Most advice on getting cited by AI answers is a list of tactics with no comparison group behind it. A study published by SISTRIX on October 22 tries to fix that, and some of the popular tactics do not survive.
The team took pages that ChatGPT, Google AI Mode and AI Overviews actually cited, and compared each one with a page from the same website that was not cited. Because both pages share a domain, the brand's size and authority cancel out. What remains is the page itself.
How the study worked
SISTRIX pulled 48,527 cited URLs from its AI Index for the German-speaking market (Germany, Austria, Switzerland), collected in the second quarter of 2026. Roughly 16,000 came from each of the three engines. Each was paired with an uncited page from the same domain. The full data set came to 96,418 URLs from 17,380 domains.
For every URL they recorded 43 features: headings, schema types, lists, tables, author details, dates, link counts and more. They then used conditional logistic regression, with a correction for testing many features at once. The result is a factor for each feature. A factor of 1.0 means cited and uncited pages look the same on it.
The authors are clear about the limits. These are correlations, not proof. The data covers one region, and Perplexity, Claude and Gemini chat are not in it.
What the cited pages had
The strongest single signal was a current year in the title, with a factor of 2.7. A title that carried only an older year scored 0.7, so a stale date seems to hurt.
Structure came next. A table of contents scored 1.7, and only 14% of pages had one. Article schema scored 1.6, as did five or more H2 headings. FAQPage schema, a table with a header row and ten or more bullet lists each scored 1.5. A question used as an H2 scored 1.4, and a short summary at the top scored 1.3.
Trust signals mattered too. An author box with a photo scored 1.4. A visible last-updated date scored 1.3, yet only 5% of pages showed one. Ten or more outbound links also scored 1.3.
Taken together, pages that met many of eleven combined features were cited 2.1 times as often as barely optimized pages. Only 10% of cited pages were in that group.
What did not help
Some favourites came out flat. Breadcrumb, HowTo and QAPage schema all scored 1.0. Readability scores made no significant difference.
llms.txt showed no measurable effect. In fact, 11.7% of cited URLs came from domains that had the file, against 14.0% of all domains in the sample. Product schema scored 0.8 overall and 0.5 in AI Overviews, which leans toward informational pages over product pages.
There is also a warning about small studies. FAQ schema looked harmful in a test run of 833 URLs, helpful at 4,600 and mildly helpful at the full 96,418. SISTRIX says anything under about 5,000 URLs is unreliable. That covers a lot of the GEO studies passed around on LinkedIn.
Each engine has its own taste
The engines do not reward the same things. ChatGPT likes questions: question-style titles scored 1.9, header-row tables 1.8 and question H2s 1.6. AI Mode rewards plain structure, with a table of contents and five or more H2s each at 1.8. AI Overviews leans on Article schema at 2.0.
Site type changes the picture as well. On SaaS sites, Product schema scored 1.7, the opposite of its overall result. On editorial sites, a table of contents reached 2.5 and a last-updated date 1.9. For a B2B software company, that suggests a clear product page and a separate set of dated, well-structured explainer pages, rather than one page trying to do both.
The bigger picture
The useful thing about this study is its design, not any single number. Comparing pages from the same domain asks the question a company can act on: given the site we already have, which of our pages gets picked, and why? It also explains why rankings in AI answers can feel random. Each engine is choosing between pages on slightly different grounds.
The practical reading is short: put the year in the title and keep it current, and show both the date you last updated the page and who wrote it. Long pages get a table of contents and question-style headings. Do not expect llms.txt to do the work. Then measure it on your own buyer questions, because a German-market correlation is a starting point, not a result for your category.
ClawWorld runs AI visibility work for B2B startups. Start with a free baseline report: 10 buyer questions, who ChatGPT and Claude name instead of you, and the pages they cite.