Does llms.txt actually work?
Short answer: there is no evidence it does. We run an AI visibility audit and we check for llms.txt, so this is not a comfortable thing for us to publish. But the data is what it is, and you should know what you are spending an afternoon on.
What the evidence actually says
Four independent things point the same way, and none of them are ambiguous. Google's Gary Illyes confirmed in July 2025 that Google does not support llms.txt and has no plans to; John Mueller compared it to the keywords meta tag, which is about as damning as Google gets. OpenAI's crawler documentation controls its bots through robots.txt and never mentions llms.txt at all. Ahrefs looked at 137,000 sites in May 2026 and found that 97% of llms.txt files received zero traffic. SE Ranking checked roughly 300,000 domains and found no statistically significant correlation between having the file and how often a domain gets cited in AI answers. Crawler log analysis rounds it out: GPTBot, ClaudeBot and PerplexityBot almost never request the file, and they do not probe for it on sites that lack it.
Why we still check for it
Because it costs almost nothing and the downside is zero. If the convention gains adoption, the file is already there. What we changed is how we report it. A missing llms.txt is not a critical finding and we no longer treat it as one. It is a cheap early-mover bet, and calling it anything more than that would be selling you something the evidence does not support.
Structured data does not move AI citation either
This one surprised us more. Ahrefs ran a matched difference-in-differences study on 1,885 pages that added JSON-LD schema between August 2025 and March 2026, compared against roughly 4,000 control pages with similar prior citation levels. Google AI Overviews went down 4.6%, a small but statistically significant decline. AI Mode went up 2.4% and ChatGPT up 2.2%, both indistinguishable from zero. No platform showed a meaningful uplift. The likely mechanism is simple: AI retrieval systems extract the visible HTML when deciding what to cite, so markup hidden in a script tag is not in front of the model at the moment the decision gets made. Keep schema. It is still what makes the rich results that survive possible, and it helps search engines interpret your pages. Just do not buy it as an AI citation strategy.
Some rich results you may still be chasing no longer exist
Google deprecated FAQ rich results on 7 May 2026. Search Console reporting for them ends in June 2026 and API support in August 2026. FAQPage remains a valid schema type and is harmless to keep, but it no longer produces the expandable question-and-answer listing that made it worth adding. HowTo rich results were dropped for desktop earlier. Speakable was never broadly active. If a tool is still telling you to add FAQPage markup to earn a rich result, it is working from a world that ended in May.
What the same research says does correlate
Brand mentions. Across the studies, how often a brand is mentioned across the web correlates with AI visibility more strongly than backlinks or domain authority do. Worth separating two things that sound alike: being mentioned in an AI answer and having your own site cited as the source are different outcomes, and on Gemini the overlap between mentioned brands and cited domains can be as low as 30%. The ranking connection has also weakened considerably. Only 38% of AI Overview citations now come from pages ranking in Google's top ten, down from 76% in mid-2025, and ranking first gives roughly a one in three chance of being cited. Earned presence across sources an AI engine can cross-check against is doing the work, not markup on your own domain.
What to do with an afternoon
Check that you are not blocking the crawlers that matter, because that is the one technical signal with a guaranteed effect: if GPTBot or ClaudeBot or PerplexityBot cannot fetch your page, you cannot be cited from it, and robots.txt rules for named user-agents are genuinely honoured. Then put the remaining time into being mentioned somewhere you do not control. Directory listings, review platforms, industry press, a Wikidata entry if you qualify. That is slower and less satisfying than adding a file to your web root, which is probably why the file got so popular.
Frequently asked questions
Should I delete my llms.txt file?
No. It costs nothing to keep and it is already there. If the convention gains adoption you are covered. Just do not expect it to be doing anything for you today, and do not prioritise creating one over work that has evidence behind it.
Does Google read llms.txt?
No. Gary Illyes confirmed in July 2025 that Google does not support it and has no plans to, and John Mueller compared it to the keywords meta tag. Creating the file neither helps nor harms your Google rankings or your visibility in AI Overviews.
If schema does not help AI citation, should I remove it?
No. Schema still helps search engines interpret your pages and it is what makes the rich results that still exist possible, like star ratings and product listings. The finding is narrower than it sounds: adding JSON-LD does not increase how often AI systems cite you. Keep it for what it actually does.
What is the single highest-value thing for AI visibility?
Being mentioned across sources you do not control, so an AI engine has something independent to cross-check you against. Before that, confirm you are not blocking AI crawlers in robots.txt, because that is a hard gate: blocked means uncitable regardless of everything else.
Why is an audit tool publishing this?
Because we check llms.txt in our AI visibility suite and we would rather tell you what the evidence says than let the finding imply more than it should. We corrected our own check descriptions in September 2026 for exactly this reason.