SEO and GEO: how AI search engines choose sources to cite

On October 10, 2026, Mahe Chen and Nick Koudas published a research paper examining Generative Engine Optimization (GEO).

The study covers 158,099 AI-generated answers and 468,037 citations across two US commercial markets.

In the ChatGPT sample, retrieval information explains 32% of citation-count variation between pages on the same website. Observed page features explain 2.8%.

This does not mean content quality accounts for only 2.8% of citations. It shows that the retrieval process plays a substantial role in determining which documents become sources.

AI search engines can break a question into multiple searches, retrieve documents and select sources for the final answer. Google calls this process query fan-out.

In the experiments, ChatGPT retrieves an average of 26 pages per question and cites approximately three. When researchers repeated the same question on the same day, only 15% of cited pages overlapped on average.

Content still matters. Pages with titles and paragraphs that semantically match relevant questions are associated with more citations.

FAQ blocks, summary boxes and structured data show no robust advantage when pages from the same website are compared.

The researchers identified another factor: knowledge already present in the model.

In the dataset, 46% of brand mentions introduced by ChatGPT without being requested were unsupported by any cited page mentioning that brand.

A model may therefore name a company based on prior knowledge rather than retrieved evidence. For smaller, less familiar businesses, relevant documentation and third-party sources can play a more significant role.

Search Engine Optimization (SEO) remains essential: accessible pages, useful content and clear internal links. Google confirms that AI Overviews and AI Mode require no special optimization beyond established search practices.

GEO adds another question: how do AI systems retrieve and reference company information?

Documentation that clearly describes products, workflows, integrations and operational limitations gives retrieval systems more specific information than generic marketing claims.

Measurement tools are also changing. Google Search Console provides generative AI impression reporting, while Bing Webmaster Tools exposes citations and topic-level visibility.

The study is an observational preprint covering specific markets and model configurations. It does not establish a guaranteed technique for earning AI citations. It does, however, provide a useful distinction between available content, retrieved pages and sources actually cited.