--- title: "Does llms.txt Work? 96,418 URLs Say No. Your Table of Contents Beat It." description: "Does llms.txt work for AI search? A 96,418-URL study says no. Here is what actually earns citations in ChatGPT, AI Mode and AI Overviews, and what to fix." canonical: https://www.561media.com/blog/llms-txt-ai-citations-study source: 561 Media (https://www.561media.com) author: "561 Media" published: "2026-10-25" category: "SEO" tags: "llms.txt, generative engine optimization, AI Overviews, AI search citations, SISTRIX" readingTime: "6 min read" --- # Does llms.txt Work? 96,418 URLs Say No. Your Table of Contents Beat It. > SISTRIX tested 96,418 URLs against ChatGPT, AI Mode and AI Overviews citations. llms.txt showed no measurable effect, while a plain table of contents earned a 1.7x lift. On October 22, SISTRIX published the largest public test of generative engine optimization tactics anyone has released: 96,418 URLs, 17,380 domains, 43 page features each, measured against whether ChatGPT, Google AI Mode and AI Overviews cited the page. The headline result is the one the industry will not enjoy. llms.txt, the file a year of conference talks told you to install, showed no measurable effect. A table of contents, which only 14% of pages have, showed a 1.7x lift. So does llms.txt work? On this evidence, no. That is not the story. The story is in the methodology, and it explains why you got the wrong list all year. ### TL;DR - SISTRIX compared 48,527 cited URLs against non-cited pages on the same domains, stripping out brand strength to isolate the page itself. - llms.txt showed no measurable benefit. Domains with the file supplied 11.7% of cited URLs while making up 14.0% of all domains in the sample. - The winners are unglamorous: table of contents 1.7x, Article schema 1.6x, five or more H2s 1.6x, a table with a header row 1.5x, a TL;DR block 1.3x. Stacked, 2.1x. - The buried finding: run at 833, 4,600 and 96,418 URLs, the same analysis made FAQ schema reverse direction. SISTRIX's rule of thumb is to distrust any GEO study under 5,000 URLs. - Plenty of the GEO tactic lists circulating this year were built below that line. ### Why the Study Design Matters More Than the Results Most AI visibility research asks which sites get cited. That question answers itself. Wikipedia, Reddit and Forbes get cited because they are Wikipedia, Reddit and Forbes. That measures brand strength and calls it optimization. SISTRIX did something harder. For every cited page they pulled a non-cited page from the same domain. Same brand, same authority, same backlinks. The only variable left is the page. That makes this the first GEO study whose results could plausibly apply to a 40-page home services site in Boca Raton rather than only to a national publisher. Two limits up front. The sample comes from SISTRIX's AI Index for the DACH region (Germany, Austria and Switzerland) in the second quarter of 2026, with no English-language replication yet, and the results are correlation, not proof. They come as odds ratios, where 1.0 means no difference. ### Does llms.txt Work? The Data Says No, and Google Said So in June llms.txt came in at no measurable effect. Of all cited URLs, 11.7% came from domains carrying the file, while those domains made up 14.0% of the sample. SISTRIX is careful to say this does not show the file does any harm, and so are we. Sites that install experimental files tend to be sites that chase tactics. None of this should be a surprise, which is the uncomfortable part. Google's John Mueller called llms.txt "purely speculative for now" in a Reddit reply reported by Search Engine Journal on June 2, suggesting none of the AI systems use it. Two weeks later, on Google's Search Off the Record podcast, he explained why the design cannot work: "So in LLM system, it basically, by design, can't trust what is here as a way of differentiating between different websites." A file where you tell the machine what to think of you cannot be an input. Knowable in June. Sold anyway. ### The Tactics That Won Are Free, Boring, and Already in Your CMS Every figure below is from the SISTRIX study: - Table of contents: 1.7x, and only 14% of pages have one - Article schema: 1.6x - At least five H2 headings: 1.6x - A table with a header row: 1.5x - Author box with a photo: 1.4x - At least one question used as an H2: 1.4x - Visible last-updated date: 1.3x, and only 5% of pages show one - TL;DR at the top: 1.3x Pages combining many of these hit 2.1x against barely structured pages. Not one requires a new vendor, a new platform, or code your developer has not written before. That is not a moat. That is a to-do list, and it is template work any Boca Raton SEO program should finish in weeks, not quarters. ### The Finding Nobody Covered: Most GEO Advice Came From Samples Too Small To Mean Anything SISTRIX ran their analysis three times. At 833 URLs, FAQ schema was 23% less common on cited pages. At 4,600 URLs it was 31% more common. Only from around 5,000 URLs did the result reliably point the same way. Their rule of thumb: don't trust a GEO study with fewer than 5,000 URLs. Sit with that for a second. Plenty of the GEO tactic lists circulating this year were built on a few hundred pages, or a few thousand at best. By the standard of the people who ran the biggest version of the test, those were coin flips presented as findings. So ask how many URLs. Under 5,000 and the tactic is a hypothesis, with you as the test. ### If You Sell Products, the Answer Changes Product schema came in at 0.8x and Recipe schema at 0.1x. AI Overviews actively disfavored product pages at 0.5x while rewarding Article schema at 2.0x. For online shops a visible last-updated date scored 0.7x, the reverse of editorial sites, where a table of contents hit 2.5x. So polishing product pages for AI citation is the wrong project. The citation-earning asset is advice content marked up as an article: buying guides, comparisons, sizing and fit, the questions your support inbox answers twelve times a week. ### What To Do This Week - Count your headings. Open your top 20 organic landing pages and count H2s. Anything under five gets restructured into five or more, with at least one phrased as a question your customer would say out loud. A headings pass, not a rewrite. - Fix the template, not the posts. Add a table of contents and a visible last-updated date to your blog template, so every post inherits both. Then confirm Article schema is firing on posts rather than Product or WebPage, using Google's Rich Results Test. - Audit your AEO invoice against the 5,000-URL rule. Find the research your vendor or playbook is based on and check the sample size. Under 5,000 URLs, or no study at all, and that budget belongs on the items above. ### FAQ #### Should I delete my llms.txt file? No. It costs nothing to leave in place and the specification could be adopted later. Just stop counting it as work completed, and stop paying anyone to maintain it. #### Does this mean AI search optimization is worth doing at all? Yes, and it got cheaper. The tactics with real evidence behind them are structural: headings, tables, a table of contents, Article schema, a dated byline. A template project with an end, not an open-ended retainer. #### My pages already rank well in Google. Do I need to do anything differently? Probably less than you think. The features that correlate with citations are the ones that make a page easy to skim and parse. If your pages are 600 words of unbroken prose under one heading, that is the gap. #### How long until I know whether this worked? Change ten pages, leave ten comparable pages alone, and compare over 60 to 90 days, tracking AI referral traffic separately from organic. Then make sure that traffic lands somewhere built to convert it, which is the job of your funnel architecture. The AI search era has produced an enormous amount of advice and very little evidence. This is the first study large enough to argue with, and what it found is almost disappointing: structure your pages well, date them, name the author, use real headings and real tables. The file you were told to install does nothing. The table of contents you skipped does something. That is not a technology problem. It is a discipline problem, the same gap behind the finding that most businesses use AI but under 9% actually deploy it. Book a 30-minute strategy call. No pitch, just an honest read on whether your pages are built for how people find you now. Get a free strategy call. --- Contact 561 Media: https://www.561media.com/contact