Should I have an llms.txt file, and how do I generate one?
Probably, but not for the reason it is usually sold. In May 2026, 97% of llms.txt files received zero AI-bot requests, and Google has said the file affects nothing on its side. It is a fifteen-minute, zero-risk exercise in writing down what your site is, and that exercise is the actual benefit.
What goes in, and what comes out
In: Site name and one-line description
In: A short summary of what the business does and who it serves
In: Your most important pages, grouped into sections, with a one-line note each
In: Optional sections for documentation, policies and contact
Out: A valid, correctly structured llms.txt in Markdown
Out: A copy button and a download
Out: Placement instructions and what to check afterwards
Out: A shareable link that reproduces the form state
The method, stated
This generator formats what you type into the llms.txt convention: an H1 site name, an optional blockquote summary, prose sections, and H2-grouped Markdown link lists with one-line notes. It validates structure only. It makes no claim that publishing the file will change how any system treats your site.
The formulas, in evaluation order
line 1 = '# ' + site_name
line 3 = '> ' + one_line_summary
body = optional prose paragraphs, plain Markdown
each section = '## ' + section_name, then '- [title](url): note' per entry
output is served at /llms.txt as text/plain or text/markdown
Every assumption this model bakes in
The convention is a Markdown document, not a robots.txt-style directive file. It describes; it does not instruct, allow or deny anything.
Link notes should be one line and factual. The file is a map, and a map with marketing copy on it is a worse map.
Fewer entries beat more. Twenty well-chosen pages that describe what the site is beat two hundred that reproduce the sitemap.
Nothing in the file grants or restricts crawler access. Access control is robots.txt, your CDN and your terms, and none of them are affected by this file.
The generator does not fetch your site. Everything in the output is what you typed.
What this model cannot tell you
It cannot make anything read the file. Adoption is the entire open question and it is not one a generator can influence.
It cannot substitute for the pages themselves. A beautiful index over thin content indexes thin content.
It does not affect Google, which has said publicly that the file has no effect on its systems.
It is not an access-control mechanism and must never be used as one.
What the evidence on llms.txt actually shows
Observed 2026 evidence, published here because the case for this file is usually made without any. Neither line is a TrainYourAgent claim about outcomes.
llms.txt files receiving zero AI-bot requests, May 2026: 97% — The overwhelming majority of published files were never fetched by an AI crawler at all.
Effect on Google: None, per Google — Google has stated the file does not affect its systems. Treat any claim to the contrary as unsupported.
What is llms.txt and what is it supposed to do?
llms.txt is a proposed convention: a Markdown file at the root of your domain that gives a language model a curated map of your site. An H1 with the site name, a one-line summary in a blockquote, some prose about what the site is, and then grouped lists of your most important URLs with a short note on each. The stated theory is that a model with a limited context window benefits from a hand-curated index far more than from a raw sitemap of ten thousand URLs, and that the site owner is best placed to say which twenty pages matter. It is worth being precise about what it is not. It is not robots.txt. It does not allow or deny anything, it grants no permissions and it withholds none. It is a description, and the only power it has is the power any description has, which is to be read.
Does llms.txt actually do anything?
On the current evidence, mostly no, and anyone selling you a service to implement it should be saying so. In May 2026, 97% of llms.txt files received zero AI-bot requests. Not few requests. Zero. The overwhelming majority of these files, carefully written and correctly placed, were simply never fetched by anything. Google has stated that llms.txt does not affect its systems. That is a direct statement from the largest consumer of web content in the world, and it forecloses the most commonly implied benefit. That leaves a small set of honest possibilities: a handful of tools and agents do read it, adoption may increase, and the file costs almost nothing. Those are real, and they are considerably smaller than the claims usually made for it. If you have been told that publishing llms.txt will get you cited by ChatGPT, you have been told something the evidence does not support.
So why does TrainYourAgent ship one anyway?
Three reasons, and none of them is that it will get us cited. First, the exercise is the benefit. Writing an llms.txt forces you to answer, in one blockquote line, what your site is, and then to choose the twenty pages that matter most out of the several hundred you have. Most organisations cannot do either of those things without an argument, and having the argument is worth more than the file. Ours changed how we wrote our service pages. Second, the cost is genuinely near zero. It is one static file, fifteen minutes of writing, no runtime cost, no maintenance burden beyond a quarterly re-read, and no risk. Optionality at that price is worth buying even at low probability. Third, adoption asymmetry. If the convention goes nowhere, we lost fifteen minutes. If a meaningful set of agents start reading it in eighteen months, we already have one and it is already accurate. That is a reasonable bet, and it is a different argument entirely from claiming the file works today. What we will not do is charge anyone for implementing it as though it were a growth lever, or list it on a proposal next to work that actually moves something.
What should actually go in the file?
Start with one H1 line: the name of the site. Then a blockquote with a single sentence saying what it is. If you cannot write that sentence, stop and write it, because it is the single most useful output of this exercise and it belongs on your homepage too. Then a short paragraph or two of orienting prose. What the business does, who it is for, what is distinctive about how it works. Plain language, no marketing register, because the reader in this scenario is a machine assembling an answer and there is nobody to persuade. Then grouped link lists under H2 headings. Group by what a reader would be looking for rather than by your internal org chart: services, pricing, documentation, tools, policies. Each entry is a Markdown link with a colon and a one-line factual note about what is on that page. Keep it short. Twenty to forty links is plenty for most sites. If your llms.txt has three hundred entries you have written a sitemap, and a sitemap already exists at a well-known location that things actually read.
What should actually go in the file? — in detail
H1: the site name, nothing else
Blockquote: one sentence on what the site is
Prose: two paragraphs at most, plain language
H2 sections: grouped by what a reader wants, not your org chart
Links: twenty to forty, each with a one-line factual note
Where does the file go and what about llms-full.txt?
The file goes at the root of your domain, at /llms.txt, served as plain text or Markdown. Not in a subdirectory, not behind a redirect chain, and not gated by anything. Deploy it, then fetch the URL yourself and confirm you get the raw Markdown rather than your application's HTML shell, which is the most common mistake on single-page applications. llms-full.txt is a companion convention holding the full text content of your key pages rather than links to them. The theory is that a reader gets your content without fetching anything else. In practice it is a large file that must be regenerated whenever content changes, and given that 97% of the basic file gets no traffic at all, generating and maintaining an expanded version is difficult to justify on current evidence. If you want to do it anyway, generate it from your CMS rather than by hand. A stale llms-full.txt containing last year's prices is worse than no file, because unlike your website nobody will notice it is wrong.
Does llms.txt work?
On current evidence, mostly no. In May 2026, 97% of llms.txt files received zero AI-bot requests, and Google has said the file does not affect its systems. Anyone claiming it will get you cited is going beyond the evidence.
Then why publish one?
Because writing it forces you to state what your site is and pick the twenty pages that matter, which is useful independently. It costs fifteen minutes, carries no risk, and gives you cheap optionality if adoption ever increases.
Is llms.txt the same as robots.txt?
No, and confusing them is dangerous. robots.txt is a directive file that requests crawler behaviour. llms.txt is a description. It grants nothing and blocks nothing, and it must never be used as an access-control mechanism.
Does Google use llms.txt?
No. Google has stated publicly that the file does not affect its systems. Treat any service that prices llms.txt implementation as a Google ranking lever as mispriced.