Product

Solutions

More

llms.txt Example: A Copyable File, a Real One, and How to Validate It

A copyable llms.txt example, the real llms.txt on polydraft.ai with our validation results (including a missing charset), what changed in version 2, who reads the file today, and a validation checklist.

The parts of an llms.txt file: a heading with the site name and summary, sections of links to answering pages, and validation of the response, links and facts.

A useful llms.txt example is a short Markdown map of reliable website information: a heading with the site's name, a one-line summary, and a few sections of links to the pages that answer the questions people actually ask. Below is a copyable llms.txt example, the real file we publish on polydraft.ai with what we found when we checked it, and the validation steps we use. The file is optional; it doesn't replace a readable website.

Updated November 6, 2026: re-checked against version 2 of the proposal (August 10, 2026), and added who reads llms.txt today, our own file as a real example with validation results, and common mistakes.

For Google Search specifically, llms.txt is unnecessary. As of November 6, 2026, Google's guide to optimizing for generative AI features says Search doesn't use these files, and that having one neither helps nor hurts your visibility or ranking there. Build one because a tool or agent you care about can use it, not because a checklist promises an SEO boost.

  • Keep a curated index; do not copy the entire website into one file.
  • Link only to information you can verify and maintain.
  • Serve it as UTF-8 and say so in the Content-Type header.
  • Test retrieval separately from answer quality and public search visibility.
  • Keep essential content on accessible pages for people and crawlers.

Understand the current llms.txt proposal

llms.txt is a community proposal by Jeremy Howard, not an internet standard. Its current text is version 2, dated August 10, 2026. Only an H1 with the site or project name is required. After it come a blockquote summary, optional paragraphs of context, and H2 sections that each hold a Markdown list of links, each with optional notes after a colon.

Version 2 made three changes that matter for a business site. A file can sit at the root or under a path, and it covers the pages under that path; where several apply, the most specific wins, so /docs/llms.txt covers /docs/. Pages can point to their Markdown version with rel="alternate" type="text/markdown" and to the llms.txt that covers them with rel="describedby", as HTML <link> elements or an HTTP Link header. And a section named Optional no longer has a special meaning for tools; it's still a useful convention for secondary links.

Those are proposal conventions, not a promise that every assistant discovers or follows them. Maintain a simple file your intended tools can actually consume. Do not advertise Markdown copies that your site does not serve, or treat a passing syntax check as proof of adoption.

ArtifactPractical purposeDoes not establish
robots.txtState crawler access preferencesPrivate access control or guaranteed indexing
XML sitemapProvide discoverable site URLsGuaranteed inclusion in search
llms.txtOffer a curated reading map to supporting toolsGoogle ranking or AI citation improvement
llms-full.txtSome documentation sites publish their full docs in one file; it isn't part of the proposalThat any assistant reads it
The actual pageProvide the information a person or system needsAccuracy merely because it is online

Google's sitemap overview and its robots.txt introduction explain the separate discovery and crawling roles. Keeping those responsibilities separate makes a broken setup easier to diagnose.

A pin board where one hub pin is tied to three source pins while the other pins stay untied: a compact index points to maintained source pages instead of copying them.
Keep current facts in maintained source pages rather than copying every detail into the index.

Who reads llms.txt today

As of November 6, 2026, the clearest readers of llms.txt are coding agents and tools you point at a file, not search engines. Here is what each party says in its own documentation:

  • Google Search: doesn't use llms.txt. Google added a note saying so to its generative AI optimization guide on June 15, 2026.
  • Chrome Lighthouse: its Agentic Browsing category includes an llms.txt audit that looks for the file at the site root. Lighthouse marks a 404 as not applicable, because the file is optional. That audit checks that the file exists and has basic structure, such as an H1 and links; it isn't evidence that any assistant reads it.
  • OpenAI, Anthropic, Perplexity and Microsoft: we found no official statement that their crawlers or assistants read llms.txt on other people's sites. The crawler pages from OpenAI and Anthropic describe robots.txt controls and don't mention llms.txt.
  • Publishers: the same companies publish llms.txt files for their own developer docs: OpenAI, Anthropic and Google's Gemini API docs. Cloudflare's docs publish one file per product. Documentation platforms such as Mintlify, and Wix for eligible sites, generate one automatically.

