Stop Keyword Stuffing: Princeton's GEO Paper Unpacks the New Rules of Generative Search

Original paper: 《GEO: Generative Engine Optimization》

Published in KDD 2024 (Proceedings of the 30th ACM SIGKDD Conference on Knowledge Discovery and Data Mining) · 12 pages

Authors: Pranjal Aggarwal, Vishvak Murahari, Tanmay Rajpurohit, Ashwin Kalyan, Karthik Narasimhan, Ameet Deshpande

Institutions: Indian Institute of Technology Delhi (IIT Delhi), Princeton University

Open-source Project: generative-engines.com/GEO · GitHub: GEO-optim/GEO

For three decades, the core mechanism of web content discovery remained anchored to the same paradigm: ten blue links on a search results page.

Site owners, marketing teams, and independent creators built an entire industry around Search Engine Optimization (SEO). They calibrated keyword densities, engineered title tags, and acquired thousands of backlinks to appease PageRank and inverted indices.

That paradigm is now fragmenting under the rapid rise of Generative Engines (GEs) such as Perplexity, Google AI Overviews, and Bing Copilot.

When a user submits a query, an LLM retrieves relevant web sources behind the scenes, synthesizes a single, cohesive natural-language response, and annotates individual statements with inline citations [1][2]. Users increasingly consume the synthesized answer directly without visiting every search result. In the era of generative discovery, visibility is determined not by where your link ranks on a page, but by how prominently and frequently an LLM cites your content in its synthesized response.

A research team from Princeton University and IIT Delhi formalized this paradigm in their KDD 2024 paper, GEO: Generative Engine Optimization.

By establishing GEO-bench—a comprehensive benchmark of 10,000 diverse search queries across 25 domains—the authors evaluated 9 content optimization strategies. Their findings deliver a stark reality check to digital creators: traditional SEO heuristics fail completely against LLMs, while factual optimization techniques allow a 5th-ranked website to overtake top incumbents by up to 115% in generative citation share.

---

Redefining Visibility: How Do LLMs Attribute Attention?

In traditional search engines, impression and visibility are straightforward: rank on SERP (Search Engine Results Page) and click-through rates (CTR).

In generative engines, multiple sources are synthesized into a single composite response. One source might back an entire opening paragraph, while another only provides a parenthetical detail. A simple ordinal list rank fails to capture actual reader exposure.

The authors introduced formal metrics designed specifically for generative engines:

1. Position-Adjusted Word Count ($Imp_{pwc}$)

This metric measures objective citation prominence by combining word count with reader attention decay:

$$Imp_{pwc}(c_i, r) = \frac{\sum_{s \in S_{c_i}} |s| \cdot e^{-pos(s)/|S|}}{\sum_{s \in S_r} |s|}$$

Where:

• $S_{c_i}$ is the set of sentences citing source $c_i$;
• $|s|$ represents sentence length in words;
• $pos(s)$ is the 0-indexed position of sentence $s$ within the generated answer;
• The exponential decay term $e^{-pos(s)/|S|}$ reflects the empirical power-law of user reading habits: earlier sentences carry exponentially higher cognitive weight than footnotes at the end.

A source that is cited early and backs substantial portions of the answer achieves a significantly higher visibility score.

2. Subjective Impression (7-Dimension G-Eval)

To account for qualitative user perception, the authors deployed LLM-as-a-judge across seven subjective dimensions:

1. Relevance: Alignment between the cited statement and the user query;
2. Influence: The degree to which the final synthesis depends on this source;
3. Uniqueness: Whether the source provided non-redundant information;
4. Subjective Position: Prominence of the citation in the user visual field;
5. Subjective Count: Perceived volume of content attributed to the source;
6. Click Likelihood: Probability that a reader will click through the citation;
7. Diversity: Multiplicity of perspectives offered.

---

9 GEO Strategies Tested: What Wins and What Fails

Testing across 10,000 queries on GEO-bench using top-5 Google Search results as source documents, the authors evaluated 9 distinct content transformation methods against an unoptimized baseline:

Takeaway 1: Keyword Stuffing Backfires

The classic SEO staple—stuffing high-volume keywords throughout headers and body text—resulted in an 8.7% drop in position-adjusted citation visibility.

Because LLMs operate on semantic attention rather than lexical token counts, unnatural keyword repetition dilutes information density and impairs sentence coherence. When synthesizing concise summaries, models routinely skip disjointed or spammy passages in favor of well-structured prose.

Takeaway 2: Verifiable Numbers and Direct Quotes Dominate

The highest-performing strategies were Quotation Addition (+41%) and Statistics Addition (+30-40%).

