The latest GEO citation research: 7 studies on what AI models cite (October 2026)

The latest GEO citation research points one way. Whether an AI model cites your page is decided mostly by whether its search retrieves the page at all. Once a page is retrieved, the on-page feature that matters most is a paragraph that answers the question in the words a searcher would use. Lists, tables and headings help in some content types and not others. Freshness is a 12-month question, not a three-month one. And Google’s AI Mode has cut the number of sources it cites per answer by about three quarters, with Google rank deciding who survived.

Seven studies published between April and October 2026 make that case, from a University of Toronto paper, SISTRIX, AirOps, Ahrefs and The Search Spot. This is a plain-English reading of what each one tested, how big it was, what it found and where they disagree. The section at the end covers what it changed in the Botworthy Score™, the research-weighted score I maintain for this.

Key points

• Within a single site, page features explain about 2.8% of the variation in which pages get cited by ChatGPT, Gemini and Perplexity. The engine’s own retrieval rank explains about 32% (Chen and Koudas, University of Toronto, October 2026).
• The page feature that holds up best under rank control is a paragraph that matches the question. Title match is next. Headings, tables, lists and word count had no effect within a site in that study.
• SISTRIX’s 96,418-URL study found tables, lists, question headings, a TL;DR, an author box and a visible update date all positive. It matched pages on domain but not on retrieval, which is why it disagrees with Toronto on structure.
• Format effects depend on content type. Among pages ChatGPT had already retrieved, three tables lifted comparison pages by 25.7% and did little for awareness pages (AirOps).
• Over 70% of ChatGPT-cited pages were updated within a year, and the median cited page is about 500 days old. Age acts as a tie-breaker, not a gate.
• AI Mode’s sources per answer fell from 16.3 to 4.3 in late August. Pages in Google’s top three kept their citation 46% of the time. Pages outside the top 20 kept it 8% of the time. No on-page feature predicted survival.

Retrieval decides who gets cited

The largest and best-designed study of the year is by Mahe Chen and Nick Koudas at the University of Toronto. They collected 158,099 answers and 468,037 citations from ChatGPT, Gemini and Perplexity across two 90-day periods, in two commercial markets. Then they did the thing most citation studies skip: they compared pages on the same site against each other, so brand strength and domain authority could not explain the result.

Within a site, retrieval rank explained 32% of the variation in citations. Page properties explained 2.8%. That is not a small gap. It means most of the advice about formatting a page for AI is working on the part of the problem that matters least, and the part that matters most is whether the engine’s search returns the page in the first place.

Two vendor datasets put numbers on the same curve. AirOps and Kevin Indig ran 16,851 queries through ChatGPT three times each in April. Pages in first place in ChatGPT’s search were cited 58.4% of the time. Pages in tenth place were cited 14.2% of the time. In a separate AirOps study of 548,534 retrieved pages, only about 15% of what ChatGPT retrieved made it into an answer.

Then came the change that made the point for everyone. Maeva Cifuentes at The Search Spot tracked 50,527 prompt-and-page pairs for eight B2B SaaS brands through Google’s late-August change to AI Mode. Sources per answer fell from 16.3 to 4.3, a 74% cut. AI Overviews followed in mid-October with a 54% cut. Only 9.3% of previously cited pages kept their citation for the same prompt. Pages in Google’s top three kept it 46.2% of the time. Pages outside the top 20 kept it 8.3% of the time. On a scrape of 497 of those pages, no on-page feature predicted survival: not FAQ sections, not schema, not comparison tables, not question headings, not word count. Brands were still mentioned at the same rate. They were cited far less.

The on-page feature that survives: matching the question

If retrieval explains most of it, what is left for the page? In the Toronto study, one thing stood out. The strongest page feature was how closely the best paragraph on the page matched the query, at +15.5% citations per standard deviation. Title match was next, at +14.1%. A summary box at the top was a small positive. Headings, tables, lists and word count had no measurable effect once rank was held constant.

Two more large datasets agree at the title level. Ahrefs ran 1.4 million ChatGPT prompts and measured how closely the page title matched the prompt. Cited pages scored 0.602 on similarity. Uncited pages scored 0.484. Pages with natural-language URL slugs were cited 89.8% of the time against 81.1% without. AirOps found that pages whose title shared half or more of its words with the query were cited at 20.1%, against 9.3% for pages sharing less than a tenth.

The AirOps April study went a level down, to headings. Pages whose headings closely matched the query were cited 41% of the time. Pages with weak heading matches were cited 30% of the time. At the same retrieval rank, that gap was about 19 points. The same study found that covering every sub-topic the engine fanned out to did not help: pages covering a quarter to a half of the sub-topics were cited more often than pages covering all of them. The lesson is to match the question you want to be cited for, not to cover everything near it.

