TL;DR
- Do FAQ blocks, schema markup and summary boxes get a page cited by ChatGPT or Gemini? In one 90-day study the answer was no, once pages of the same website were compared with each other. Across websites, each of those markers went with 7.5 to 20.9 percent more citations. Within a website, none had a significant effect.
- The study is a University of Toronto preprint (arXiv 2610.15033, version 1, 10 October 2026, not peer reviewed). It read 158,099 answers to English commercial questions about two US markets, collected through the ChatGPT, Gemini and Perplexity APIs on their small models. Its results are associations, not proven causes.
- What held up on the page was match to the question: a title and a paragraph that answer, in the question's own terms, something buyers ask.
- For ChatGPT, which pages its own search returned, and how often in the top five, explained 32 percent of the citation differences between pages of the same website. Everything measured on the page explained 2.8 percent. Every one of the 22,315 pages ChatGPT cited in the authors' replay was a page its own search had returned, and 97 percent of the pages it retrieved were absent from Google's top 50 for the same queries.
- For a business the model does not already know, the route the study measured was a cited third-party page that names it. For any business, the first check is the same: run your buyers' questions through the engines and write down what comes back, before anyone audits your markup.
Most GEO checklists handed to a business owner carry the same items: add an FAQ block, add schema.org markup, put a summary box at the top, show an author, speed up the page. We have recommended some of them ourselves. A study published in October 2026 is the first we have read that tests those items by comparing pages inside the same website, and the first that shows the search an engine runs before it answers. Two of our earlier posts drew on parts of it. This post is the fuller reading: what it measured, what it did not, and what we changed in our own work.
A word on the source first. It is a preprint, version 1, by two authors, and it has not been peer reviewed. The authors say they are releasing the data, replay logs and code. We read the paper in October 2026 and have not re-run it. The answers were collected by a monitoring platform the authors operate, and the questions were written for the organizations that use that platform, not sampled from real user logs. The answers came through each engine's public API with web search on, from the small models (GPT-5.4 nano for ChatGPT, Gemini 3.1 Flash Lite, Perplexity Sonar), not from the consumer apps. Every figure below, except the ones we label as ours, is the authors' own and describes English questions about the US market, in 2026.
What did the study measure?
The study measured which pages ChatGPT, Gemini and Perplexity cited over 90 days, and then what separated the pages cited often from the pages cited rarely. Mahe Chen and Nick Koudas of the University of Toronto sent 651 questions about US home internet service to the three engines every day from 24 June to 21 September 2026, and collected 143,249 answers (Gemini answered only part of the questions until 25 July, and accounts for 26,069 of the answers). A second, smaller set of 55 questions about proposal software for architecture, engineering and construction firms added 14,850 answers. The two sets hold 468,037 citations.
The design differs from the studies behind most GEO advice:
- Pages are compared only with pages of the same website. Anything constant about a website, its size, its reputation, the links pointing at its domain, drops out of the comparison. The main analysis covers 12,038 pages cited at least twice, on 1,120 websites, scored on fourteen page features.
- A feature counts only if it passes five separate statistical tests. One significant coefficient among fourteen features is weak evidence, so the authors required each result to survive five checks built to catch different errors.
- The engines' own searches were observed in a replay. After the collection, the 651 questions went through the ChatGPT and Gemini APIs (the access developers use, not the apps) with web search on, 13 times on 11 separate days, the last on 7 October 2026. For ChatGPT, the API returns the searches the model ran and the ranked list of pages those searches brought back. Gemini's returns its searches and sources without a rank. Perplexity could not be replayed.
The engines did not behave alike. ChatGPT cited 2.0 pages per answer and sent 64 percent of its citations to the internet providers' own websites. Gemini cited 10.1 pages per answer, 83 percent of them on third-party websites. Perplexity cited nothing in 95.8 percent of its answers. In this data, a Perplexity citation count near zero describes how the study reached Perplexity (its Sonar model through the API), not your website, and not necessarily the Perplexity app.