The paper illustrates this through concrete transformations:

• On the query "What is the secret of Swiss chocolate?", adding the factual citation "(According to a survey conducted by The International Chocolate Consumption Research Group, per capita annual consumption averages between 11 and 12 kilos)" increased source visibility in the LLM answer by 132.4%.
• On the query "Should robots replace humans in the workforce?", inserting the quantitative metric "...with a staggering 70% increase in robotic involvement in the last decade" yielded a 65.5% boost in attributed citations.

LLMs tasked with generating grounded responses naturally gravitate toward passages that offer high-precision, unambiguous facts that can directly back declarative sentences.

---

The Democratization Effect: Rank-5 Sites Gain +115%

In traditional search engines, high-ranking incumbents benefit from powerful backlink moats and historical domain authority that smaller sites cannot easily overcome.

The GEO paper explored what happens when all five retrieved websites on a SERP apply GEO optimizations concurrently:

SERP Position          Cite Sources             Statistics Addition       Authoritative Tone
─────────────────────────────────────────────────────────────────────────────────────────────
Rank 1 (Top Result)      -30.3%                    -20.6%                    -6.0%
Rank 2                   +2.5%                     -3.9%                     +4.1%
Rank 3                   +20.4%                    +8.1%                     -0.6%
Rank 4                   +15.5%                    +10.0%                    +12.6%
Rank 5 (Bottom Result)  +115.1%                    +97.9%                    +6.1%

Websites ranked 5th in Google search results experienced a +115.1% surge in citation visibility through source citations and a +97.9% boost through statistics. Conversely, the 1st-ranked website lost 30.3% of its citation share.

This happens through a two-stage retrieval-and-synthesis pipeline:

1. In the retrieval stage, the search engine brings the top five results into the candidate set; once they are retrieved, the LLM does not weigh how many backlinks each page has or how old its domain is.
2. In the synthesis stage, the LLM places the five documents side by side in its context window and builds the answer from the quality, clarity, factual density, and credibility of the text itself.
3. If the 5th-ranked page offers clearer explanations, precise statistics, and sharper citations while the top-ranked page is padded with repetitive marketing copy, the model will cite the 5th result more heavily.

Generative engines flatten the backlink and domain-authority moat, shifting competition back toward the quality of the content itself.

---

Domain Nuance and Combinations

The effectiveness of GEO methods varies meaningfully across domains:

• Law & Government, Opinions: Statistics Addition delivers the highest uplift. Empirical metrics transform subjective claims into defensible arguments.
• People & Society, History, Explanations: Quotation Addition proves most potent by providing authentic perspectives.
• Fact-Seeking Queries: Cite Sources provides the necessary grounding for verification.
• Debates: An Authoritative Tone performs best when reconciling competing arguments.

Multi-Strategy Synergies

Evaluating pairwise combinations of the top four methods revealed clear synergies:

Relative Visibility Gain Matrix:
                 Fluency Opt.      Statistics Addition   Cite Sources   Quotation Addition
Fluency Opt.         22.4%               35.8%              34.4%             33.0%
Statistics           35.8%               27.0%              30.3%             35.4%
Cite Sources         34.4%               30.3%              19.1%             20.1%
Quotation Addition   33.0%               35.4%              20.1%             30.3%

• Fluency Optimization + Statistics Addition achieved the highest overall score at +35.8%.
• While Cite Sources had a lower standalone score (+27.5%), it functioned as a universal force multiplier, boosting every other strategy with an average pairing uplift of 31.4%.

---

Validating in the Wild: Live Tests on Perplexity.ai

To verify that these dynamics hold on commercial platforms, the authors tested GEO transformations on Perplexity.ai using real file-grounded queries:

• Keyword Stuffing: Dropped word-count visibility by 10% (24.1 down to 21.9).
• Quotation Addition: Boosted position-adjusted word count by 22% and subjective score by 30%.
• Statistics Addition: Surged subjective impression by 37% (24.7 to 33.9).

The findings confirm that the mechanics of generative optimization reflect fundamental properties of retrieval-augmented generation rather than artifacts of a single benchmark model.

---

Actionable Guidelines for Modern Technical Writing

Writing for generative search engines requires a structural shift in how content is drafted and formatted:

1. Eliminate Artificial Keyword Padding: Focus on topical depth and logical structure. Keyword stuffing degrades syntax and signals low-quality content to language models.
2. Maximize Fact and Data Density: Replace vague generalizations ("rapidly growing", "significant improvements") with concrete numbers, percentages, and benchmark results.
3. Embed Direct Quotes and Source Attributions: Back key assertions with named studies, institutions, and quoted experts. LLMs prioritize verifiable source anchors.
4. Optimize for Syntactic Fluency: Keep paragraph structures clean and explanations unambiguous. Clear prose makes information extraction effortless for generative models.
5. Seize the Level Playing Field: Independent publications and technical blogs no longer need to match multi-million dollar SEO backlink budgets to achieve visibility in AI search. Superior factual precision and readability can outperform incumbent domains directly inside the AI answer.