An llms.txt file is a plain Markdown file served at /llms.txt on your domain that lists your most useful pages, each with one factual line saying what the page answers. It is cheap to add, and one platform, Shopify, now serves it by default. None of the AI crawler documentation from OpenAI, Anthropic, Perplexity or Google that we track says the file is read when an answer is composed, so treat it as housekeeping you do after crawler access and readable HTML are fixed, not as a lever on its own.

Below: a copyable example, how to write the descriptions so the file is worth having, where the file lives on Shopify, WordPress, Webflow and HubSpot, and a two-command check.

What llms.txt is, and what nobody has promised

robots.txt is a control file the vendors document: OpenAI, Anthropic, Perplexity, Google and Apple each name their agents and say what blocking them does. llms.txt is different. It is a convention for summarising a site to a language model, and the crawler documentation we check for every vendor does not mention reading it. That is not a reason to skip it. It is a reason not to sell it internally as the thing that will get you cited, and not to measure your AI visibility programme by whether the file exists.

The strongest signal of adoption so far comes from a platform rather than an engine. Shopify serves /agents.md on every store by default and mirrors it to /llms.txt and /llms-full.txt, even with no theme file present. That tells you the format is worth getting right if you run a store. It does not tell you which assistants read it.

There are still two reasons to ship the file that hold whatever adoption turns out to be. The cost is one file. And writing it forces a useful exercise: deciding which pages actually represent what you do, and saying in one sentence what each one answers. You may find in the process that the page you would most like to list does not exist yet.

It also helps to be clear about how the file differs from the two control files it is usually mentioned alongside, because the three are often confused:

robots.txtsitemap.xmlllms.txt
JobAllow or block named agentsList every URL you want crawledSummarise your most useful pages for a language model
FormatGroups of user-agent and path rulesXML, one entry per URLMarkdown, one curated line per page
Vendor documentationEvery major AI vendor names its agents and how they obey itStandard for search enginesNo AI engine we track documents reading it
What goes wrongBlocking a search agent while meaning to block trainingStale or missing URLsStale facts, dead links, served as HTML
Replaceable by the others?NoNoNo, and it replaces neither

A working llms.txt example and its format

Shopify names its template agents.md.liquid, and the extension tells you what it expects: Markdown. The layout we recommend is the simplest one a person and a parser can both read: a title line, a short summary as a quote block, then sections of links with one line each.

# Example Co
> Example Co makes inventory software for independent retailers.
> This file lists the pages that describe the product, how it is
> priced, and where to get help. Every link is a public page.
## Product
- [Pricing](https://example.com/pricing/): The plans, what each one includes, and how billing per store location works.
- [Integrations](https://example.com/integrations/): The point-of-sale and accounting systems Example Co connects to, and what syncs.
- [Security](https://example.com/security/): Where customer data is stored and who can access it.
## Documentation
- [Getting started](https://example.com/docs/start/): Creating an account, importing products and connecting a till.
- [Returns and stock adjustments](https://example.com/docs/returns/): How a return changes stock counts across locations.
## Company
- [About](https://example.com/about/): Who Example Co is, which countries it serves, and how to contact the team.

Treat this layout as a sensible default rather than a specification. No engine documents a grammar for the file, so the value is in what the lines say, not in matching a particular parser.

llms.txt# Example Co> Who you are, in two or three> plain sentences.## Product- [Pricing](url): what it answers- [Security](url): what it answers## Documentation- [Getting started](url): …Titlethe site or brand name, onceSummarya quote block a model can lift wholeSectionsgroup links the way a buyer thinksEntriesone public link + one factual line
Figure 1. The four parts of an llms.txt file in the layout this article recommends. The layout is a sensible default, not a published grammar; what matters is that every line says what its page answers. Source: the example file in this article.

Each part has a job. The title names the site once, so nothing below has to repeat it. The summary is the one passage a model might lift whole, so write it as two or three plain sentences that would still make sense quoted on their own: what you are, who you serve, what the file contains. Sections group links the way a buyer would look for them, not the way your navigation happens to be organised. And every entry pairs one public link with one sentence of fact.

How to write each line so the file is worth reading

  • Select, do not dump. A file generated from your sitemap, hundreds of bare URLs with no descriptions, adds nothing a sitemap does not already provide. The value is in the selection and the one-line summaries.
  • Say what the page answers, with the subject named. "Pricing: the plans, what each includes, and how billing per location works" can be lifted on its own. "Learn more about our flexible options" cannot. The same discipline applies to the pages themselves, as described in writing for extraction.
  • Keep volatile facts out unless you generate them. A file that states a price you changed months ago is a liability, because it is a second, forgotten copy of the fact. If you want prices or plan names in it, build the file from the same data source as the pages.
  • List only public pages. Anything behind a login, a password or a form returns nothing useful to an automated reader.

