Local SEO blog
What Makes ChatGPT Cite Your Pages, Updated October 2026
Last updated
Last week I was doing my usual doomscroll through a few thousand LinkedIn posts about SEO. That is genuinely most of what my feed is. Send help. Somewhere in the scrolling, a study stopped my thumb. A SISTRIX analysis of 96,418 URLs tested which page features come with higher citation odds in AI search, and it read like a checklist of what I have been building for over a decade. Tables of contents. Article schema. H2s on a rhythm. Visible update dates. The boring stuff, vindicated.
Two findings stopped me cold, though. The current year in the title carried a 2.7x citation multiplier, and FAQ schema, which SEOs spent much of 2025 telling each other to stop worrying about, measured 1.5x. The study also found no measurable effect from llms.txt, which tracks with my own choice never to install it on client sites.
Honestly? I care more about showing up inside LLM answers than about another green arrow in a rank report. This study is the kind of evidence that decides where the effort goes. So let me walk through it. And yes, I checked the sample size before trusting any of it. More on that at the end, because it is the best part.
What SISTRIX studied and why the comparison matters
SISTRIX analyzed 48,527 URLs cited by ChatGPT, Google AI Mode, and Google AI Overviews in the DACH region during Q2 2026. It paired each cited page with a non-cited page from the same domain, then analyzed 96,418 URLs across 17,380 domains using 43 page-level features and conditional logistic regression.
That paired design is useful. Comparing pages on the same domain helps control for site-level authority and brand strength. The study is asking what cited pages have in common with uncited siblings on the same site.
The reported multipliers describe associations, not causes. A page with a table of contents was cited 1.7 times as often as its same-domain counterpart without one. That does not prove adding a table of contents will produce more citations on any individual site.
The research was a DACH-region snapshot and did not include every AI platform. I first saw it alongside other recent findings in this weekly AI search roundup from PromptMarketing.
Which page features were associated with more citations
SISTRIX reported these citation multipliers:
- Table of contents: 1.7x
- Article schema: 1.6x
- Five or more H2 headings: 1.6x
- FAQPage schema: 1.5x
- Table with a header row: 1.5x
- A question used as an H2: 1.4x
- Visible last-updated date: 1.3x
- Current year in the title: 2.7x
An older year alone in the title was associated with a 0.7 multiplier. The current-year result is an early signal: SISTRIX added title-year features retrospectively, so it recommends treating them as initial indications.
Most of that list is stuff I have shipped on client sites for years without needing a study to justify it. Headings on a rhythm, schema that matches the content, dates that show the page is alive. Which raises a fair question: if I already do all of it, why did this study still teach me something? The year. The current year in the title. I had never been deliberate about it, and now I am, starting with my next post. A 2.7x association from a feature that costs one word is the cheapest test on the list. Cheap tests are my favorite kind.
Only 14% of pages had a table of contents. Just 5% showed a visible last-updated date. Pages combining many of the studied features were cited 2.1 times as often, and only 10% of cited pages had that combination.
That is the part I find most useful. A single tactic rarely carries the whole job. A well-structured page gives an answer engine multiple clear ways to identify, extract, and attribute the information.
What showed no measurable effect
SISTRIX found no measurable citation effect for HowTo, QAPage, or breadcrumb schema. Each scored 1.0. Readability scored 1.03, which the study treated as no significant effect. Product schema scored 0.8 overall and 0.5 in AI Overviews.
Those results do not make readability or product schema pointless. They describe citation patterns in this dataset. Readability still matters to people, and Product schema can support other search and shopping uses.
The FAQPage one deserves its own paragraph, because the SEO conversation on this flip-flopped hard. Not long ago the consensus was to stop bothering with FAQ schema. I watched that take travel through the same feed where I found this study, repeated with total confidence by people who clearly had not checked a single sample size. A 1.5x association on a sample this big says the consensus outran the data again. FAQ schema stays in my toolkit, and I do not feel weird about it.
