Plumbing
What makes AI engines cite a page? Not your FAQ block.
By Arnav Mukherjee, founder of TofuBofu · October 27, 2026
We published a copy-paste guide to FAQ schema on this blog, and it called FAQ schema one of the highest-return things you can do for AI visibility. Its evidence was a famous number: most pages ChatGPT cites carry structured data. Every vendor in this category quotes some version of it. So did we.
On 10 October two University of Toronto researchers published the test that number never had. They compared pages on the same website, so the website's own strength couldn't take the credit. The FAQ block's advantage disappeared. So did schema's. One plain thing held on every engine they tested, and it needs no developer.
How was this measured?
The paper is "What Drives AI Citations? Content, Retrieval and Model Knowledge in Generative Engine Optimization (GEO)" by Mahe Chen and Nick Koudas, arXiv 2610.15033. They're the same pair who showed this month that ChatGPT changes its vendor picks by city. I read this one at source too.
- Sample: 158,099 answers carrying 468,037 citations, across two 90-day collections.
- Main market: US home internet. 651 questions, 143,249 answers, 24 June to 21 September 2026.
- B2B replication: proposal software for architecture, engineering and construction firms. 55 questions, 14,850 answers, a market of small vendors.
- Engines: ChatGPT, Gemini and Perplexity, asked once a day through their APIs with web search on.
- The design that matters: each page is compared only with other pages on its own website.
The limits are theirs, and I'm putting them before the findings. In their words, "the results are associations," not proven causes. The replay ran on the engines' APIs, which "need not behave like their consumer applications," and the models were small ones. The questions were English and US-only, and about a third of cited pages couldn't be downloaded for analysis. Read it as the strongest evidence published so far, and expect it to move.
Why did FAQ blocks and schema lose their advantage?
They sit on websites that get cited for other reasons. Here's every recommended practice the authors tested, measured both ways:
| Practice | Comparing websites | Within one website |
|---|---|---|
| FAQ block | +18.1% | −3.4% |
| Summary box | +20.9% | +6.9% |
| Author box | +14.9% | −4.6% |
| schema.org markup | +7.5% | −5.3% |
| Page speed | +11.7% | −1.1% |
All five look good when you compare websites. None of them passed the authors' tests when you compare pages on one site. The summary box's +6.9% didn't hold up across their checks. The authors name the cause: "an instance of Simpson's paradox." Strong websites use these markers more. FAQ blocks appear on 43% of pages from the most-cited quarter of websites and 23% from the least-cited quarter. Count citations across the web and the FAQ block takes credit for the website's authority.
It goes a step further. Holding ChatGPT's own ranking equal, the authors found FAQ blocks and schema went with slightly lower odds of being cited, 0.93 and 0.92 times. That's an association, not proof of harm. It's enough to retire the claim our guide made, and I've added a correction to it today.
What went with more citations within a website?
Words on the page, and the engines differ on which words. The authors' summary is "a title and a paragraph that answer a question in its own terms." Split by engine, the two don't behave the same:
| Match with the question | ChatGPT | Gemini | Perplexity |
|---|---|---|---|
| Best-matching paragraph | +9.5%* | +7.8%* | +4.8%* |
| Title | −0.2% | +16.8%* | +4.6%* |
| First paragraph | +0.7% | +8.6%* | +0.2% |
- The paragraph is the one lever that held everywhere. A paragraph answering the question went with more citations on all three engines.
- The title went with more Gemini citations and no extra ChatGPT ones. In the authors' words, title match goes with more citations from Gemini and Perplexity "but not from ChatGPT (−0.2%)."
- The paragraph doesn't need to come first. The best-matching paragraph sat at a median of 45% of the page length. "It need not be at the top."
- No association within a site: word count, headings, lists, tables and outbound links.
My read on the ChatGPT gap: ChatGPT rewrites your buyer's question into its own searches before it reads anything, so the words that matter are the ones its search and its reading land on, the paragraph. That's a hypothesis, and the paper doesn't test it. Why Gemini rewards titles isn't explained either; its cited pages mostly aren't Google's top results.
The B2B replication is the part I'd act on. In the AEC software market the title association rose to 26.4% per standard deviation, against 14.1% in the main dataset, both pooled across the three engines. The paragraph association was a little weaker there. My read is that thin markets reward the one page that answers the buyer's exact question, and most B2B service firms sell in thin markets. That's a judgement, drawn from 55 questions about one software niche.
A page has to come back from ChatGPT's own search before anything on it can matter.
Why does retrieval matter more than the page?
