How AI reads this page
ChatGPT, Claude, Gemini and Perplexity don't look at colours, images or animations. They receive text, headings, links and structured data. Below is exactly what they read when they open “Which pages does AI cite? SISTRIX's 96,000-URL study on what matters”.
The AI receives 2,154 words in 9 sections, 16 links and 3 blocks of structured data.
Text
The page content in Markdown, without menus or graphics. It is the format closest to how a language model reads text.
# Which pages does AI cite? SISTRIX's 96,000-URL study on what matters > SISTRIX compared 48,527 pages cited by ChatGPT, AI Mode and AI Overviews with uncited pages from the same sites. The factors that count and the ones that don't. URL: https://geosnap.ai/en/blog/quali-pagine-cita-ai-studio-sistrix-96000-url-cosa-conta-davvero Language: English Version IT: https://geosnap.ai/blog/quali-pagine-cita-ai-studio-sistrix-96000-url-cosa-conta-davvero Author: Rinald Sefa, CMO Geosnap · Category: Research and data Publisher: Geosnap (Maind Group S.r.l.), https://geosnap.ai ## In short On October 22, 2026 SISTRIX published the largest study so far on what sets AI-cited pages apart: 48,527 URLs cited by ChatGPT, Google AI Mode and AI Overviews, each compared with an uncited page from the same domain, 96,418 URLs in all. Structure matters (table of contents, at least five H2s, tables with a header row, Article and FAQPage schema), and so do trust and freshness signals (author box with photo, visible update date, current year in the title). llms.txt, readability and length on their own do not. These are correlations, not causes, but they are the first measured on a serious sample. - Pages with many of the eleven winning features are cited 2.1 times as often as barely structured pages, yet only 10% of cited pages actually have them. - Strongest factors: current year in the title (2.7x), table of contents (1.7x), Article schema and at least five H2s (1.6x), FAQPage schema and tables with a header row (1.5x). - No measurable effect: llms.txt (11.7% of cited URLs vs 14.0% of all domains), Breadcrumb, HowTo and QAPage schema, Flesch readability. Product and Recipe schema even score below 1. - ChatGPT rewards question titles and tables, AI Mode a table of contents and H2s, AI Overviews Article schema: the differences are real but small, structure pays everywhere. - The study is a snapshot of the German-speaking market in Q2 2026 and shows correlations: use it to decide what to fix first, not to promise citations. Written by the Geosnap AI agent, reviewed and approved by Rinald Sefa. Translated from the Italian original. [Read the original](https://geosnap.ai/blog/quali-pagine-cita-ai-studio-sistrix-96000-url-cosa-conta-davvero) For two years GEO has run on trial and error: schema or no schema, llms.txt or not, longer or shorter copy. On October 22, 2026 [SISTRIX published a study](https://www.sistrix.com/blog/citation-worthiness-study/) that measures those choices on a large sample for the first time: 48,527 pages cited by ChatGPT, Google AI Mode and AI Overviews, each compared with an uncited page from the same domain. Here is what was measured, which features carry weight, which do not, and what differs across the three engines, with exact figures and the limits the authors themselves state. ## What did the SISTRIX study measure? It compared cited and uncited pages from the same site, to separate page features from domain strength. The sample comes from the [SISTRIX AI Index](https://www.sistrix.com/blog/citation-worthiness-study/) for the German-speaking market (Germany, Austria and Switzerland) in the second quarter of 2026: around 16,000 cited URLs for each of the three engines, 96,418 URLs analysed in total from 17,380 domains, 43 features per page. The method is conditional logistic regression with a correction for multiple testing; the "factors" below are odds ratios, which at a citation rate of about 1.4% are practically the same as "how many times more often a page is cited". Two details help you read the results. 