GEO Generative Engine Optimization Paper: The Citation Lineage

Introduction: Why the GEO Origin Story Needs a Correction

Search for the “GEO generative engine optimization paper” and a familiar story appears. A team of researchers published a study, found that certain content tweaks boost visibility in AI-generated answers by up to 40%, and a new discipline was born. Most coverage ends there, treating the 2023/2024 paper by Aggarwal et al. as a closed origin story.

That framing is incomplete. The original paper was the first data point in an evolving research trajectory, and the work published since then has tested, narrowed, and in places contradicted its headline claims.

This article constructs a citation-accurate lineage from the original GEO-bench findings through the 2025 and 2026 corrective literature: E-GEO, C-SEO Bench, AutoGEO and related systems, SAGEO Arena, and a 2026 critical survey of 45 studies. Along the way, it highlights a finding almost entirely absent from practitioner content: optimizing content to be cited can actively harm its chances of being retrieved in the first place.

This is a source-grounded synthesis, not a marketing recap. It is written for researchers, practitioners, and anyone deciding whether to spend budget on GEO tactics. By the end, readers will be able to separate what the academic record supports from what trade-press mythology has added.

The Founding Paper, Correctly Attributed

The paper is “GEO: Generative Engine Optimization,” arXiv:2311.09735, first submitted on November 16, 2023. It was peer-reviewed and published at KDD 2024, the 30th ACM SIGKDD Conference on Knowledge Discovery and Data Mining, under DOI 10.1145/3637528.3671900.

Competitor content frequently gets the basics wrong. Some articles call it simply “the Princeton study,” which overlooks the multi-institution collaboration behind it. Others misname the authors entirely, citing names such as “Muralidhar and Nagar” that do not appear on the paper.

The accurate author list is:

• Pranjal Aggarwal
• Vishvak Murahari
• Tanmay Rajpurohit
• Ashwin Kalyan
• Karthik Narasimhan
• Ameet Deshpande

Affiliations span Princeton University, IIT Delhi, the Allen Institute for AI, and Georgia Tech.

The paper’s academic standing is real. KDD is among the most prestigious venues in data mining, and the paper had accrued 39 Scopus citations as of late 2025. That standing is exactly why its limits deserve careful attention.

What GEO-Bench Actually Measured

The paper’s core contribution was GEO-bench, a benchmark of roughly 10,000 queries drawn from nine datasets, including MS MARCO, ORCAS-1, and Natural Questions.

The researchers tested nine optimization strategies applied to source content. Three stood out: adding citations, adding quotations from relevant sources, and adding statistics, each producing significant boosts in visibility. Crucially, the paper also reported that efficacy varied by query domain, a detail that rarely survives summarization.

The headline result was that content optimization could boost visibility in generative engine responses by up to 40%. That figure became the most-quoted statistic in the GEO conversation, usually stripped of the conditions under which it was produced.

The Black-Box Framing That Defines the Paper’s Limits

The paper explicitly frames GEO as a black-box optimization problem. Because the internal workings of generative engines are unknown, strategies are inferred purely from observable outputs.

That framing is methodologically honest, and it also sets a hard ceiling on what the paper can claim. It was designed to show that content changes correlate with visibility changes in a controlled setup. It was never designed to establish durable, real-world causal effects across live search ecosystems.

This self-imposed scope boundary explains why later studies found the original claims did not generalize. Practitioner summaries almost never acknowledge it, which is where the corrective literature begins.

The 40% Statistic: What It Proves and What It Doesn’t

The “40% visibility boost” spread because it is simple, large, and attributable to a prestigious venue. It fits neatly into a slide deck or a sales pitch.

The 2026 critical survey, “Optimizing Visibility in Generative Engines” (arXiv:2607.14035), supplies the qualifier that should accompany it. The foundational paper’s gains are valid within its experimental setting, but they are conditional on a source already being present in a fixed context. They do not establish organic discoverability or durable traffic effects.

The gap between lab and field is substantial:

Lab conditions (GEO-bench)	Real-world content ecosystems
Fixed set of candidate sources	Millions of competing pages
Source is guaranteed to be in context	Source must first be retrieved and reranked
One optimized document among unoptimized ones	Many competitors applying the same tactics
Static snapshot	Engines, indexes, and answers change constantly

If every competitor adds statistics and quotations, the relative advantage shrinks toward zero. If a page never makes it into the retrieval set, generation-stage optimization is irrelevant. The 40% number measures something real, but that something is narrow.

The Corrective Literature: Mapping 2025-2026 GEO Research

This section is the core of the lineage. Each study below narrowed, tested, or contradicted the original paper’s claims, and read in sequence, they form a clear research trajectory rather than a list of disconnected papers.

