On October 10, 2026, two University of Toronto researchers published a paper that challenges a good deal of the advice offered under the GEO (generative engine optimization) label. Mahe Chen and Nick Koudas spent 90 days putting shopping questions to ChatGPT, Gemini and Perplexity through their APIs, then looked at what separates a cited page from an ignored one. Their answer: a page gets cited mainly because the engine found it during its own search, far more than because of its FAQ block, its markup or its load time. This article walks through the findings and their limits, then turns them into practical recommendations for small businesses and agencies.
Key takeaways
- For ChatGPT, the engine's own search explains roughly ten times as much of the citation gap between pages on the same site as all the measured page features combined, content included.
- FAQ blocks, summary boxes and schema.org markup are associated with more-cited sites, but they don't decide which page on a given site gets picked.
- How closely a paragraph answers the question remains the content factor most strongly tied to citation, ChatGPT included.
- ChatGPT readily names brands it picked up in training, sometimes with no cited page to back the mention up.
- For a small B2B vendor, the third-party pages that name it are far more closely tied to showing up in ChatGPT answers than they are for a consumer brand.
A 90-day study across two US markets
The study covers 158,099 answers collected daily between June 24 and September 21, 2026. Titled "What Drives AI Citations? Content, Retrieval and Model Knowledge in Generative Engine Optimization (GEO)", it was posted to arXiv on October 10, 2026. The same authors co-wrote a January 2026 study comparing the sources AI engines cite with Google's results, and this paper picks up where that one left off.
The first market is US fiber home internet, a category dominated by a handful of big brands. It accounts for 651 questions and 396,188 citations across the three engines. The second market serves as a cross-check: software vendors for architecture, engineering and construction firms, a niche made up of small players. It contributes 55 questions and 71,849 citations.
What sets the method apart is that the researchers don't just correlate page features with citation counts. They also compare pages within the same site, which strips out the effect of a domain's overall reputation or authority. The answers are collected through the engines' public APIs with web search turned on, using lightweight models: GPT-5.4 nano for ChatGPT, Gemini 3.1 Flash Lite for Gemini and Sonar for Perplexity. The researchers then replay the questions through the same APIs to see which pages each engine's search actually retrieves.
Why does one page on your site get cited instead of another?
Mostly because the engine's search retrieved it, and secondarily because its text closely matches the question. In the fiber market, 87% of the variation in citations happens between pages on the same site rather than between sites. For ChatGPT, all the page features the researchers measured, content included, explain just 2.8% of that within-site variation. The set of pages retrieved by the engine's own search, observed during the replay, explains 32%. Allowing for statistical error, that share falls between 24% and 42%, so there is real uncertainty, but not enough to change the order of magnitude.
Among content factors, question match stands out. The authors express effects per standard deviation, which here means a page that is noticeably better aligned with the question than average. On that basis, and across all three engines, a title closer to the question is associated with 14.1% more citations. A page whose best paragraph closely answers the question is associated with 15.5% more.
On third-party pages, those published by someone other than the brands themselves, the effects are stronger still. Engine by engine, the best-paragraph link holds for ChatGPT too, at +9.5%. The title, however, has no measurable effect there (−0.2%), which the authors connect to the way ChatGPT rewrites the question into searches of its own.
Technical best practices don't pick between your own pages
The features often recommended to win over AI engines travel with the most-cited sites, but they don't explain why one page beats its neighbor. The table below shows this for FAQ blocks, schema.org markup, page speed and author boxes. Their effect is positive when you compare sites, then stops being statistically significant when you compare pages on the same site.
| Page feature | Effect across sites | Effect within a site |
|---|---|---|
| Title matches the question | not reported | +14.1% |
| Best paragraph matches the question | not reported | +15.5% |
| FAQ block | +18.1% | −3.4% |
| Summary box | +20.9% | +6.9% |
| Schema.org markup | +7.5% | −5.3% |
| Page speed | +11.7% | −1.1% |
| Author box | +14.9% | −4.6% |
Change in citations: per standard deviation for question match and page speed (measured on 466 pages), with vs. without for FAQ blocks, boxes and markup. All three engines combined, US fiber market. None of the within-site effects for these five practices is statistically significant. Source: Chen and Koudas, arXiv, October 2026.
