Retrieval, not FAQ blocks or schema, tracked AI citations in a 158,099-answer preprint

Analysis · Preprint, not peer-reviewed

What happened: Mahe Chen and Nick Koudas of the University of Toronto posted “What Drives AI Citations? Content, Retrieval and Model Knowledge in Generative Engine Optimization (GEO)” to arXiv on Oct. 10, 2026 (version 1, 16:18 UTC). The 11-page paper is a preprint; its arXiv listing gives no journal or conference publication. It tests page changes that the authors describe as commonly recommended in generative engine optimization (GEO) advice, using 90 days of daily answers from the APIs behind three AI search engines.

What did the study measure?

The authors collected answers once a day from three public APIs with web search on: OpenAI’s GPT-5.4 nano with the web search tool, Google’s Gemini 3.1 Flash Lite with Google Search grounding, and Perplexity’s Sonar. The paper labels these “ChatGPT,” “Gemini” and “Perplexity,” but they are API models, not the consumer apps; this article uses the API names.

The main dataset covers 651 questions written for a US fiber internet provider, collected from June 24 to Sept. 21, 2026: 143,249 answers with 396,188 citations. A second, smaller dataset of 55 questions about software for architecture, engineering and construction firms added 14,850 answers with 71,849 citations over the same 90 days. A citation is a URL linked in an answer, counted once per answer.

The design compares each page only with other pages on the same website. The authors argue this matters because a page on a popular website is cited more often for reasons unrelated to the page itself. The authors counted a feature as real only if it passed five separate robustness checks.

What did it find about common page tactics?

Across websites, each of these five tactics was associated with more citations. Pages with an FAQ block had 18.1% more citations, a summary box 20.9%, an author 14.9%, schema.org markup 7.5%, and one standard deviation faster page speed 11.7%, all statistically significant in that comparison.

Within websites, none was significant: the estimates were −3.4%, +6.9%, −4.6%, −5.3% and −1.1%. The authors attribute the gap to where the tactics sit: FAQ blocks appeared on 43% of pages in the most-cited quarter of websites and on 23% in the least-cited quarter. They call this “an instance of Simpson’s paradox.”

What did pass all five tests was relevance to the questions. Comparing two pages of the same website, a one-standard-deviation better best-paragraph match was associated with 15.5% more citations (95% CI 11.1–20.1%), and title match with 14.1% (10.7–17.6%). The paper reports that length, outbound links, headings, lists and tables “show no effect.” For the OpenAI model alone, title match showed no statistically significant association (−0.2%), while best-paragraph match did (+9.5%).

Why did retrieval explain more than page features?

The authors resubmitted the 651 questions through the OpenAI and Gemini APIs 13 times on 11 days, through Oct. 7, 2026, to see which searches each model ran and which pages came back. Perplexity could not be replayed.

For the OpenAI model, 100.0% of 22,315 cited URLs were pages its own searches had returned. On average, per question, it consulted 26 pages and cited 3. It cited 43.8% of the pages ranked first in its search results and 3.5% of those at ranks 11 to 20. Its rewritten queries for the same question on the same day overlapped by 2.7%.

Those retrieved pages were rarely in classic search results: 97% did not appear in Google’s top 50 results, and 99% not in Bing’s top 30, for any of the answer’s queries. Retrieval explained 32% (95% CI 24–42%) of the within-website variation in the OpenAI model’s 90-day citation counts, against 2.8% for all page features. The authors’ summary: “Pages are thus cited when searched; brands are also named when remembered.”

Brand mentions without a cited page

The paper keeps mentions and citations apart. Without web search, the OpenAI model still named brands and websites; the authors call this its “prior.” For 46% of the OpenAI model’s mentions of a brand the question did not name, no cited page named that brand. In constructed tests, revealing brand names changed each product’s top-three rate 2.1 times as much as repeating the identical prompt did.

