Correction 3 November 2026. The count of four external links covers editorial links to sources. Pages also carry share buttons, which are template chrome marked rel=”nofollow”; those are not what the study measures.
SISTRIX compared 48,527 cited pages against a non-cited page from the same domain and scored 43 on-page features. We ran their checklist against our own 27 articles. We clear eight of the eleven headline items and fail three, and all three are things our own editorial rules forbid.
Table of contents
- What this is about
- What the study says
- What we measured
- Where it matters and where it does not
- How to verify on your own site
- The numbers
- FAQ
Key takeaways
- SISTRIX scored 96,418 URLs from 17,380 domains, pairing each cited page with a non-cited page from the same domain so brand strength drops out of the comparison. Published 22 October 2026.
- The largest associations are structural: a visible table of contents 1.7x, Article schema 1.6x, FAQPage schema 1.5x. Pages meeting many of the eleven headline features are cited 2.1 times as often.
- llms.txt shows no measurable effect. Breadcrumb, HowTo and QAPage markup show none either, and Recipe schema is associated with a tenth of the citation chance.
- Their own sample-size test is the most useful part. At 833 URLs, FAQ schema looked 23% less common on cited pages; at 4,600 URLs, 31% more common; at 96,418, 19% more. Results only settled above about 5,000 URLs.
- On our 27 articles we clear eight of eleven. The three we fail — outbound links, questions as headings, the year in the title — are consequences of our own section rules, not oversights.
What this is about
A large study tested which on-page features go with being cited by ChatGPT, AI Mode and AI Overviews. The findings are correlations within a domain, not proven causes. If your pages already carry a table of contents and Article schema, there is nothing to do. If they do not, the fix is template work, not strategy.
What the study says
The analysis covers the second quarter of 2026 in the DACH region, drawn from the SISTRIX AI Index: 48,527 cited URLs, roughly 16,000 each from ChatGPT, AI Mode and AI Overviews. Each cited page is paired with a non-cited page from the same domain, which is the design decision that makes the rest worth reading. In the authors’ words, that pairing means “the strength of the domain and brand plays no role, leaving only what comes down to the page itself.”
Figures are reported as factors: a factor of 1.7 means that of two pages on the same domain, the one with the feature is cited 1.7 times as often. The authors state these are odds ratios, practically equivalent to the frequency ratio at a citation rate of about 1.4%. The model is conditional logistic regression with Benjamini-Hochberg correction across the 43 features.
The structural signals lead, and the table below carries the full list. A visible table of contents is the largest of them at 1.7x, and only 14% of pages have one. Article schema is 1.6x. FAQPage schema is 1.5x, holding across all three platforms and every domain type except encyclopaedias — after Google dropped FAQ rich results. The trust signals are smaller and consistent: an author box with a photo 1.4x, a visible last-updated date 1.3x on the 5% of pages that show one.
The highest single figure, the current year in the title at 2.7x, is also the least certain. The authors added that feature after the analysis was planned, call it an initial indication, and say plainly that the year most likely stands in for content that was actually refreshed.
Combined, pages meeting many of eleven features are cited 2.1 times as often as barely optimised ones, mid-range pages 1.6 times. Only 10% of cited pages already meet many of them.
The negative results carry as much weight. llms.txt: no measurable advantage. Breadcrumb, HowTo and QAPage markup: no effect. Recipe schema: 0.1x. Word count above 1,000 words scores 1.5x, which the authors attribute to longer pages carrying more structure rather than to length itself.
What we measured
We ran their eleven headline features against every article on this site: 27 pages, fetched 3 November 2026, parsed from the rendered HTML rather than from the CMS, with schema read out of the JSON-LD blocks.
Eight come out at or near 100%: Article schema, five or more H2 headings, a table, ten or more list items, a thousand words or more, an author in the markup, an H1 of five words or more, and a TL;DR at the top — which in our template is the Key takeaways block. A table of contents is on 24 of 27, FAQPage schema on 26 of 27, a visible last-updated date on 22. BreadcrumbList is on all 27, which the study says does nothing for citations; we keep it because it is for readers and for the results page, not for this.
Three fail, and none of them is an accident.
Outbound links: zero pages of 27 clear the ten-link threshold. The entire site carries four editorial links to sources; the other outbound links on a page are share buttons. This is not neglect — our section rules say “No outbound links. Name the publisher and author in the sentence that makes the claim.” We chose that to keep attribution in the prose where a reader sees it. The study puts ten or more external links at 1.3x, which is at the low end of its range, and it cannot tell us whether the links cause anything or merely mark the kind of page that cites its sources. We are keeping the rule and noting the cost.
Questions as H2 headings: 3 of 27. Our section prompts fix the heading sequence — “What this is about”, “What the source says”, “What the data shows” — so a question only appears inside the FAQ block, which sits in its own markup. The study puts a question H2 at 1.4x overall and 1.6x on ChatGPT specifically, and unlike the heading-count figure it does not depend on text length. This one is cheap to fix and we have not decided to.
The year in the title: 3 of 27. Our titles avoid the year unless the data is tied to it. This is the study’s largest figure and its least certain one, and the authors say outright that they cannot tell whether updating the year alone does anything. Writing 2026 into a headline about a 2026 dataset is honest; writing it into everything is the oldest trick in the catalogue, and we are not doing that.
What this check cannot tell us is whether any of it matters for this site. We have four pages in Google’s index out of 41 submitted and 475 impressions in 90 days, which is the state the section reference describes as the starting line, not the finish. The conditions under which a 1.3x factor would become visible here do not exist yet — the threshold for measuring any of this properly is higher than most sites clear.
