If you've read anything about AI search this year, you've probably met the argument: websites now need a second version written for machines. Some people say an llms.txt file. Some say a markdown copy of every page. Some say it's all noise. We ship an llms.txt on every site we build, so we have a stake in the answer, and the honest version is less exciting than either side makes it sound.
What is llms.txt?
An llms.txt is a plain-text file that sits at the root of a website, at yoursite.com/llms.txt. It's a short brief written for AI systems: what the business is, where it is, and which pages matter, with a one-line description of each. The idea was proposed in 2024 as a way to give a language model the essentials without making it wade through menus, cookie banners and scripts. You can read ours to see what one looks like — it's a page of plain text.
A "markdown version" is the bigger cousin of the same idea: a stripped-down text copy of each page, served alongside the normal one, so an assistant can read the words without the design.
What it does, and what it doesn't
Here is the part people skip. Nobody can promise you that any particular assistant reads llms.txt today. The major AI companies don't document it as something their crawlers use, and Google has said publicly that it doesn't rely on it. It's a proposal that some sites have adopted, not a standard the assistants have committed to. It's early, and anyone who tells you otherwise is guessing.
So why do we ship one? Because on our platform it costs nothing and can't go wrong. The file is generated from the site's own pages every time it's requested, so it always says what the site says. If assistants start reading it widely, every site we run is ready. If they don't, nothing was spent.
An llms.txt is a good cover letter. It's no substitute for the application.
That "generated, not written" detail matters more than it sounds. Earlier this month we found our own llms.txt, sitemap and feed had been quietly serving an old version of the site, because they were built once when the site deployed instead of being read live. Nothing looked broken. We only caught it by checking what a crawler actually received. A hand-maintained llms.txt fails the same way, just more slowly: someone updates the services page, nobody updates the text file, and now you're telling machines something that's no longer true.
The markdown copy is where we'd be careful
A separate text version of every page is a bigger commitment. It doubles what has to stay accurate. If it's written separately it will drift, and a page that tells machines something different from what it tells people is exactly the pattern search engines have penalised for years. If your site is built so the real page is already clean, readable text, you don't need the copy. If it isn't, the copy is treating a symptom.
What actually decides whether an assistant understands you
When we ran our free scan across a dozen Fort Worth HVAC and accounting websites, the common gaps weren't missing llms.txt files. Four of the twelve had no structured data at all — nothing on the page that tells a machine what kind of business this is, where it is, or how to reach it. Three had no sitemap. One had a robots.txt that welcomed some assistants by name and, by leaving others off the list, told them to stay out. None of that is exotic, and all of it matters more than a brief at the root of the site.
In order, this is what we'd look at first:
Can the assistants get in? Check that robots.txt and your firewall aren't turning the answer engines away. We wrote about how that happens by accident in Is your website blocking AI bots?, and about choosing to let AI search in while keeping AI training out.
Is the real page readable? Your services, hours, area and prices should be text on the page, not baked into images or hidden behind a script that only runs in a browser.
Does the page say what the business is? An Organization or LocalBusiness record in the page's structured data — name, address, phone, what you do — is the thing assistants can cite with confidence.
Does it all agree? Your site, your Google Business Profile and your reviews should describe the same business. Assistants cross-check, and contradictions cost you.
Then the llms.txt. Generated from the site itself if your platform can do it, short and accurate if you write it yourself. It's the last step, not the first.
Where this is heading
The honest reason to care about any of this isn't a text file. It's that assistants are moving from reading websites to acting on them — checking availability, starting a booking, answering a specific question about your service area. We wrote about that shift in how AI agents will use your website. Sites that are already clean, structured and consistent will make that jump easily. Sites held together by a parallel machine-only copy will have two things to fix instead of one.
If you'd like to know where your site stands today, our free AI visibility scan checks all of the above in about twenty seconds — including whether you have an llms.txt, and whether it matters yet.
