GEO: Generative Engine Optimization — The Foundational Paper, Simplified
Based on the paper "GEO: Generative Engine Optimization" by Aggarwal, Murahari, Rajpurohit, Kalyan, Narasimhan & Deshpande (Princeton University / IIT Delhi). Published at KDD 2024.
Paper: arXiv:2311.09735 | Code: github.com/GEO-optim/GEO
The Problem: AI Search Is Killing Web Traffic
For 25 years, when you searched for something on Google, you got a list of blue links. You clicked on one, visited a website, and the site owner received traffic. Everyone was happy.
Then came generative search engines — tools like ChatGPT, Perplexity, Google AI Overviews, and Bing Copilot. Instead of giving you links, they read multiple websites, synthesize the information, and give you a direct answer. You get what you need without ever clicking through to the original source.
It's great for users. But it's a disaster for the millions of websites, blogs, small businesses, and content creators who depend on that traffic to survive.
The researchers behind this paper identified three stakeholders in this new world:
- Users — they win (faster, better answers)
- Generative engine providers — they win (more users, more revenue)
- Content creators — they lose (less traffic, less visibility)
The paper asks a simple question: Can content creators do anything about it?
What Is GEO?
Generative Engine Optimization (GEO) is the new equivalent of SEO, but for AI-powered search engines.
- SEO = optimizing your content so Google ranks it higher in its list of links
- GEO = optimizing your content so AI search engines cite it and use it more prominently in their generated responses
The key difference: in traditional search, visibility means "appearing higher in a ranked list." In generative search, visibility means "how much of the AI's response comes from your content, and are you credited."
How Generative Search Engines Actually Work
The paper formalizes what they call a Generative Engine (GE). Here's the simplified pipeline:
- You ask a question (e.g., "What's the secret of Swiss chocolate?")
- The engine reformulates your query into simpler sub-queries that are easier to search
- A traditional search engine (like Google) retrieves the most relevant web pages
- An LLM summarizes each source
- Another LLM generates a final response, weaving together information from multiple sources with inline citations
The result is a rich, structured answer — not a list of links. Your website may be one of the sources used, but the user may never visit it.
The Big Challenge: How to Measure Visibility?
On Google, visibility is straightforward: what's your position on the results page? Position 1 gets ~36% of clicks, position 10 gets almost nothing.
In generative engines, it's much more complex. Your content can be:
- Cited once at the top of the response (high visibility)
- Cited five times but buried at the bottom (medium visibility)
- Used to inform the response but never explicitly cited (zero visibility)
The researchers proposed three ways to measure visibility:
1. Word Count
How many words in the AI's response are attributed to your source, relative to the total response? More words = more visibility.
2. Position-Adjusted Word Count
Same thing, but words appearing earlier in the response count more (because people read the beginning more carefully). This uses exponential decay — being cited first is worth much more than being cited last.
3. Subjective Impression
A multidimensional score covering: relevance to the query, influence on the response, uniqueness of your contribution, likelihood that the user clicks your citation, and diversity of the material presented.
The 9 Optimization Strategies Tested
The researchers tested 9 different ways to modify a website's content and measured whether each improved visibility in generative engine responses.
What Doesn't Work
| Strategy | What it does | Result |
|---|---|---|
| Keyword stuffing | Adding more query keywords to your content | Worse than doing nothing. Traditional SEO's favorite trick is useless here — LLMs understand meaning, not just keywords. |
| Unique words | Adding rare/unusual vocabulary | No significant improvement. |
What Works
| Strategy | What it does | Improvement |
|---|---|---|
| Adding quotations | Adding relevant quotes from credible sources | +40% visibility (best result) |
| Adding statistics | Replacing vague claims with precise numbers and data | +30% visibility |
| Citing sources | Adding references and citations to authoritative sources | +27% visibility |
| Fluency optimization | Improving the readability and flow of your writing | +25% visibility |
| Technical terms | Using precise, domain-specific terminology | +15% visibility |
| Easy to understand | Simplifying language for broader accessibility | +14% visibility |
| Authoritative tone | Writing in a more persuasive, confident style | +10% visibility |
The top three strategies share a common thread: they add credibility signals. Quotations, statistics, and references make content more trustworthy — and generative engines reward that.