Do FAQ blocks, schema and summary boxes earn AI citations?
FAQ blocks, schema markup and summary boxes did not go with more AI citations in the Toronto preprint, once pages were compared with other pages of the same website. Across websites, pages with an FAQ block, a summary box, an author box or schema.org markup were cited more, and so were faster pages. Within the same website, none of the five showed a significant effect.
| Recommended practice | Across websites | Within the same 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 (one standard deviation faster) | +11.7% | -1.1% |
The explanation is a statistical trap called Simpson's paradox. The markers sit on websites that get cited more for other reasons. In the data, FAQ blocks appeared on 43 percent of the pages of the most-cited quarter of websites and on 23 percent of the pages of the least-cited quarter. Compare every page on the web and the FAQ block gets credit that belongs to the website it sits on. Compare a website's FAQ pages with its other pages and the credit disappears.
The smaller software market showed the same pattern: an FAQ block went with 22.5 percent more citations across websites and 1.9 percent fewer within them. And in the replay, among pages ChatGPT had retrieved at the same rank, the markers went with lower odds of being cited, not higher: 0.93 times the odds with an FAQ block, 0.92 with schema markup, 0.77 with an author box. (Odds are the chance of being cited divided by the chance of not being cited. A figure under 1 means lower odds. It is not a percentage change in citations.)
This does not make schema or page speed useless. Structured data still tells a search engine who published a page, and a slow page still loses visitors. What the study found no support for is the claim that these markers, by themselves, get a page cited in an AI answer. Google's own guidance says structured data is not required for its AI features, which we covered in how AI chooses what to cite. The Toronto preprint is the first measurement from outside the engine makers that we have seen point the same way for ChatGPT and Gemini.
What on a page does go with more citations?
Within a website, one page property went with more AI citations in the Toronto preprint: how well the page matches the questions people ask. Of the fourteen features, only two passed all five tests, and both measure match: the title and the best-matching paragraph, scored by how close their meaning is to the questions in the study's set. Word overlap, first-paragraph match and page age passed four.
| Page feature | All pages | Third-party pages | Tests passed (all pages) |
|---|---|---|---|
| Best paragraph matches a question | +15.5% | +25.7% | 5 of 5 |
| Title matches a question | +14.1% | +23.6% | 5 of 5 |
| Word overlap with the questions | +9.1% | +12.8% | 4 of 5 |
| First paragraph matches a question | +5.9% | +3.1% | 4 of 5 |
| Page age (older page) | -5.1% | -4.2% | 4 of 5 |
| Word count | +0.9% | -3.0% | 1 of 5 |
| Headings | -2.7% | +1.9% | 2 of 5 |
| Lists | -1.0% | +0.5% | 1 of 5 |
| Tables | -1.4% | -4.0% | 1 of 5 |
Some of this cuts against common advice.
- The paragraph can be anywhere. The best-matching paragraph sat at a median of 45 percent of the way down the page. It was in the top quarter for 32 percent of cited pages and in the bottom quarter for 28 percent. A matching first paragraph went with more citations too, but fewer (5.9 percent, 4 of 5 tests). What counted was that some paragraph on the page matched a question closely, wherever it sat.
- The title matters less to ChatGPT. Title match went with more Gemini citations (16.8 percent), a smaller rise for Perplexity and no change for ChatGPT (-0.2 percent), which the authors put down to ChatGPT rewriting the question into its own searches before it looks.
- Match decides which questions, not how often. For a given page, the questions it was cited for matched it better in 88 percent of comparisons. But knowing how many questions a page matches added one point, from 4.9 to 5.9 percent, to the share of citation differences the authors' within-website model explained.
That last point is where most of the study sits. About 10 percent of the variation in citation counts was day-to-day chance. About 75 percent was repeatable, sat between pages of the same website, and was explained by no page feature. ("Explained" is the statistical sense: accounted for, not caused.)
Why does the engine's own search matter more than the page?
