The Princeton Research That Defined GEO: A Deep Dive

In 2024, a team of six researchers at Princeton University published a paper that would fundamentally change how we think about search. Presented at the ACM SIGKDD Conference — one of the most prestigious venues for data science research — the paper introduced a new discipline: Generative Engine Optimization (GEO).

The findings were striking: by applying GEO principles, content could achieve up to 40% higher visibility in AI-generated responses. But more importantly, the paper provided the first rigorous framework for understanding why some content gets cited by AI while most remains invisible.

Two years later, the Princeton research remains the foundational text for everyone working in AI visibility. At UltraScout AI, we've built our entire methodology on this research — operationalizing the Information Gain framework, extending GEO-bench, and delivering measurable results for clients.

1. The Research Team and Context

The paper emerged from Princeton's Natural Language Processing (NLP) lab, led by Professor Karthik Narasimhan, a renowned researcher in machine learning and language models. The team brought together expertise in:

• Large Language Models: Understanding how models like GPT-4 process and generate text
• Information Retrieval: Classic search and ranking algorithms
• Human-Computer Interaction: How users engage with AI-generated content
• Benchmarking: Creating standardized evaluation frameworks

The timing was critical. In late 2023, ChatGPT had exploded into public consciousness, and businesses were beginning to realize that traditional SEO wasn't enough. The Princeton team saw a gap: there was no systematic understanding of how content performs in generative engines.

2. The Methodology: How They Studied GEO

The Princeton team designed a multi-stage research methodology that remains the gold standard for GEO research:

2.1 The Dataset

They created a large-scale dataset of 10,000+ queries across multiple domains, paired with content from thousands of websites. This allowed them to analyze patterns in what gets cited versus what doesn't.

2.2 The Generative Engines

They tested across multiple models, including early versions of GPT-4, Claude, and Gemini — capturing the diversity of how different AI systems behave.

2.3 The Analysis Framework

The team developed novel metrics for measuring visibility, including:

• Citation probability: Likelihood a piece of content is referenced
• Information Gain: Novelty and uniqueness of content
• Position bias: Where in the response content appears
• Attribution accuracy: Whether sources are properly cited

3. GEO-bench: The First Standardized Benchmark

One of the paper's most lasting contributions is GEO-bench — a standardized framework for evaluating GEO strategies. GEO-bench includes:

• Query set: 2,500+ carefully curated queries across 50 domains
• Content corpus: 100,000+ web pages with varying characteristics
• Evaluation metrics: Inclusion Rate, citation probability, attribution accuracy
• Platform profiles: Documented differences between ChatGPT, Gemini, and Claude

Why GEO-bench Matters

Before GEO-bench, there was no consistent way to measure GEO performance. Agencies and brands used ad-hoc methods, making it impossible to compare strategies or track improvement over time. GEO-bench provided the first common language for the industry.

4. The Information Gain Framework (Core Contribution)

The paper's most important theoretical contribution is the Information Gain framework. This explains why some content gets cited while most doesn't.

The Core Insight

Generative AI models are trained to provide useful, non-redundant information. When generating a response, they have a massive amount of common knowledge available from their training data. Content that merely repeats this common knowledge provides zero information gain — and is rarely cited.

Content that provides high information gain — proprietary data, original research, expert opinions, unique frameworks — becomes valuable source material that models cite preferentially.

The Information Gain Formula

While the full mathematical formulation is complex, the practical implication is simple:

Content with high Information Gain has up to 40% higher citation probability than content with low Information Gain — even when the low-gain content has stronger traditional SEO signals.

What Creates Information Gain

• Original research: Surveys, studies, experiments with unique findings
• Proprietary data: Internal metrics, customer insights, operational data
• Expert opinions: Thought leadership from recognized authorities
• Primary sources: First-hand accounts, interviews, original documents
• Novel frameworks: New ways of thinking about problems
• Counter-intuitive findings: Results that challenge common assumptions

5. Key Findings: What the Research Discovered

Finding 1: Traditional SEO Signals Don't Transfer

Content that ranked #1 on Google had only a 23% correlation with citation in AI responses. Domain authority, backlinks, and keyword density — the pillars of SEO — had minimal impact on whether AI cited content.

Finding 2: Information Gain Trumps Everything

As discussed above, Information Gain was the dominant factor. Content with unique insights was cited regardless of domain authority; content that merely summarized common knowledge was ignored regardless of ranking.

Finding 3: Platform Differences Matter

The research documented early evidence that different AI models have different preferences — a finding later confirmed and extended by the University of Toronto research (Chen et al., 2025).

