Skip to content
LLMO Framework

GEO Paper: What the Science Says

The GEO (Generative Engine Optimization) paper is the first academic framework for optimizing content visibility in AI-powered search engines. Published at KDD 2024 (ACM SIGKDD), it provides empirical evidence for content optimization strategies that the LLMO Framework builds upon.

Paper Details

Section titled “Paper Details”
FieldValue
TitleGEO: Generative Engine Optimization
AuthorsAggarwal, Murahari, Rajpurohit, Kalyan, Narasimhan, Deshpande
InstitutionPrinceton University, IIT Delhi, Adobe Research
ConferenceKDD 2024 (ACM SIGKDD)
arXiv2311.09735
Published2024

Research Setup

Section titled “Research Setup”

The researchers built GEO-Bench, a benchmark of 10,000 search queries across multiple domains. They tested 9 content optimization strategies against a generative search engine to measure which approaches improved source visibility.

The 9 Strategies Tested

Section titled “The 9 Strategies Tested”
  1. Cite Sources
  2. Quotation Addition
  3. Statistics Addition
  4. Fluency Optimization
  5. Unique Words
  6. Technical Terms
  7. Authoritative Tone
  8. Easy-to-Understand Language
  9. Keyword Stuffing

Key Findings

Section titled “Key Findings”

Strategy Effectiveness

Section titled “Strategy Effectiveness”
StrategyVisibility ImprovementLLMO Component
Statistics Addition+115.1%Citation Signals
Cite Sources+77.0%Citation Signals
Quotation Addition+72.2%Authority Signals
Authoritative Tone+21.5%Knowledge Clarity
Fluency Optimization+15.2%Knowledge Clarity
Technical Terms+5.8%Knowledge Clarity
Easy-to-Understand+2.4%Knowledge Clarity
Unique Words-3.1%—
Keyword Stuffing-10.2%—

The Top Three

Section titled “The Top Three”

The three most effective strategies share a common trait: they provide verifiable, external evidence.

  1. Statistics Addition (+115.1%): Adding specific numbers and data points made content more than twice as visible. Example: “Revenue grew 34% YoY” vs “Revenue grew significantly.”

  2. Cite Sources (+77.0%): Referencing specific papers, reports, or documentation increased visibility by 77%. AI systems prefer content they can cross-reference.

  3. Quotation Addition (+72.2%): Including direct quotes from experts or authoritative sources added credibility that AI systems recognized and cited.

What Doesn’t Work

Section titled “What Doesn’t Work”
  • Keyword Stuffing (-10.2%): Traditional SEO tactics actively hurt AI visibility. AI systems can detect and penalize artificial keyword density.
  • Unique Words (-3.1%): Using unusual vocabulary did not improve visibility. Clarity beats cleverness.

Implications for LLMO

Section titled “Implications for LLMO”

1. Citation Signals are the highest-leverage component

Section titled “1. Citation Signals are the highest-leverage component”

The GEO data shows that Citation Signals (statistics, sources, quotes) account for the largest visibility improvements. This is why the LLMO Framework places Citation Signals as Component 5 — the capstone that multiplies the effect of all other components.

2. Content clarity matters, but less than evidence

Section titled “2. Content clarity matters, but less than evidence”

Strategies related to Knowledge Clarity (authoritative tone, fluency, easy language) all showed positive but modest improvements (2–22%). Good writing is necessary but not sufficient. The multiplier comes from adding verifiable facts.

3. SEO tactics are counterproductive for AI

Section titled “3. SEO tactics are counterproductive for AI”

Keyword stuffing, the cornerstone of early SEO, actively reduced AI visibility. This confirms that LLMO requires a fundamentally different approach from traditional SEO.

Domain-Specific Variations

Section titled “Domain-Specific Variations”

The GEO paper found that strategy effectiveness varies by domain:

  • Factual/scientific queries: Statistics Addition was most effective
  • Opinion/subjective queries: Quotation Addition performed best
  • Technical queries: Cite Sources had the highest impact

This suggests that LLMO implementation should be tailored to your content domain. A research site benefits most from statistics, while a thought leadership blog benefits more from expert quotations.

How LLMO Builds on GEO

Section titled “How LLMO Builds on GEO”

The LLMO Framework extends GEO in three ways:

  1. Broader scope: GEO focuses on generative search engines. LLMO covers all LLM interactions including direct queries, RAG, and AI agents.
  2. Implementation focus: GEO identifies what works. LLMO provides how to implement it with specific file formats (llms.txt), structured data (JSON-LD), and content design patterns.
  3. Retrieval layer: GEO assumes content is already retrieved. LLMO adds the Retrieval Signals component to ensure content is discoverable in the first place.

Further Reading

Section titled “Further Reading”