GEO briefing: a new study says retrieval, not page tweaks, drives most ChatGPT citations
A new arXiv study of 468,037 ChatGPT citations finds that which pages the engine retrieves explains far more of the variation in citation than page features do. The authors report about 32% from retrieval against 2.8% from page properties. It is a preprint and focuses mainly on ChatGPT.
Window checked: 23 September to 23 October 2026. Labels: Proven means the primary source we opened states it. Practice means it comes from secondary reporting or an author’s own data. Experimental means a proposal or draft. We list at the end what we could not verify.
What the study found
- Proven “What Drives AI Citations? Content, Retrieval and Model Knowledge in Generative Engine Optimization (GEO)” by Mahe Chen and Nick Koudas was submitted to arXiv on 10 October 2026 (arXiv:2610.15033).
- Proven The authors analysed 158,099 answers containing 468,037 citations to commercial questions, collected over two 90-day periods in two contrasting markets. They released their data and code, and say the findings held in a market of small B2B vendors.
- Proven Within a single website, pages whose titles and paragraphs closely match the questions people ask are cited more often. Other features, including FAQ blocks, structured data, summary boxes and page speed, appear to help only when comparing across websites.
- Proven Which pages the engine’s own search retrieves explains about 32% of the variation between pages. Page properties explain about 2.8%. ChatGPT rewrites a question into sub-queries, and these often differ from Google or Bing results.
- Proven Model knowledge matters as well: ChatGPT includes brands it knows in its searches and favours sites it recognises. In controlled tests, brand names alone changed how the model ranked products with identical facts.
A small vendor study points the same way on sources
- Practice Proofsource, a vendor, studied 1,541 buyer-style answers from four engines for 14 brands, with data collected 4 to 8 October. It reports that only 0.8% of 11,831 citations pointed to a brand’s own site. The most-cited domains were youtube.com (10.2%), reddit.com (7.9%), en.wikipedia.org (7.8%) and linkedin.com (7.5%). Of brands named on days when all four engines answered, only 4.6% were named by all four.
- Practice The sample is non-random and the page labels its unit as answers in places and citations in others. Treat it as an illustration, not a benchmark.
What we could not verify
This window was thin for GEO. We found one new arXiv paper dated inside it. We could not verify any new tool launch, funding round or pricing change, and several circulating statistics (including large “percent of citations” claims from vendor blogs) could not be traced to an original study, so we left them out.