Three independent research streams — academic, practitioner, and platform — have studied how AI search systems choose which websites to cite. Their findings overlap more than they disagree, and the conclusions challenge several popular assumptions about how to optimize for AI.
This post synthesizes the full research landscape. If you want to skip straight to implementation, see the GEO guide for the complete framework. If you want the deep dive on one specific finding, Part 1 of our LLM Visibility series covers answer capsules in detail.
Quick Navigation
- What GEO Research Exists Today
- The Princeton GEO Study
- The Answer Capsule Research
- Platform-Specific Citation Behavior
- What the Research Agrees On
- What the Research Disagrees On
- What This Means for Small Businesses
- GEO Glossary
- Frequently Asked Questions
What GEO Research Exists Today
GEO research falls into three streams, each approaching the problem from a different angle.
Academic research treats GEO as a formal optimization problem. Researchers run controlled experiments where they modify content using specific strategies and measure whether AI systems cite it more or less often. The Princeton GEO study (Aggarwal et al., 2024) is the most cited example.
Practitioner research studies what already works in the wild. Instead of experimenting with modifications, researchers analyze pages that AI systems are already citing and look for shared structural traits. Adam Gnuse's answer capsule research (published via Search Engine Land, 2025) is the primary example, analyzing 15 domains and 7,500 ChatGPT referral sessions.
Platform research examines how different AI systems behave when selecting sources. This includes tracking studies like Paul DeMott's work on measuring LLM visibility (also published via Search Engine Land) and Authoritas's data on citation patterns across platforms.
None of these streams alone tells the complete story. Together, they form a surprisingly consistent picture.
The Princeton GEO Study
The Princeton study (Aggarwal et al., 2024) was the first academic paper to define "Generative Engine Optimization" as a discipline. The researchers tested nine content optimization strategies and measured their effect on AI citation rates.
The Nine Strategies Tested
| Strategy | What They Did | Citation Impact |
|---|---|---|
| Cite Sources | Added authoritative citations | +30-40% |
| Add Statistics | Included specific data points | +30-40% |
| Expert Quotes | Added quotations from experts | +41% |
| Fluency Optimization | Improved writing quality alone | Negligible |
| Technical Terms | Added domain-specific terminology | Moderate |
| Authoritative Tone | Rewrote in authoritative voice | Moderate |
| Unique Wording | Used distinctive phrasing | Moderate |
| Easy to Understand | Simplified language | Moderate |
| Keyword Stuffing | Added extra keywords | Negative |
Key Findings
Expert quotes had the single highest impact at +41% citation improvement. This was the most surprising result — simply adding a relevant expert quotation to content made AI systems significantly more likely to cite it.
Citations and statistics tied for second at +30-40%. Content that included specific data points or referenced authoritative sources saw consistent citation improvements across AI platforms.
Fluency alone did not help. Improving writing quality without changing substance had negligible impact. This is a critical finding because many GEO guides recommend "writing better" as a primary strategy. The Princeton data says better writing only helps when combined with structural changes.
Keyword stuffing actively hurt. Adding extra keywords to content reduced citation rates, suggesting AI systems can detect and penalize low-quality optimization attempts.
What This Means
The Princeton research establishes that GEO is real and measurable. Content modifications can meaningfully change whether AI systems cite you. But the modifications that work are substantive (adding data, quotes, citations) rather than cosmetic (rewriting for fluency, adding keywords).
The Answer Capsule Research
Adam Gnuse's practitioner research (Search Engine Land, 2025) analyzed pages that ChatGPT was already citing and identified structural patterns that the cited pages shared.
The core findings:
- 72.4% of blog posts cited by ChatGPT contained answer capsules — concise, self-contained explanations placed immediately after headings
- 91% of cited passages contained no outbound links within the capsule text
- 52.2% of cited posts contained original data (proprietary statistics, original research, first-party case studies)
- 38% of pages ranking on Google's first page received zero AI referral traffic
For a full breakdown of answer capsules — what they are, how to write them, and examples — see Part 1: Answer Capsules.
How This Connects to Princeton
The Gnuse findings and Princeton findings are complementary. Princeton tested modifications and found that adding statistics and expert quotes improved citations. Gnuse studied existing cited content and found that cited pages already contained original data and self-contained answer passages.
Both point to the same conclusion: AI systems prefer content that provides complete, extractable answers with supporting evidence.
Platform-Specific Citation Behavior
Different AI platforms choose sources differently. Understanding these differences matters because a page optimized for one platform may not perform identically on another.
ChatGPT
ChatGPT draws from its training data and, when browsing is enabled, from real-time web searches. Its citation behavior tends to favor:
- Pages with clear, extractable passages
- Content with original data or unique framing
- Established domains — but domain authority is not the only factor
Perplexity
Perplexity searches the web in real time for every query and cites sources prominently. It tends to cite:
- Multiple sources per response (typically 5-10 citations)
- Pages that rank well in traditional search
- Content with specific, quotable statements
Google AI Overviews
Google AI Overviews draw primarily from Google's existing search index. Pages that rank well in traditional search results are more likely to be cited in AI Overviews. This creates the strongest overlap between traditional SEO and GEO — Authoritas research found 62% overlap between organic rankings and AI Overview citations.
Claude
Claude uses search-augmented generation when connected to web search tools. Its citation patterns are similar to ChatGPT's browsing mode — structured content with clear answers tends to be cited more frequently.
The Practical Implication
There is no single "AI search algorithm" to optimize for. Each platform has different citation behaviors. But the structural principles (clear answers, original data, extractable passages) work across all of them.