The authors put this down to Simpson's paradox, where a trend that shows up across a whole population disappears or reverses inside each of its subgroups. In their reading, these practices cluster on sites that already get cited for other reasons, such as brand recognition. They signal an established site, but the study gives no grounds for treating them as the cause of its citations.
That lines up with Google's own position. Its documentation on AI features in Search, last updated December 10, 2025, says no special schema.org structured data is needed to appear in AI Overviews or AI Mode. None of this makes markup pointless. Structured data helps engines parse a page and identify a business or product, and it is still associated with the best-cited sites. In light of these results, though, it would be unwise to pitch it as a lever that wins citations for a specific page.
We draw a lesson from this for our own audits. The technical checks we run describe whether a site can be accessed and understood. Which page gets cited depends first on the engine's search, whose ranking is only partly driven by how well the text matches its queries. The two layers complement each other, and the study is a useful reminder not to confuse them.
ChatGPT only cites what its own search found
In the study, ChatGPT never cited a page that its own search hadn't retrieved first. All 22,315 URLs cited by ChatGPT and analyzed in the replay appeared in the results of that internal search. The engine then chooses among those pages with a clear preference for the top ones: the page it ranks first gets cited 43.8% of the time.
That search looks very little like Google's or Bing's. ChatGPT rewrites the user's question into about four narrower queries. When the researchers run those same queries on Google, 97% of the pages ChatGPT retrieves are still missing from the top 50. The gap is even wider with Bing, where 99% of those pages don't appear in the top 30. Yet half of these pages come from sites that do appear in those same Google results, which suggests the engines pick different pages, often from the same sites.
Google describes a similar mechanism for its own generated answers. According to the same documentation, AI Overviews and AI Mode may use a "query fan-out" technique, issuing several related searches across subtopics before writing a response. A page can therefore be cited for a sub-question the user never typed, which may help explain why each AI engine cites different sources.
Those rewrites change with every answer, so two runs of the same question on the same day share only 15% of their cited pages. The researchers also split the 90 days into two-week windows to measure how stable citations are over time. For ChatGPT, 40% of the citation differences between pages on the same site still carry over two weeks later, but only 14% after ten weeks. That churn extends the pattern we described in our piece on why AI citations decay.
How does the model's memory shape which brands appear?
The model feeds what it learned in training into both its searches and its answers, which favors brands it already knows. The researchers call this the model's prior, meaning its existing knowledge. They measure it by asking the model, with search turned off, which brands and websites it would recommend for each question. When a brand is among its picks for a question, ChatGPT writes the brand name into its search queries 67% of the time, against 11% otherwise.
That memory also shows up in which sites get cited. Nearly half of ChatGPT's citations, 47% to be precise, come from sites the model names for that question without searching, against 25% when using the sites it names for a different question. The study also finds unsupported mentions: when ChatGPT brings up a brand the question didn't mention, 46% of those mentions have no cited page naming that brand.
A controlled experiment confirms how much a name weighs. The researchers gave the lightweight models used for data collection, search turned off, a set of products with identical specs, revealing or hiding the brand names. For ChatGPT's model, simply revealing brand names shifts the top three of the ranking 2.1 times as much as ordinary run-to-run variation does. Gemini is more sensitive still, at 3.3 times, while the name of a page's publisher plays a smaller role.
These results build on what we found when looking at whether AI engines cite brands or the press. A well-known brand starts with an advantage that its pages alone can't create. The authors are careful to frame the idea of shaping a model's prior as a hypothesis they did not test.
Small B2B vendors lean on third-party pages
In the niche market studied, whether ChatGPT mentions a vendor is closely tied to the outside pages that name it. When a third-party page cited by ChatGPT names a software vendor, the odds of the answer mentioning that vendor are 39.7 times higher. In the fiber market, where every brand is a household name, the same figure is only 4.2.
This follows from the mechanism described above. The model barely knows these small vendors and so doesn't add them to its searches unprompted. A page found online is then the main way they make it into an answer. Trade press and comparison pages that name these vendors can become their primary point of entry, even though ChatGPT often consults comparison pages but rarely cites them.
In this market, overall site authority shows no measurable effect, at −1.3% per standard deviation. Title-to-question match, on the other hand, is even stronger than in the consumer market, at +26.4% per standard deviation, a figure the study doesn't break down by engine. The sample is small, though, and the authors note that with 629 pages the advantage of search over page features is not statistically significant.