In the replay, ChatGPT cited a page only after its own search had returned it, and the pages it ranked highest were the ones it cited most often. When ChatGPT searches before answering a commercial question, it writes its own search queries, runs them, and reads what comes back. The replay made that step visible.
| What the replay showed for ChatGPT | Figure |
|---|---|
| Cited pages that its own search had returned | 100.0% of 22,315 |
| Pages consulted and pages cited, per question | 26 and 3, on average |
| Share of retrieved pages cited, by rank: 1st, 2nd, 6th to 10th, 11th to 20th | 43.8%, 35.5%, 12.2%, 3.5% |
| Overlap between two runs of the same question on the same day | 2.7% of search queries, 15% of cited pages |
| Retrieved pages absent from Google's top 50 and from Bing's top 30 | 97% and 99% |
| Differences in citation counts within a website explained by retrieval | 32% (95% interval 24 to 42%) |
| The same, explained by all page features | 2.8% |
The 32 percent needs its method beside it. The authors summarized the replay for each page as two numbers: how often ChatGPT's searches retrieved it, and how often it landed in the top five of a result list. Those two numbers explained 32 percent of the differences in 90-day citation counts between pages of the same website (6,178 pages on 362 websites), against 2.8 percent for every page feature combined. When the authors shuffled the replay numbers between pages of the same website, the figure fell to 0.05 percent. The replay ran after the collection, so it explained 14 percent of the differences in the first two weeks' citation counts and 42 percent in the last days. In the small software market the figures were 30 percent against 10.0 percent, on 629 pages, and that gap was not statistically significant.
For anyone who treats a Google ranking as a stand-in for AI visibility, the replay showed two problems.
- The engine does not search for what the buyer typed. ChatGPT split each question into about four keyword-style queries, kept 43 percent of the question's words, added the current year to 23 percent of its queries and a brand the user had not named to 28 percent. It rewrote them from scratch on every answer.
- It does not bring back Google's pages. For half of ChatGPT's retrieved pages (51 percent), the website did appear in Google's top 50 for the same queries, with a different page. Even for Gemini, which is grounded on Google Search, 93 percent of cited pages were absent from Google's top 50.
A good position in Google is still worth having for the clicks it brings. It does not tell you whether an AI engine's search retrieves your page. All of this sits before the step where an engine picks among the pages it has in hand, which is the step we looked at in GEO tricks, measured.
How does a business the model does not know get named?
In the Toronto preprint, a business the model did not already know was named mostly when a third-party page that the engine cited named it. The authors asked each engine's model, with search switched off, which brands it would recommend for every question. That list is the model's memory (its prior, in the paper's terms), and it showed up at every step they could observe. ChatGPT wrote a brand into its own searches 67 percent of the time when the brand was in its memory for that question, against 11 percent otherwise. And 46 percent of its mentions of a brand the question had not named had no cited page naming that brand.
The small software market is the opposite case. ChatGPT's model named the tracked vendors for 2 to 16 percent of questions with search off, against 61 to 72 percent for the large internet providers. Gemini's model named the same vendors for 60 to 96 percent of questions, so "unknown" describes one engine's model here. In that market, a cited third-party page that named a vendor went with 39.7 times the odds that ChatGPT's answer named it (17.7 times for Gemini), against 4.2 and 6.3 times for the large providers. Those are odds ratios, not "39.7 times more likely". Website authority, which went with more citations in the consumer market, showed no effect in the small one.
On our reading, most Quebec SMBs are closer to the small vendors than to the national brands. The study did not measure that. When we tested what models know about Quebec roofing with search off in July 2026, 5 of 70 responses named a local business we could verify. We went through what that means for a late arrival in the rich get richer, the fit get rich, so we will not repeat it here. Our probe was one trade, twelve French prompts and nine models, and it measured memory, not citations.
How do you check this on your own questions?