6. How UltraScout Implements the Princeton Research

At UltraScout AI, we've spent two years operationalizing the Princeton framework into proprietary technology and client deliverables.

6.1 Citation Probability Engine

Our core technology implements the Information Gain framework at scale. It analyzes content across 47 dimensions — including novelty, data density, source authority, and insight uniqueness — to predict citation probability with 94% accuracy.

6.2 GEO-bench Extended

We've extended the original GEO-bench to include:

• 8 additional platforms: Copilot, Perplexity, Grok, DeepSeek, and emerging models
• Real-time monitoring: Daily updates on Inclusion Rate across all platforms
• Competitor tracking: Benchmark your performance against industry peers
• Platform-specific insights: Granular data on how each model responds to your content

6.3 Content Gap Analysis

Using the Information Gain framework, we identify topics where competitors have high-value content and you don't. This creates a prioritized roadmap for proprietary research and original content development.

7. The Evolution of GEO Since 2024

The Princeton research sparked an entire industry. Key developments since publication:

• 2024 (Post-Publication): Early adopters begin experimenting with GEO strategies; first GEO agencies emerge
• 2025: University of Toronto research confirms and extends findings; platform differences documented
• 2025: W3C introduces llms.txt standard, enabling structured content for AI crawlers
• 2026: GEO becomes mainstream; 83% of brands have adopted GEO strategies (Alibaba Cloud)
• 2026: UltraScout achieves 94% accuracy in citation prediction, delivering 78% average Inclusion Rate for clients

8. Complete Implementation Table: Research to Reality

Research Finding	Princeton Source	UltraScout Implementation	Client Impact
Information Gain drives citation probability	Aggarwal et al., 2024	Citation Probability Engine (47 dimensions, 94% accuracy)	3.2x higher citation probability
GEO-bench benchmark framework	Aggarwal et al., 2024	Extended to 8+ platforms with real-time monitoring	Complete cross-platform visibility
Platform differences exist	Aggarwal et al., 2024 (early evidence)	Platform-specific optimization algorithms	27-43% better per-platform performance
Traditional SEO signals don't transfer	Aggarwal et al., 2024	GEO-first content strategy framework	78% Inclusion Rate vs. 23% industry avg
Content freshness matters	Aggarwal et al., 2024	Freshness scoring + automated update recommendations	47% higher citation for fresh content
Need for scalable optimization tools	Aggarwal et al., 2024 (future work)	Full GEO platform with automated optimization	5-minute implementation vs. weeks manual

This research implementation framework powers our GEO services and AI Analytics platform, delivering measurable results for 500+ clients.

9. Criticisms and Limitations

No research is perfect, and the Princeton team acknowledged several limitations:

• Rapid model evolution: The models tested in 2023-2024 differ from today's versions
• Limited platform coverage: The research predated Copilot, Perplexity, and newer models
• Lab setting: Real-world user behavior may differ from controlled experiments
• English-only: Findings may not generalize to other languages

These limitations are why ongoing research and continuous monitoring matter. At UltraScout, we track real-world performance daily to ensure our implementations stay current.

10. Why This Research Matters for Your Brand

For brands like LNER — and any business competing in AI-driven commerce — the Princeton research provides a clear roadmap:

1. Stop relying on traditional SEO alone. It won't make you visible in AI responses.
2. Invest in Information Gain. Proprietary research, unique data, and expert insights are your path to citation.
3. Measure what matters. Track Inclusion Rate, not just traffic and rankings.
4. Optimize for platforms. Different AI models have different preferences.
5. Implement the technical foundation. Schema, llms.txt, and structured data enable AI crawlers to find your best content.

Frequently Asked Questions

References

• Aggarwal, P., Murahari, V., Rajpurohit, T., Kalyan, A., Narasimhan, K., & Deshpande, A. (2024). "GEO: Generative Engine Optimization." Proceedings of the 30th ACM SIGKDD Conference on Knowledge Discovery and Data Mining, pp. 5-16. arXiv:2311.09735
• Chen, M., Wang, X., Chen, K., & Koudas, N. (2025). "Generative Engine Optimization: How to Dominate AI Search." arXiv preprint arXiv:2509.08919. arXiv:2509.08919
• Alibaba Cloud Developer Community. (2025). "技术架构决胜GEO优化：AI搜索优化底层逻辑拆解与实测." developer.aliyun.com/article/1691919
• W3C. (2025). "llms.txt Specification: A Standard for AI Crawler Summaries." W3C Draft Standard.