Scientific research › Aggarwal et al., KDD 2024
The Founding GEO Study: What Aggarwal et al. (KDD 2024) Actually Proved
Aggarwal et al. formalized Generative Engine Optimization (GEO) and showed, on a 10,000-query benchmark and on live systems like Perplexity, that black-box content changes can lift a site’s visibility in AI-generated answers by up to 40% — with the biggest gains going to lower-ranked sites. This is the paper that turned “being cited by AI” from folklore into a measurable discipline. Everything CapstonAI measures traces back to the questions it opened.
The study in numbers
| Item | Value |
|---|---|
| Paper | GEO: Generative Engine Optimization — arXiv:2311.09735, accepted to KDD 2024 |
| Authors | Aggarwal, Murahari, Rajpurohit, Kalyan, Narasimhan, Deshpande (Princeton, Georgia Tech, IIT Delhi, Allen Institute for AI) |
| Benchmark | GEO-bench — 10,000 queries across multiple domains |
| Headline result | Up to 40% relative visibility improvement in generative engine responses |
| Metrics introduced | Position-adjusted word count, subjective impression scores |
What worked — and what did not
The paper tested nine optimization strategies. The ranking is uncomfortable for classic SEO habits:
- Winners: quotation addition, statistics addition, citing sources. Content that carries verbatim quotes, numbers and named sources gets reused more by generative engines.
- Loser: keyword stuffing. The tactic that once moved rankings does little or nothing for answer visibility.
- Tone alone is weak. Sounding authoritative without evidence produced far smaller gains than adding actual evidence.
In short: generative engines reward evidence density, not persuasion. A claim with a number, a source and a date is worth more than three paragraphs of confident prose.
The democratizing finding
The largest visibility gains went to lower-ranked websites. In ranked search, position compounds: the big get bigger. In generative answers, a small site with highly extractable, well-evidenced content can be absorbed into an answer alongside — or instead of — a market leader. That window is the opportunity. It will not stay open once your competitors read the same research.
What this means for your business
- Rewrite money pages so every key claim carries a number, a source and a date.
- Add verbatim quotations from named experts or documented cases — the top-performing tactic in the study.
- Stop investing in keyword density. It does not buy answer visibility.
- Then measure: an optimization you cannot verify per engine is a hope, not a strategy.
CapstonAI operationalizes exactly this loop: it tracks whether ChatGPT, Gemini and Perplexity mention and cite you, and pinpoints which pages lack the evidence signals this research identifies. It never promises a citation — it makes you eligible for one, and proves whether it happened.
Frequently asked questions
What is the GEO paper by Aggarwal et al.?
The first academic formalization of Generative Engine Optimization (arXiv:2311.09735, KDD 2024). It introduced GEO-bench and showed content-side changes can lift visibility in AI answers by up to 40%.
Which optimization tactics performed best?
Adding quotations, adding statistics and citing sources. Keyword stuffing was among the weakest tactics tested.
Does the 40% gain apply to every site?
No. It is the upper bound observed under benchmark conditions, and gains were strongest for lower-ranked sites. Your result depends on your baseline, market and how far your content is from evidence-dense structure.
Is this still relevant in 2026?
Yes — it is the reference baseline. Later work (Zhang et al. 2026 on citation absorption, Kim et al. 2026 on structural optimization) refines it: evidence matters, but structure and selection eligibility gate everything.
Check these findings on your own site — free AI visibility audit →
Related: All GEO scientific research · Citation selection vs absorption · Evidence-container design · The five pillars of GEO
