llms.txt and AI citations: what the evidence says, and what it does not

Two posts this week asked the same question from opposite ends: does anything you do on your own site decide whether an AI assistant quotes it? One maintainer says the file everyone recommends, llms.txt, did very little for him. One vendor says a site built entirely on invented statistics kept getting fetched by ChatGPT long after Google stopped showing it. Neither is a study. Both are worth reading carefully, and both point at the same uncomfortable gap between being found and being right.

A maintainer's report: the file was rarely read

Cristian Tabacitu maintains Backpack for Laravel, an admin panel package that is about ten years old. He wrote that they shipped an llms.txt and he doubts it does much: agents rarely look for it, and when they do it is a table of contents for documentation they could already read. He is considering removing it. An MCP server they built for searching their docs gave mixed results for a similar reason: the agent has to decide to call it, and usually does not. His rule of thumb is that anything an agent has to remember to go and fetch is weak, and what works is what is already in front of it or already in its training data.

What he credits instead is mostly old-fashioned: an API that stayed stable for years, so the correct code is what ended up in training data; deprecating instead of renaming, because a rename splits what a model learned into an old version and a new one; ten years of open issues and tutorials as examples; and code that sits readable in vendor/, where agents go to answer their own questions.

A vendor's experiment: retrieval is not ranking

The second post is from the blog of Peec AI, a company that sells AI-visibility tracking, so read it as a vendor describing its own experiment. The author says that in March 2026 he generated 50,000 statistics pages in one day with GPT-4.1, every number invented and never checked, for about $50. By his account Google indexed the site within about a week, then visibility collapsed after two Google updates. Yet he reports that ChatGPT kept retrieving it: about 72,000 requests from ChatGPT-User, the agent that fetches pages in real time, in the last 30 days alone. He also says the invented numbers turned up in an academic paper and in a bank's investment research.

His conclusion is that ranking on Google and being retrievable by an assistant are different problems. An assistant rewrites a prompt into several sub-queries, collects candidate pages, reranks them and opens a few. Nothing in that pipeline, as he describes it, checks whether a number is true.

What this site knows about its own llms.txt

This site serves /llms.txt and /llms-full.txt. Both are built from real data: the tool category pages, every enabled tool with its description (the full file adds each tool's how-to steps and FAQ), and published main-blog posts. Both are English only. robots.txt explicitly allows four AI crawlers by name: GPTBot, ClaudeBot, PerplexityBot and Google-Extended.

What the site does not know matters more. A visit from one of those four is counted separately in its analytics, but only when the page's own tracking beacon reports it, which a crawler that reads the HTML without running the script never does, and a plain text file has no script at all. ChatGPT-User is not on the list. So there is no measurement here of who reads llms.txt, and none of whether it has led to a single citation. This post makes no claim either way, because there is nothing to base one on.

The house rule: only numbers that can be sourced

The vendor's experiment shows that an invented number can travel a long way before anyone checks it. The rule on this site is the plain response to that: every number and every quote is real, with no invented metrics. If a figure appears, the sentence says whose it is and a link leads to where it came from. That costs some reach, because a sourced claim is always narrower than a made-up one. It is still the only kind of claim worth being quoted for.

If you are deciding what to do this week: keep llms.txt if it is cheap for you, but do not expect it to do the work. Spend the time where both posts agree the effect is real: stable URLs and names you do not rename, clear facts with their sources in the sentence, structured data that matches the visible page, and documentation an agent can read without being told to. Sources: Backpack for Laravel and the Peec AI experiment.