GEO: Generative Engine Optimization by Pranjal Aggarwal — The Authoritative Paper Breakdown and Business Application Guide for 2026

Introduction: Why One Academic Paper Redefined How Content Gets Found in 2026

Something fundamental broke in the way content gets discovered, and most marketers are still using tools built for the old world. By 2026, AI-referred web traffic grew 527% year over year. At the same time, the overlap between Google’s top-10 organic results and the sources cited in AI-generated answers collapsed from roughly 75% to somewhere between 17% and 38%. In plain terms: ranking well on Google no longer guarantees that AI engines will cite the content. Traditional SEO, on its own, has become insufficient.

The problem is that “GEO” has become a vague marketing buzzword, thrown around in strategy decks without any connection to its empirical origin. That origin is not a blog post or a vendor whitepaper. It is a specific peer-reviewed paper with exact data, novel metrics, and reproducible findings: “GEO: Generative Engine Optimization” by Pranjal Aggarwal and colleagues, first uploaded to arXiv on November 16, 2023 (arXiv:2311.09735) and published at KDD 2024. That paper formally coined the term.

This article delivers two things. First, a rigorous breakdown of what the paper actually found, including the exact percentage lifts per strategy and the two novel visibility metrics it introduced. Second, a direct translation of those findings into content strategy decisions a business can act on. The stakes are concrete: AI search converts at roughly 4 to 5 times the rate of traditional organic, and lower-ranked websites benefit disproportionately from GEO tactics. That last point makes this a democratizing opportunity, not just a technical curiosity.

The Origin Story: Who Wrote the GEO Paper and Why It Matters

The paper was authored by Pranjal Aggarwal (lead author, IIT Delhi), Vishvak Murahari, Tanmay Rajpurohit, Ashwin Kalyan, Karthik R. Narasimhan, and Ameet Deshpande. It is a multi-institutional collaboration spanning Princeton University, IIT Delhi, Georgia Tech, and the Allen Institute for AI.

This is worth clarifying because the paper is frequently called the “Princeton GEO paper,” owing to its Princeton-affiliated co-authors. That attribution is not wrong, but it is incomplete: lead author Pranjal Aggarwal’s primary affiliation is IIT Delhi. The work is genuinely cross-institutional.

The publication venue matters enormously for credibility. The paper appeared at KDD 2024, the 30th ACM SIGKDD Conference on Knowledge Discovery and Data Mining, held in Barcelona from August 25 to 29, 2024, on pages 5 through 16, with DOI 10.1145/3637528.3671900. KDD is one of the most prestigious venues in data science and machine learning. Acceptance there means the research survived rigorous peer review.

The impact has been substantial: 35 citations on the ACM Digital Library and more than 16,392 downloads as of mid-2026. It has spawned an entire subfield of follow-up research, including AutoGEO (accepted at ICLR 2026), IF-GEO, FeatGEO, E-GEO, and a 45-study critical survey covering 2023 through 2026. For practitioners, the takeaway is simple: this is not a vendor’s marketing claim. It is peer-reviewed empirical research, which makes it a reliable foundation for business decisions.

The Three-Stakeholder Problem: The Core Theoretical Contribution Most Guides Miss

The paper defines Generative Engines (GEs) as AI-powered search systems that synthesize information from multiple sources and summarize it using large language models. These systems are rapidly replacing traditional search engines like Google and Bing.

The paper’s core theoretical contribution is its three-stakeholder framework. First, there are users, who benefit from synthesized, conversational answers. Second, there are the search engines and generative engine providers, who benefit from increased utility and engagement. Third, and critically, there are content creators and website owners: the “third stakeholder” who faces an existential challenge.

The third-stakeholder problem can be stated plainly. When a generative engine answers a query, it may draw from a website’s content without ever sending the user to that site. The traffic and visibility that content creators depend on simply evaporate. Traditional SEO offered a clear mechanism: rank higher, get more clicks. Generative engines break that mechanism entirely, because in the new paradigm the answer itself is the destination.

As the paper notes, content creators have “little to no control over when and how their content is displayed” in generative engine responses. GEO is the paper’s proposed solution to restore some of that control.

One honest caveat is worth surfacing. A 2026 position paper warns that GEO also creates risks of information concentration and manipulation of AI-generated answers. Acknowledging this builds credibility rather than undermining it: GEO is a tool, and like any optimization discipline, it carries governance concerns.

GEO Defined: The Paper’s Formal Definition and What It Actually Means

The paper introduces GEO as the first novel paradigm to help content creators improve content visibility in generative engine responses, through a flexible black-box optimization framework.