48.7% of cited pages are advice and how-to content, and 38.7% come from brand websites, not publishers: companies are in the game, not watching it. And citation is a rare event, 1.4% of pages. A factor of 1.5 does not mean "half the pages with that feature get cited"; it means that between two similar pages on the same site, the one with the feature is 50% more likely to be cited. ## Which features do AI-cited pages have most often? Structure a machine can parse and trust signals a person can read. Here are the main factors, where 1.0 means no difference from pages without that feature. | Page feature | Factor | | --- | --- | | Current year in the title (2025 or 2026) | 2.7 | | Table of contents | 1.7 | | Article schema | 1.6 | | At least five H2 headings | 1.6 | | FAQPage schema | 1.5 | | Table with a header row | 1.5 | | Bullet lists (ten or more per page) | 1.5 | | At least 1,000 words | 1.5 | | Author box with photo | 1.4 | | At least one H2 phrased as a question | 1.4 | | Visible last-updated date | 1.3 | | At least three schema types | 1.3 | | At least ten external links | 1.3 | | TL;DR at the start | 1.3 | | Meta description present | 1.2 | | Scientific source cited | 1.2 | | Images in the main content | 1.1 | The features add up. Pages that meet many of the eleven the authors combined are cited [2.1 times as often](https://www.sistrix.com/blog/citation-worthiness-study/) as barely structured ones; mid-range pages 1.6 times. The more useful number for you is a different one: only 10% of cited pages actually have many of these features, only 14% have a table of contents and only 5% show an update date. Competition on these elements is still low. The findings match what we described in [answer-format content](https://geosnap.ai/en/blog/contenuti-in-formato-risposta-come-scrivere-pagine-che-chatgpt-e-ai-mode-riprendono-prima-e-dopo): the cited page is the one a model can lift a clean passage from, with a title that says what it covers, an answer up front and the rest clearly separated. ## Why is the year in the title worth 2.7x, and why is it not a trick? Because it measures how current the content is, not the presence of four digits. Pages with "2025" or "2026" in the title are cited 2.7 times as often; pages with only an older year drop to 0.7, which means they are cited less than pages with no year at all. The authors note this analysis was added after the fact and is an initial indication, and they write plainly that it ["is not a title trick"](https://www.sistrix.com/blog/citation-worthiness-study/): the data cannot tell whether simply updating the year produces the effect. The careful reading is this: AI engines prefer sources that claim to be current and prove it. Build the freshness signal on several layers, meaning a visible last-updated date (factor 1.3), a title consistent with that date, and content that has actually changed. A "2026 guide" headline over a 2023 text is exactly the kind of page that lands in the 0.7 group the moment 2027 arrives. ## What does not help: llms.txt, schema for its own sake, readability and length Four common beliefs find no support in the data, and one runs the other way. - **llms.txt.** 11.7% of cited URLs come from domains with an llms.txt file, against 14.0% of all domains in the sample. SISTRIX concludes there is [no measurable effect](https://www.sistrix.com/blog/citation-worthiness-study/) on AI citations and that it should not be a priority. This lines up with what we found in [does llms.txt actually work?](https://geosnap.ai/en/blog/llms-txt-serve-davvero-dati-2026). - **"Schema can't hurt."** Breadcrumb, HowTo and QAPage score 1.0: no difference. Product sits at 0.8 (0.5 in AI Overviews) and Recipe at 0.1. Keep them, they serve other purposes, but do not expect citations from them. This is consistent with [Google's documentation](https://developers.google.com/search/docs/appearance/ai-features), which says explicitly that there is no special schema and no dedicated file to add in order to appear in AI Overviews and AI Mode. To decide which markup to keep, start from [which schema types GEO actually needs](https://geosnap.ai/en/blog/dati-strutturati-schema-geo-quali-servono). - **Readability.** The Flesch index scores 1.03, not significant. Writing simply serves the reader; it does not make the page more citable. - **Length.** Pages with at least 1,000 words score 1.5, but the authors call it a "symptom, not cause": long texts have more H2s, more tables and more lists. Padding a text without adding structure does not reproduce the effect. ## Does anything change between ChatGPT, Google AI Mode and AI Overviews? Little, and in the same direction. The three engines are described as close together, with ChatGPT slightly more sensitive to structure. On ChatGPT the title phrased as a question (1.9), the table with a header row (1.8) and the question H2 (1.6) weigh more. On AI Mode the table of contents and the five H2s reach 1.8. In AI Overviews Article schema is worth 2.0, while product pages drop to 0.5. The study does not cover Gemini, Perplexity or Claude, so for those engines you can only reason by analogy. The differences between site types matter more. On brand sites, Article and FAQPage are worth 1.6 each. On