llms.txt in 2026: A Reality Check Before You Ship One

llms.txt is the most confidently recommended file on the internet that almost nothing reads.

It was proposed in September 2024 by Jeremy Howard of Answer.AI as a way to give language models a clean, markdown map of a site: the original proposal is still the clearest statement of intent. The idea is reasonable. Context windows are finite, HTML pages are full of navigation and scripts, and a curated markdown index of your documentation is genuinely easier for a model to work with than your rendered site.

The problem is what happened next, which is that the file became an SEO deliverable before any consumer of it existed.

What The Spec Actually Proposes

Two files at the root of a domain. /llms.txt is a short markdown document: a title, a paragraph of background, and a linked list of the pages that matter with one-line descriptions of each. /llms-full.txt is an optional companion containing the full content in a single markdown file.

Nothing about it is a directive. It does not grant or deny access, unlike robots.txt, which people constantly assume it resembles. It is a suggestion to a reader that has to choose to look.

Who Reads It

As of mid-2026, the honest answer is: documentation tooling, some coding assistants when you paste the URL yourself, and not much else at retrieval time.

Google stated in 2026 that llms.txt has no effect on Search rankings or on AI Overviews, and that Search simply ignores the file. That is about as unambiguous as platform guidance gets. Independent crawl studies through mid-2026 have not found evidence that publishing one improves your odds of being cited by ChatGPT, Claude, Gemini, or Perplexity in their answer surfaces.

Adoption numbers are also less meaningful than they look. Overall percentages jumped sharply in 2026, but a large part of that came from Shopify pushing a default file to every store on the platform rather than from publishers choosing to adopt it. A default is not a signal.

Where It Genuinely Helps

Three cases, all of them real, none of them about rankings.

Developer documentation. If people paste your docs into an assistant to build against your API, a clean markdown index is a direct quality improvement for them. Anthropic, Stripe, and several docs platforms publish one for exactly this reason.

Your own retrieval. If you feed your site into an internal assistant or a support bot, a curated index saves you writing a scraper.

Agent-facing products. If your users are running agents against your site on purpose, giving those agents a map is basic hospitality.

Notice that none of those are marketing sites. If your site is a blog and a pricing page, an llms.txt file is a tidy artifact that nothing consumes.

The Cost Of Shipping One Anyway

It is small but not zero, and it is mostly maintenance. A stale llms.txt that lists pages you deleted is worse than none, because the one category of tool that does read it will follow the links. If you publish one, it needs to be regenerated on deploy, not written once by hand.

The bigger cost is opportunity. An afternoon spent generating llms.txt is an afternoon not spent making the twelve pages that matter specific and quotable, which is the work that is actually supported by evidence. We go through that work in generative engine optimization: what actually works.

If You Do Publish One, Write The Descriptions Properly

The linked list is the whole value, and most generated files fill it with the page title twice.

Before:

• Pricing: Pricing
• Docs: Documentation for our product
• Blog: Blog

After:

• Pricing: Free and Pro tiers, monthly and annual figures, and what each tier includes.
• Local mode setup: Downloading a model, hardware requirements, and offline behaviour.
• Privacy boundary: Exactly what leaves the device in each mode, and what does not.

The second version is useful to a model deciding which page to fetch. The first is a sitemap with extra steps.

A Wrivio Context for writing those descriptions could say:

Rewrite each of these link descriptions into one sentence that states what a reader will find on the page and what question it answers. Do not repeat the page title. Keep every product name, figure, and page path exactly as written. Do not invent features or sections that are not described in the source.

Press Ctrl+Shift+Space, paste the list, and check the diff. Watch for descriptions that became more confident than the source justified, which is the usual failure when compressing a page into one line.

The Pattern Worth Recognising

llms.txt is the third file in five years to be sold as the thing that makes AI systems notice you, and it follows the same arc: a sensible technical proposal, a wave of tooling, a wave of consultants, and then platform guidance clarifying that it does not do the thing it was being sold for.

The recurring lesson is that file-level signals almost never substitute for content quality, because the thing doing the retrieving is reading your sentences either way. That is also the reason nobody needs to guess which sources get cited: the patterns are observable and they are about content, not configuration.

Common Questions

Does Google use llms.txt?

No. Google stated in 2026 that the file has no effect on Search rankings or AI Overviews and that Search ignores it. Publishing one is neither a benefit nor a penalty there.

Is llms.txt an official standard?

No. It is a proposal published by Answer.AI in September 2024. As of mid-2026 it has no W3C, IETF, or schema.org recognition.

Will it stop AI companies training on my site?

No. It contains no access controls of any kind. Crawler access is governed by robots.txt, WAF rules, and your infrastructure provider’s controls.

Should I remove the one I already have?

Keep it if it is generated automatically and stays current, especially for developer docs. Remove it if it is hand-maintained and already out of date.

Download Wrivio for Windows to turn a page list into descriptions a reader can actually use, without the rewrite inventing sections you never wrote.