Does GEO Work? 468K AI Citations: Retrieval Explains 32%, Page Properties 2.8% (2026)

Key takeaway: A University of Toronto team analyzed 158,099 answers and 468,037 citations from ChatGPT, Gemini, and Perplexity collected between June and September 2026. Comparing pages of the same website, 32% of the variation in ChatGPT citation counts was explained by retrieval — whether the AI's own search fetched the page — and only 2.8% by page properties. FAQ blocks, structured data, and summary boxes had no association with citations within a website; what did matter was titles and paragraphs written in the words of the question. In RanketAI's October 2026 measurement of Korean companies, 62% of Gemini answers came without a web search, so no page could be cited at all. GEO does work, but the effect sits in questions, retrieval, and outside sources rather than in page add-ons.

---

Summary in three lines

• Within a website, retrieval explained 32% of the citation differences between pages; all 14 page features combined explained 2.8%.
• FAQ blocks and schema.org markup looked effective only when comparing websites. Well-cited websites simply use them more often; the paper found no evidence that the add-ons create citations.
• ChatGPT named brands without a supporting page in 46% of its mentions. For lesser-known brands, a third-party page that names the brand sharply raised the odds of being named.

Within a website, retrieval explained 32% of citation differences

Comparing pages of the same website, retrieval explained 32% of the variation in ChatGPT citation counts, while page properties explained only 2.8%.

In What Drives AI Citations?, published on October 10, 2026, Mahe Chen and Nick Koudas of the University of Toronto tested how generative search engines choose sources, using real answers. An academic monitoring platform sent the same questions to each service's public API (web search on) every day from June 24 to September 21, 2026.

	US home internet	B2B software (architecture, engineering, construction)
Questions	651	55
Answers	143,249	14,850
Citations	396,188	71,849

The authors first split where citation differences come from. 13% of the variation in citation counts lies between websites and 87% between pages of the same website. Even inside one company's site, some pages are cited often and others almost never.

To explain that 87%, the authors replayed the same questions after collection and measured how often each page appeared in ChatGPT's search results and how often it ranked in the top five. These retrieval variables explained 32% of the within-website variation. All 14 page features the authors computed — length, heading structure, lists, tables, FAQ, structured data, and more — together explained 2.8%.

"explains far more of the variation than page properties (32% against 2.8%)" — Chen and Koudas, arXiv 2610.15033 abstract

In numbers, a polished page has no chance of being cited if the AI's search never returns it. Improving citations starts with checking whether the AI retrieves your page for that question, before adding anything to the page.

FAQ blocks and structured data showed no association within a website

FAQ blocks, schema.org markup, summary boxes, author bylines, and page speed had no significant association with citation counts when comparing pages of the same website.

Comparing websites instead makes the same add-ons look helpful. Section 4.2 reports that across websites, pages with an FAQ block went with 18.1% more citations. The table shows how citation counts changed per standard deviation of a feature, or when the add-on is present.

Page add-on	Within the same website	Across websites
FAQ block	-3.4% (not significant)	+18.1%
Summary box	+6.9% (not significant)	+20.9%
Author byline	-4.6% (not significant)	+14.9%
schema.org markup	-5.3% (not significant)	+7.5%
Page speed	-1.1% (not significant)	+11.7%

The authors explain the gap as Simpson's paradox. Highly cited websites are cited for other reasons, and those same websites also tend to use add-ons like FAQ blocks, so a cross-site comparison credits the add-ons. FAQ blocks appeared on 43% of pages in the most-cited quarter of websites and on 23% in the least-cited quarter.

"The across-website comparison credits the markers with the effect of the website, an instance of Simpson's paradox" — Chen and Koudas, arXiv 2610.15033, Section 4.2

Page features are not meaningless. They explained 42% of the between-website part, but only 3.4% of the within-website part. Well-run websites are associated with more citations, yet this study found no evidence that adding FAQ blocks and markup to an existing site raises a specific page's citations.

Titles and paragraphs that match the question went with 14–16% more citations

The only page features consistently tied to citations within a website were how closely the title and paragraphs matched the questions people ask.

The authors checked whether each feature passed five robustness tests. Only best-paragraph match and title match passed all five.