Our longer llms.txt guide covers how to structure the file for a large site and what to leave out.

Where llms.txt lives on each platform

PlatformHow to serve itThe catch
ShopifyServed by default. Lookup order for /llms.txt is llms.txt.liquid, then agents.md.liquid, then Shopify's managed default.Adding agents.md.liquid changes /agents.md, /llms.txt and /llms-full.txt together; adding llms.txt.liquid changes only /llms.txt.
WordPress (self-hosted)Upload llms.txt to the document root. The recommended server config serves existing files before routing to index.php.Does not apply to WordPress.com hosted plans, where filesystem access is restricted.
WebflowA dedicated llms.txt upload serves the file from the root of a custom domain.Other arbitrary root-level files remain unsupported. This is reported from Webflow's help centre rather than verified by us.
HubSpotNo documented way to serve an arbitrary file at the root. A reverse proxy in front of HubSpot is the reliable option.A page whose slug is literally llms.txt resolves but is served as text/html, not text/plain.

The sources are Shopify's llms.txt template docs and agents.md template docs, WordPress's nginx guidance, Webflow's llms.txt upload article and HubSpot's supported file types. Platform-by-platform detail is on our Shopify and HubSpot pages.

Request for /llms.txttemplates/llms.txt.liquid in the published theme?yesServed. Only /llms.txt changes.notemplates/agents.md.liquid in the published theme?yesServed. /agents.md and /llms-full.txt change too.noShopify's managed default (also served as /agents.md)
Figure 2. The order in which Shopify resolves /llms.txt. Only templates in the published theme count. Source: Shopify's llms.txt.liquid and agents.md.liquid template documentation, checked 2026-09-09.

On Shopify, write your own agents.md.liquid rather than leaving the managed default, so the file describes your catalogue in your own words. On any platform, do not serve the file through a redirect from /llms.txt: programs that fetch control files commonly do not follow one. And if you edit the template in a theme you have not published yet, nothing changes on the live store until you publish it.

How to check it in two commands

# 1. Status, redirects and content type of the file itself
curl -sI https://example.com/llms.txt | grep -iE "^(HTTP|location|content-type)"
# 2. Every link inside it should resolve
curl -s https://example.com/llms.txt \
| grep -oE "https://[^) ]+" \
| while read -r u; do
printf "%s %s\n" "$(curl -s -o /dev/null -w "%{http_code}" "$u")" "$u"
done

What a healthy result looks like:

  1. The file returns 200 directly, with no location header, and a content type of text/plain rather than text/html. An HTML page that happens to live at that path is a different thing.
  2. Every listed link returns 200. A file full of 404s is a worse signal than no file. Re-run the check after any routing or URL change, because that is when links break silently.
  3. You tested the real domain. On Webflow, the webflow.io staging subdomain can serve different control files from your custom domain.

If the first command times out, record that as "could not fetch", not "file missing". Collapsing those two states is how an audit tells you to fix something that is already fine.

What to fix before llms.txt matters

If your pages return 403 to search agents or arrive as an empty JavaScript shell, a well-written llms.txt changes nothing, because the pages it points to are unreadable. Check, in order:

  1. Access. The search and user-triggered agents can fetch your pages. Test with the real user agent rather than reading robots.txt, since a CDN rule or firewall that returns 403 is a block in practice whatever the file says (see AI crawler access).
  2. Readable HTML. The content is in the HTML without scripts. Most AI crawlers do not execute JavaScript, and a page that returns 200 with an empty body is a failure, not a pass.
  3. Answer-first pages. Each page opens with a passage that answers its question, so whatever reads it has something to quote.
llms.txta curated index of pages that already workAnswer-first passageseach page opens by answering its questionContent in the HTMLreadable without running JavaScriptCrawler accesssearch and user agents get a 200, not a 403Optional, cheap,not documented as readFix these first: withoutthem the file points at nothing
Figure 3. Where llms.txt sits among the things that make a page usable to an AI engine. The bottom three are preconditions; the file on top only helps once they hold. Source: the checks described in this article.

Then add the file, keep it short, and keep it true. A small file that lists ten honest pages is more useful than a long one nobody maintains, and it is far easier to keep in step with the site when prices, plans or URLs change.

Sources