Key takeaways
- llms.txt is an optional note for agents. It is not a ranking lever, and Google Search does not use it as a ranking signal.
- Treat it as a README: it must match the public site, its links must resolve, and it must never outrank the HTML.
- The work that matters is honest public pages — the site, the copy, and SEO — not a file that invents a better version of the business.
llms.txt is a short Markdown file at the root of a website. It is a note for agents. It is not a ranking lever.
That is the first thing a "what is llms.txt" search should hear. The file is optional. Many systems ignore it. Google Search does not use it to rank pages. A well-made one is a README: it points at public pages that already exist, in language those pages already support. A badly made one invents a cleaner company than the HTML will admit.
People also type llm.txt. Same question. The proposed filename is `llms.txt`, with the *s*. The extra letter is not a product. It is the file the proposal named.
What llms.txt actually is
In September 2024, Jeremy Howard published a proposal at llmstxt.org: put a Markdown file at `/llms.txt` so a language model that is trying to use your site at inference time can read a curated orientation instead of guessing from a homepage full of chrome.
The idea is modest. An agent that fetches a marketing page also fetches the cookie banner, the nav, the footer, the newsletter modal, and three competing headlines. A short file can say: this is the project, these are the pages that matter, start here.
The proposed shape is a README, not a schema:
- One H1 with the site or project name. That is the only required section.
- An optional blockquote with a short summary.
- Optional prose — paragraphs and lists, not more headings — that explains how to read the rest.
- Zero or more H2 sections. Each section is a list of Markdown links — a name, a URL, optionally a colon and a note.
- An H2 titled `Optional` has a conventional meaning: skip these URLs if context is tight.
A companion file at `/llms-full.txt` appears in the same proposal: a longer dump for systems that want the pages concatenated, not just the map. Most sites do not need it. A map that lies is worse than no map.
The file is served as plain text. A person can open it in a browser and read it. That is the point. If a human cannot tell what the site is from the file, an agent will not either.
What the file can do, when something actually reads it:
- Name the organization in one sentence that matches the homepage.
- Point at the pages that define the entity, the offers, and the way to start work.
- Separate essential URLs from skippable ones.
- Restate a fact that is already visible in HTML.
What it cannot do is make an untrue page true, or make a thin page look like a finished presence.
What it is not
llms.txt is easy to oversell because it looks like `robots.txt` and sits in the same mental slot as a sitemap. It is neither.
| File | Job |
|---|---|
| `robots.txt` | Permission. What a crawler may request. |
| `sitemap.xml` | Discovery. Which URLs you want a search engine to know about. |
| `llms.txt` | Orientation. A curated note for agents that choose to read it. |
Permission is not orientation. Discovery is not ranking. A note is not a lever.
llms.txt is also not:
- A Google ranking signal. Search ignores it. Presence neither helps nor harms Google visibility.
- A guarantee that ChatGPT, Claude, Perplexity, or any other system will cite you.
- A replacement for crawlable HTML, titles, internal links, or useful copy.
- A place to publish prices, offers, or proof that the public pages do not carry.
- A required GEO tactic. Generative engine optimization, when the phrase means anything, is whether a page is safe to quote — not whether a text file exists at a fashionable path.
If a vendor sells "AI crawler documentation" as if the file were a backlink or a featured snippet, they are selling a filename. The filename does not rank.
It is a README
Write one the way you would write a README for a public repo.
Match the public site. The summary should be a compression of the homepage, not a second brand. If the site says you design and build brands, websites, and products, the file should say that. If the site does not claim a metric, the file does not invent one.
Links must resolve. Every URL in the list should be a real, current page. Dead links, redirects to the homepage, preview URLs, and "coming soon" paths do not belong. An agent that follows a broken link learns that the map is untrustworthy. A person who opens the same file learns it faster.
Never outrank the HTML. The HTML is the source. The file is a pointer. If the two disagree, the page wins — or should. Structured data has the same rule: it may clarify a visible fact; it may not introduce a claim the visitor cannot read. llms.txt is the same discipline in Markdown.
Keep it current. A README that lists last year's offers is worse than silence. When you retire a URL, remove it. When you publish a page you want an agent to prefer, add it. When the company description changes on the homepage, change the blockquote the same week.
Do not put secrets in it. The file is public. Private previews, internal tools, draft pages, and staff-only routes have no place in a document that any client can fetch.
Those rules are not a growth hack. They are how you avoid teaching a model a fictional company.
An example: CLICK.BLUE's /llms.txt
CLICK.BLUE publishes llms.txt. Use it as a format example, not as a ranking claim. We do not say the file makes the site rank. We do not say it gets the company cited. We say: if an agent looks for a note, this is the note, and it agrees with the site.