Structure: where the studies disagree

Nine days after the Toronto paper, SISTRIX published the largest matched-pair study of page features so far: 96,418 URLs from 17,380 domains in the German-speaking market, each cited page paired with an uncited page from the same domain, 43 features tested, cited by ChatGPT, AI Mode and AI Overviews in the second quarter of 2026.

Almost everything structural is positive here. A table with a header row and ten or more bullet lists are both at 1.5x. A question as an H2 is at 1.4x, which is the first large test of the question form itself rather than of heading-to-query match. A TL;DR at the top is 1.3x. An author box with a photo is 1.4x. A visible last-updated date is 1.3x, a current year in the title is 2.7x and only an older year is 0.7x. Readability makes no difference at all.

Tables and lists are where SISTRIX and Toronto part company, and the difference is the design. SISTRIX matched each cited page against an uncited page on the same domain. That removes brand strength, but it says nothing about whether the uncited page was ever retrieved. Toronto held the engine’s own retrieval rank constant. So SISTRIX’s odds ratios sit somewhere between the raw 2025 gaps and the rank-controlled studies, and some of what they measure is “this page got retrieved” rather than “this page got quoted”. Both are true in their own frame. A well-structured page is more likely to be retrieved and cited. Among pages that are already retrieved, structure buys you less than the question-matching paragraph does.

SISTRIX also found two things most scoring systems do not measure: a table of contents at 1.7x and five or more H2 headings at 1.6x. The authors note that both partly track page length. Toronto found heading count null within a site, and the AirOps April data has citation rate drifting from 30% at zero subheadings to 33% at four to ten and back to 32% at 21 or more. One study, no retrieval control, and a length confound: worth watching, not yet worth acting on.

Format depends on content type

The study that does most to reconcile the two sides is AirOps’ From Retrieved to Cited. It took 7,500 commercial queries and 217,508 pages ChatGPT had retrieved, extracted 25 on-page signals and compared cited pages with retrieved-but-uncited pages stage by stage through the buying journey. Because every page in it was retrieved, the retrieval question is removed and only the page is left.

Finding	Where it held	What it means
Sentences of 11 to 14 words raised citation likelihood by about 7%. Ten words or fewer raised it 18.8% on research-stage pages.	All commercial stages	The first retrieval-controlled figure for sentence length. Short answer sentences help. Page-wide readability scores do not.
Three tables raised comparison pages by 25.7%. Eight list sections raised validation pages by up to 26.9%. Neither did much for awareness pages.	Comparison and validation stages	A table helps when the content compares things. On an explainer it is close to neutral.
Five to seven supporting statistics raised awareness pages by 20.3%. Seven price points raised comparison pages by 15.7%.	Awareness and comparison stages	Figures that answer the question count. Prices and specifications count on commercial pages.

That is the reconciliation. Lists and tables are not a universal lift. They lift the content types where the reader wants a comparison or a checklist, and they do little where the reader wants an explanation. Any rule that scores “has a table” regardless of what the page is for is measuring the wrong thing.

Freshness is a 12-month question

Much GEO advice says to refresh content every quarter. The evidence says a year.

AirOps looked at more than 4,000 pages ChatGPT had cited: 35% had been updated in the last three months, 53% within six, and over 70% within 12. For commercial queries it was 83% within 12 months. Pages left for more than a year were more than twice as unlikely to be cited. You will see a “three times more likely to lose citations” version of that figure quoted around the web. It is not in the primary report.

Ahrefs’ 1.4 million prompts put the median age of a cited page at about 500 days, with age acting as a tie-breaker between otherwise similar pages rather than a gate. For news, cited pages had a median age of about 200 days against 300 for uncited, with title relevance nearly identical, which is what a tie-breaker looks like. SISTRIX’s 0.7x for only an older year in the title says the same from the other direction: a stale date is a signal against you, and a visible recent date is a small signal for you.

Where ChatGPT is going for its sources

Ahrefs analysed 15 million ChatGPT fan-out queries from 66.6 million logged-out responses collected in August and September. Only 14.4% of answers ran a web search at all. When ChatGPT answers a fact question about a company, it cites an official source 84.4% of the time. Of the searches it runs, 6.8% go straight to a single site with a site: operator, and three quarters of those are brand sites. “Official”, “price” and the current year are all rising in its queries. Reddit citations fell 88% between March and September, and Wikipedia fell 80%.

Put that next to the Search Spot numbers and the picture for a content team is clear. The pages ChatGPT is going to for facts are increasingly product, pricing and support pages rather than articles. A brand that has a clean, current, official page for every fact a buyer asks about is cited for those facts. A brand that only has articles is not.

