The GEO Research Paper
the World Is Citing
If you've ever wondered why some websites consistently appear in ChatGPT or Perplexity answers while yours gets ignored — it's not luck, it's craft. Princeton University and IIT Delhi published GEO: Generative Engine Optimization at KDD 2024, validating 9 actionable optimization strategies across 10,000 queries — turning "AI citation probability" into a measurable engineering problem for the first time.
What Actually Drives AI Citation
Traditional SEO is about ranking — but AI search no longer gives you a list of links. It gives one synthesized answer with a few cited sources. "Can I be cited?" is now the core success metric.
GEO-bench Dataset
10,000 queries · 25 domains · 9 data sources. 80% informational queries modeling real AI assistant usage patterns.
PAWC Metric
Position-Adjusted Word Count: the closer cited text appears to the start of the answer, and the more words cited, the higher the score.
Subjective Impression
An LLM-judged score of how much a site contributed to the final answer — catches "short but pivotal" citations that PAWC alone would miss.
The research challenges a long-held intuition: keyword density is not the core driver — what actually moves citation likelihood is the content's "credibility density" and "phrasing pattern."
9 Strategies: Effectiveness by the Numbers
The team applied each of 9 optimization strategies to GEO-bench sites and measured before/after visibility improvement.
| Strategy | PAWC Lift | Subjective Lift |
|---|---|---|
| Quotation Addition | +41% | +28% |
| Statistics Addition | +33% | +23% |
| Cite Sources | +30% | +15% |
| Fluency Optimization | +28% | +18% |
| Technical Terms | +20% | +7% |
| Easy-to-Understand | +14% | +10% |
| Authoritative | +12% | +14% |
| Unique Words | +7% | +5% |
| Keyword Stuffing | -10% | +5% |
Source: Aggarwal et al. (KDD 2024), arXiv:2311.09735
Phrasing Over Keywords, Every Time
Look at the top three (Quotation +41% / Statistics +33% / Cite Sources +30%). Their common pattern: none of them are "add what words" problems — they are "switch how you phrase it" problems:
- Change "many studies indicate X" → "Smith et al. (2023) found X" — turning a vague claim into a citable quotation.
- Change "the effect is substantial" → "the effect is a +41% improvement" — turning prose into a citable statistic.
- Add a source link after each claim — turning assertions into verifiable facts.
AI is not fooled by word inflation. It judges whether the passage itself is worth quoting. "Keyword Stuffing" at -10% versus "Unique Words" at +7% makes this crystal clear.
Source: Aggarwal et al., KDD 2024 — arXiv:2311.09735Lower-Ranked Sites Gain the Most
How Generative Engines Cite
AI synthesizes the top-N results into a single answer — not just the rank-1 result. A lower-ranked site that wins on "citability" can earn a spot in the AI's answer regardless of its search rank.
The SMB Opportunity
You don't need to outspend industry leaders for Google rank 1. Restructuring your site to match generative-engine preferences puts you directly in AI's answer — without the big budget.
How Sense Applies These 9 Strategies
Understanding the strategies is step one. The hard part is systematically applying all 9 to a real website.
AEO Revamp
For existing sites with rich content but structure not yet optimized for AI
New AEO Site
Strategies baked into HTML semantics and information architecture from day one
Start with an AEO Health Check
Audit your site against the 9 research-validated strategies, identify the weakest links, and get a PDF report with a concrete optimization roadmap.
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 (KDD '24). arXiv:2311.09735