- The GEO paper (Aggarwal, Murahari, Rajpurohit, Kalyan, Narasimhan, and Deshpande, presented at KDD 2024) is the actual academic source behind a claim already cited three times elsewhere on this site: that adding statistics, citations, and quotations measurably improves how often a source gets cited in a generative answer.
- It tested nine distinct methods, not just one general "add more evidence" approach, and measured them against two specific metrics (word-position-weighted visibility, and a separate subjective-impression score), not a single vague "visibility" number.
- Reported improvements across secondary sources cluster around 27–40% on the paper's visibility metrics for the top-performing methods. Statistics, citations, and quotations consistently rank among the strongest, though different summaries disagree on the exact order between them. Worth knowing too: secondary write-ups describe 40% as a per-method, per-domain ceiling, not an average, and gains skew toward sources that started lower-ranked.
- When the paper simulated every competitor optimizing at once, the effect was redistributive, not additive: a lower-ranked source gained roughly 115% relative visibility while the top-ranked source lost roughly 30%, as the gap between them closed rather than the pie growing evenly for everyone.
- Keyword stuffing was one of the nine methods actually tested, and it didn't just underperform. Secondary reporting puts it 8–10% below the unoptimized baseline, an active regression. The paper's own results argue against exactly the tactic most old-SEO instinct would reach for first.
- The paper's own most-dropped finding: results are domain-dependent. What works in factual, statistics-friendly categories underperforms in opinion-driven ones. There's no single universal winner across every content category.
This site has cited the GEO paper three times already, in AI visibility strategy, the five stages for winning in the AI search era, and how to write content that AI models can actually quote, each time as shorthand for "adding citations, statistics, and quotations improves visibility in generative answers." That's accurate, but it's also the smallest possible summary of a paper that actually tested nine distinct methods, built its own benchmark to do it, and found results that are considerably more specific, and more conditional, than the one-line version suggests.
What did the paper actually build?
"GEO: Generative Engine Optimization" (Pranjal Aggarwal, Vishvak Murahari, Tanmay Rajpurohit, Ashwin Kalyan, Karthik Narasimhan, and Ameet Deshpande) was first posted in November 2023 and presented at KDD 2024, the ACM SIGKDD Conference on Knowledge Discovery and Data Mining, one of the more established venues in the field. The paper's contribution isn't just an argument that content optimization for generative engines should exist. It's a benchmark (commonly referred to as GEO-bench) and a black-box testing framework built specifically to measure it, spanning a large set of real queries across multiple content domains, each paired with source documents a generative engine would need to draw from to answer.
That distinction matters: this wasn't a survey of opinions about what might work. It was a controlled test of specific, named interventions against a measurable outcome.
What are the nine methods, and the two metrics?
The paper tested nine distinct methods for editing existing content, roughly split into two categories:
Methods that add genuinely new information: citing external sources, adding statistics, adding direct quotations from authoritative figures.
Methods that change style or structure without adding new information: improving fluency/readability, simplifying language, adding technical/domain-specific terminology, writing in a more authoritative tone, keyword stuffing, and increasing the uniqueness of the wording relative to other sources.
Results were measured on two separate metrics: one weighting how prominently and how much a source's specific wording contributed to the generated answer (a position- and word-count-weighted visibility score), and a separate score built from subjective human or model judgments of how impressive/credible the citation appeared. Reporting a single "visibility went up 40%" figure, as most secondary summaries do, collapses two different measurements into one number, worth knowing if you ever see the paper's results quoted with more decimal-point precision than that.
What actually worked, and what's the honest caveat on exact ranking?
Across the sources that discuss the paper's actual results tables, statistics, citations, and quotations consistently land as the strongest-performing category, each landing in roughly the 27–40% improvement range on the paper's visibility metric depending on which specific method and which metric is being read. Different secondary summaries disagree on the exact order between these three (some rank quotation addition highest, others rank statistics addition highest), which is a real inconsistency worth flagging rather than resolving with false precision. What isn't in dispute across any source describing the paper: this cluster of methods (the ones that add genuinely new, checkable information) outperformed the purely stylistic changes, and combining methods (the paper specifically tested fluency optimization plus statistics addition together) outperformed any single method alone.
One precision worth adding to that range: secondary write-ups describe 40% as a per-method, per-domain ceiling, not an average effect across the board, and the gains reported skew toward sources that started lower-ranked rather than lifting every source by a flat amount. That distinction turns out to matter a lot once you look at what happens when competing sources all apply the same methods at once, below.
What happens when everyone optimizes at once?