E-GEO: Testing Heuristics in E-Commerce

E-GEO (Bagga et al., MIT/Columbia, 2025) built a testbed for generative engine optimization in e-commerce contexts, where product discovery queries behave differently from the informational queries that dominate GEO-bench.

The key finding was blunt: 10 of 15 common hand-crafted GEO heuristics were neutral or negative in e-commerce settings.

A more promising result pointed in a new direction. A learned, optimization-based approach reliably outperformed fixed heuristics, and the optimized prompts revealed a stable, domain-agnostic pattern. The authors suggested that a “universally effective” GEO strategy may exist, but it is not the static checklist from the original paper.

C-SEO Bench: The Empirical Corrective

C-SEO Bench (Puerto et al., 2025), published under the title “Does Conversational SEO Work?”, is the field’s most important empirical check on generalizability. It was built on approximately 1,900 queries and 16,360 documents.

The result was stark: only 3 of 54 method-domain combinations were significantly positive, and none were positive for question-answering.

The implication is direct. Heuristics presented as broadly effective in the original paper do not generalize across domains or query types. Question-answering, the very use case most people associate with AI search, showed no reliable gains at all.

SAGEO Arena: Exposing the Retrieval Trade-off

SAGEO Arena (Kim et al., 2026) introduced a realistic environment for evaluating search-augmented GEO, modeling the full pipeline instead of assuming a source is already in context.

Its pivotal finding: optimizing content bodies for citation, the classic GEO approach, reduced retrieval presence by about 9%, post-reranking presence by about 16%, and final citation by about 6%. The authors summarize the dynamic plainly: optimizing for ranking alone (SEO) or citation alone (GEO) causes failure at either the retrieval/reranking stage or the generation stage.

This is the clearest academic evidence that GEO rewrites can backfire, hurting the earlier retrieval stage before generation ever happens.

AutoGEO and the Shift to Learned, Query-Adaptive Optimization

A newer wave of systems, including AutoGEO, FeatGEO, IF-GEO, MAGEO, and AgenticGEO, moves the field from fixed heuristics toward learned, query-adaptive optimization. AutoGEO, for example, frames its work around “what generative search engines like and dislike,” learning preferences rather than assuming them.

These approaches generally outperform the original paper’s static strategies in controlled benchmarks. That represents methodological maturation: the discipline is moving away from “checklist GEO” toward adaptive, context-sensitive systems. Practitioner content rarely covers this shift, yet it is arguably the most important direction in the field.

The 2026 Critical Survey: Synthesizing 45 Studies

“Optimizing Visibility in Generative Engines: A Critical Survey of Generative Engine Optimization (2023-2026)” (arXiv:2607.14035) is the most comprehensive meta-analysis to date, reviewing 45 studies.

Its core conclusion: the foundational paper’s widely cited gains are valid within its experimental setting, but they are conditional and do not generalize to organic discoverability.

The survey functions as the capstone of the corrective literature. It is the point at which the field formally reassessed its founding claims, without discarding the discipline those claims launched.

The Retrieval-vs-Generation Split: The Overlooked Core Insight

Generative search runs on a two-stage pipeline:

1. Retrieval and reranking: the system selects which sources enter the model’s context.
2. Generation: the model writes an answer and decides which in-context sources to cite.

Most practitioner content treats “getting cited” as one monolithic goal. In reality, these are separate objectives, and they can conflict.

SAGEO Arena’s data makes the conflict concrete. Citation-optimized content can suppress retrieval presence, which means a page may become less likely to be seen by the system at all, even if it would perform well once seen. Optimizing purely for generation-stage citation, without regard for retrieval, can be self-defeating.

This aligns with Google’s own 2026 framing that GEO and AEO (Answer Engine Optimization) are “still SEO.” A well-structured, retrievable page is the prerequisite for any citation strategy to work.

Google’s Official Position: Narrowing the Gap Between Myth and Method

On May 15, 2026, Google published “Optimizing your website for generative AI features on Google Search,” its first official consolidated statement on the topic. On June 5, 2026, Google clarified that from its perspective, optimizing for generative AI search is optimizing for the search experience, and thus “still SEO,” not a separate discipline.

The guide explicitly states that sites do not need the following to appear in generative AI search:

• llms.txt files
• Content chunking
• Special schema markup
• AI-specific rewriting

Cross-referencing this guidance against the academic record reveals a striking pattern. Popular practitioner tactics are unsupported by both sources. C-SEO Bench found fixed heuristics rarely work; SAGEO Arena found citation-focused rewriting can hurt retrieval; Google says AI-specific rewriting is unnecessary. The mythology sits in the gap between peer-reviewed research and the platform’s own statements, a gap that few published resources address.

