llms.txt vs agents.md: What Every Brand Needs to Know

If you have spent any time reading about AI search lately, you have probably seen two new file names come up again and again: llms.txt and agents.md. Both sit quietly in the root folder of a website or codebase. Both are meant to help AI systems understand things faster. And both are often confused with each other, even though they solve very different problems.

For brands trying to win visibility in AI search results, understanding this difference is not just a technical detail. It affects how AI Overviews, ChatGPT, Claude, and Perplexity find, read, and cite your content. This post breaks down what each file actually does, where the industry stands on them in 2026, and what your brand should actually do about it.

What is llms.txt?

llms.txt is a plain-text file websites can publish at a set location (yoursite.com/llms.txt) that gives AI models a clean, direct summary of a site’s key pages and content, so an AI doesn’t have to dig through menus, ads, pop-ups, and other clutter to find the real information.

It was first proposed in September 2024 by Jeremy Howard, co-founder of Answer.AI and fast.ai. The problem he was solving is simple: AI models can only read a small amount of text at a time, and most websites are cluttered with menus, ads, pop-ups, and code that get in the way. A person can skip past a cookie banner or sidebar without thinking; an AI can’t skip it as easily. It has to search through all the clutter to find the real content, which uses up its limited reading space.

A simple, real-life way to picture it: say you run a small bakery website with pages for your menu, your story, your locations, and a blog. Without llms.txt, an AI has to open each page and dig through the header, footer, and pop-ups just to find your opening hours. With llms.txt, you leave a short note at sunrisebakery.com/llms.txt. Now an AI trying to answer “what time does Sunrise Bakery open” can go straight to that one page instead of guessing its way through the whole site.

This is exactly the same logic real companies use at scale. Anthropic, Stripe, Cloudflare, and Vercel all publish live llms.txt files today, for example at docs.anthropic.com/llms.txt, just built for documentation sites with hundreds of pages instead of four.

How the file is actually structured

• The spec is deliberately narrow. A valid llms.txt follows one set order:
• Optional section: a special ## Optional heading for links that are useful but safe to skip if an AI is short on space. A simple example looks like this:
• H1 heading: the name of the project or site. This is the only required part.
• Blockquote: sits right under the H1, a one-sentence summary of the site.
• Optional context: a few paragraphs or lists with more background, if needed.
• H2 sections: each one holds a bullet list of links, written as [Page Name](URL): short description.

A simple example of llms.txt looks like this:

# Acme Docs

> Acme is a project management tool for engineering teams. This file helps AI tools find the most useful pages fast.

Acme's documentation covers setup, the API, and integration guides.

## Docs
- [Getting Started](https://acme.com/docs/start): Install and configure your first project
- [API Reference](https://acme.com/docs/api): Full list of endpoints and parameters

## Optional
- [Changelog](https://acme.com/changelog): Release history, safe to skip for most tasks

What it is meant to be used for

The original goal was developer documentation, so coding assistants and AI agents could quickly find API references and setup guides without wading through a full website. Since then, some marketing and ecommerce teams have adopted it more broadly, hoping it would also help their product pages and blog content get pulled into AI Overviews and chatbot answers.

What is agents.md?

agents.md solves a different problem. It started with OpenAI Codex, and tools like Cursor, Amp, Google’s Jules, and Factory helped shape it from there. In December 2025, it moved to the Agentic AI Foundation, a group under the Linux Foundation, so it is no longer tied to just one company.

Think of agents.md as a README, but written for an AI agent instead of a new employee. A normal README tells a new hire what a project does. agents.md tells an AI agent how to actually work on it: the exact commands to build and test the code, style rules that are not obvious just from reading the code, and which files or folders it should never touch.

Example: say a new developer joins your team. Before they touch any code, a teammate probably tells them things like “run npm install first,” “we use 2 spaces, not tabs,” and “never edit the /legacy folder.” agents.md is that same conversation.

How the file is actually structured

Unlike llms.txt, agents.md has no required format at all. It is just standard markdown, and teams write whatever sections make sense for their project. The sections that show up most often across real-world repositories are project overview, setup or dev environment commands, code style guidelines, testing instructions, security considerations, and pull request or commit message conventions. A simple example looks like this:

# Acme Web App

## Setup commands
- Install dependencies: `pnpm install`
- Start the dev server: `pnpm dev`

## Code style
- TypeScript strict mode, no use of `any`
- Functional components with hooks, no class components

## Testing
- Run all tests: `pnpm test`
- Run a single file: `pnpm test path/to/file.test.ts`

## PR instructions
- Title format: [component] short description
- Run lint and tests before committing

Some tools also support layered versions of the file, so a large project can have one general agents.md at the root and more specific ones inside subfolders, with the more specific file taking priority for agents working in that area.