The paper doesn't stop at testing methods in isolation. It also simulated the more realistic condition where multiple competing sources apply GEO to the same query simultaneously, since a tactic that only works while competitors aren't using it isn't a durable advantage on its own. The result is redistributive, not additive. In that simultaneous-optimization scenario, a source originally ranked 5th gained roughly 115% relative visibility using the Cite Sources method, while the source originally ranked 1st lost roughly 30% of its visibility in the same scenario. The gap between them closed rather than the total pie growing evenly for everyone. That's directly relevant if you're already ranking well for a given query: the paper's own data suggests GEO tends to close gaps between competitors more than it protects an existing lead, which is worth knowing before treating this as a one-sided advantage rather than a competitive reset.
Did keyword stuffing work?
This is the detail that gets dropped most often in summaries built for an SEO audience specifically, and it's the one most worth knowing: keyword stuffing was one of the nine methods the paper actually tested, not an assumption left untested. It didn't just fail to help. Secondary reporting on the paper's results puts it roughly 8–10% below the unoptimized baseline, including in the paper's own live Perplexity.ai validation pass, a real regression rather than a null result. The paper that's frequently cited to justify "optimizing for AI search" tested the classic old-SEO instinct directly and found it actively worked against the source using it.
What finding do most summaries drop entirely?
The paper is explicit that method effectiveness varies by content domain: what performs well in more factual, statistics-friendly categories underperforms in opinion-driven or debate-style categories. There is no single method, ranked once, that holds as "the best" across every domain the paper tested. This is the reason the practical writing guide built on this research doesn't prescribe one universal template. The paper's own data argues against that, and any summary that hands you a fixed top-3 list without that caveat is oversimplifying its own source.
A 2025 follow-up: does this hold on live engines, two years later?
The original paper's testing pipeline used a fixed corpus and a GPT-3.5-based simulation, not live production answer engines. That's a reasonable way to isolate variables, but it leaves open whether the same pattern holds when auditing actual citations happening in production. A September 2025 preprint, Kumar and Palkhouski's "AI Answer Engine Citation Behavior: An Empirical Analysis of the GEO-16 Framework" (arXiv:2509.10762), takes that different approach directly: 70 product-intent prompts, 1,702 real citations, and 1,100 audited URLs across Brave Summary, Google AI Overviews, and Perplexity. It finds citation-quality signals vary sharply by engine (a mean quality score of 0.727 for Brave, 0.687 for Google AI Overviews, and 0.300 for Perplexity) and that pages clearing a quality threshold get cited roughly 78% of the time regardless of which engine. It's a different methodology answering a related question, not a direct replication of the original paper's specific numbers, but it's the most relevant available check on whether the underlying idea generalizes to what's actually happening on live engines now.
Frequently asked questions
What is the GEO paper actually about?
"GEO: Generative Engine Optimization" (Aggarwal, Murahari, Rajpurohit, Kalyan, Narasimhan, and Deshpande) is an academic paper, presented at KDD 2024, the ACM SIGKDD conference, that built a benchmark and black-box optimization framework to test which content-editing methods actually improve how often a source gets cited in a generative AI answer, rather than assuming which ones would work.
What methods did the GEO paper test?
Nine, spanning content changes that add real information (citing sources, adding statistics, adding quotations) and stylistic or structural changes (improving fluency, simplifying language, adding technical terms, using authoritative language, keyword stuffing, and increasing uniqueness of the wording). Not all nine performed equally, and that gap is the paper's most useful finding for practitioners.
Did the GEO paper test keyword stuffing?
Yes. It was one of the nine methods tested, alongside the ones that measurably worked. Secondary reporting on the paper's results puts it roughly 8-10% below the unoptimized baseline, not just failing to help, an active regression, consistent with the broader pattern across the study: methods that added real, checkable information outperformed methods that changed the text's style or density without adding anything new.
Does the GEO paper's improvement percentage apply to every industry equally?
No. This is the finding most secondary summaries drop entirely. The paper is explicit that its methods' effectiveness is domain-dependent: what works well in more factual categories underperforms in others, such as opinion-driven or debate-style content. There is no single universal ranking of "best" method that holds across every category the paper tested.
Sources referenced
- Aggarwal, Murahari, Rajpurohit, Kalyan, Narasimhan & Deshpande, "GEO: Generative Engine Optimization" (arXiv 2311.09735, presented at KDD 2024)
- Kumar & Palkhouski, "AI Answer Engine Citation Behavior: An Empirical Analysis of the GEO-16 Framework" (arXiv 2509.10762, Sep 2025)
This piece is the deep dive behind a citation already used in AI visibility strategy and the five stages for winning in the AI search era. For the practical writing guide built on this research, see how to write content that AI models can actually quote.