The phrase “black-box optimization” deserves unpacking. Content creators cannot access the internal workings of a generative engine: not the model weights, not the retrieval logic, not the ranking signals. GEO strategies therefore have to work by modifying content in ways that demonstrably influence outputs, without the creator knowing the exact internal mechanism.

At the paradigm level, this distinguishes GEO sharply from SEO. SEO optimizes for rank position in a list of links. GEO optimizes for “share of cited text” inside an AI-generated answer. Those are fundamentally different success metrics.

It is also worth addressing the terminology confusion that much competitor content perpetuates. “GEO” as a coined term originates specifically from this paper. Related terms like AEO (Answer Engine Optimization) and LLMO (Large Language Model Optimization) are adjacent concepts, but as of early 2026, no consensus definition distinguishing these terms had been established in academic literature. The business implication of the black-box framework is empowering: content teams can define and optimize custom visibility metrics relevant to their specific goals, giving them more strategic control than traditional SEO’s one-size-fits-all rank metric ever offered.

GEO-BENCH: The Research Infrastructure Behind the Findings

Rigorous findings require rigorous infrastructure, and the paper delivered it in the form of GEO-BENCH: a large-scale benchmark of 10,000 diverse queries drawn from nine different sources and the first benchmark specifically designed for evaluating GEO strategies.

The dataset is structured for reproducibility. Queries are categorized by domain, difficulty level, query intent, and other dimensions. The public release on HuggingFace (GEO-Optim/geo-bench) includes 8,000 training queries and 1,000 test queries. Drawing from nine sources, both real-world and synthetically generated, ensures that findings are not artifacts of a single query distribution.

For testing, the paper used a simulated generative engine: GPT-3.5-turbo applied to the top-5 Google search results for each query. This is a methodologically transparent choice, and it introduces a key limitation the paper is honest about. The testbed used a simulated engine, not a live production system. Real-world engines like Perplexity, Google AI Overviews, and ChatGPT may behave differently. To address this, the authors validated findings on Perplexity.ai as a real-world check.

GEO-BENCH has since become the foundation for the entire follow-up field. AutoGEO, IF-GEO, E-GEO (which added 7,000-plus e-commerce queries), and the 45-study critical survey all build directly on this benchmark.

The Two Novel Visibility Metrics: What the Paper Actually Measures

New metrics were necessary because traditional SEO measurements are meaningless in a generative context. Rank position, click-through rate, and impressions assume a ranked list of links to click. In an AI-generated answer, there is no such list.

Position-Adjusted Word Count (PAWC) is the first metric. It is a normalized count of the words in sentences within the AI-generated response that cite or draw from a specific source, with exponential decay applied by position. Words cited earlier in the response count for more than words cited later. The causal logic is elegant: PAWC captures both the quantity of a source’s contribution and the prominence of that contribution, reflecting how users actually consume AI answers, with front-loaded attention.

Subjective Impression is the second metric. It is a seven-sub-metric, LLM-based score covering relevance, influence, uniqueness, diversity, citation likelihood, and related dimensions. It is designed to capture the qualitative aspects of visibility that a pure word count cannot.

Two metrics matter because a source can contribute many words of low quality, or few words of high influence. PAWC captures quantitative presence; Subjective Impression captures qualitative influence. The business implication is to track both dimensions: the volume of AI citation and the quality of that citation. The 2026 critical survey found that terminology and metrics across the 45 GEO studies reviewed remain heterogeneous, which makes the paper’s two original metrics the most consistently validated baseline available.

The Nine GEO Strategies: Causal Mechanics, Not Just a List

Most practitioner guides list the nine strategies without explaining why each works mechanically. This section fills that gap. All percentage improvements are measured against unmodified content, providing a clear reference point.

Strategy 1: Quotation Addition — The Highest Performer at +41% PAWC

Quotation Addition delivered +41% on Position-Adjusted Word Count and +28% on Subjective Impression, making it the single highest performer across both metrics.

The mechanism: generative engines are trained to synthesize and cite credible sources. Content that already contains attributed quotations from authoritative figures signals to the engine that the content is itself a credible synthesizer, making it more likely to be cited. A secondary mechanism is that quotations provide specific, extractable text the engine can incorporate directly, reducing its need to paraphrase.

Quotation Addition excels particularly in “People & Society,” “Explanation,” and “History” domains. The business application: identify two to four authoritative, domain-relevant quotations per piece and attribute them explicitly with name, title, and source. On Perplexity.ai, Quotation Addition showed a +22% PAWC improvement, confirming the finding beyond the simulated testbed.