The proposal's author writes in the v2 change log that thousands of sites publish the file and that coding agents use them reliably. That is his account, not a measurement we can check. The practical reading: publish one if your buyers or their agents work with your documentation or product facts, and don't expect it to change AI search results.

A copyable llms.txt example for a business website

The fictional company below sells scheduling software. The names, products and URLs are examples, not customer evidence. Replace every entry with a real, public, maintained destination before publishing. This example uses ordinary page URLs; use verified Markdown alternatives where your site provides them.

# Example Scheduling

> Scheduling software for independent service teams.

This index describes public product information. For current plan
limits, use the pricing page. For support policies, use the help center.

## Product

- [Features](https://example.com/features): Supported scheduling workflows.
- [Pricing](https://example.com/pricing): Current plans, limits and billing terms.
- [Getting started](https://example.com/docs/start): Account setup and first booking.

## Company and support

- [About](https://example.com/about): Company identity and contact information.
- [Help](https://example.com/help): Troubleshooting and support routes.

## Optional

- [Release notes](https://example.com/updates): Dated product changes.

The last section is called Optional by convention, for links an agent can skip when it needs a shorter context. Since version 2, tools aren't expected to treat it specially, so don't hide anything there that an answer depends on.

Notice what this file leaves out: every blog post, promotional superlatives, unpublished roadmap items and copied prices. Linking to the page that lists your current plans is easier to maintain than storing a second price list in a text file that someone may forget to update.

The short descriptions distinguish which page answers which question. “Learn more” repeated six times would be less useful. An agent asking about support should not have to guess whether a release note, company page or landing page is authoritative.

A real example: polydraft.ai/llms.txt

We publish an llms.txt on our own site. On November 6, 2026, polydraft.ai/llms.txt had an H1, a one-line summary, two paragraphs of company context, and 10 H2 sections holding 29 links. Here is an excerpt; the omissions are marked.

# PolyDraft

> PolyDraft is an AI website builder that researches a business and its competitors, then plans, writes, builds and quality-checks a complete SEO- and AI-search-ready website in one session — deployed to the customer's own cloud account with a built-in admin panel and support for 100 content languages.

[… two paragraphs of company context omitted …]

## Core pages

[…]
- [Features](https://polydraft.ai/features): Research and planning, SEO foundation, GEO/AI-search readiness, built-in admin, design and motion, localization, ownership and content integrity.
- [How it works](https://polydraft.ai/how-it-works): The 4-step brief wizard, the autonomous build, the automated quality gate, live preview and one-click deploy.

[…]

## Free tools

- [llms.txt generator](https://polydraft.ai/llms-txt-generator): Build a spec-shaped llms.txt file for any website, in the browser, with no sign-up.

[…]

## Key facts

[…]
- Billing runs on Stripe in USD. Refund requests are accepted within 14 days of purchase via support@polydraft.ai, with a first human reply within 5 business days.

[…]

## Optional

[…]
- [Terms of service](https://app.polydraft.ai/terms)
[…]

The sections are grouped by the job a page does: core pages, capabilities, trust and policies, audiences, free tools, the blog, key facts, optional links, and release checks. Each link has a short note that says which question the page answers. A "Key facts" section repeats a few facts as plain bullets with no links, such as the operating company, plan limits and the refund window.

Terminal output of curl on polydraft.ai/llms.txt: status 200 with content type text/plain and no charset, the H1 PolyDraft and ten H2 headings from Core pages to Website rebuilding and release checks, a count of 29 linked entries, and 6 lines containing an em dash.
polydraft.ai/llms.txt, real curl output run November 6, 2026. Every line is real command output.

We ran the validation steps below against this file on November 6, 2026:

  • Response: HTTP 200 with Content-Type: text/plain, and no charset in the header.
  • Links: all 26 links to polydraft.ai returned 200. The 3 links to app.polydraft.ai (sign-in, terms, privacy) didn't respond from our test network, so we couldn't check them there.
  • Descriptions against sources: the plan prices and credit amounts under Key facts matched our plans page that day. The summary line still uses our older description of the product; we now describe PolyDraft as an AI-native website marketing system, so that line is due for an update. We're also reviewing the two context paragraphs we left out of the excerpt against our sources.
  • Discovery: our pages don't carry a describedby link and we don't serve Markdown versions, so the file points to ordinary HTML pages.

