The file nobody reads just got added to the test everybody runs

What’s the point of an llms.txt file? The short answer

If you’re deciding whether to add an llms.txt file for AI visibility, the honest answer is that it won’t magically get your site surfaced more frequently by Chatty G and the other robots, in fact it will do very little at all. The data is one-sided: almost nobody's file gets read, and Google has said plainly that it doesn't use it. What changed last month is smaller and stranger. Google's own Lighthouse, the engine behind PageSpeed Insights, started checking whether you have the file present. So this file nobody reads just got added to the test everybody runs. That doesn't make it a ranking lever. It makes it a checkbox. I still include llms.txt files on sites I work on and I'll explain why that isn't a contradiction doing so.

What the data actually says

On 15 June 2026 Ahrefs published the largest look at llms.txt in the wild so far: 137,210 domains. The headline number is hard to argue with. 97% of llms.txt files received zero requests in May. Nothing fetched them. Not a human, not a bot.

Of the 3% that did get touched, it gets worse for the believers. 77% of the bots fetching the file weren't AI tools at all. The biggest single category was SEO audit software at around 22%, then general crawlers and tech-profiling tools. The actual AI retrieval bots, the PerplexityBots and OAI-SearchBots that are meant to be the whole point, showed about as much interest in this new standard as a giraffe does in applied trigonometry.

Adoption sat near 28%, and Ahrefs was honest that this is an upper bound, because the sites in its sample skew technical and SEO-aware. The real number across the web is lower. So the picture is a file that a minority of sites publish and that its intended audience almost never reads.

Google doesn't use llms.txt, and says so

This isn't a case of reading the tea leaves. Google has been unusually direct. At Search Central Live in July 2025, Gary Illyes said Google doesn't support llms.txt and isn't planning to. John Mueller, asked on Bluesky whether using the file counts as an endorsement, gave a flat no, and called it "more of a temporary crutch, perhaps to save some tokens." Google's own guidance on showing up in AI search says you "don't need to create new machine readable files, AI text files, markup, or Markdown" to appear there.

OpenAI and Anthropic haven't committed to it either. So the standard that was meant to be how you talk to the big models is, as of now, used by none of the big models. I've argued before that LLM SEO is a bit of a contradiction, and this is the same shape of thing: a tactic that sounds like control but mostly sells reassurance.

So why did Google's own browser just start grading it?

Here is the part that made me write this. Lighthouse 13.3.0, released on 7 May 2026, moved a new Agentic Browsing category out of experimental and into the default config. That means PageSpeed Insights, the tool your client runs and forwards to you with the red bits highlighted, now runs three agentic checks. One of them is llms.txt.

Worth being precise, because the detail matters and most write-ups get it wrong. It is not a simple have-it-or-fail. If the file is absent, the audit is marked Not Applicable. You are not penalised for skipping it. The flag only fires if you have one and it's bad: under 50 characters, or not following the recommended format. The category isn't scored out of 100 either, it's a fractional "you pass this many of these checks".

So the contradiction is real, but narrow. Google Search doesn't use llms.txt and tells you not to bother. Google's browser tooling, looking ahead to agents that read pages on a person's behalf, now nudges you to have a good file if you have one at all. Two different Googles, two different time horizons.

Is this the meta keywords tag all over again?

This is the right instinct if you remember the 2000s. The meta keywords tag was the thing every SEO stuffed and every tool checked, until Google confirmed in 2009 that it ignored the tag for ranking. It died of abuse and irrelevance at the same time, and nothing ever picked it up again.

The neat part is that the comparison isn't mine. Mueller himself likened llms.txt to the old keywords meta tag. When the platform reaches for its own most famous dead signal to describe your new one, you should take the hint.

But I don't think the meta keywords ending is settled here, and the Lighthouse move is why. Meta keywords had no second act. llms.txt has a second surface waiting: agentic browsing, where a model genuinely is walking your site on someone's behalf and a clean summary could save it real work. That future may not arrive. But Google's own tooling just placed a small bet that it might, and that's a difference the keywords tag never had. My honest read: for search, llms.txt is already as dead as meta keywords. For agents, the jury is out, and I would rather be early than re-tooling in a panic.

I build llms.txt files anyway, and that isn't a contradiction

I have one on this site. I list llms.txt as part of my GEO work, and I add it to client builds. After all of the above, that deserves an explanation.

It comes down to cost against option value. A good llms.txt takes an hour and costs nothing to maintain. The downside of having one is roughly zero. The upside, if agentic browsing becomes a real way people reach your site, is that you are already there. That is a bet I'll take every time the stake is an hour.

There are two smaller reasons. One is that it's now a check in a report clients see, so doing it properly keeps a needless red mark off the page and saves a conversation. The other is plainer: I want to be among the sites showing up when people search for help with this, which means having a view on it and, yes, doing it. This post is part of that and I’d be lying trying to pretend otherwise.

What I won't do is sell it as a visibility lever, because it isn't one. The difference between a consultant and a vendor is whether they tell you that.

What to actually do

If you don't have an llms.txt, you can skip it with a clear conscience. Lighthouse won't dock you, and the AI bots aren't waiting for it.

If you want one anyway, for the future-proofing, do it once and do it right: a genuine summary of what your site is and where the important pages live, over 50 characters, kept honest. Then forget about it.

And if your actual goal is to get cited by AI, spend the hour where it pays. The same Lighthouse category that checks llms.txt also checks a clean accessibility tree and a stable layout, both of which help real machine readers today rather than hypothetical ones tomorrow. The accessibility tree matters enough that I built it into my free agent readiness test, which scores how well an AI agent can actually perceive and navigate your page. Clear schema, a legible entity, content that answers the question first: that is what gets you quoted. I've written about why GEO is a bad name for that work, and the free LLM visibility audit and the schema sniffer will show you where you actually stand. The file was always going to be too simplistic to be a catch-all for "getting found". Just like you can't hand Google a list of preferred keywords and watch as your page-one visibility soars, so too does spoon-feeding LLMs this way yield little tangible return. But it's not redundant, or it wouldn't be getting measured in PageSpeed tests. Reports of llms.txt's death, it seems, have been greatly exaggerated.