What the Research Agrees On
Despite different methodologies and data sources, the three research streams converge on several points.
Structure Matters More Than Authority
Princeton found citation improvements of +30-40% regardless of site authority. Gnuse found that cited content shared structural traits (answer capsules, link-free formatting) independent of domain strength. This does not mean authority is irrelevant — it means formatting and substance can overcome authority gaps.
Self-Contained Answers Win
AI systems need to extract a passage and present it as part of a response. Content that provides complete, self-contained answers within a few sentences is easier for AI to cite than content that spreads an answer across multiple paragraphs or requires clicking through to understand.
Original Data Is a Moat
Both Princeton (+30-40% for statistics) and Gnuse (52.2% of cited posts had original data) found that original data significantly increases citation likelihood. This makes intuitive sense — AI cannot fabricate your proprietary data, so it must cite you as the source.
Cosmetic Changes Are Not Enough
Princeton's finding that fluency optimization alone had negligible impact is important. GEO is not about "writing better" in a general sense. It is about structuring content so AI can extract, attribute, and present it.
What the Research Disagrees On
The research is not unanimous on everything.
How Much Authority Matters
Princeton suggests structure can partially overcome authority gaps. Authoritas data shows that pages already ranking well in traditional search get cited more in AI Overviews (62% overlap). The truth is likely both — authority helps, but it is not the only factor, and structural optimization can meaningfully improve citation rates for lower-authority sites.
Whether llms.txt Files Help
The llms.txt specification allows websites to provide machine-readable summaries for AI systems. No published research has measured whether having an llms.txt file directly increases citations. Some practitioners recommend it as part of a GEO strategy; others consider it unproven. We include it in the 9-factor audit but weight it as the lowest-priority factor.
How Quickly GEO Changes Take Effect
AI platforms update their indexes on different schedules. ChatGPT's training data updates periodically (months). Perplexity searches live (immediately). Google AI Overviews reflect index changes (days to weeks). No research has established a reliable timeline for when content modifications translate into citation changes.
What This Means for Small Businesses
The research consistently shows that GEO favors substance and structure over raw authority. This is good news for small businesses.
Format and substance can overcome backlink profiles. A small business with a well-structured page containing original data can get cited over a large competitor with a poorly structured page. Princeton measured this directly — citation improvements were driven by content modifications, not site authority.
The investment is content, not tools. The most effective GEO strategies (answer capsules, original data, expert quotes) require content work, not expensive software. You need to know what to fix, and then you need to fix it.
Traditional SEO still matters. GEO is additive, not a replacement. Pages that rank well in traditional search are more likely to be cited in AI Overviews (62% overlap). The best strategy is solid traditional SEO plus GEO-specific structural optimization.
For the complete GEO implementation framework, see the Generative Engine Optimization guide. To audit your own pages, see Part 3: How to Audit Any Page for AI Citability.
GEO Glossary
The industry has not settled on a single term for optimizing content for AI search. Here are the key terms and how they relate to each other.
| Term | Full Name | What It Means |
|---|---|---|
| GEO | Generative Engine Optimization | Optimizing content so generative AI systems (ChatGPT, Perplexity, Claude) cite it. Coined by Princeton researchers (Aggarwal et al., 2024). |
| AEO | Answer Engine Optimization | Broader term that includes GEO plus traditional answer features like featured snippets and Google AI Overviews. |
| LLM Visibility | Large Language Model Visibility | Whether your content appears in LLM-generated responses. Used interchangeably with GEO in practitioner contexts. |
| AI Citability | — | How likely a page is to be cited by AI systems, based on structural and content factors. What Brass-SEO audits measure. |
| AI Overviews | Google AI Overviews | AI-generated answer summaries shown at the top of Google search results. Draws from Google's existing search index. |
| Answer Capsule | — | A self-contained explanation (1-2 sentences) placed immediately after a heading. The most common structural trait of AI-cited content (72.4%, Gnuse research). |
| AI Search | — | Umbrella term for any search experience powered by AI, including ChatGPT, Perplexity, Google AI Overviews, and Claude. |
| Zero-Click Search | — | A search where the user gets their answer directly in the results without clicking through to a website. AI Overviews are a type of zero-click search. |
Frequently Asked Questions
Where can I read the original Princeton GEO study?
The paper is "GEO: Generative Engine Optimization" by Aggarwal et al. (2024). It was published through Georgia Tech and Princeton researchers and is available on arXiv. Search for "GEO Generative Engine Optimization Aggarwal" to find the full paper.
How reliable is the answer capsule research?
The Gnuse research analyzed 15 domains and 7,500 ChatGPT referral sessions. It was published via Search Engine Land, a major industry publication. The sample size is meaningful for practitioner research, though it focused specifically on ChatGPT citations and may not perfectly generalize to other platforms.
Does this research apply to all industries?
The Princeton study tested across multiple topic categories. The Gnuse research covered 15 different domains. While no research is perfectly universal, the structural principles (clear answers, original data, expert quotes) apply broadly because they address how AI systems work, not industry-specific content preferences.
Is GEO research still evolving?
Yes. GEO as a discipline is less than two years old. The research base is growing but still limited compared to traditional SEO research, which spans decades. The findings summarized here represent the best available evidence, but new research may refine or update these conclusions.
How does Brass-SEO use this research?
Brass-SEO's AI Citability audit is built on the 9-factor framework derived from this research. The audit checks each page for answer capsules, original data, expert quotes, and the other factors identified by Princeton and Gnuse. See the GEO guide for the full framework.
Top comments (0)