Strategy 2: Statistics Addition — Data That Engines Can Extract

Statistics Addition ranked among the top performers, working in tandem with citations (Cite Sources produced up to 34.4% improvement on factual queries). The mechanism: generative engines prioritize extractable, verifiable information, and specific numerical data points serve as precise anchor facts. Statistics also signal research depth.

Statistics Addition is most effective in “Law & Government” and “Opinion” domains, where quantitative evidence carries particular authoritative weight. The business application: replace vague claims like “many businesses report improvement” with sourced specifics like “X% of businesses in a Y study reported Z outcome.”

Strategy 3: Cite Sources — Building the Citation Chain

Cite Sources produced up to 34.4% improvement on factual queries, making it one of the most reliable strategies for information-dense content. The mechanism: content that explicitly cites its own sources mirrors the synthesis behavior the engine is performing, creating a “citation alignment” that increases the engine’s confidence. Content that cites Nature, PubMed, or government databases becomes part of a traceable credibility network. The business application: add explicit inline citations, not just a references section, to key factual claims.

Strategy 4: Authoritative Tone — Signaling Expertise Through Language

Authoritative Tone means rewriting content with confident, declarative language rather than hedged phrasing. The mechanism: generative engines are trained on human feedback that rewards confident, clear answers, so content in an authoritative register resembles the output style the engine is trying to produce. This is about voice, not keyword density. The business application: audit for hedging language (“it might be,” “some experts suggest”) and replace it with confident, evidence-backed assertions where evidence supports them.

Strategy 5: Easy-to-Understand — Clarity as a Visibility Signal

Stylistic changes including Easy-to-Understand yielded a 15% to 30% visibility boost. The mechanism: engines must extract and re-present information, and clearly structured content requires less transformation. The business application: use shorter sentences, define technical terms on first use, prefer active voice, and structure paragraphs around single ideas. Crucially, this does not mean shallow content. Clarity of presentation and depth of content are not in tension; the goal is accessible expertise.

Strategy 6: Fluency Optimization — The Combination Multiplier

Fluency Optimization improves grammatical correctness, coherence, and natural flow. The mechanism: poorly written content introduces noise into the extraction process, reducing accurate citation. The standout finding is that Fluency Optimization combined with Statistics Addition outperformed any single strategy by more than 5.5%, the most powerful combination in the paper. Fluency makes statistics easier to extract accurately; statistics give the engine high-value data to cite. Treat fluency as a baseline requirement, and note the practical implication: AI-generated content often has fluency issues, so human editing for fluency is a high-ROI intervention.

Strategy 7: Unique Words — Vocabulary Differentiation

Unique Words means incorporating domain-specific vocabulary that distinguishes content from generic treatments. The mechanism: engines retrieve partly based on semantic relevance, and precise domain vocabulary matches the engine’s own language patterns. This is the opposite of keyword stuffing; it is about semantic richness, not repetition. The business application: analyze authoritative sources in the domain and use the same precise terminology rather than synonyms.

Strategy 8: Technical Terms — Depth Signals for Specialized Queries

Technical Terms means incorporating field-specific terminology that signals deep expertise. The mechanism: for technical queries, engines must satisfy expert users, and accurate technical terminology reduces the risk of an oversimplified answer. This is most effective in specialized domains like medicine, law, engineering, and finance, and less useful for general consumer queries. The business application: identify the 10 to 15 most important technical terms in the domain and use them accurately, with brief contextual definitions. Technical Terms and Easy-to-Understand appear contradictory but are complementary: use the term accurately, then explain it clearly.

Strategy 9: Keyword Stuffing — The Strategy That Actively Hurts Visibility

The most important negative finding in the paper is this: Keyword Stuffing, the cornerstone of traditional SEO, scored 8% to 10% below the unmodified baseline. It was the only strategy that actively harmed visibility.

The mechanism of failure: generative engines use language models sensitive to unnatural repetition, and keyword-stuffed content reads as low-quality, reducing both retrieval probability and citation likelihood. In traditional search, keyword density was a positive signal because crawlers counted term frequency. In generative search, the “crawler” is a language model that penalizes unnatural language. The same tactic that once helped now actively hurts. Existing content optimized with high keyword density may need revision, both for GEO performance and because that content is likely being evaluated by AI systems for AI Overview inclusion. This finding directly validates KOZEC’s SCO (Search Compliance Optimization) framework, which prioritizes genuine best practices over algorithmic tricks.

The Domain-Specificity Finding: The Insight Most Practitioner Guides Completely Miss

The paper demonstrates that the same GEO strategy performs very differently across content domains. This fundamentally changes how content teams should prioritize their efforts.

