What is llms.txt, and does it do anything?
llms.txt is a proposed plain-text file at a site’s root that summarises it for language models. It is a community proposal, not a standard — and no major engine has publicly confirmed that it reads one.
We publish one, at bungad.com/llms.txt. We also have no evidence it has changed a single answer. Both of those things are true and this page is here to say so, because almost every other page about llms.txt is written by someone selling llms.txt generation.
What it is
A Markdown file served at /llms.txt, proposed in September 2024 by
Jeremy Howard of Answer.AI. The shape is simple: a title, one blockquote saying what
the site is, then sections of links with a sentence of description each. The idea is
that a model retrieving your site gets a curated summary instead of having to infer
one from your navigation.
It is worth being exact about its status. It is not a W3C or IETF standard, not published by any engine, and not something any engine has committed to consuming. It is a proposal that a number of sites have adopted.
How it differs from the files it gets confused with
| File | Status | Purpose | Controls access? |
|---|---|---|---|
robots.txt |
Long-established convention, widely honoured | Tells crawlers what they may and may not fetch | Yes — it is a permission file |
sitemap.xml |
Established standard | Lists your URLs so crawlers can find them all | No — discovery only |
llms.txt |
Community proposal, 2024 | Describes and prioritises your site for a model | No — it grants and withholds nothing |
The misrepresentation to watch for: llms.txt is sometimes sold as
a way to control whether your content is used for AI training. It does nothing of
the sort. It has no permission semantics at all. If you want to express crawling
preferences, that is robots.txt and the engines’ own documented
user-agent strings.
Does it work?
The honest answer is that nobody outside the engines knows, and the engines have not said. What we can state precisely:
-
No public confirmation. OpenAI, Google, Anthropic and Perplexity
have not publicly documented reading
llms.txt. - No credible published evidence that adding one changes how often a brand is named or cited. The case studies that circulate tend to change several things at once.
- Our own data point, such as it is. We publish one. Our audited visibility score on 19 September 2026 was 0 out of 100, with 0 of 90 cited sources pointing at our domain. That is one site, one month, and many other things about us are also new — so it is not proof the file does nothing. It is simply not evidence that it works, and we are not going to dress it up as any more than that.
We will keep publishing ours and keep publishing the monthly number. If the score moves, we will say what else changed at the same time, because a single-variable claim from a multi-variable month is exactly the sort of thing this field is already full of.
Should you publish one?
Probably, and last. It takes under an hour, costs nothing and carries no risk provided the file agrees with your visible pages. The cost is not the hour — it is the attention it takes from work that demonstrably matters.
If writing an llms.txt is the AEO work you did this month, you did the
wrong work. The things with evidence behind them are on
how to get mentioned, and the short version
is: publish pages that answer real questions, put your specifics in plain text, keep
your structured data matching the page, and earn mentions on the sources the answers
already cite.
How to write one that is actually useful
- Start with one sentence saying what you are. A blockquote under the site title stating plainly what you do, including the limits of what you cover. This is the line most likely to be read and reused.
- Put the facts a buyer asks for next. Prices, contact address, business model, coverage. Plain text, no adjectives. If a number is on your pricing page it belongs here too, and the two must agree.
- Link your key pages with a description each. A short list of URLs, each with a sentence saying what question that page answers — written for someone deciding whether to fetch it.
- State what you do not do. Coverage you lack, engines you do not support, markets you do not serve. Everyone omits this, and it is the part that stops an engine describing you wrongly.
- Keep it consistent with the visible site. Every fact in the file must also appear on a page as text. A file that contradicts the site is worse than no file, for the same reason mismatched schema markup is worse than none.
-
Update it whenever a fact changes. A stale
llms.txtquietly publishes last quarter’s prices. Put it on the same checklist as the pricing page and review it monthly.
What ours contains
Ours follows the structure above, and step four is the part we would point at. It states that Perplexity is the only engine we hold an API key for, that ChatGPT and Claude are built but unkeyed, that Gemini is not built, and that Google AI Overviews and Copilot have no API and are reviewed by hand. It also carries our own audited score of 0, the exact prices, and the paths to the stored run records.
Whether any engine reads it, we do not know. But if one does, we would rather it read an accurate description of what we can and cannot do than a flattering one — that is the whole argument for the file, and it is the only part of it we are confident about.
Common questions
What is llms.txt?
A proposed plain-text file, in Markdown, placed at the root of a website to summarise it for large language models: what the organisation is, its key facts, and links to its most important pages with a description of each. It was proposed in September 2024 by Jeremy Howard of Answer.AI. It is a community proposal, not a web standard, and it is not published or endorsed by any engine.
Does llms.txt actually work?
There is no public confirmation from OpenAI, Google, Anthropic or Perplexity that they read llms.txt, and no credible published evidence that adding one changes how often a brand is cited. We publish one at bungad.com/llms.txt and our own audited visibility score is still 0 out of 100. That is not proof it does nothing, but it is the opposite of evidence that it works, and we would rather say so than sell the file.
How is llms.txt different from robots.txt?
robots.txt is an established convention that tells crawlers what they may fetch — it grants or withholds permission. sitemap.xml lists your URLs so crawlers can find them. llms.txt does neither: it is a summary written for a reader, intended to describe and prioritise rather than permit or enumerate. Publishing llms.txt does not control AI training or crawling in any way, and anyone selling it as a consent mechanism is misrepresenting it.
Should I publish an llms.txt file?
Probably yes, but rank it last. It takes under an hour, costs nothing, and carries no risk provided it agrees with your visible pages. What it must not do is displace the work that demonstrably matters: publishing pages that answer real questions, putting your specifics in plain text, keeping structured data matching the page, and earning mentions on the sources engines already cite. If writing llms.txt is the AEO work you did this month, you did the wrong work.
Do the work with evidence behind it
Enter your domain. We ask Perplexity three of your buyers’ questions and show you which questions you are losing — a better starting point than a text file.
3 questions on Perplexity, free. A domain checked in the last week is served from that stored run rather than asked again, and the daily free allowance resets at midnight UTC.