The short version
- Perplexity reads it. PerplexityBot parses
llms.txtwhen it indexes a site, and its engineering team has referenced the spec publicly. - Nobody else has confirmed they do. OpenAI, Anthropic and Google have never stated that their crawlers parse it, and no provider documents it as a citation signal.
- StayFound still generates one, because it costs nothing and Perplexity is real traffic. We’d rather say that plainly than imply more.
- If you only have time for one thing this week, it is not this file. It is getting mentioned on the sources engines already trust.
Why we are writing this against ourselves#
StayFound generates an llms.txt for you. It sits in the Actions tab next to the schema block and the crawler rules, as a paste-ready artefact. Our own comparison post listed it as part of what you get.
Then we went looking for evidence that it works, and the evidence is much thinner than the volume of advice telling you to publish one. So here is what we actually found, including the part that makes our own feature look smaller.
What llms.txt is supposed to do#
llms.txt is a proposed convention: a markdown file at the root of your domain that gives language models a clean, curated map of your site — what you do, which pages matter, where the documentation lives. The pitch is that instead of a model guessing its way through your marketing nav, you hand it the summary.
It is a genuinely good idea. robots.txt and sitemap.xml both started as conventions before anyone honoured them. The question is not whether the idea is sound. It is whether anything reads the file today.
Who actually reads it#
| Engine | Reads llms.txt? | Basis |
|---|---|---|
| Perplexity | Yes | PerplexityBot parses it at index time; the spec has been referenced publicly by their engineers. [1] |
| ChatGPT / OpenAI | Not confirmed | No OpenAI statement that GPTBot or OAI-SearchBot parse it or treat it as a special input. [2] |
| Claude / Anthropic | Not confirmed | No published commitment that the file is read at crawl or inference time. [3] |
| Gemini / Google | Not confirmed | Google has been publicly non-committal; it is not documented as a ranking or citation input. [4] |
Verified August 2026 from public sources: Alejandro Rioja, llmtxt.info, AEO Engine, Webyes on Mueller’s comments, We-Optimizz. “Not confirmed” means exactly that — an absence of any public commitment, not proof of absence. These are fast-moving products and any of them could start honouring the spec without announcing it.
The distinction that gets lost#
Most arguments about this file are really two arguments wearing one coat.
- Crawl time. Does the bot fetch
llms.txtwhile indexing? For Perplexity, yes. For the others, there is no evidence it is treated differently from any other file. - Inference time. When an assistant answers a live question, does it load your
llms.txtfirst to orient itself? No provider has claimed this, and it would be a strange design. The model retrieves and reads pages relevant to the question.
Advice that promises the second thing is overselling. The first thing is real, narrow, and worth having if Perplexity matters to you.
So why do we still ship one?#
Three honest reasons, none of them dramatic:
- Perplexity is not a rounding error. It is one of the five engines we track, and it demonstrably reads the file. That alone justifies twenty minutes of work.
- The cost is close to zero and the risk is zero. It is a static markdown file. It cannot slow your site, leak anything, or confuse a search engine that ignores it.
- Conventions get adopted. If OpenAI or Anthropic start honouring it, the sites that already have one are done. That is a cheap option to hold, as long as nobody has told you it is a strategy.
What we will not do is put it at the top of your action list. In the product it now sits below the schema block and the third-party listings, which is where the evidence puts it.
What to do instead, in order#
- Get mentioned off your own domain. One analysis of 21,311 brand mentions across ChatGPT, Claude and Perplexity found 85% came from external domains, with brands roughly 6.5× more likely to be surfaced through third-party sources than their own site. [5] This is the whole game, and it is the thing least under your direct control — which is exactly why it is worth starting now.
- Add schema to the pages you want quoted. Pages carrying structured data see around 2.8× the citation rate. This one is fully in your control and takes an afternoon.
- Keep the page fresh. Content updated within 30 days gets roughly 3.2× more citations than stale pages. A quarterly refresh of your best pages beats a new one nobody finds.
- Publish a number that exists nowhere else. Original data is the strongest citation magnet there is: if an engine wants to use your figure, it has to name you.
- Then write the llms.txt. Fifth. Not first.
The honest summary#
llms.txt is a reasonable file to have and a bad thing to believe in. It buys you something concrete with Perplexity and an option on everyone else. It does not get you into ChatGPT’s answers, and any tool — including ours — that lets you think otherwise is selling you the easy version of a hard problem.
If you want to know which sources are actually feeding the answers about your category, that is measurable, and it is a better place to spend the afternoon.