Back to the blog
AI Visibility4 Nov 2026· 7 min read· By The AI Product Builder and MarTech Strategist

What 468,037 AI Search Citations Reveal about GEO and Technical SEO

A study of 468,037 AI citations found retrieval and brand memory outweigh FAQ blocks and schema. What that means for Technical SEO, GEO, AEO, and SXO.

AI answer engines cite the pages their own searches find, and name the brands their models already know. That is the conclusion of a study posted to arXiv on 10 October 2026 by Mahe Chen and Nick Koudas of the University of Toronto, who tracked 158,099 answers and 468,037 citations from three AI answer engines between 24 June and 21 September 2026. For teams adding FAQ blocks and schema markup in the hope of being cited, the useful finding is how little those features explained once pages on the same website were compared.

What the Study Measured

The researchers asked the same commercial questions every day in two US markets: 651 questions about home Internet service, and 55 about proposal software for architecture, engineering, and construction firms, a niche B2B market. They compared pages from the same website against each other, which separates what a page does from how strong its website already is. That design matters, since 87% of the variation in how often pages were cited lay between pages of the same website, and 13% between websites. The paper is a preprint with no peer-review venue noted, and the authors describe their results as associations.

Which Page Features Made a Difference

Within a website, the page that matched the question was cited more often. A page whose best paragraph matched the question one standard deviation more closely went with 15.5% more citations, and a closer title match went with 14.1% more. Across websites, an FAQ block went with 18.1% more citations, a summary box with 20.9%, an author box with 14.9%, schema.org markup with 7.5%, and faster page speed with 11.7%. Compared within the same website, none of those five was significant. In the paper's words, these practices "appear effective only across websites", a pattern consistent with stronger websites doing many of these things at once. For one engine, all page features together explained 2.8% of the within-website variation in citations, against 32% for what the engine's own search retrieved.

Why Retrieval Comes Before Content

An engine cites a page only after its own search has found it. For one engine, every one of the 22,315 cited URLs in the main dataset was among the pages its searches returned. That engine cited 43.8% of the pages it ranked first, and 3.5% of those it ranked 11 to 20. Of the pages it retrieved, 97% were absent from Google's top 50 results. On this evidence, a strong Google ranking says little about whether an AI engine will find the page. The engine rewrites each question into searches of its own, and those searches decide which pages compete.

What Brand Memory Adds to the Picture

What a model already knows shapes what it searches for. When a brand was part of the model's prior knowledge, the same engine wrote that brand into its searches in 67% of cases, against 11% otherwise. In 46% of its mentions of a brand the question did not name, no cited page named that brand.

In Search Engine Journal on 26 October 2026, Pedro Dias proposed "AI visibility = SEO + GEO" as a working shorthand. In that framing, SEO makes a source discoverable and retrievable, and GEO deals with what a model learnt in training. The piece discloses that Dias works at geoSurge, a firm researching this area. In a DEV post on 31 October, Ali Farhat describes GEO as emerging as a memory layer for AI search visibility, and cautions that memory here does not mean an AI system permanently stores a particular page or brand. What other websites say about a brand counts in the answer too. In the B2B market, a cited third-party page naming a vendor multiplied the odds of the answer naming that vendor by 39.7 on one engine, and 17.7 on another. That is consistent with this blog's earlier piece on earned media as what AI answer engines quote.

Why Technical SEO Should Be Split into Tests

Findings like these argue for splitting Technical SEO into segments that each get tested on a firm's own website: does schema markup, an FAQ block, or a faster template change citations for these pages and these questions? The measurement has to run for weeks. After 10 weeks, the share of citation variation still shared with the starting period was 14%, 29%, and 0% across the three engines. A single snapshot of AI answers therefore proves little.

The direction this points to is a single GEO framework that takes the retrieval checks from Technical SEO, the question-and-answer structure from AEO, and the after-the-click experience from SXO, and organises them around pillar pages. The SXO part rests on judgement rather than this study, which measured citations and nothing after the click. Dias describes a buyer who asks an AI assistant, shortlists a supplier, and only then searches for the company. That buyer arrives with the answer's claims in mind, and the page they land on has to confirm them. Since the page that matches the question is cited more often inside a website, a pillar page can carry the broad topic while each supporting page answers one question people ask, in its title and in a paragraph that answers it directly. Page type mattered less than the match: comparison, how-to, pricing, product, and support pages were not cited significantly differently from articles.

The Case for Keeping Technical SEO as its Own Discipline

The strongest counterargument comes from the same study. For one engine, every cited page had been retrieved first, and retrieval depends on the fundamentals Technical SEO exists to inspect: whether a crawler can reach a page, render it, and index it. For 11.5% of the pages the researchers downloaded, the main text required JavaScript, and within websites those pages went with 10.6% fewer citations from one engine. A fintech or web3 product site built as a JavaScript application, where key text appears only after scripts run, is exposed to exactly that. Google's guidance on its AI features sets no additional technical requirements, and it still requires a page to be indexed and eligible to be shown in Google Search with a snippet before it can appear as a supporting link.

Two further findings argue for inspection over assumption. Blocking a crawler did not reliably keep pages out of answers: 7.1% of the pages one engine cited disallowed its search crawler, and 19.9% of those another engine cited disallowed Googlebot, checked when the researchers downloaded the pages. A regulated firm relying on robots.txt to keep pages out of AI answers should test that assumption. And since 97% of one engine's retrieved pages sat outside Google's top 50, a team that checks its visibility only through Google rankings is measuring a different list from the one the AI engine works from. The July piece on Technical SEO, GEO, and AEO as one discipline put Technical SEO first, and this evidence keeps it there.

Where the Discussion Goes from Here

If page features explain under 3% of within-website citation variation for one engine, how much of a Technical SEO audit should run as a standalone discipline, and how much should move into a GEO framework measured by citations? If a brand's place in a model's memory decides whether an engine searches for it at all, does that work belong to SEO, to PR, or to a discipline still being named? For regulated brands, who owns the test when a robots.txt rule no longer settles whether a page appears? The study covers two US markets in English, neither of them financial services, and its authors say the findings describe 2026 since the engines keep changing. Fintech teams have good reason to run these tests on their own pages, and share the results.

FAQs

What makes an AI search engine cite a page? In a 2026 study of 468,037 citations, every page one engine cited had first been retrieved by its own search. Within a website, pages whose titles and paragraphs matched the question more closely were cited more often.

Do FAQ blocks and schema markup help AI citations? They went with more citations across websites and showed no significant effect when pages on the same website were compared. Treat them as part of a well-run site, and test their effect on your own pages.

Does Technical SEO still matter for AI search? Yes. An engine has to retrieve a page before citing it, pages whose main text needed JavaScript were cited less often by one engine, and Google requires a page to be indexed and eligible for a snippet before it can appear as a supporting link in its AI features.

Testing which parts of your Technical SEO still earn their place in AI search, or drawing the line somewhere else? Reach out at hello@marcomfin.tech.

#MarcomFintech#FractionalCMO#MarTech#FinTech#AISearch#GEO#TechnicalSEO#AIVisibility#ContentStrategy

This article was first published on MarcomFintech at https://marcomfin.tech/blog/ai-search-citations-geo-technical-seo. © MarcomFintech. All Rights Reserved.