What to do with this

• Get retrieved first. Classic rankings, crawler access and brand mentions decide most of it. No amount of formatting rescues a page the engine never returns.
• Write the answer in the searcher’s words. One sentence per section that answers the question as it would be asked, with the title and the heading carrying the same words. This is the one on-page feature that survives every study design.
• Match the question, do not cover the topic. Pages covering a quarter to a half of the related sub-topics were cited more often than pages covering all of them.
• Use tables where the content compares. Comparison and validation pages gain a quarter more citations from tables and lists. Explainers gain almost nothing.
• Keep answer sentences short, and stop worrying about page readability. Eleven to fourteen words is the band that lifted citations among retrieved pages. Flesch scores made no difference anywhere.
• Update substantively inside 12 months, and show the date. Take old years out of titles and URLs.
• Build official fact pages. Prices, specifications, company details and support answers, on your own domain, kept current.

The studies

Study	Sample	Design
Chen and Koudas, University of Toronto, What Drives AI Citations? (October 2026)	158,099 answers, 468,037 citations, ChatGPT, Gemini and Perplexity	Within-site regression with retrieval rank held constant
SISTRIX, Citation Worthiness (October 2026)	96,418 URLs, 17,380 domains, DACH	Cited pages paired with uncited pages on the same domain, conditional logistic regression
AirOps with Kevin Indig, The Fan-Out Effect (April 2026)	16,851 queries, 353,799 pages, ChatGPT	Citation rate by retrieval rank, heading-to-query match and sub-topic coverage
AirOps, From Retrieved to Cited (April 2026)	7,500 commercial queries, 217,508 retrieved pages	Cited vs retrieved-but-uncited, 25 on-page signals, by buying stage
Ahrefs, Why ChatGPT Cites One Page Over Another (April 2026)	1.4 million prompts	Cited vs retrieved-but-uncited, title similarity and page age
AirOps, The Impact of Stale Content on AI Visibility (August 2025)	4,000+ cited pages, 900 queries	Cited pages only, by update date
Ahrefs, How ChatGPT Does Query Fan-Out (October 2026)	66.6 million responses, 15 million fan-out queries	Logged-out ChatGPT, August to September 2026
The Search Spot, AI Mode Cut 70% of Its Sources (October 2026)	50,527 prompt-page pairs, eight B2B brands	Before and after the AI Mode change, matched to Google rank

What this changed in the Botworthy Score

The Botworthy Score is a 0 to 100 score for how quotable a page is, across ten criteria weighted by this research. It moved to version 3 this month on the strength of the studies above, and it now says plainly that it measures how quotable a page is once an engine has retrieved it, not whether it will be retrieved.

Criterion	Version 3	Why
Answer-first sections	Up	A query-matching passage is the strongest within-site content predictor. From v3.1 it is scored on match to the question, not just position.
Question-led headings	Up	Heading-to-query match lifted citation rate by about 19 points at constant rank. Question form on its own now has one positive study, SISTRIX, and no rank-controlled one.
Summary at the top	Slightly down	Small positive in Toronto, 1.3x in SISTRIX.
Self-contained passages	Up	Every passage-level dataset shows engines lifting complete, stand-alone units.
Short sentences	Down, and redefined	Page-level readability is null everywhere. The rule now scores section opening sentences only.
Lists and tables	Down	Null within site once rank is controlled, positive by content type. Tables now carry more of the criterion than lists.
Statistics and specifics	Held	Figures that answer the question help. Figures sprinkled in for their own sake lowered rank in 19 of 24 test settings.
Named sources	Down	Support is lab-only. Null in the replication studies and for outbound links within a site.
Named expertise	Down	Author signal null within site. A quoted expert scores full marks, a byline alone half.
Recently updated	Up, and redefined	A substantive update within 12 months, with a penalty for a stale year in the title or URL.

Still ten criteria, still 100 points. The table of contents and H2 count findings from SISTRIX are now reported by the member skill as unweighted watch metrics on every library it scores. If a second retrieval-controlled study agrees with SISTRIX, they become criteria in version 4. The exact weights, the scoring rules and the study behind every number are on the member scoreboard, with a changelog of what moved and why. The free skill on GitHub stays on version 2.

• Join Botworthy for £20 a month if you want the live score, the monthly research brief and the current skill.
• Get the free scoring skill and run it on a handful of your own pages.
• Read the original insurer study for the method and the eight writing habits that bury the answer.

Botworthy is a service of Work That Limited. If you have questions about the score, the membership or the Index, email jessica@didgeheads.com.