Common llms.txt mistakes, including ours

The most common llms.txt mistakes are encoding, stale descriptions and second copies of facts. Our own file had all three on November 6, 2026:

  1. No charset in the header. The file is UTF-8, but the server sends Content-Type: text/plain without charset=utf-8. Chromium then decodes it as windows-1252, and every em dash turns into —; we counted 8 of them on screen. A program that reads the bytes as UTF-8 is unaffected, but a person checking the file in a browser sees mojibake. Send Content-Type: text/plain; charset=utf-8, as OpenAI's and Cloudflare's files did that day.
  2. A summary that drifted from the product. Positioning changes faster than anyone remembers to edit a text file. Put llms.txt on the same checklist as the homepage, or run it as one check in a GEO audit.
  3. A second copy of prices. Our Key facts section repeats plan prices. They were correct on November 6, but they're one more place to update when prices change, which is the risk the example above avoids.
  4. Hiding important links under Optional. Since version 2 it's only a convention, and an agent short on context may skip it.
  5. Advertising Markdown versions that don't exist. If you add rel="alternate" links, every one has to resolve.
polydraft.ai/llms.txt opened in Chromium: the H1 PolyDraft and the summary line, where the em dash before 'deployed to the customer's own cloud account' is shown as the garbled characters —.
polydraft.ai/llms.txt in Chromium, captured November 6, 2026. The page reports its encoding as windows-1252; we enlarged the font for legibility and didn't change the text.

Select sources by responsibility, not traffic

Start with the questions your buyers or users actually ask. For each question, nominate one page that owns the answer. If several pages contradict one another, resolve that inconsistency before adding them to the map. Curation cannot make conflicting source material reliable.

QuestionGood source candidateReason to exclude or fix
What does the product do?Current features or product documentationPage describes planned capabilities as available
What does a plan include?Maintained plans and billing termsOld launch announcement with obsolete limits
How do I perform a task?Verified step-by-step documentationUnpublished instructions or broken assets
Who operates the service?About and contact informationConflicting company names or unsupported credentials
A four-drawer cabinet with one drawer open and a card raised for checking: include a source only if it is authoritative, public, maintained and distinct.
Exclude stale, private or redundant destinations until their underlying issues are resolved.

Our suggested inclusion test has four questions: Is the page authoritative for this topic? Is it public? Is it maintained? Does it add something the other selected pages do not? These are editorial criteria, not a score specified by the proposal. A short file that passes them is more useful than a long inventory of indistinguishable URLs.

For a multilingual site, make the language context explicit and choose the appropriate regional source. Do not mix a US plan description with another market's terms. When a translation is incomplete, link to an accurate source and describe its language instead of pretending the localized answer exists.

Validate the response, links and facts

Run a validation pass whenever the file changes and when important source pages change. Save the date and the final URL so the result is reproducible. A successful check today does not establish that a future release will preserve the same paths.

  1. Fetch the file: open its public URL without an admin session, or run curl -s -o /dev/null -w '%{http_code} %{content_type}\n' https://example.com/llms.txt. Expect a 200, a text content type that names charset=utf-8, and the intended Markdown rather than a login page or application shell.
  2. Inspect the structure: check the H1, the summary and readable link sections. grep -E '^#' lists the headings. Remove broken Markdown, duplicated entries and accidental private notes.
  3. Follow every selected link: record the final response, destination and whether the relevant text is available.
  4. Compare descriptions with sources: verify product scope, plan and price references, company identity and language.
  5. Check discovery if implemented: confirm any advertised Markdown alternative or describedby link resolves to the resource it claims.
  6. Assign maintenance: name the role responsible for updates when selected pages move or change.

If you implement version 2 discovery, the links look like this on a page under /docs/. The header form comes from the proposal and also works for non-HTML files:

<link rel="alternate" type="text/markdown" href="/docs/page.html.md">
<link rel="describedby" href="/docs/llms.txt">

Link: </docs/page.html.md>; rel="alternate"; type="text/markdown", </docs/llms.txt>; rel="describedby"
A four-level tray, one level per check from the file response up to a bounded answer test: validate each layer instead of relying on a single valid badge.
An answer test supplied with llms.txt does not demonstrate independent search discovery.

Use a small worksheet rather than a single “valid” badge:

File URL | Checked at | Final status | Content type and charset | Expected content present
Linked URL | Final destination | Description accurate | Owner
Question tested | Source used | Answer correct | Missing information
Change required | Responsible role | Retest date

If the index points to a redirected page, decide whether that redirect is intentional and update the link to the stable destination where appropriate. If the source is private, remove it from the public index and design a separate authenticated workflow. Never put secrets or customer-specific URLs into a public file.