Page feature (per SD)	All pages	Brand's own pages	Third-party pages
Best-paragraph match to the question	+15.5%	+5.2%	+25.7%
Title match to the question	+14.1%	+7.4%	+23.6%
Page age	-5.1%	-10.6%	-4.2%

The paragraph-match effect had a 95% confidence interval of 11.1–20.1%. The B2B software data replicated the result, with title match at +26.4% per SD.

In practice, when there is a question you want to appear for, the most direct step is to use that question's wording in the title, subheadings, and first paragraph. How AI splits one question into several sub-queries is covered in query fan-out and sub-query coverage. The larger effect on third-party pages is worth reading together with the brand knowledge results below.

Every page ChatGPT cited came from its own search, and almost none were in Google's top 10

All 22,315 URLs ChatGPT cited were returned by its own searches, but only 0.1–0.6% were in Google's top 10 for the question as asked.

ChatGPT consulted 26 pages on average per question and cited 3. It cited 43.8% (95% CI 42.6–45.1%) of the pages it ranked first. Yet those search results barely overlapped with Google or Bing.

"97% of ChatGPT's retrieved pages are absent from Google's top 50 and 99% from Bing's top 30" — Chen and Koudas, arXiv 2610.15033, Section 6.3

The search queries themselves changed constantly. Two runs on the same day shared 2.7% of ChatGPT's queries (Jaccard) and 15% of cited pages. Over weeks the change grows: within a 14-day window, 63–67% of the variation repeated, but ten weeks later the shared share was 14% for ChatGPT, 29% for Gemini, and 0% for Perplexity.

A high Google ranking alone does not predict a ChatGPT citation. Whether you appear in AI answers can only be known by measuring AI answers directly, and by repeating the same questions rather than measuring once. RanketAI's own case on repeat-measurement variance is in when measurement noise outweighed monthly change.

ChatGPT named brands without a supporting page in 46% of its mentions

When ChatGPT named a brand the question did not mention, 46% of those mentions had no cited page naming the brand — the model drew on brands it already knew.

The paper also measured how what a model already knows from training — brand knowledge — shapes citations. 47% of ChatGPT's citations came from websites its model names without searching. In product descriptions with identical facts, revealing brand names changed top-three rates 2.1 times (ChatGPT) and 3.3 times (Gemini) as much as simply repeating the question.

"the same profile still reaches ChatGPT's top three in 80% of the answers under one name and 39% under another" — Chen and Koudas, arXiv 2610.15033, Section 6.6

For lesser-known brands, third-party pages mattered a great deal. In the B2B software data, where the model knew the vendors poorly, a cited third-party page that named a vendor raised the odds of the answer naming it by 39.7 times on ChatGPT and 17.7 times on Gemini. Polishing your own site alone rarely gets a brand the model does not know into the answer.

Securing third-party pages falls into three groups.

• Places you can publish directly — communities, reviews, and Q&A sites where a brand can post. Use the question's wording in the title and first paragraph.
• Places reached through press or contributed articles — industry media, comparison articles, and listicles. Approach them with press releases, contributed pieces, or pitches to editors.
• Places you structurally cannot enter — sources such as a competitor's own blog. Replace them with your own article or another third-party page that answers the same question.

The procedure has four steps: identify the sources the AI read for that question → sort them into the three groups → execute → re-measure the same question. How to find those sources is in the source gap playbook, and how to check how well models know your brand is in the six-axis entity footprint audit.

In Korean measurement too, answers without a search cited no pages at all

In RanketAI's October 2026 measurement of 12 Korean B2B SaaS companies, 62% of Gemini answers came without a web search, and none of those answers had a source list.

In its October 2026 benchmark, RanketAI measured the same 12 companies three times on the same day and checked whether a web search ran for each answer. Full results are in 62% of Gemini answers came without a web search.

Service	Answers that ran a web search	Answers citing the company's own domain
ChatGPT	100%	82%
Perplexity	100%	37%
Gemini	38%	19%

Looking only at Gemini answers that did search, the own-domain citation rate was 52%. This points the same way as the paper: without a search, no page can be cited regardless of its condition, and even with a search, a page has to be in the results to be cited. RanketAI's measurement covers 12 mostly well-known brands on a single day with a different unit of count, so the size of the numbers is not compared directly with the paper.

A sensible order: measure questions, fix titles and paragraphs, secure third-party pages, then check the site

Combining this study with RanketAI's measurement, improving citations works better as question-level measurement followed by matching articles and outside pages than as adding page features.

