower in traditional search, a democratization result especially relevant for growth-stage businesses and challenger brands that lack the domain authority of established competitors.

From KDD 2024 to 2026: What the Follow-On Research Confirms and Complicates

The Aggarwal et al. paper is now cited as the foundational anchor of the entire GEO academic field. It appears across E-GEO (e-commerce), IF-GEO (multi-query optimization), SafeGEO (recommendation agent risks), and a 2026 critical survey covering 45 GEO studies published between 2023 and 2026.

The follow-on work confirms the core framework. Across subsequent research and real-world deployment, citation-oriented, fluency-optimized content has consistently outperformed keyword-dense content in generative engine responses.

The follow-on work also complicates the picture. E-GEO notes that the original paper centers on heuristics rather than systematic optimization and that its Subjective Impression score does not translate directly into commercial value. The 2026 critical survey and SafeGEO both flag heterogeneous terminology and inconsistent evidence standards across the field.

A 2026 citation behavior study surfaced a striking paradox: only about 1% of users who encounter AI summaries click the cited sources, compared with 15% who click traditional search results. Yet generative search visitors convert at roughly 23 times the rate of traditional organic visitors. This reframes what GEO success actually looks like. The goal is not raw traffic volume; it is citation-gated conversion quality. Understanding how to measure SEO content performance in this new environment requires updating the metrics that matter.

A 2026 position paper adds three underexamined risks worth taking seriously: concentrated influence from low contestability, undisclosed commercial influence embedded in AI answers, and academic-industry blind spots created by evaluating offline rather than in deployed systems. These are legitimate considerations for anyone building a long-term GEO strategy.

Finally, Wikipedia’s GEO entry confirms that as of early 2026, no consensus definition distinguishes GEO from adjacent terms such as AEO, LLMO, AIO, and AI SEO. In that still-forming landscape, Aggarwal et al. remains the primary academic anchor.

Translating the Paper Into Execution: How KOZEC’s SCO Methodology Maps to the Aggarwal Findings

The Aggarwal paper proves what works in a controlled benchmark. The missing layer for most businesses is an operational system that executes those findings at scale, and that is the bridge KOZEC’s Search Compliance Optimization methodology is designed to be.

The mapping is direct. SCO’s emphasis on useful content aligns with the Fluency Optimization plus Statistics Addition combination. Its focus on clear page structure aligns with the paper’s finding that well-organized, citable content increases PAWC. Its use of smart internal links aligns with the source-citation strategy that produced recursive credibility signals — a practice KOZEC implements through automated internal linking built directly into its publishing workflow. Its insistence on consistent publishing aligns with the finding that GEO benefits lower-ranked sites, because sustained volume builds the topical authority that amplifies individual GEO tactics. The importance of why content consistency matters for SEO is not merely a best practice; it is a structural requirement the paper’s data supports.

On keyword stuffing, SCO’s rejection of density-based optimization is no longer just a philosophical stance. It is validated by the paper’s controlled benchmark data showing that keyword stuffing performs at or below baseline.

KOZEC’s agentic AI execution operationalizes the combinatorial insight rather than treating GEO as a flat checklist. Its content workflow is built to layer fluency, citation anchors, and domain-appropriate statistics into each piece systematically. Configurable settings and industry-specific workflows apply the right combination of strategies for each vertical, echoing the paper’s domain-specificity finding.

There is also an equity dimension. The paper’s finding that GEO disproportionately benefits lower-ranked websites aligns precisely with KOZEC’s target market of growth-stage businesses. Companies without established domain authority can use GEO-optimized content to compete for generative engine citations on a more level playing field.

The 2026 Execution Checklist: Applying Aggarwal et al. to Your Content Strategy

The following checklist translates the paper’s findings into weekly content decisions.

1. Audit for citation density. Does each piece contain at least two or three directly quotable, attributable statements, whether expert quotes, sourced statistics, or regulatory data? If not, Quotation Addition and Statistics Addition are the highest-leverage edits available.
2. Apply the combination rule. Ensure Fluency Optimization and Statistics Addition appear together, not in isolation. A statistic-dense but poorly written piece, or a fluent but data-free piece, underperforms the combination by more than 5.5%.
3. Match strategy to domain. Use the paper’s Table 3 findings as a vertical guide. Legal and financial content should prioritize Statistics Addition; healthcare and educational content should prioritize Quotation Addition.
4. Eliminate keyword stuffing. This is not a stylistic preference. Any workflow that raises keyword density without adding citable substance is actively reducing GEO visibility.
5. Measure for citation, not just traffic. Given the 1% click-through rate on AI summaries against the 23x conversion advantage for generative search visitors, success metrics should include AI Overview citation tracking and conversion quality, not organic traffic volume alone.
6. Build topically interconnected content. The advantage GEO gives lower-ranked sites is amplified when content lives inside a structured topical ecosystem rather than as isolated pages, consistent with KOZEC’s interconnected content approach.

Conclusion: The Paper Proved It. The Question Is Whether You Execute It.

Aggarwal et al. did not publish a speculative framework. They ran a 10,000-query benchmark and produced empirical lift data. The strategies that outperform keyword stuffing by 30 to 40% are not opinions; they are measured results.

The 2026 evidence base has added nuance, including domain specificity, combination effects, and the citation behavior paradox, but it has not invalidated the foundational findings. The paper remains the primary academic anchor of GEO as a discipline.

That reframes the practitioner’s decision. The question in 2026 is not whether GEO matters; EMARKETER’s 31.3% adoption figure and Gartner’s 25% search-volume decline projection settle that. The question is whether content strategy is being built on the paper’s empirical evidence or on legacy SEO intuitions that the paper’s data has now contradicted.

Understanding the findings is the starting point. Systematically applying them across verticals, at publishing scale, with the right strategy combinations, is where the competitive advantage is built. For growth-stage businesses that cannot win on domain authority alone, the paper’s finding that GEO disproportionately benefits lower-ranked websites is the single most important data point in the study. GEO is not merely a visibility tactic; it is a structural opportunity to compete against established players in the generative engine environment.

Ready to Execute What the Research Proves? See How KOZEC Applies GEO at Scale.

Understanding the Aggarwal et al. paper is the academic foundation. Operationalizing it is where most businesses stall.

To see how KOZEC’s SCO methodology operationalizes these findings for a specific industry vertical, schedule a demo at kozec.ai/schedule-a-demo/. Growth-stage teams that want to discuss how the paper’s domain-specific findings apply to their content strategy can reach KOZEC directly at (888) 545-7090 or through the contact page.

KOZEC delivers 15 to 60 or more content pieces per month at $600 to $1,500 per month, the execution infrastructure that makes applying GEO strategy at the volume and consistency the paper’s findings require economically viable for lean marketing teams. The Aggarwal et al. paper is the proof. KOZEC is the bridge between that proof and the content a business publishes next week.