Evaluate usefulness without claiming search results

Choose a few representative questions, such as “Which plan includes this capability?” and “Where are the setup instructions?” Give a compatible agent the file as its starting reference, ask it to identify the source for each answer, and compare the response with the current pages. This is a bounded retrieval exercise, and it's the test the proposal itself suggests.

Record the tool, model if available, date, exact question and supplied context. Count an answer as successful only if it is accurate and points to a supporting source. A correct answer produced after you explicitly supplied llms.txt does not prove that an unrelated public search assistant would discover the file or cite your site. To track whether assistants cite you over time, see the GEO tools we ran on our own brand.

Separate four observations: the file exists; a tool fetched it; a tool answered correctly from its linked sources; an independent search experience cited the site. Each needs its own evidence, and our dated AI search test shows how to record the last one. This prevents a deployment check from being reported as a visibility result.

Frequently asked questions

Does llms.txt improve Google rankings?

No. As of November 6, 2026, Google says Search doesn't use llms.txt, and that having one neither helps nor hurts visibility or ranking in Google Search. Treat the file as an optional resource for other systems that use it, and invest first in accurate, accessible pages.

Which AI assistants read llms.txt?

We found no official statement from OpenAI, Anthropic, Perplexity or Microsoft that their crawlers or assistants read llms.txt on other sites, as of November 6, 2026. Coding agents and tools you give the file to can use it. Chrome's Lighthouse checks whether the file exists and is well formed, which isn't the same as an assistant reading it.

Should the file list every page?

Use a curated set that helps the intended reader or agent choose the right source. An exhaustive list can make that choice harder. Keep the sitemap for broad URL discovery, and give llms.txt entries specific descriptions that explain why a destination matters.

Do I need a Markdown copy of every page?

The proposal recommends clean Markdown resources, but do not invent URLs or duplicate an entire site without a maintenance plan. Start with useful, reachable sources. If you offer alternate representations, validate that they remain accurate and equivalent as the original pages change.

What is llms-full.txt?

It's a single file with a site's full documentation, which some docs sites publish next to llms.txt; OpenAI's API docs have one. It isn't part of the llms.txt proposal, and we found no assistant that says it reads one. A business site rarely needs it.

Build a map you can maintain

Use the PolyDraft llms.txt generator to prepare a starting file in your browser, then run the source and link checks above. If tools can't fetch your pages at all, start with our manual AI crawler access check. The meaningful deliverable is a trustworthy map of useful pages, with evidence of how it was checked.

We make PolyDraft, an AI-native website marketing system. Every plan includes an llms.txt, along with a sitemap and an AI-crawler-friendly robots.txt. Check the generated file the same way as above before you rely on it, and edit any description that no longer matches your site.

Review access decisions separately; a curated link file does not grant access. Read the robots.txt examples for search and AI crawlers.

Use the broader guide to distinguish site readiness from observed citations. Read the generative engine optimization guide.

Sources

Sources were checked on November 6, 2026. We read the llms.txt proposal from its source files in the project's public GitHub repository, and read the OpenAI, Anthropic and Cloudflare pages and files directly. Google's and Chrome's pages couldn't be opened from our tools that day, so we checked those points against search excerpts of the publishers' own pages.

About this article: Prepared with AI assistance for PolyDraft, checked by an automated QA script and a separate AI fact-check, and approved by Vincent Chen. The scheduling-software example and the diagrams are illustrative, not customer results. The polydraft.ai findings come from checks we ran on our own file on November 6, 2026; the terminal image is real curl output and the browser image is a screenshot from that day. The hero is an AI-generated editorial illustration. For corrections, contact PolyDraft with the page URL and supporting source.

Founder of PolyDraft

Vincent Chen is the founder of PolyDraft and CEO of BlackMonolith, Inc. He has worked on technical SEO for multiple products and websites, and runs polydraft.ai on PolyDraft.

Put your site’s SEO and GEO on autopilot

PolyDraft builds your website, then finds the keywords you can win every week and writes the content.

Start your site

Start with a brief.
Publish when it’s right.

Build and preview your whole site for free. Every week after launch, PolyDraft finds the keywords you can win and writes the content.

Free to build and preview. Pay only to publish.