Google told you llms.txt does nothing. Then it shipped an audit for it.

You have probably seen both screenshots this week. One is John Mueller saying llms.txt is not done for search. The other is Chrome's Lighthouse, now checking whether your domain has an llms.txt file at the root. Somebody posts them together, adds a caption about Google contradicting itself, and it does numbers.

I get why. It looks like a contradiction. It is not one, and the reason it is not one turns out to be the clearest thing Google has said about agents in months.

The two statements everyone is putting side by side

Here is what was actually said, in full, because the screenshots always crop it.

Read the right-hand column. Nothing in there conflicts. Two of those statements are about whether a file helps you get found. Two are about whether a file helps a machine move around once it is already there. Those are different products, different teams and different problems.

What the llms.txt Lighthouse audit actually checks

The llms.txt check does not sit in Performance or SEO. It sits in a new experimental category called Agentic Browsing, alongside three other checks: WebMCP integration, accessibility tree integrity, and layout stability measured through CLS.

That is a genuinely different claim from the one the SEO discourse has been arguing about for a year. Nobody at Google said llms.txt gets you into an answer. They said it makes you cheaper to operate. Those are not the same sentence and only one of them is about your traffic.

It is also worth noting what the category is labelled. Experimental. Google ships experimental audits regularly and quietly retires a fair number of them. Treating a pass ratio in an experimental category as a compliance requirement is how teams end up with a backlog full of work nobody asked for, prioritised above things that demonstrably move revenue. Read it as a signal about direction, not as a specification you owe somebody.

Discovery is not functionality

Mueller drew the line himself, and it is the most portable idea in the whole story: discovery is finding websites through search engines. Functionality is helping a user accomplish a task once they are there. llms.txt is a functionality file that got adopted by a discovery industry.

“A file that makes an agent faster is not a file that makes an engine cite you. The first is plumbing. The second is authority.”

This is why the last year of llms.txt debate went nowhere. One camp tested it for ranking lift, found none, and declared it useless. The other camp pointed at the growing list of sites adopting it and declared it essential. Both were measuring the wrong axis. The file was never going to move rankings, and its adoption curve was never evidence that it did.

The distinction generalizes past this one file. Every technical AI-readiness decision you make sits on one side of that line or the other, and the honest question is always which one you are buying. Making your pages parseable and your facts extractable is discovery work. Making your site operable by software is functionality work. We keep them in separate workstreams for exactly this reason, and it is the same split behind AI crawlers versus AI agents — two different machines with two different failure modes.

Why this argument keeps happening

Because our whole industry was trained on one question, and it is the wrong one here.

For twenty years, every technical change on a website could be evaluated against a single yardstick: did it help you rank. Schema, sitemaps, canonical tags, page speed — all of it ultimately cashed out in visibility. That produced a reflex. A new file appears, and the first thing anyone asks is whether it moves rankings. When the answer comes back no, the conclusion is that it is worthless.

That reflex is now misfiring, and it will keep misfiring, because a second category of work has appeared that does not cash out in visibility at all. It cashes out in whether a piece of software can complete a task on your site without a human. Nothing about your ranking changes when an agent successfully books an appointment. Something about your revenue might.

The uncomfortable part is that functionality work has no clean measurement story yet. There is no agent equivalent of a rank tracker, no dashboard that shows you how many automated sessions abandoned because your checkout confused a model. Google shipping a pass/fail category rather than a score is an admission of exactly that: they do not have a good number either. Measurement always lags the surface, which is a pattern the reporting side of AI search has been living through for two years.

So should you ship an llms.txt file

Probably, and for boring reasons. It costs an afternoon. It cannot hurt you. And if agents do become a meaningful traffic source for your category, you will already be legible to them.

Write it like a table of contents for someone who has never seen your site and has thirty seconds. What the company does, where the main sections live, which pages are canonical for which topic. Skip the marketing copy — nothing is reading it for persuasion. The file is a map, and most of the bad ones fail because somebody treated it as a brochure.

But do it with the right expectation, and do not let it jump the queue. Mueller's actual advice was blunter than the screenshots suggest: you have much more important things to do for SEO than prepare for a potential future situation. That is not a dismissal of agents. It is a sequencing argument, and he is right about the sequence.

That last row is the one I would underline. A meaningful share of the sites asking about llms.txt still serve their primary content through JavaScript that AI crawlers never execute, which is a much larger and much less fashionable problem — the one we walk through in why AI crawlers don't render JavaScript. Shipping a tidy manifest for a site the crawler cannot read is decorating a locked door.

The part worth actually worrying about

Here is what the Agentic Browsing category really tells you, and it has nothing to do with a text file. Google is now shipping developer tooling that measures whether software can use your website. Not read it. Use it.

That is a bet on a web where a lot of the interaction is machine-mediated, and it is being placed in the most mundane way possible: an experimental Lighthouse category that most people will never open. Tooling arrives before the traffic does. It always has. The robots.txt spec predated crawler-driven commerce by years, and the sites that had their access rules straight early were not visionaries — they just had less to fix later. The same logic applies to how you handle AI crawler access today.

You are not behind. The audit is experimental, the traffic is small, and the file takes an afternoon. What you should take from this week is not a to-do item, it is a frame: stop asking whether a given technical change helps you rank, and start asking whether it helps you get found or helps you get used. Those need different budgets, different owners and different success metrics. If you want help drawing that line across your own stack, that is most of what a technical audit is for.

See where you are cited today

A free snapshot audit of your rankings and AI citations before we ever talk.