llms.txt Doesn't Work: What Actually Gets You Cited by AI

Publishing an llms.txt file became the trendy ritual for “getting ready for AI” back in 2025: a Markdown list at the domain root, meant to give ChatGPT, Claude, or Perplexity a shortcut before they read the rest of the site. Ahrefs just analyzed 137,000 domains to check whether anyone actually uses it, and the answer is blunt: 97% of llms.txt files got zero requests in the entire month studied. No bots, no humans, nothing. At Evicron, an AI and custom software studio based in Barcelona, that doesn’t fully surprise us — but it does change what we tell clients to prioritize if they want to show up in an AI assistant’s answers.

What llms.txt is, and why it caught on

The idea, proposed in late 2024, was simple: just as robots.txt gives search engines instructions, an llms.txt would hand language models a curated summary of a site — what the company does, which pages matter, what to skip — sparing them from parsing the full HTML. Tens of thousands of sites adopted it without any AI lab formally requesting it or confirming it’s used for training or live answers. That’s exactly the assumption Ahrefs’ study puts to the test with real traffic data instead of marketing promises.

The number: 137,000 domains, almost nobody reads it

Of the domains analyzed, around 38,000 had a valid llms.txt. Of those, 97% logged zero requests in May 2026. Of the remaining 3% that did see traffic, the breakdown is telling: 96% of those requests came from bots that aren’t generative AI at all. Retrieval bots tied to ChatGPT and Perplexity — the ones that, in theory, should be using it — made up barely 1% of requests. The rest split between SEO audit tools (21%), unidentified bots (14%), classic crawlers like Googlebot (13%), and tech-profiling tools like BuiltWith (11%). Adding up all four AI-bot categories gets you to 19% of total traffic to the file — and the heaviest readers aren’t conversational assistants at all, but coding agents (10.5%) and training crawlers (5.3%). Another 12% of traffic came from the GEO/AEO industry itself: validators and consultants checking whether it works, much like we’re doing here.

Why almost nobody requests it, even though almost everyone publishes it

The reason is straightforward, and the labs themselves confirm it: ChatGPT, Claude, and Perplexity don’t advertise llms.txt as a priority source for live answers, and there’s no public evidence it feeds model training. Publishing one doesn’t hurt, but treating it as the centerpiece of an AI-visibility strategy means betting your effort on the wrong channel. As we explained in our GEO guide, generative engines don’t work like a classic search engine you simply “notify” — they crawl, compare sources, and write, and they do it on a site’s actual content, not a separate summary they rarely even open.

What actually moves the needle, according to the data itself

Don’t block the crawlers that matter

Before optimizing anything, check that your robots.txt isn’t accidentally blocking GPTBot and OAI-SearchBot (OpenAI), ClaudeBot (Anthropic), PerplexityBot, or Google-Extended (Gemini and AI Overviews). It’s the cheapest check to run and the one that costs the most visibility when it fails: if the crawler can’t get in, no amount of content work matters.

Self-contained, verifiable content

A generative engine favors pages it can cite without ambiguity: concrete claims, sourced data, clear H2/H3 structure, content that doesn’t depend on the reader having seen a previous page. It’s the same criterion we detailed in how to check whether ChatGPT and Google already cite you: generative AI doesn’t rank by backlinks, it selects by how easy it is to extract a reliable answer from your page.

Real authority, not a self-declared file

Other reputable sites in your industry mentioning or linking to you carries more weight than any self-declared file — an llms.txt is, after all, a company telling the world what matters about itself, and no engine gives that the same weight as an independent external source.

How we handle this at Evicron

In our web development and AI SEO projects we don’t skip llms.txt — we generate one too, same as on this site — but the real priority sits with technical crawler accessibility, content structure, and external authority signals, which is exactly where the data says visibility is actually won or lost. Auditing this takes under a week and saves months of effort aimed at the channel that moves the least traffic.

Want to know whether your company shows up in ChatGPT, Gemini, or Claude answers — and what to fix if it doesn’t? Get in touch: we start with a real visibility audit, not a checklist of files to publish.