Where it matters and where it does not
It matters most for editorial sites, where a table of contents goes with 2.5x and a visible last-updated date with 1.9x — both well above the all-site figures.
It inverts for shops. A visible last-updated date is associated with fewer citations there, 0.7x, Product schema with 0.6x, and AI Overviews cite product pages at 0.5x. The authors’ own reading is that a shop’s advice section, marked up as articles, is a better target than its product pages.
It is close to meaningless for some types. Product schema runs 1.7x for SaaS providers and 0.1x for government sites; for encyclopaedias the combined optimisations show no effect at all. These subgroup figures rest on far fewer URLs than the headline numbers, which the authors flag.
The platforms differ less than expected. The authors say they expected AI Overviews to reward optimised pages most and found all three close, ChatGPT narrowly ahead. ChatGPT favours question-shaped structure, AI Mode plain basics, AI Overviews Article schema at 2.0x.
And the method check. Of the seven points we require of vendor research, this study clears four. It names the three engines, states the period as the second quarter of 2026, states the region as DACH, and documents the statistics down to the correction for multiple testing. It does not say whether citations were collected through each product’s interface or an API, it does not publish the prompt set behind the index, and it reports a single snapshot rather than repeated runs. It is also explicit that these are correlations and that Perplexity, Claude and Gemini chat were not included. That is more disclosure than most of what we read, and it is why the figures are worth quoting at all — though connected markup across sites looked thinner when we checked it ourselves.
How to verify on your own site
Check the schema that the study found does something, and ignore the rest. Fetch the rendered page and read the JSON-LD types:
curl -sL https://example.com/page/ | grep -o '"@type": "[^"]"' | sort | uniq -c
Article and FAQPage are the two associated with more citations. BreadcrumbList, HowTo and QAPage showed no effect in this dataset, so finding them tells you nothing either way about citations.
For the table of contents, confirm it is in the HTML rather than built by script after load, because a crawler that does not execute your JavaScript will not see it:
curl -sL https://example.com/page/ | grep -c 'href="#'
Count the in-page anchors. If the number is zero in the raw HTML but the page shows a table of contents in your browser, the markup is being generated client-side.
For the last-updated date, check that it is visible text and not only a dateModified value in the JSON-LD. The study measured a visible date, which is a different thing.
The numbers
| Feature | Citation factor | Share of pages that have it |
|---|---|---|
| Current year in the title | 2.7x | — |
| Table of contents | 1.7x | 14% |
| Article schema | 1.6x | — |
| At least five H2 headings | 1.6x | — |
| FAQPage schema | 1.5x | — |
| Table with a header row | 1.5x | — |
| Ten or more bullet list items | 1.5x | — |
| At least 1,000 words | 1.5x | — |
| Author box with photo | 1.4x | — |
| Question as an H2 | 1.4x | — |
| Visible last-updated date | 1.3x | 5% |
| Three or more schema types | 1.3x | — |
| Ten or more external links | 1.3x | — |
| TL;DR at the beginning | 1.3x | — |
| llms.txt | no measurable effect | — |
| Recipe schema | 0.1x | — |
SISTRIX, published 22 October 2026. 96,418 URLs from 17,380 domains, DACH region, second quarter of 2026; 48,527 cited pages each paired with a non-cited page from the same domain. Conditional logistic regression with Benjamini-Hochberg correction. The year-in-title and H1 figures were added after the analysis was planned and the authors label them initial indications. The study does not state whether citations were collected through each product’s interface or an API, and does not publish its prompt set.
| Our 27 articles against the eleven headline features | Pages |
|---|---|
| Article schema | 27 of 27 |
| Five or more H2 headings | 27 of 27 |
| Ten or more list items | 27 of 27 |
| At least 1,000 words | 27 of 27 |
| Author in the markup | 27 of 27 |
| TL;DR at the top | 27 of 27 |
| FAQPage schema | 26 of 27 |
| Table | 25 of 27 |
| Table of contents | 24 of 27 |
| Visible last-updated date | 22 of 27 |
| Question as an H2 | 3 of 27 |
| Year in the title | 3 of 27 |
| Ten or more external links | 0 of 27 |
Every article on this site, fetched 3 November 2026, parsed from the rendered HTML; schema read from the JSON-LD blocks. Counting editorial links only: the whole site carries four, because our section rules forbid them. Share buttons are excluded — they are template chrome and carry rel=”nofollow”.
FAQ
Is a table of contents worth adding just for AI citations?
It is the largest structural association in the study at 1.7x, rising to 2.5x on editorial sites, and only 14% of pages have one. It is also template work: build it once, every page gets it. The caveat is that this is a correlation within a domain, and a table of contents tends to accompany pages that are long and organised anyway. If your pages are long enough to need one, add it. If they are not, adding one changes nothing.
Should I add FAQ schema even though Google dropped FAQ rich results?
The study found FAQPage schema at 1.5x across all three platforms and every domain type except encyclopaedias — measured after the rich result went away. The condition the authors state is that the questions be real rather than token ones. Nothing here says the markup causes the citation; it may simply mark pages that answer questions directly. Adding genuine questions and marking them up costs little. Adding fake ones is exactly what the caveat is aimed at.
Why do these studies disagree with each other so often?
Sample size, and this study documents it better than most. In their own test runs, FAQ schema looked 23% less common on cited pages at 833 URLs, 31% more common at 4,600, and 19% more common at 96,418. The direction only became stable above roughly 5,000 URLs. That gives a usable filter for everything else you read: a GEO study built on a few hundred pages can land on either side of zero by chance, and the authors apply the rule to their own subgroup figures too.