The Most Surprising Finding: GEO Helps the Little Guys
When researchers tested what happens when all sites optimize simultaneously, they found something remarkable:
- Sites ranked #5 on Google (the underdogs) saw visibility gains of up to +115%
- Sites ranked #1 on Google (the big players) saw their visibility drop by 30%
Why? Traditional Google rankings depend heavily on backlinks, domain authority, and brand recognition — things that favor large companies. But generative engines care more about content quality. When a small blog adds solid quotations and statistics, the AI treats it as seriously as a Fortune 500 company's website.
This is potentially a great equalizer for independent creators and small businesses.
Domain Matters: One Size Doesn't Fit All
The most effective strategies depend on your topic:
| Strategy | Works best for |
|---|---|
| Authoritative tone | Debate topics, History, Science |
| Fluency optimization | Business, Science, Health |
| Citing sources | Factual claims, Law & Government |
| Adding quotations | Society, Explanations, History |
| Adding statistics | Law & Government, Debate, Opinion articles |
A legal blog should focus on adding statistics and citations. A history site should add direct quotations. A health site should improve fluency. There's no universal "best" strategy.
Combining Strategies Works Even Better
The researchers also tested pairs of strategies together. The winning combination:
Fluency optimization + Adding statistics = +35.8% visibility
This outperforms any individual strategy by more than 5 percentage points. The lesson: don't pick just one approach — layer them.
Adding references in combination with other methods was also particularly powerful. While "Citing sources" alone wasn't the top performer, it consistently boosted results in combination with other strategies (average of +31.4% in combination).
Real-World Validation: It Also Works on Perplexity.ai
The researchers didn't just test on their own simulated engine. They also ran experiments on Perplexity.ai, a real generative search engine with millions of users.
The results hold:
- Adding quotations: +22% on position-adjusted visibility
- Adding statistics: +37% on subjective impression
- Keyword stuffing: -10% (even worse than doing nothing on a real engine)
This confirms that the strategies generalize beyond the lab.
Practical Advice for Content Creators
If you create content on the web — whether you run a blog, a corporate website, or an e-commerce store — here's what to do:
- Stop keyword stuffing. It doesn't work for AI search. LLMs understand context, not keyword density.
- Add real data. Replace "sales increased significantly" with "sales increased 47% year-over-year (Source: 2024 Industry Report)."
- Quote experts. Include direct quotations from recognized authorities in your field.
- Cite your sources. Reference and link to credible, authoritative sources.
- Write clearly. Improve the flow and readability of your content.
- Adapt to your domain. Choose strategies that match your content type.
- Layer strategies. Combine 2-3 approaches for maximum impact.
The Big Picture
This paper, published at KDD 2024 (one of the largest data science conferences), essentially created a new field. Before GEO, content creators had no framework for thinking about visibility in AI search. Since its publication, the field has exploded — with over a dozen follow-up papers in 2025-2026 covering agentic GEO systems, e-commerce applications, multimodal optimization, and more.
The fundamental insight remains: generative engines reward content quality, credibility, and clarity over the traditional SEO playbook of backlinks and keywords. For content creators willing to adapt, this is both a challenge and an opportunity.
GEO-bench: The Benchmark
The paper also published GEO-bench, a benchmark of 10,000 diverse queries covering 25 domains, 9 query types, and multiple difficulty levels. It includes:
- Real user queries from Bing (MS MARCO), Google (Natural Questions, ORCAS)
- Demanding academic questions (All Souls College, Oxford)
- Trending queries from Perplexity.ai
- ELI5 (Explain Like I'm 5) questions from Reddit
- GPT-4 generated queries for diversity
This benchmark is publicly available and has become the standard testbed for GEO research.
Paper: Aggarwal, P., Murahari, V., Rajpurohit, T., Kalyan, A., Narasimhan, K., & Deshpande, A. (2024). GEO: Generative Engine Optimization. Proceedings of KDD 2024. arXiv:2311.09735