Pages carrying a contents list earned citations from ChatGPT, AI Mode, and AI Overviews at 1.7 times the rate of same-site pages lacking one, according to a SISTRIX study published Oct. 22. That was the strongest structural signal among the 43 page features tested. SISTRIX sells SEO and AI-visibility software, so this is a vendor study, and the post closes by pointing readers to the company’s own citation-checking tool.

The design is what sets it apart from most GEO advice. SISTRIX drew 48,527 cited URLs from its AI Index, limited to the DACH markets (Germany, Austria, and Switzerland) and gathered during Q2 2026, with roughly 16,000 per platform. Each was paired with an uncited page on the same domain, giving 96,418 URLs across 17,380 domains. Because both pages share a site, brand strength drops out and only page-level traits remain.

The company reports its figures as odds ratios from conditional logistic regression, with a Benjamini-Hochberg correction for testing 43 features at once. At a citation rate near 1.4 percent, SISTRIX says, these behave almost like plain frequency ratios. A factor of 1.0 means no difference.

The features that scored best:

Treat the year-in-title figure cautiously. SISTRIX says it analysed the title and H1 traits retroactively, outside the original plan, calls them initial indications, and cannot say whether merely changing the year is enough. Combined, eleven features produced a 2.1 factor for heavily optimised pages against barely optimised ones, and 1.6 for mid-range pages. Just one in ten cited pages currently satisfies a large share of the eleven.

Several popular tactics fared poorly. For llms.txt, the one test run across domains instead of within them, 11.7 percent of cited URLs came from domains with the file, against a 14.0 percent share of all domains. SISTRIX concludes there is no measurable citation effect, while noting that those domains differ in other ways. Breadcrumb, HowTo, and QAPage markup each scored 1.0. Recipe schema scored 0.1. Product schema scored 0.8, and 0.5 for AI Overviews. A good Flesch readability score was not significant at 1.03.

Length also gets a discount. Pages of 1,000 words or more scored 1.5, but SISTRIX attributes that mostly to the extra headings, tables, and lists that long pages tend to carry.

Platform differences are modest. ChatGPT favored question-shaped titles at 1.9 and header-row tables at 1.8. AI Mode rewarded a table of contents or five or more H2 headings, each at 1.8. AI Overviews gave Article schema 2.0 and product pages 0.5. SISTRIX had expected AI Overviews to react most to optimized pages, and says the data showed all three close together.

Site type changes the picture more. Editorial sites saw 2.5 for a table of contents and 1.9 for a last-updated date. Online shops saw 0.7 for the date and 0.6 for Product schema. Product schema scored 1.7 for SaaS providers and 0.1 for government sites. The subgroup counts are smaller, and SISTRIX warns they deserve extra caution.

The limits are stated plainly. The data shows correlation, not cause, covers one region, and leaves out Perplexity, Claude, and Gemini chat. SISTRIX also shows how sample size can mislead: in an 833-URL run, FAQ schema appeared 23 percent less often among cited pages. At 4,600 URLs it appeared 31 percent more often, and across the full set 19 percent more often. Its rule is to distrust GEO studies under 5,000 URLs.

What the source does not address is direction. A cited page may have been better built to begin with, so adding a table of contents to a weak page may not move it. Teams can test that cheaply: add a table of contents and Article schema to one set of templates, hold back a matched control group, and compare citation counts over the next ninety days before rolling it out sitewide.

Source: the SISTRIX blog, “Citation Worthiness: Which GEO Tactics Hold Up Across 96,418 URLs,” by Johannes Beus, Julia Weissbach, and the SISTRIX Content Team, published Oct. 22, 2026.