What the study cannot show

• Preprint. arXiv lists no journal or conference publication, and the version 1 PDF still carries a conference template placeholder rather than a named venue.
• API models, not apps. The authors write that the replay uses the engines’ APIs, “which need not behave like their consumer applications.” The models are small ones: GPT-5.4 nano and Gemini 3.1 Flash Lite. Perplexity’s API cited nothing in 95.8% of its answers, which limits what the data says about it.
• US English, client-written questions. Both datasets are English questions about the US market, “written for the organizations that use our monitoring platform rather than sampled from user logs.”
• Associations, not causes. The within-website design does not remove unmeasured differences between pages, and pages were downloaded once, after the collection. In the main dataset, 63.2% of cited URLs returned usable HTML.
• No Google AI features. The study collected answers from the three APIs only, not from AI Overviews or AI Mode.
• Drift. Two weeks apart, 40% of the OpenAI model’s within-website variation was still shared; ten weeks apart, 14%.
• Data and code. The abstract says “We release data and code,” and the paper refers to “the released repository,” but neither the version 1 PDF (text or embedded links) nor arXiv’s HTML version gives its address, and we found none in a web search on Oct. 23, 2026. We re-checked arXiv on Oct. 28, 2026; it still listed only version 1, with no repository address.
• Acknowledgments. The paper thanks “the team at ktau.ai for their assistance with the data in our analysis.” ktau.ai’s website describes a Toronto-based company that sells AI-visibility services. The paper says the answers were collected on an academic monitoring platform operated by the authors, which it does not name, for anonymity.

How does this fit with Google’s guidance?

Google’s documentation says “There are no additional requirements to appear in AI Overviews or AI Mode, nor other special optimizations necessary,” and that “There’s also no special schema.org structured data that you need to add.” The preprint did not test those Google features, so it does not directly test that statement.

Our view: the paper’s most useful contribution is the method, not a verdict on any single tactic. Comparing pages within a website and replaying retrieval control for more than a simple across-website comparison does, and the paper states its limits (small API models, two US client markets, no peer review yet) plainly. It does not show that FAQ blocks or schema hurt. In this data, they had no statistically significant association with citations once the website was held constant.

What we don’t know yet

• Whether the results hold for consumer apps, for Google’s AI Overviews and AI Mode, for non-US markets, or for non-commercial questions.
• Whether peer review or a later arXiv version changes the estimates.
• Whether the released data, once linked, lets others reproduce the replay.

What to do now

1. Treat single-tactic checklists with caution: in this data, the across-website association did not hold when pages within the same website were compared.
2. The authors’ own advice is to “write pages for the questions people ask” and to monitor citations “over time and per question.”
3. If you evaluate a tactic, compare pages within your own site and repeat measurements over several weeks, as the paper’s observational design did.

Background

The term GEO comes from a paper first posted to arXiv in November 2023. It framed GEO as rewriting a page to be more visible in answers generated from a given set of sources.

Update log: Oct. 28, 2026 — First published.

Sources

[1] Chen, M., & Koudas, N. (2026, Oct. 10). “What Drives AI Citations? Content, Retrieval and Model Knowledge in Generative Engine Optimization (GEO).” arXiv:2610.15033v1 [cs.IR]. Preprint, not peer-reviewed. Accessed Oct. 23, 2026. https://arxiv.org/abs/2610.15033

[2] (PDF of [1], version 1, used for section and table locations.) https://arxiv.org/pdf/2610.15033v1

[3] Google Search Central. “AI features and your website.” Google for Developers. Last updated Dec. 10, 2025. Accessed Oct. 23, 2026. https://developers.google.com/search/docs/appearance/ai-features

[4] Aggarwal, P., et al. (2023; rev. 2024). “GEO: Generative Engine Optimization.” arXiv:2311.09735 (v1 Nov. 16, 2023; published in Proceedings of KDD 2024, per [1]’s reference list). https://arxiv.org/abs/2311.09735

[5] ktau (vendor). Home page. Accessed Oct. 23, 2026. https://ktau.ai/

[6] ktau (vendor). About page. Accessed Oct. 23, 2026. https://ktau.ai/about

How this was made

Prepared with AI assistance from the sources cited above. Editorial review by Nhi Phan (editor) and fact-checking by Jackies continue after publication; any change will be recorded in the update log and in our Corrections Policy and Log at https://aivisibilityjournal.com/corrections/.