What SISTRIX's 96,418-URL Study Means for SaaS Page Structure and AI Citations

SISTRIX's study of 96,418 URLs found that pages with a clear structure, such as a table of contents, five or more H2 headings, tables and a visible author, were cited more often by AI search than plainer pages from the same site. llms.txt showed no effect. The results are correlations from the DACH region.

Cekpr covers SEO and AI search for B2B software teams. This guide is for content leads who want evidence, not folklore, when they decide how to structure product, comparison and help pages. It explains the study, separates what held up from what did not, and turns the findings into a short test plan.

What did SISTRIX study?

The SISTRIX team published its citation worthiness study on October 22, 2026. It analyzed 96,418 URLs from 17,380 domains and measured 43 features on each page. Of those pages, 48,527 were cited by ChatGPT, AI Mode or AI Overviews in the DACH region during the second quarter of 2026, roughly 16,000 per platform.

The design is the smart part. Each cited page was paired with a non-cited page from the same domain. That isolates page-level choices from domain strength, so a big brand's authority does not explain the gap.

SISTRIX used a Gemini model to classify domain and content types. A second model agreed 84% of the time, and the team checked 100 pages by hand. Results are shown as a citation factor, where 1.0 means no difference between cited and uncited pages.

Which tactics held up?

Several structure and clarity features scored above 1.0. A current year in the title scored 2.7, though the authors describe that as an initial, retrospective indication. A table of contents scored 1.7, and only 14% of pages in the sample had one.

Article schema and five or more H2 headings each scored 1.6. FAQPage schema and a table with a header row each scored 1.5. An author box with a photo scored 1.4, as did at least one question phrased as an H2.

Pages that combined many of these features were cited 2.1 times as often as barely optimized ones. The study stresses that word count's apparent benefit came from better structure, not from length itself.

Which tactics did nothing?

The null results matter as much as the wins. The file llms.txt showed no measurable effect. About 11.7% of cited URLs came from domains that had it, against 14.0% across all domains. Breadcrumb, HowTo and QAPage schema scored 1.0, meaning no difference.

Some features scored below 1.0. Recipe schema scored 0.1, and Product schema scored 0.8, so pages with those types were cited slightly less often. Readability, measured by the Flesch index, scored 1.03, which the authors call not significant.

This lines up with Google's guidance. The company's AI features documentation says you do not need new machine-readable files, AI text files or special structured data to appear in AI Overviews or AI Mode. Our earlier piece on whether your SaaS site needs llms.txt reached the same practical conclusion.

Do platforms differ?

Yes, and the differences are useful. ChatGPT favored question-style titles, with a factor of 1.9, and tables with header rows, at 1.8. AI Mode rewarded clean basics, with a table of contents or five or more H2s each scoring 1.8. AI Overviews favored Article schema at 2.0 and cited product pages less often, at 0.5.

Domain type changed things too. On e-commerce sites, a visible last-updated date was linked to fewer citations, at 0.7. That is a reminder that a pattern in one context can reverse in another. B2B software pages are closer to the informational end, but your own data should decide.

Sample size affects trust as well. The authors showed that FAQ schema results flipped, negative at 833 URLs and positive at 4,600 and 96,418. They advise against trusting GEO studies with fewer than 5,000 URLs.

What are the limits?

There are four. The findings are correlations, so a table may travel with better writing rather than cause citations. The data covers only the DACH region. Perplexity, Claude and Gemini chat were left out. Subgroup results rest on fewer URLs and need more caution.

Cekpr editors treat the study as a source of hypotheses. That is a respectable use. A good hypothesis tells you what to test next, and it costs little to try on a handful of pages.

It also helps to compare it with other evidence. Our write-up of Ahrefs' 1,885-page test of schema markup and AI citations found a mixed picture too, and the two studies point in the same cautious direction.

How should a SaaS team use the findings?

Start with structure, since it is cheap and good for readers. Add a short table of contents to long guides and comparison pages. Use at least five clear H2 headings with descriptive names. Phrase at least one heading as a real question, and answer it in two or three sentences directly below.

Add a table where you compare things. Give it a header row and the same columns for each product or plan. Show an author with a name, photo and role, and keep the bio specific.

Next, handle dates. A visible last-updated date appeared on only 5% of pages and scored 1.3. For software content, where prices and features change, a visible date is also honest. Pair it with a real review process, as our guide on running a content review on a quarterly and annual schedule describes.

Google's Article structured data guidance lists no required properties and recommends author, datePublished, dateModified, headline and image. It says the markup helps Google understand a page and affects how titles, images and dates may appear. It does not claim Article markup changes rankings. Add it if you want clean date and author data, and keep it consistent with the visible page.

Should you add a TL;DR to every page?

The study found that a short summary at the start scored 1.3. That is a modest link, and it fits how people read. A buyer scanning a comparison page wants the answer in the first few lines.

Write the summary as plain sentences, not as a teaser. Name the product, the audience and the main fact. For a pricing guide, that might be the plan names, the starting price and the limit that forces an upgrade.

Keep it honest and dated. A summary that says "as of October 2026" gives a reader and a machine a clear anchor. When the facts change, update the summary first, then the sections below it, so the top of the page never contradicts the rest.

What should you not do?

Do not stuff a year into every title. The 2.7 factor is a retrospective hint, and an out-of-date year hurts trust. Use the year only where the page truly covers that year, and update it when you update the content.

Do not add schema types you cannot support. Markup must match the visible page, according to Google's guidance. Do not buy a tool that promises citations through a text file or a special tag.

Do not rewrite your whole site on one study. A change that helps one page type can hurt another. Our guide on fact-checking AI-assisted content under Google's 2026 guidance is a reminder that accuracy beats formatting every time.

How do you test a change?

Pick ten pages that are similar in type and traffic. Change five and leave five alone. Record the starting state: which queries trigger an overview, who is cited and whether your page appears.

Keep the test small and honest. If the changed pages gain and the controls do not, you have a signal. If both move together, something else is going on, and you should not credit your edits.

What is the bottom line?

Structure that helps readers also correlates with citations, and tricks that target machines show nothing. That is a useful message for a content team with limited time. Spend effort on clear headings, honest dates, specific tables and accurate claims.

Treat the SISTRIX numbers as a starting point for your own tests. Cekpr will report back when new studies cover the United States and other platforms, so you can compare the results.