Run the questions your buyers ask through the engines, more than once, and record what comes back for each one. The authors tell publishers to monitor citations per question and over time, and they name the replay as the way to see which pages are retrieved, at what rank, and which are retrieved but not cited. A page audit shows none of the three.
| What to record, per question and per engine | What it tells you | What it points to |
|---|---|---|
| The searches the engine wrote | The wording the engine searches with, which is not the wording your buyer typed | The questions and terms your pages need to match |
| Whether one of your pages was retrieved, and at what rank | Whether you were a candidate at all | If you were not: a page that answers that question, or access for the engine's crawler |
| Whether a retrieved page was cited | Whether the page was read and passed over | If it was passed over: a fuller, more direct answer on that page |
| Which third-party pages were cited, and whether they name you | Where the engine gets its names in your market | The pages worth earning a mention on |
| Whether the answer names you | The outcome a buyer sees | The count to track over time |
Before you trust the record:
- Repeat every question. Two same-day ChatGPT runs shared 15 percent of their cited pages. One run tells you little about the next.
- Date it and run it again. Compare two two-week windows ten weeks apart: 14 percent of the differences in ChatGPT citations between pages of the same website carried over, and 29 percent for Gemini. A score from last quarter describes last quarter.
- Say which surface you tested. The replay used small models through the API, and the authors note that the APIs need not behave like the consumer apps.
You can do a rough version by hand: five buyer questions, three engines, two runs a week apart, with the cited links copied into a spreadsheet. It will not show you rank, but it shows who is named and which pages are cited. The six things any such measurement should state are in our post on reading AI visibility scores, and the stages around this table are in the AI visibility funnel. That post uses "retrieved" for a crawler reaching the page. Here it means the engine's own search returned it.
What does this not tell you?
This study does not tell you what causes a citation, and it measured nothing in French or in Quebec.
- Cause. The results are associations. Comparing pages of the same website removes what is constant about the website, not the unmeasured differences between its pages.
- Anything about French-language questions, Quebec markets, or Google's AI Overviews and AI Mode. Both datasets are English questions about the US market. Gemini was tested through its API, not through Google's search results page.
- How the consumer apps behave. The engines were gpt-5.4-nano and gemini-3.1-flash-lite through their APIs, the small models, and Perplexity's search could not be observed.
- What will hold next quarter. The cited pages turned over within weeks, and the engines change continuously. These are 2026 measurements.
- The full set of cited pages. About a third of the cited URLs could not be analysed (blocked to the authors' crawler, or not HTML), and pages were downloaded once, after the collection.
- That the small-market results are firm. That dataset has 55 questions, and the gap between retrieval and page features there was not statistically significant.
What we changed
We changed four things in how we scope GEO work. None of them guarantees a citation.
- The replay comes first. From this post on, a GEO mandate opens with the client's question set run through the engines, recorded per question as in the table above. We have not published one run this way yet. The measurements we have published so far are at the answer level: in our heavy civil contractor case study, four buyer questions put to Perplexity and Gemini, noting who was named and which pages were cited. The retrieval rows of that table are the part this paper added.
- We no longer present markup as a citation lever. Our own post on generative engine optimization from May 2026 lists structured data among its five moves and ranks a short direct answer under the H1 first. We still do both, for search engines and for readers (this post opens with a summary for that reason). We no longer tell a client that either one earns AI citations.
- Pages are written to questions. A page for each question a buyer really asks, with the question's own terms in the title and a paragraph that answers it. That split is our choice: the paper measured match to the questions and did not test one page per question against one long page. An FAQ block that spreads ten questions thinly over one page is not the same thing.
- For a business the engines do not know, third-party mentions come early. We look at which pages the engines already cite for the client's questions, and work on getting the client named there accurately.
If someone offers to audit your site for AI visibility, ask whether the audit includes the questions, the engines and what each engine retrieved. In this study, the page features an audit grades explained 2.8 percent of ChatGPT's citation differences within a website, and the markup among them showed no significant effect. We sell audits, and ours include a page audit, so we have an interest here. A page audit tells you whether search engines and crawlers can reach and read your pages, which is worth knowing. It does not tell you whether an AI answer cites you. Ours are described on our SEO audits page.