Study: retrieval, not GEO tricks, decides AI citations

Three numbers, in order of how much they ought to worry the GEO consultancy business: 158,099, 468,037, and the one the paper doesn't state as a percentage but might as well: close to zero. That last figure is how much of the variation in citation success across pages of the same website is explained by the on-page tweaks that an entire cottage industry now sells to CMOs: FAQ schema, summary boxes, structured data, page speed.

The study, posted to arXiv, tracked two 90-day collections of answers to commercial questions across contrasting markets, 158,099 answers carrying 468,037 citations in total. Its authors did something most GEO vendors don't bother with: they compared pages within the same website, holding domain authority, backlink profiles and brand trust roughly constant, to isolate what page-level content actually does. The answer is narrower than the industry's sales decks suggest. One thing works: pages whose titles and paragraphs closely match the phrasing of the question asked. Everything else on the GEO checklist, the researchers find, correlates with citation success only when comparing across different websites, which means it is probably picking up a proxy for domain quality, not a causal lever any single publisher can pull.

The retrieval layer, not the page, decides

Here is the finding that should reframe how financial-services, multilateral and industrial communications teams think about this channel entirely. The researchers replayed the same questions directly through the engines' own APIs and found that every page ChatGPT cites is a page its own search retrieved first. That sounds tautological until you sit with the mechanism: ChatGPT does not search the web the way a person does, or the way Google Search Console assumes. It rewrites the user's question into its own sub-queries, generates them freshly on every single answer, and the resulting search rarely overlaps with what Google or Bing would retrieve for the same words. That rewriting step, the paper finds, explains far more of the variation in which pages get cited than anything measurable about the page itself.

This is the part the GEO industry has been selling around rather than through. If citation depends on whether a model's self-generated sub-query happens to surface your page in its own index, then optimising your H2 tags or adding a summary box at the top of the article is tinkering with a layer of the system that the retrieval step has already bypassed. A pharmaceutical company or a development bank can spend a quarter rewriting its FAQ schema and see no durable change in citation share, not because the work was done badly but because the mechanism it was targeting isn't the one deciding outcomes.

For institutions whose public communications strategy runs through named reports, policy papers and technical documentation, rather than consumer landing pages, this matters in two directions. First, the one content lever that does work inside this study, close linguistic match between page content and the phrasing of likely user questions, rewards specificity over polish. A paper titled "Climate Risk Disclosure Framework 2024" written in the register of a UN technical annex will lose to a competitor's page that happens to restate the exact question a user typed into Perplexity. That is uncomfortable for institutions that write for precision and peer review, not query matching, but it is the actionable half of the finding. Second, and more structurally, if the sub-queries a model generates rarely coincide with Google's own ranking of the same terms, then an organisation's existing SEO investment, built over a decade against Google's retrieval logic, offers only partial cover for this channel. CGAP's position on microfinance, or Holcim's on cement decarbonisation, could rank first on Google and still be invisible to ChatGPT's rewritten sub-query, because the model asked a different question than the one Google was asked.