editorial sites the table of contents reaches 2.5 and the update date 1.9. Online shops are the reverse: update date 0.7 and Product schema 0.6, and the authors' advice is to build a guide section with Article schema next to the catalogue, because that is where citations come from. A hiking-gear store, say, gets cited for "how to choose hiking boots" long before it gets cited for a single product page. ## What to do on your site, and what not to promise Start with the pages that already answer questions and add structure, in the order the study measures it. On every important page: a table of contents, at least five H2s with one phrased as a question, Article schema, an author box with name and photo. Where the content allows: an FAQ block with FAQPage schema, a table with a header row, a TL;DR up top, linked sources and a real last-updated date. The authors warn that [none of these should be added artificially](https://www.sistrix.com/blog/citation-worthiness-study/): an FAQ with three made-up questions is not an FAQ. The stakes have grown. According to [Semrush](https://www.semrush.com/blog/ai-overviews-commercial-modifiers-study/), in August 2026 AI Overviews appeared on 88% of 400 US commercial keywords (97.08% when weighted by search volume), up from 45.75% a year earlier: "best", "reviews", "price" and "alternatives" searches almost all carry an AI answer above the results. Being the source of that answer is now part of the commercial job, not just the editorial one. In Geosnap's analysis you can see which pages of your site are cited by ChatGPT, Gemini and Google AI Mode and for which questions: that is the starting point for deciding where to apply this list first. Three caveats, finally. The study shows correlations, not causes: well-structured pages may be cited because they are also the best-maintained ones. It is a snapshot of one language area and one quarter. And the authors themselves show how misleading small samples are: in their test runs FAQPage schema came out at minus 23% with 833 URLs, plus 31% with 4,600 and plus 19% with all 96,418, hence the rule they propose, [don't trust a GEO study with fewer than 5,000 URLs](https://www.sistrix.com/blog/citation-worthiness-study/). That applies to single-page case studies too. ## Sources - [Citation Worthiness: Which GEO Tactics Hold Up Across 96,418 URLs](https://www.sistrix.com/blog/citation-worthiness-study/) (SISTRIX, 2026-10-22) - [AI features and your website](https://developers.google.com/search/docs/appearance/ai-features) (Google Search Central, 2025-12-10) - [AI Overviews have nearly doubled across five commercial search modifiers](https://www.semrush.com/blog/ai-overviews-commercial-modifiers-study/) (Semrush, 2026-10-21) ## Frequently asked questions ### Does the SISTRIX study say FAQPage schema earns citations? It says pages with FAQPage schema are cited 1.5 times as often as similar pages without it, in the DACH sample from Q2 2026. It is a correlation: the authors' advice is to add an FAQ only where the content has real questions to answer, with markup that matches the visible text. ### Should I remove llms.txt if it does not help citations? No. The data says it does not help: 11.7% of cited URLs come from domains with llms.txt against 14.0% of all domains. It does not say it hurts, because the comparison is at domain level. Leave it if it is there, but do not invest time in it or expect results. ### Is putting "2026" in the title enough to be cited 2.7 times as often? The data does not show that and the authors rule it out explicitly. The current year in the title correlates with citation because it signals up-to-date content; a page with only an older year drops to 0.7. Update the content for real, show the update date, and only then align the title. ### Do the results apply to sites in other languages and to Gemini? The study covers pages cited in the German-speaking market by ChatGPT, Google AI Mode and AI Overviews; Gemini, Perplexity and Claude are not included. The features measured (structure, schema, trust signals) do not depend on language, so applying them elsewhere is reasonable, but verify on your own pages by monitoring which ones get cited. ### What is the single most important thing if I can only do one? Structure. That is the authors' own summary: a table of contents, clear and numerous H2s, tables with a header row, and lists. These are the elements with the highest factors and the ones fewest pages already have, so it is where a change shows first.
Structured data
Schema.org data (JSON-LD) embedded in the page code. It tells the AI explicitly what the page is about and who published it.
