# What is Generative Engine Optimization (GEO)?

> The discipline of getting your brand cited and recommended inside AI-assistant answers — ChatGPT, Claude, Perplexity, Gemini, DeepSeek, Qwen, Doubao.

Published: 2026-05-09
Site: https://www.eastbound.ai/what-is-generative-engine-optimization/

## One-sentence definition

GEO is the practice of optimising web content, infrastructure and off-site source presence so generative AI engines surface, cite and recommend a brand when answering user prompts. The term was coined in the 2024 KDD paper "GEO: Generative Engine Optimization" (Aggarwal et al., Princeton + IIT Delhi). AEO, AI search optimization, LLM SEO and AI visibility are used interchangeably in industry practice.

## The mechanism: two stages, not one

Zhang Kai & Yao Jingang's 2026 framework (arXiv:2604.25707v1) separates generative search into **citation selection** (which pages enter the source pool) and **citation absorption** (which pages actually shape the answer). Tw93's 2026 instrumentation of ChatGPT: ~100 pages retrieved per query, ~15% absorbed. Selection ≠ absorption ≠ user-visible mention.

## The citation pyramid

1. Search-engine indexed (Bing for ChatGPT search; Google for AI Overview).
2. AI-crawler reachable (granular robots.txt for retrieval bots, not just training).
3. AI-parseable (clean HTML, semantic URLs, llms.txt, Markdown alternates). Vercel 2025: AI crawlers don't execute JS.
4. Selection-worthy (relevance, specificity, dated facts, named entities).
5. Absorption-worthy (quotable phrasing, evidence density, structured comparisons).
6. Third-party validated (~6.5× more effective than self-citation).

## What the research has measured

Aggarwal et al. (KDD 2024, n=10,000 queries):

- Authoritative citations: +115%
- Direct quotes from credible sources: +43%
- Relevant statistics with named sources: +33%

Length sweet spot: 1,000–3,000 words. Encyclopedia-style explainer pages: ~3× influence per citation vs news pages. Specificity (real numbers, dated comparisons, named entities) > fluff (~50%+ citation lift).

## Anti-patterns (do NOT)

- FAQ format / FAQPage schema. SE Ranking 129K-domain study: 3.6 vs 4.2 ChatGPT citations. Aggarwal KDD 2024 didn't test FAQ. Williams-Cook DUCKYEA: no extraction advantage.
- JSON-LD as universal AI signal. SearchVIU 2025: 0/5 systems extracted JSON-LD-only data. Bing/Copilot is the confirmed exception (Canel, SMX Munich March 2025); not generic ChatGPT/Claude/Perplexity.
- UA sniffing to serve bot-only content (cloaking; Google penalises).
- Speculative meta tags (`<meta name="ai-content-url">` etc.) — no spec, no support.
- Generic "increase length" without each chunk being independently useful.

## China is a separate surface

Mainland-Chinese consumers use DeepSeek, Qwen, Doubao (+ Yuanbao, Kimi, ERNIE Bot) — not ChatGPT. Top-15 source overlap (Jaccard) between the three Chinese engines: 0.20–0.30 in our 540-call panel. Generic GEO frameworks cannot be ported without rebuilding the source-substrate model from scratch.

## Read next

- [China AI visibility for global brands](https://www.eastbound.ai/china-ai-visibility/)
- [AI visibility vs SEO](https://www.eastbound.ai/ai-visibility-vs-seo/)
- [How to get cited on Chinese AI](https://www.eastbound.ai/how-to-get-cited-on-chinese-ai/)
- [Why isn't my brand on DeepSeek](https://www.eastbound.ai/why-isnt-my-brand-on-deepseek/)
- [Methodology](https://www.eastbound.ai/methodology/)

## Run the audit

- Multi-engine: https://www.eastbound.ai/ai-visibility-audit/
- DeepSeek-only: https://www.eastbound.ai/deepseek-rank-tracker/