The study's limitations
These are statistical associations from English-language questions about the US market, not proven cause and effect. Comparing pages within a site removes differences between domains, but not every difference between pages. About a third of cited URLs could not be analyzed.
Both the collection and the replay ran through the engines' APIs on lightweight models, which may behave differently from the consumer apps, and Perplexity could not be replayed at all. The questions were written for clients of the authors' own monitoring platform, which shapes their selection. Engines also change constantly, and the authors say their results describe 2026.
None of this overturns earlier GEO research. The foundational paper by Aggarwal and co-authors, posted to arXiv in November 2023, measured the effect of content rewrites on a fixed set of sources. The new findings concern an earlier step, when the engine decides which pages to read in the first place. Our review of the GEO techniques that actually work already pointed to that step.
What we recommend for small businesses and agencies
Treat the engine's search as the first hurdle to clear, before any fine-tuning. That starts with making sure AI search crawlers can read your pages. OpenAI's crawler documentation states that sites blocking OAI-SearchBot won't be shown in ChatGPT search answers, except as plain navigational links. It also says a robots.txt change takes about 24 hours to take effect.
Next, build each important page around a real customer question, with a paragraph that answers it directly. It is the content factor the study ties most clearly to citations, ChatGPT included, and it costs nothing but editorial effort. A title that echoes the question is associated with more citations from Gemini and Perplexity, though the study finds no such effect for ChatGPT. For an agency, that means starting from the questions prospects raise in sales calls or support tickets rather than from a keyword list.
If your brand isn't well known, put part of your effort into outside pages that can mention you. A trade press article or a place in an industry comparison can get you into answers even when the model barely knows you. The study does find, however, that at equal rank in its search results, ChatGPT prefers a concise page from the original source over a roundup or a forum. Keep your FAQ blocks and markup for what they do for readability, without expecting them to win citations page by page.
Finally, measure over time and question by question, since the set of cited pages changes from week to week. An AI visibility audit checks that engines can access and understand your pages, and citation tracking shows which ones actually get picked up over the following weeks. Given the variability the study documents, a single snapshot proves very little.
Frequently asked questions
What did Mahe Chen and Nick Koudas's October 2026 study find about AI citations?
Over 90 days, the two researchers put commercial questions to ChatGPT, Gemini and Perplexity through their public APIs and logged the answers, collecting 468,037 citations. They conclude that the engine's own search explains which pages get cited far better than the properties of those pages. For ChatGPT, the pages its search retrieves explain 32% of the citation differences between pages on the same site, versus 2.8% for all their measured features, content included. The model also draws on what it learned in training, particularly about brands it already knows.
Do FAQ blocks or structured data help a page get cited by ChatGPT?
Not within a single site, according to this study. Sites that use FAQ blocks, summary boxes or schema.org markup are cited more often than others, but those features don't separate one page from another on the same site. The authors attribute this to a statistical paradox: these practices cluster on sites that are already cited for other reasons. They remain useful for making your pages easier to understand, but they aren't a citation lever on their own.
Is ranking well on Google enough to be cited by ChatGPT?
No. In this study, 97% of the pages ChatGPT's search retrieved were missing from Google's top 50 results, even for the queries ChatGPT itself had written. ChatGPT rewrites the question into several searches of its own and only cites pages that come back from those searches. Half of those pages still sit on sites that appear in those Google results, so Google rankings matter, but they don't predict which page will show up in a ChatGPT answer.
How can a small B2B company show up in ChatGPT answers?
In the B2B market studied, a cited third-party page that names a vendor goes with 39.7 times the odds of ChatGPT mentioning it, against 4.2 in the consumer market. For a small company the model barely knows, trade press coverage and other sites that name it are therefore an important way into the answers. Writing pages around real customer questions is still worth it, because a paragraph that closely answers the question is associated with more citations, in ChatGPT as in the other engines.
Do these findings apply outside the US?
The authors don't claim they do. Both datasets cover English-language questions about the US market, one in consumer home internet and the other in software for architecture, engineering and construction firms. Nothing suggests that the underlying mechanism, an engine-specific search followed by a selection among the pages it finds, works differently elsewhere. The magnitudes may vary, though, which is a good reason to measure your own questions rather than borrow these numbers.