The file follows the proposed order. It opens with `# CLICK.BLUE` and a blockquote that restates the homepage fact: a creative and technical network founded in 2014 that designs and builds brands, websites, digital products, campaigns, software, AI agents, automation, and ongoing systems. The same sentence carries the public record — 12+ years in operation, 50+ businesses supported, 100+ websites launched. Those numbers appear on the site. They are not invented for the file.
Then the file does the README job:
- Canonical facts a model should not have to reconstruct: founded 2014, distributed network, remote and international work, project intake at get-started, contact by email.
- A short key for how to read the site: company pages, service pages, and offer pages are labeled so an agent does not treat a location page as a second company.
- H2 sections of links — essential pages, services, fixed-scope offers, then optional URLs that can be skipped.
After the first H2, every non-blank line is a Markdown link with an absolute `https://www.click.blue` URL and a one-line note. That is the proposal's file-list grammar. The notes restate what the destination page already says. They do not add a second offer.
The file is a map of the public site. It is not the site.
Never outrank the HTML
This is the rule that keeps the file honest.
An agent that only reads llms.txt can be misled. An agent that reads the file and then fetches a page can check. A person who lands from search will only see the page. If the file promises a service the page does not describe, or a proof the page does not show, you have published two companies. The cheaper one to write is the file. The one that gets judged is the HTML.
So the order of work is not "write llms.txt, then decorate a site around it." The order is:
1. Decide what the business actually sells. 2. Put that on public pages a stranger can finish. 3. Make those pages crawlable, fast enough, and internally linked. 4. If you want a note for agents, summarize the pages you already shipped.
Skip step four if you have not finished step two. A precise file about a vague site is still a vague site.
The same rule applies to answer engines. A model may quote a short, supported sentence from visible HTML. It may ignore you. Schema and llms.txt may repeat that sentence. They cannot mint a citation. Anyone who tells you otherwise is selling a file.
Google Search does not use it
Google's published guidance on optimizing for generative AI features in Search is blunt: you do not need new machine-readable files, AI text files, markup, or Markdown to appear in Google Search, including those generative features, because Google Search itself does not use them. A later note adds the practical line site owners actually needed: it is fine to create and maintain llms.txt for other services or systems that use the file. Doing so will not harm, and will not help, visibility or rankings in Google Search, because Google Search ignores it.
Read that twice. Ignores is the verb. Neutral is the outcome. Optional is the status.
That does not make the file illegal, stupid, or forbidden. It makes it the wrong place to spend the week if the homepage is still a slogan and the service pages still disagree with each other. Google may discover and crawl many kinds of files. Discovery is not special treatment. A crawled text file is not a ranking document.
If you maintain llms.txt, maintain it for the agents that might read it. Do not maintain it for Google. Do not report it to a client as an SEO deliverable that moved rankings. If rankings move, they moved because pages, links, demand, and time moved — the ordinary search work, not a filename.
What actually moves the work
The expensive part is still the presence.
Someone has to hold the brief. Someone has to write and structure pages a skeptical buyer will finish. Someone has to build a site that loads, states the offer, and does not hide the next step in a carousel. Someone has to launch without broken titles, orphan URLs, or a form that goes nowhere. Someone has to keep search honest after the first publish: technical gaps, thin pages, claims that drifted from the work.
llms.txt does none of that. It can point at the pages that did.
This is the same fork as the other tools. Webflow is software. A $20 AI plan is software. llms.txt is a text file. The unfinished presence is the expensive part: the designer, the writer, the marketer, the evenings, the site that does not yet look, read, or rank like a business.
CLICK.BLUE is a delivery team for that presence. A project manager scopes the work. Specialists design, write, build, and launch. We have done this since 2014 — more than twelve years, more than fifty businesses, more than a hundred websites. Sometimes the build includes an llms.txt that matches the public site. The file is a byproduct of pages we are willing to stand behind. It is not the product.
The full website is the scoped site: designed, written, launched. SEO is how search engines and buyers keep understanding the thing that shipped. One relationship. Written scope. You keep running the business.
When to write one
Write an llms.txt when most of these are true:
- The public pages already agree with each other.
- You can list the essential URLs without padding.
- Every link resolves to the page the note describes.
- You will update the file when the site changes.
- You want a note for agents that look for one — not a ranking project.
Skip it when the homepage is still mush, the offers are still being invented, or someone is asking you to "add GEO" by dropping a file on a site that cannot yet explain itself. In that case the next step is not Markdown at the root. It is the pages.
If you already know the offer and need the site to say it, that is a full website and the SEO that follows. If you want to master the file yourself, read the proposal, open our llms.txt, and copy the discipline — not the claims. Match your pages. Do not copy ours.
A filename will not rank you. Honest public pages might be found. That is the work we take.
The estimator at get started turns the current setup into a scoped proposal — the site and the search work, not a file for sale.