The specific mappings: Statistics Addition is most effective in “Law & Government” and “Opinion” domains. Quotation Addition excels in “People & Society,” “Explanation,” and “History” domains. The causal logic: different domains have different conventions for credible evidence. Legal content is validated by precedent and statistics; historical content is validated by primary source quotations; opinion content is validated by expert attribution. Generative engines learn these conventions from their training data and preferentially cite content that follows them.

The business application is twofold. First, identify the primary domain of each content piece before selecting strategies; do not apply a uniform approach. Second, map the content portfolio to the paper’s domain categories and assign domain-appropriate primary strategies, with universal strategies (Fluency Optimization, Cite Sources) as baseline requirements everywhere.

A 2025 follow-up, E-GEO, extended GEO-BENCH to e-commerce with 7,000-plus consumer product queries and found a “universally effective” strategy pattern through prompt optimization, suggesting e-commerce may have different dynamics than the original domains. The 2026 critical survey confirmed that domain effects are real and persistent across follow-up research, not an artifact of the original methodology.

Combination Strategies: When Stacking Methods Outperforms Any Single Approach

The key combination finding: Fluency Optimization plus Statistics Addition outperformed any single strategy by more than 5.5%. Combinations work because each strategy addresses a different dimension of evaluation. Fluency reduces extraction friction; Statistics increases the value of what is extracted.

Quotation Addition and Cite Sources both operate through the attribution mechanism and can be layered to create a multi-layered credibility signal. Over-stacking, however, is a mistake. Applying all nine strategies at once is not the goal; some interact awkwardly (Technical Terms and Easy-to-Understand require careful balance).

A practical framework proceeds as follows: (1) Fluency Optimization as a universal baseline; (2) a domain-appropriate primary strategy, either Quotation Addition or Statistics Addition; (3) Cite Sources as a secondary universal layer; (4) Authoritative Tone as a voice-level enhancement. The ICLR 2026 AutoGEO framework automates the identification of optimal combinations for specific content and queries, building directly on these findings.

The Democratizing Effect: Why Lower-Ranked Websites Benefit More from GEO

One of the most motivating and most overlooked findings is that lower-ranked websites benefit disproportionately more from GEO tactics than already-authoritative sites.

The mechanism: in traditional SEO, domain authority compounds. High-authority sites rank higher, earn more links, and gain more authority. In generative search, the engine evaluates content quality at the document level, not the domain level. A small website with a highly optimized, quotation-rich, statistics-dense article can achieve significant PAWC even with low overall domain authority, because the engine evaluates the specific document.

For SMBs and challenger brands, this is a genuine opportunity to compete with established players in AI-generated responses without the backlink profiles traditional SEO demands. The value is amplified by conversion: AI-sourced traffic converts at 4 to 5 times the rate of traditional organic. This finding directly supports systematic GEO implementation for growth-stage businesses with lean marketing teams, where ROI potential is highest precisely because they are currently under-represented in AI answers. For a deeper look at how content marketing ROI for small businesses stacks up in this new environment, the dynamics of GEO’s democratizing effect are particularly relevant.

Real-World Validation: The Perplexity.ai Findings and What They Confirm

Because the primary testbed was simulated, real-world validation matters. On Perplexity.ai, a live deployed generative engine, Quotation Addition showed a +22% PAWC improvement and Statistics Addition showed a +37% Subjective Impression improvement.

The validation confirms the direction of the findings: Quotation Addition and Statistics Addition remain top performers. The magnitude differs, which suggests real-world engines carry additional signals that moderate the effects. The 2026 critical survey confirmed that the paper’s three key findings held across follow-up research: keyword stuffing fails, extractable information (figures, definitions, quotations) facilitates citation, and domain effects are real.

The honest caveat for practitioners: GEO is an empirical discipline in active development. The paper’s findings are the best available evidence but should be treated as strong directional guidance, not precise engineering specifications. The July 2026 critical survey reviewed 45 studies published between November 2023 and July 2026. The field moves fast, and practitioners should monitor it.

From Paper to Practice: A GEO Implementation Framework for Content Teams

The following steps translate the research into a repeatable workflow.