1. Measure by question. Record which questions mention your brand and which pages are cited. Cited pages change within weeks, so repeat the same questions every two weeks.
2. For each missing question, write the title and paragraphs in that question's words. Start by checking whether existing titles, subheadings, and first paragraphs match the question.
3. If your brand is not well known, secure third-party pages. Follow the three groups and four steps above.
4. Check the site's basic conditions. Confirm that AI crawlers can read your pages and that search can return them. FAQ blocks and structured data serve other purposes, such as how results display in search, so there is no reason to remove them — but this study gives no basis for making them the first citation task.

In RanketAI, AI Brand Visibility Analysis measures ChatGPT, Gemini, and Perplexity answers by question, and "This week's task" on the results screen turns missing questions into content, outside-source, and site-check tasks. Page Structure Diagnostics checks the site conditions in step 4 with no login. Because results are diagnosed and analyzed question by question, those per-question results — not a single overall score — decide the actual work. The procedure for turning a diagnosis into tasks is in how to improve when your brand is missing from AI answers.

Caveats before applying the study elsewhere

This is an observational study of English questions about the US market, collected through APIs, so it does not transfer directly to other languages, consumer apps, or Naver AI Briefing.

"Both datasets consist of English questions about the US market" — Chen and Koudas, arXiv 2610.15033, limitations

• Association, not causation. Comparing within a website controls for website differences, but the authors note that unmeasured differences between pages cannot be ruled out.
• APIs may differ from consumer apps. The retrieval replay used APIs, which the authors say "need not behave like their consumer applications." Perplexity could not be replayed.
• Only part of the pages were analyzed. Of 39,059 cited URLs, 63.2% returned usable HTML.
• A snapshot of 2026 engines. The engines change continuously, so the findings describe 2026, and the paper is a preprint not yet peer reviewed.

These results do not contradict lab studies that found GEO techniques effective. Lab studies measure the effect of rewriting text while holding competing pages fixed; this study measures how page add-ons relate to citations in real answers. The conditions behind 45 lab studies are covered in the conditions behind a 40% GEO effect.

Frequently asked questions

Does this mean GEO doesn't work?

No. What showed little effect in this study were page add-ons such as FAQ blocks and markup, while question-matching titles and paragraphs and third-party pages were clearly associated with citations. Defined narrowly as adding page features, GEO has a small effect; defined to include questions, retrieval, and outside sources, the work becomes clear.

Should I remove FAQ blocks and structured data?

There is no reason to. Within a website, FAQ blocks and schema.org markup simply had no significant association, and there is no evidence they reduce citations. Their other purposes, such as search result display, still hold, so keep them — but put them after question measurement and title and paragraph work in your citation priorities.

If I rank first on Google, will ChatGPT cite me?

Not necessarily. In this study, only 0.1–0.6% of the pages ChatGPT cited were in Google's top 10 for the question as asked. ChatGPT rewrites questions into its own search queries, and its results barely overlapped with Google or Bing.

Do the results hold for non-English questions or Naver?

That has not been tested. This study covered only English questions about the US market. Naver AI Briefing mostly cites sources inside Naver, such as Naver Blog, so its structure differs. RanketAI measures ChatGPT, Gemini, and Perplexity and does not include Naver AI Briefing.

How can I tell whether AI already knows my brand?

Ask questions that do not include your brand name (for example, "recommend a service for X in Korea") several times in new chats, and record whether your brand is named and whether a cited page names it. A mention with no cited page means the model already knows you; no mention means third-party pages come first. The procedure is in five steps to check whether ChatGPT mentions your company.

Can I check this for free?

Yes, by repeating the same questions in new chats and recording the results yourself. RanketAI's Page Structure Diagnostics is free with no login, and AI Brand Visibility Analysis offers a trial after login. A measurement tool becomes worth considering once the number of questions and AI services makes manual records hard to keep.

Related reading

• 62% of Gemini answers came without a web search — RanketAI's measurement comparing search runs and own-site citations across 12 Korean SaaS companies
• The conditions behind a 40% GEO effect — what 45 lab studies found and under which conditions
• One question becomes 12 searches — how AI splits a question into sub-queries and how to respond
• Half of AI citations come from outside your site — finding and sorting third-party sources
• How to improve when your brand is missing from AI answers — five steps from missing questions to tasks by cause