The llms.txt finding deserves its own line. Domains with an llms.txt file accounted for 11.7% of cited URLs, compared with 14.0% across all domains. SISTRIX found no measurable citation advantage. Since llms.txt applies at the domain level, the comparison also involves differences between domains, so the result does not establish that the file causes harm. It simply gives no evidence to prioritize it for citations. I never installed it on client sites and never planned to. No apology needed. Just data.
I have spent enough time watching SEO recommendations travel faster than their supporting data. This is a useful reminder to ask what a tactic actually changes before adding it to a client checklist.
Put the answer where a system can find it
A separate RESONEO study of 76,450 shared Gemini conversations and 222,753 Google searches looked at the location of cited passages. It found that 33% of cited passages began in the first 10% of the text, and 80% began in the first half. Half the passages were 43 words or shorter.
Put the answer near the top. Then support it. Write short passages that can stand on their own, say 40-ish words, without stripping the context that keeps them accurate. An answer buried in paragraph nine is not getting quoted.
This matches how I build content. Answer early, headings on a steady rhythm, tables when real information gets easier to scan. A table earns its space when it organizes useful facts. One built to tick a box just adds noise, and somewhere out there a tool is charging somebody for it.
Rankings are an incomplete proxy for AI citations
In a High Salience study of ChatGPT sources, 79% of the pages ChatGPT cited, 175 of 221, appeared on neither Google’s nor Bing’s first page for the searches examined. Small sites made up 26% of pages ChatGPT pulled up and 14% of pages it cited.
My first reaction to the 79% figure was a double take. My second reaction, after a glance at my own citation logs, was a shrug. That tracks. A page can get cited way beyond the first page, and smaller sites showed up among both the pages ChatGPT retrieved and the sources it used. A traditional ranking report cannot describe any of that.
For smaller sites, that creates room to compete through strong, clear pages. Structure gives an answer engine a page that is easier to interpret and quote. No study can promise a citation, but improving the page itself is work an operator can control.
The sample-size lesson is worth stealing
Here is my favorite part, and it is the nerdiest. SISTRIX ran the same FAQ test at three sample sizes and got three different answers. 833 URLs? FAQ schema 23% less common on cited pages. 4,600 URLs? 31% more common. 96,418 URLs? 19% more common. Three runs. Three stories.
Its conclusion: be cautious with GEO studies that analyze fewer than 5,000 URLs.
That methodological warning matches my own skepticism about small-sample studies. A few dozen prompts or a small collection of URLs can point toward a question worth investigating. It takes a much larger sample to make a strong claim about what works broadly.
When the evidence was thin, I said so in my own posts instead of pretending I had a slam dunk. Turning a tiny test into a universal rule is how SEO ends up with another myth. If you read the methodology section before the headline, demand a real sample size before you believe a study.
How I use these findings in my content work
I have worked in SEO since 2013 and learned through hands-on client work. This study reads like a checklist of practices I already use: clear headings, answer-first openings, tables with meaningful data, and visible freshness signals. My freshness-stamp rule predates this research.
When I include Ahrefs or Semrush data, I label the source. I want readers and systems to see where a number came from, and I want the table to help someone understand the result. My approach to SEO content strategy and content SEO strategy for AI search starts with the reader’s question and the evidence available to answer it.
I also track citations and publish an AI visibility checker. My broader work in AI search optimization connects page improvements with visibility monitoring. On agency projects, Evergreen Digital owns the work. Founder Brady Madden sets strategic direction on most accounts, and I build strategy myself on some of them.
Start with the pages buyers need
Pick a page that answers an important customer question. Check that the answer appears early, headings make the page easy to navigate, and tables have clear header rows. Add schema that fits the content, then display an update date when the page has actually been reviewed and refreshed.
For the technical side, my technical SEO page covers the site foundations that help search systems access and understand content. If you want a practical place to begin, use my free SEO audit.