Real-World Volatility: Why Lab Results Don’t Survive Contact With Production

Production data reinforces the corrective literature. A 2026 audit of 34,234 AI responses found a 46x difference in brand citation rates between ChatGPT (0.59%) and Perplexity (13.05%), with only 11% domain overlap in citation sources between the two platforms. A tactic that “works” on one engine may be irrelevant on another.

Instability compounds the problem. AirOps 2026 data shows only about 30% of brands remain visible in back-to-back AI responses for the same query, and AI Overview content changes roughly 70% of the time for the same query. When answers update, nearly half the citations are replaced.

The survey literature emphasizes the methodological consequence: AI citation measurement itself is volatile and inconsistent from run to run. Any single-study “win” should be read with that variance in mind.

The stakes, however, remain real. Brands cited in AI Overviews see a reported 23% lift in branded search volume over the following 30 days, and third-party sources are cited roughly 6.5x more often than brand-owned domains. Uncertainty about tactics does not reduce the value of visibility; it raises the cost of choosing the wrong tactics.

The Complete GEO Citation Lineage at a Glance

Year	Work	Contribution or Corrective
2023/2024	Aggarwal et al., GEO (KDD 2024)	Defined GEO as a black-box problem and showed up to 40% visibility gains in a fixed-context benchmark.
2025	E-GEO (Bagga et al.)	Found 10 of 15 hand-crafted heuristics neutral or negative in e-commerce; learned optimization did better.
2025	C-SEO Bench (Puerto et al.)	Found only 3 of 54 method-domain combinations significantly positive, none for question-answering.
2025	AutoGEO and related systems	Shifted the field from static checklists to learned, query-adaptive optimization.
2026	SAGEO Arena (Kim et al.)	Showed citation-focused rewriting can reduce retrieval, reranking, and final citation presence.
2026	Critical survey of 45 studies (arXiv:2607.14035)	Concluded the founding gains are conditional and do not establish organic discoverability.
2026	Google’s official guidance	Declared GEO and AEO “still SEO” and dismissed llms.txt, chunking, special schema, and AI-specific rewriting.

Researchers and practitioners can use this table as a direct reference when evaluating any GEO claim.

What the Lineage Means for Practitioners

The evidence points to several clear conclusions:

• Fixed heuristics are not reliable levers. Special schema, llms.txt files, and aggressive chunking are supported by neither the corrective literature nor Google’s guidance.
• Retrieval comes first. Durable visibility depends on useful content, clear structure, and internal linking before any generation-stage tactic can matter.
• Compliance beats shortcuts. Structuring content for genuine usefulness and discoverability aligns with both peer-reviewed research and Google’s stated position.
• Citation-bait carries risk. Teams investing solely in citation-optimized rewrites may see short-term generation gains offset by long-term retrieval losses.

KOZEC’s SCO (Search Compliance Optimization) framework is one example of an approach built around this evidence. It centers on Google-recommended best practices (useful content, clear pages, smart internal links, consistent publishing) rather than algorithmic shortcuts, which places it on the side of the retrieval-vs-generation split that the research supports.

Conclusion: Treating GEO as a Maturing Field, Not a Settled Fact

The GEO generative engine optimization paper was a legitimate and influential starting point. It was never a finished rulebook, and its black-box framing always limited the causal claims it could support.

The 2025 and 2026 corrective literature, from E-GEO and C-SEO Bench through SAGEO Arena, the critical survey, and Google’s own guidance, has substantially narrowed those founding claims without discrediting the underlying discipline. The field is maturing from static checklists toward adaptive, evidence-tested methods.

The central insight bears repeating: “getting cited” is not one goal but two. Retrieval and generation are separate stages, and optimizing for one can undermine the other. For anyone evaluating GEO claims or strategies in 2026, that distinction is the starting point for every sound decision.

Build Content on Evidence, Not Hype: See How KOZEC Applies It

KOZEC is an agentic AI content platform built around Search Compliance Optimization rather than unproven GEO shortcuts. Its workflow addresses both stages of the pipeline the research describes:

• Structured, interlinked content ecosystems that support retrieval and topical authority
• Metadata and structured data optimization for clear, machine-readable pages
• Consistent automated publishing to WordPress and major CMS platforms
• Performance tracking to measure what actually moves visibility over time

KOZEC reports results from client businesses including +386% AI Overview citation growth and +621% keyword visibility increase, illustrating what a compliance-first, research-informed content strategy can produce.

Teams that want a content system built for both traditional search and generative engine visibility can book a demo at kozec.ai/schedule-a-demo/.