A 96,418-page study finds no AI citation lift from llms.txt

SISTRIX took 48,527 pages that the three services cited in Germany, Austria and Switzerland in the second quarter of 2026, around 16,000 from each, and paired every one with a page from the same site that was not cited. That gave 96,418 pages from 17,380 domains, each scored on 43 features. Comparing pages within one site takes the site's overall strength out of the result, so what is left is the difference between a site's cited and uncited pages.

On llms.txt the answer was flat. Of all cited pages, 11.7% came from domains with an llms.txt, against 14.0% across all domains. SISTRIX's conclusion: "As things stand, there is no measurable effect on AI citations."

The study also shows how small samples mislead. With 833 pages, FAQ schema was 23% less common on cited pages; with 4,600 pages it was 31% more common. The authors' rule is not to trust a GEO study with fewer than 5,000 URLs.

Pages with a visible table of contents and jump links were cited 1.7 times as often as uncited pages from the same site, yet only 14% of pages had one. Pages with FAQPage schema were cited 1.5 times as often. Pages meeting many of the features together were cited 2.1 times as often as barely optimised pages, and pages in the middle 1.6 times as often. The single highest value went to the current year in the page title, which the authors read as a stand-in for freshness and call an initial indication rather than a confirmed result. Only 5% of pages showed a last-updated date.

Some features pointed the other way. Product schema went with slightly fewer citations, a factor of 0.8 overall and 0.5 in AI Overviews. Pages of 1,000 words or more were cited 1.5 times as often, but the authors say that mostly reflects better structure rather than length.

SISTRIX is clear that these are correlations and not proven causes, and that the data is a snapshot of those three countries. Google's own guidance matches the llms.txt result: its page on AI features says "You don't need to create new machine readable files, AI text files, or markup to appear in these features", and that there is no special schema.org structured data you need to add.

We took the 78 AI Overviews from our Google UK brand search reading of 6 October, which between them cited 100 different domains 314 times, and on 26 October requested the llms.txt file from each of those domains. Ten served one. Sixty-six did not, answering with a not-found page, a redirect or an empty file, and 24 blocked our request or did not answer, so we cannot say either way for them.

Several of the most cited sites have none. Wikipedia, cited in 31 of the 78 overviews, has no llms.txt. Among the price comparison sites, MoneySuperMarket has one and was cited in 10 overviews, while Uswitch, cited in 12, and GoCompare, cited in 10, have none. Of the 314 citations, 23 went to a site with an llms.txt, 207 to a site without one and 84 to sites we could not check.

This is a small, descriptive check on brand-name searches in one engine, and it does not compare cited with uncited pages the way SISTRIX did. It points the same way: most of the sites Google cited had no llms.txt.

Leave an existing llms.txt in place if you have one, but do not expect it to earn citations, and do not put it ahead of the pages themselves. Start with the pages you most want an AI answer to quote, such as a comparison, pricing or advice page. Give each one question-shaped headings, a table where you are comparing things, a table of contents on long pages and a visible date that changes when you genuinely update it. Use FAQPage schema only where the page answers real questions, and keep any structured data matching the visible text, as Google asks.

Then check what happens. Our Google AI Overview checker runs a search your customers make from a clean browser in the UK or US and shows whether an overview appears and whether it cites your website. Our Google AI Overviews tracker page explains how we measure that over time.

The cited sites come from our 6 October 2026 reading, in which we searched the names of 50 UK brands on Google UK three times each, logged out, and recorded the domains each AI Overview cited; that reading's CSV is free to download from its page. On 26 October 2026, at about 06:10 UK time, we requested https:// plus each cited domain and /llms.txt once, following redirects. A domain counted as having a file when it returned a non-empty text file at that address. Sites that answered with an access-denied page, a bot check, a login page or no response were counted as unknown. We checked each site once, 20 days after the overviews were recorded, so a site may have added or removed its file in between. The SISTRIX figures are the study's own and we have not re-run them.