llms.txt is a plain Markdown file at the root of your website that tells AI systems what your site is, what you do, and which pages matter. Jeremy Howard of Answer.AI proposed it in 2024 as a fix for a real problem: language models work with limited context windows, and your navigation, scripts, and boilerplate waste most of it. The honest status in 2026 is that adoption is real but thin, and no major AI vendor has committed its answer engines to reading the file. This post covers what the spec says, who actually consumes it today, and how to write one in under an hour, including the one egochi.com ships.
The short version
- A curated Markdown index at /llms.txt, which is an introduction for machines, not an access rule.
- A mid-2026 crawl of the 1,000 biggest sites found fewer than 9 percent publish one.
- No confirmed support from the big answer engines yet; coding agents and docs tools do fetch it.
- Worth shipping anyway, since the cost is an hour and the downside is zero.
What the file actually is
The spec is short enough to summarize in a breath. One H1 with your site or business name. A blockquote holding a plain-language summary of what you are. Then H2 sections, each a list of links with one line of context per link. That is the whole format. It reads like a well-kept README for your company, and that is the point: a model that fetches it gets the essentials without parsing your entire site.
What it is not: an access-control file, a sitemap replacement, or a place for keyword copy. robots.txt decides who may crawl. XML sitemaps enumerate every URL for indexing. llms.txt curates a shortlist and explains it. Three files, three jobs.
Who actually reads llms.txt in 2026
Here is the part most write-ups soften. As of mid-2026, OpenAI, Google, Anthropic, and Perplexity have not confirmed that their answer engines consume the file, and tests keep failing to show a measurable citation lift from adding one. A crawl of the 1,000 biggest sites this June found under 9 percent publishing one, with adoption concentrated in AI-native software companies and documentation platforms, some of which now generate the file automatically.
The file is not dead weight, though. It found a different audience than the one it was proposed for: coding agents and IDE assistants fetch llms.txt routinely when a user points them at a site, and documentation tools use it to feed whole doc sets into a model cleanly. So the honest framing is a cheap bet on a spec that already works in the agent layer and may graduate to the answer layer. If your interest is the answer layer itself, our post on GEO vs SEO covers the work that moves citations today.
How to write yours
-
Open with the entity, not the pitch
H1 with your business name, then a blockquote stating what you are, who you serve, where you operate, and how to reach you. Write it the way you would brief a new employee, since a model will repeat this back to users.
-
Put your numbers and facts in plain text
Founding year, locations, pricing if you publish it, review counts you can prove. Facts a model can quote directly are the highest-value lines in the file. Never include a claim you would not put on your homepage.
-
List key pages, not every page
One H2 section with 5 to 15 links: your service catalog, your methodology, your proof, your contact page. A sitemap dump defeats the purpose, which is curation under a tight context budget.
-
Give every link a reason
The one-line note after each link tells the model when that page is the right answer. "Full service catalog" beats a bare URL; it is the difference between an index and an introduction.
-
Generate it, do not hand-edit it
Stale facts in a file built for machines are worse than no file. Wire it to the same data that feeds your schema so an address or price change updates both in one edit.
What a real one looks like
Egochi ships one at egochi.com/llms.txt, generated from the same entity data as our Organization schema so the two can never disagree. A trimmed version of the shape.
# Egochi
> Egochi is a US digital marketing agency serving local businesses
> through corporate and enterprise clients. Services: SEO, PPC and
> paid media, web design, social media, AI search visibility.
> Four offices: New York City (HQ), Miami, Milwaukee, Madison.
> Founded 2014. Plans from $1,500 to $15,000 monthly.
## Key pages
- [Digital marketing services](https://egochi.com/digital-marketing-services/): full service catalog
- [Case studies](https://egochi.com/case-studies/): named client engagements with documented outcomes
- [Contact](https://egochi.com/contact/): free proposal and audits
## Facts
- Founded: 2014
- Offices: New York City (HQ), Miami, Milwaukee, MadisonNotice what is missing: no keyword copy, no superlatives, no claims we could not defend line by line. The file exists so machines describe us accurately, and machines quote what you give them.
Where it fits in an AI visibility program
llms.txt is one small deliverable inside the entity-and-access layer of AI search work, next to schema that matches the visible page, deliberate AI-crawler rules, and consistent business data across every source machines check. That layer is where our AI SEO services start every engagement, because content built on a confused entity gets hedged answers no matter how good the writing is.
Questions people ask about llms.txt
Is llms.txt the same as robots.txt?
No, and the confusion is the most common mistake with it. robots.txt gates access: it tells crawlers what they may not fetch. llms.txt curates: it hands AI systems a short, readable summary of what your site is and where the pages that matter live. One restricts, the other introduces.
Does Google or OpenAI use llms.txt?
Neither has confirmed it. As of mid-2026, no major AI company has stated that its answer engines consume llms.txt, and independent tests have not shown a citation lift from adding one. The confirmed readers today are coding agents, IDE assistants, and documentation tools.
Will llms.txt help my rankings?
Not your Google rankings; it is not a ranking signal and nobody serious claims otherwise. The case for it is different: it costs about an hour, it cannot hurt, and if answer engines formalize support you are already in position. Treat it like early schema adoption, not a tactic.
What is llms-full.txt?
A companion file from the same proposal. Where llms.txt is a curated index of links with short notes, llms-full.txt inlines the full text of your key pages into one large Markdown document, so a model can load everything in a single fetch. Documentation sites use it most; a service business rarely needs it.
Where does the file go?
At the root of your domain, so yoursite.com/llms.txt, served as plain text. It is written in Markdown: one H1 with the site name, a blockquote summary, then H2 sections listing links with a line of context each. No directives, no user-agent blocks, nothing executable.