Because ChatGPT only cites what its own search brings back. The authors replayed the questions through the APIs and logged every search. "For each question, ChatGPT consults 26 pages on average and cites 3." Every page it cited had come back from one of its own searches. OpenAI's developer docs draw the same line: a field called sources "returns the complete list of URLs the model consulted when forming its response." The paper found every citation inside that consulted set.
Then they measured the split. Which pages ChatGPT's search retrieved explained 32% of the difference in citations between pages on one website. Every page feature they measured, combined, explained 2.8%. Retrieval decides most of it. Page polish adjusts the margins.
And that retrieval isn't Google's. "97% of ChatGPT's retrieved pages are absent from Google's top 50." So a page on Google's first page can go uncited by ChatGPT, and a page Google buries can get cited. SEO is still the floor. ChatGPT reads a different layer.
See which buyer questions you're missing. Our free scan asks real buying questions across five AI engines and shows which ones name you and which name a rival. No card, no call.
Check your AI visibility freeWhat should a B2B firm fix first?
In this order. The first three follow from the paper; the fourth is my read.
- List the questions your buyers ask an AI engine. Write them the way a buyer types them: "best managed IT provider for a 50-person law firm in Toronto", not "Managed Services".
- Make sure each question has a paragraph that answers it in its own terms. This held on all three engines. The paper's advice is to "write pages for the questions people ask."
- Title the page with the question for Gemini and Perplexity. Gemini showed the strongest title association, 16.8%. Don't count on the title for ChatGPT, which showed none.
- Keep the markup you have, and stop adding it for citations. FAQ blocks and schema showed no within-site lift. Spend those hours on steps 1 to 3.
One note on how this applies to us. Our own Fixes list carries schema and listing jobs, because they're one-time plumbing a scan can check. The paper didn't test listings, so I'm leaving those alone. On schema, it moves our weight further toward content: our content engine drafts pages from the buying questions your scan shows you losing, which is the work this paper points at.
If you want the earlier half of this argument, we read the most-quoted schema study at source in the schema study everyone is quoting. It measured pages that were already winning. This one held the website fixed. Both point the same way.
Frequently asked questions
Does FAQ schema help you get cited by AI engines?
Not within your own website, on the best evidence published so far. A 10 October 2026 study of 158,099 AI answers found FAQ blocks went with 18.1% more citations when websites were compared with each other, and 3.4% fewer when pages on the same website were compared. The authors attribute the first number to stronger websites using FAQ blocks more often.
What makes an AI engine cite one page over another on the same site?
In the study, a paragraph that answers the question in its own terms went with more citations on all three engines: 9.5% per standard deviation on ChatGPT, 7.8% on Gemini and 4.8% on Perplexity. A matching title went with more citations on Gemini, 16.8%, and Perplexity, 4.6%, and none on ChatGPT, at minus 0.2%.
Does the answer need to be in the first paragraph?
No. In the study the paragraph that best matched the question sat at a median of 45% of the page length, and the authors write that it need not be at the top. First-paragraph match went with more citations on Gemini only. Placing the answer early is still good writing for a human reader, but the engines found it wherever it was.
Why does ChatGPT cite pages that don't rank on Google?
Because it runs its own searches and cites from what those return. In the study every page ChatGPT cited was one its own search had retrieved, and 97% of the pages it retrieved were absent from Google's top 50 for any of the searches ChatGPT ran. Your Google rank is a poor guide to whether ChatGPT finds you.
Does this apply to B2B companies?
The authors repeated the analysis in a small B2B market, proposal software for architecture, engineering and construction firms, with 55 questions and 14,850 answers. The title association was stronger there, 26.4% per standard deviation against 14.1% in the main dataset, both pooled across the three engines. The paragraph association was a little weaker.
Should I remove schema markup from my site?
The study doesn't say to. At equal rank it found FAQ blocks and schema went with slightly lower odds of being cited, 0.93 and 0.92 times, which the authors present as an association and not as harm. My read: keep markup your CMS already produces, and stop adding it in the hope of citations. Spend that time on paragraphs that answer buyer questions.
What are the limits of this study?
The authors call their results associations, not proven causes. They queried the engines' APIs, which need not behave like the consumer apps, using small models. The questions were English and about the US market, about a third of cited pages couldn't be analysed, and the B2B dataset is small. Engines also change, so the findings describe 2026.
Sources and further reading
- What Drives AI Citations? Content, Retrieval and Model Knowledge in Generative Engine Optimization (GEO), Mahe Chen and Nick Koudas, arXiv 2610.15033, 10 October 2026, read at source 27 October 2026: the design, every figure and every quotation above, including the per-engine split in Appendix Table 8
- OpenAI, web search tool guide, read 27 October 2026: the consulted-versus-cited distinction in the sources field
- Location, Location, Generation, the same authors, arXiv 2610.24081: the companion study on location, written up in our post linked below