BreadcrumbList describes the path through the site
{
"@context": "https://schema.org",
"@type": "BreadcrumbList",
"itemListElement": [
{
"@type": "ListItem",
"position": 1,
"name": "Home",
"item": "https://geosnap.ai/en/"
},
{
"@type": "ListItem",
"position": 2,
"name": "Blog",
"item": "https://geosnap.ai/en/blog/"
},
{
"@type": "ListItem",
"position": 3,
"name": "Research and data",
"item": "https://geosnap.ai/en/blog/#ricerca-e-dati"
},
{
"@type": "ListItem",
"position": 4,
"name": "Which pages does AI cite? SISTRIX's 96,000-URL study on what matters",
"item": "https://geosnap.ai/en/blog/quali-pagine-cita-ai-studio-sistrix-96000-url-cosa-conta-davvero"
}
]
}
BlogPosting describes the article: author, dates and sources
{
"@context": "https://schema.org",
"@type": "BlogPosting",
"@id": "https://geosnap.ai/en/blog/quali-pagine-cita-ai-studio-sistrix-96000-url-cosa-conta-davvero#article",
"headline": "Which pages does AI cite? SISTRIX's 96,000-URL study on what matters",
"description": "SISTRIX compared 48,527 pages cited by ChatGPT, AI Mode and AI Overviews with uncited pages from the same sites. The factors that count and the ones that don't.",
"inLanguage": "en",
"url": "https://geosnap.ai/en/blog/quali-pagine-cita-ai-studio-sistrix-96000-url-cosa-conta-davvero",
"mainEntityOfPage": "https://geosnap.ai/en/blog/quali-pagine-cita-ai-studio-sistrix-96000-url-cosa-conta-davvero",
"image": "https://geosnap.ai/img/blog/p2026-11-01-studio-sistrix-pagine-citate-ai-en.webp",
"articleSection": "Research and data",
"wordCount": 1434,
"author": {
"@type": "Person",
"name": "Rinald Sefa",
"jobTitle": "CMO",
"url": "https://geosnap.ai/en/about",
"worksFor": {
"@id": "https://geosnap.ai/#organization"
}
},
"publisher": {
"@id": "https://geosnap.ai/#organization"
},
"isAccessibleForFree": true,
"abstract": "On October 22, 2026 SISTRIX published the largest study so far on what sets AI-cited pages apart: 48,527 URLs cited by ChatGPT, Google AI Mode and AI Overviews, each compared with an uncited page from the same domain, 96,418 URLs in all. Structure matters (table of contents, at least five H2s, tables with a header row, Article and FAQPage schema), and so do trust and freshness signals (author box with photo, visible update date, current year in the title). llms.txt, readability and length on their own do not. These are correlations, not causes, but they are the first measured on a serious sample.",
"datePublished": "2026-11-01",
"dateModified": "2026-11-01",
"citation": [
"https://www.sistrix.com/blog/citation-worthiness-study/",
"https://developers.google.com/search/docs/appearance/ai-features",
"https://www.semrush.com/blog/ai-overviews-commercial-modifiers-study/"
],
"translationOfWork": {
"@id": "https://geosnap.ai/blog/quali-pagine-cita-ai-studio-sistrix-96000-url-cosa-conta-davvero#article"
}
}
FAQPage describes the frequently asked questions
{
"@context": "https://schema.org",
"@type": "FAQPage",
"@id": "https://geosnap.ai/en/blog/quali-pagine-cita-ai-studio-sistrix-96000-url-cosa-conta-davvero#faq",
"mainEntity": [
{
"@type": "Question",
"name": "Does the SISTRIX study say FAQPage schema earns citations?",
"acceptedAnswer": {
"@type": "Answer",
"text": "It says pages with FAQPage schema are cited 1.5 times as often as similar pages without it, in the DACH sample from Q2 2026. It is a correlation: the authors' advice is to add an FAQ only where the content has real questions to answer, with markup that matches the visible text."
}
},
{
"@type": "Question",
"name": "Should I remove llms.txt if it does not help citations?",
"acceptedAnswer": {
"@type": "Answer",
"text": "No. The data says it does not help: 11.7% of cited URLs come from domains with llms.txt against 14.0% of all domains. It does not say it hurts, because the comparison is at domain level. Leave it if it is there, but do not invest time in it or expect results."
}
},
{
"@type": "Question",
"name": "Is putting \"2026\" in the title enough to be cited 2.7 times as often?",
"acceptedAnswer": {
"@type": "Answer",
"text": "The data does not show that and the authors rule it out explicitly. The current year in the title correlates with citation because it signals up-to-date content; a page with only an older year drops to 0.7. Update the content for real, show the update date, and only then align the title."
}
},
{
"@type": "Question",
"name": "Do the results apply to sites in other languages and to Gemini?",
"acceptedAnswer": {
"@type": "Answer",
"text": "The study covers pages cited in the German-speaking market by ChatGPT, Google AI Mode and AI Overviews; Gemini, Perplexity and Claude are not included. The features measured (structure, schema, trust signals) do not depend on language, so applying them elsewhere is reasonable, but verify on your own pages by monitoring which ones get cited."
}
},
{
"@type": "Question",
"name": "What is the single most important thing if I can only do one?",
"acceptedAnswer": {
"@type": "Answer",
"text": "Structure. That is the authors' own summary: a table of contents, clear and numerous H2s, tables with a header row, and lists. These are the elements with the highest factors and the ones fewest pages already have, so it is where a change shows first."
}
}
]
}
Metadata
The information in the page header, which crawlers and engines read before the content.