• Step 1: Audit existing content against the nine strategies. Identify which are already present (often Cite Sources, Authoritative Tone) and which are absent (often Quotation Addition, Statistics Addition).
• Step 2: Classify by domain. Categorize each piece using the paper’s taxonomy (Law & Government, Opinion, People & Society, Explanation, History) and assign domain-appropriate primary strategies.
• Step 3: Prioritize by impact. Apply Quotation Addition first for applicable domains, Statistics Addition second, with Cite Sources and Fluency Optimization as universal baselines.
• Step 4: Implement the winning combination. Apply Fluency Optimization plus Statistics Addition, adding Quotation Addition for applicable domains to form a three-layer optimization.
• Step 5: Eliminate Keyword Stuffing. Audit for keyword-dense passages and revise them. This is risk mitigation, not just improvement, since stuffing harms visibility by roughly 10%.
• Step 6: Measure with appropriate metrics. Track share of AI citations (via tools monitoring AI Overview appearances, Perplexity citations, and ChatGPT mentions) rather than rank position alone.
• Step 7: Iterate by domain. Monitor performance differences across domains and refine assignments based on observed results.

The consistent bottleneck is execution at scale. Platforms like KOZEC that build GEO optimization into the content production workflow can systematically apply these strategies across a large portfolio, closing the gap between knowing the strategies and executing them every time.

The Evolving GEO Research Landscape: What Has Been Built on Aggarwal et al.

The paper is not a historical artifact; it is the active foundation of a growing field, with GEO-BENCH as the shared benchmark.

• AutoGEO (ICLR 2026): automatically extracts content preference rules from generative engines and rewrites documents to maximize visibility, moving from manual application to automated optimization.
• IF-GEO: extends GEO to multi-query scenarios where one piece must be optimized for multiple related queries, benchmarking against the original nine heuristics.
• E-GEO: a domain-specific extension for e-commerce with 7,000-plus product queries, finding a “universally effective” pattern. Businesses looking to scale automated blog content for ecommerce stores will find E-GEO’s findings particularly applicable to their content strategy.
• The 2026 critical survey (arXiv:2607.14035): reviewed 45 studies from November 2023 to July 2026, confirming rapid expansion while noting that terminology, metrics, and evidence standards remain heterogeneous.
• The 2026 ethics paper (arXiv:2606.12439): raises governance concerns including information concentration, manipulation risks, and the gap between academic metrics and commercial outcomes.

The consensus around the paper’s core findings is strengthening. Practitioners should watch AutoGEO’s automated approach in particular, as it may significantly change implementation economics.

Conclusion: GEO Is an Empirical Discipline, Not a Marketing Trend

GEO is not a buzzword or a rebranding of SEO. It is a formally defined, empirically validated discipline that originated in a specific peer-reviewed paper with specific, reproducible findings.

Three takeaways are immediately actionable. First, Quotation Addition is the highest-performing single strategy at +41% PAWC, so implement it first. Second, Keyword Stuffing actively harms visibility by 8% to 10%, so eliminate it from GEO-era content. Third, domain-specificity is real, so match strategies to content domains rather than applying uniform optimization.

The limitations are honest ones: the testbed was simulated, real-world engines may behave differently, and the field evolves rapidly. Yet the directional findings held up on Perplexity.ai and across 45 follow-up studies. Given that AI-sourced traffic converts at 4 to 5 times the rate of traditional organic, that lower-ranked sites benefit disproportionately, and that the overlap between traditional rankings and AI citations has collapsed, GEO implementation is a high-priority strategic decision, not an optional experiment.

For growth-stage businesses and challenger brands, GEO offers a genuine chance to compete in AI-generated responses on the basis of content quality rather than domain authority. Understanding how to increase website traffic with content in this new paradigm means embracing GEO as the core discipline. The strategies are clear. Implementation discipline is the differentiator.

Ready to Build GEO Into Your Content System, Not Just Your Strategy Deck?

The Aggarwal et al. paper provides the validated strategic foundation. The next step is building a content production system that executes these strategies consistently, at scale.

This is where most content teams stall. They understand GEO intellectually but lack the workflow infrastructure to apply Quotation Addition, Statistics Addition, Fluency Optimization, and domain-appropriate tactics to every piece, every time. KOZEC closes that gap. Its agentic AI platform builds GEO optimization, including the paper’s validated strategies, directly into an automated content strategy production workflow, ensuring every piece is structured for AI citation rather than only traditional rankings.

This approach aligns with the paper’s core insight that GEO performance is measured by “share of cited text” inside AI-generated answers. KOZEC’s content architecture is designed to maximize exactly that metric, and KOZEC clients are reportedly seeing +386% AI Overview citation growth, consistent with the paper’s finding that systematic GEO implementation produces measurable, compounding visibility gains.

To see how KOZEC implements these validated GEO strategies at scale, with setup in days rather than months, schedule a demo at kozec.ai/schedule-a-demo/. For those not yet ready for a demo, explore the platform at kozec.ai or reach the team at (888) 545-7090.