What it is meant to be used for

agents.md exists to make AI coding agents useful from the first prompt instead of needing several rounds of correction. It hands an agent the same context a human teammate would give a new hire: how to build the project, how to run the tests, and which conventions and files are off limits. Because every major tool reads the same file, a team writes it once and it works whether someone is using Claude Code, Cursor, Copilot, or Codex.

Where adoption actually stands in 2026

By early 2026, agents.md was read natively by Claude Code, OpenAI Codex, Cursor, Devin, GitHub Copilot, Gemini CLI, Windsurf, and Amazon Q, and more than 60,000 repositories were already using it. That sits inside a bigger shift: separate studies estimate 22 to 28 percent of all GitHub projects now show some trace of coding-agent assistance. Even so, context files as a category are still early, one survey of 10,000 repositories found only about 5 percent had adopted any context file format at all, agents.md included, making it the leading standard in a space that’s still mostly unclaimed.

Now that we understand what each one is, it’s necessary that we know exactly how they differ.

llms.txt vs agents.md

	llms.txt	agents.md
Proposed by	Jeremy Howard, Answer.AI, September 2024	OpenAI Codex, with Cursor, Amp, Jules, and Factory; donated to the Agentic AI Foundation in December 2025
Main purpose	Helps an AI find and understand your content	Helps an AI agent build, test, and edit your code correctly
Lives at	yoursite.com/llms.txt (public website root)	Root of a code repository, with optional versions in subfolders
Format rules	Specific structure: H1, then blockquote summary, then H2 link sections	No required structure, free-form markdown
Typical audience	AI search crawlers, chatbots, and AI-assisted browsing tools	AI coding agents like Codex, Claude Code, and Cursor
Typical content	Site summary and links to clean versions of key pages	Build commands, test steps, code style rules, PR conventions
Who actually reads it today	Mostly IDE and coding tools when pointed at docs sites, not major AI search crawlers	Coding agents across nearly every major AI development tool
2026 adoption	Roughly 5 to 9 percent of top websites	Used across more than 60,000 public repositories
Effect on AI search citations	Minimal on its own, since most answer engines skip it	Not applicable, it does not target search visibility
Effect on coding agent accuracy	Not applicable, it does not guide coding behavior	Meaningfully reduces errors and rework in AI-generated code
Best suited for	Developer documentation sites and content-heavy sites, as a supporting signal	Any codebase that AI coding agents regularly work in

Should You Actually Implement These and Do They Work?

llms.txt does not move AI search visibility on its own

• Google says so directly, in its own guide to optimizing for generative AI features: “You don’t need to create new machine readable files, AI text files, markup, or Markdown to appear in Google Search… Google Search itself doesn’t use them.” Google adds that having the file will “neither harm nor help your site’s visibility or rankings.”
• A large-scale analysis of roughly 300,000 domains, reported by Search Engine Journal, built a model to predict AI citation frequency from site signals. Removing llms.txt from that model actually made its predictions more accurate, meaning the file was adding noise, not a real signal.
• Most major AI crawlers, including GPTBot, ClaudeBot, and PerplexityBot, do not fetch the file at all during a normal crawl.

Where llms.txt does help: developer-facing sites read by coding tools

• Cursor, Claude Code, Windsurf, and similar assistants do fetch and use the file when pointed at documentation.
• A company like Stripe or Anthropic sees a genuine, if narrow, payoff here: coding agents answer questions about their API more accurately. That is a developer-experience benefit, not an SEO or AI search one.

agents.md works, but only if someone does it properly

• Independent research found a well-written, human-authored agents.md improves coding agent success rates by roughly 4 percent across tools, a real but modest gain.
• The same research found the opposite is also true: AI-generated or copy-pasted agents.md files can hurt performance, lowering success rates and raising the compute cost of each task.
• About half of the agents.md files studied had zero updates after the day they were created, so a lot of teams write one once and let it go stale as the codebase changes.

How to actually get a benefit (Step-by-Step)

For llms.txt:

1. Skip it if AI search citations or Google visibility are the goal. The evidence above says it will not move that needle.
2. Build it if you run public developer docs, an API, or an SDK that coding assistants need to reference.
3. Keep the links inside it current. A broken or outdated llms.txt actively confuses the coding tools reading it.
4. Put your real AI-search budget into content structure, quality, and third-party mentions instead, since that is what actually earns citations.

For agents.md:

1. Write it yourself, or have a senior engineer review it line by line. Do not let an AI tool generate the whole file unsupervised.
2. Lead with exact, runnable commands: build steps, test commands, and lint commands, with the real flags, not vague descriptions.
3. Include what a new hire would normally have to ask about: code style exceptions, folders to avoid, and any manual approval steps.
4. Revisit it every time your build process, folder structure, or coding conventions change. A stale file is worse than no file.

Knowing that neither file is a shortcut to AI search visibility changes what your team should actually spend time on. The fixes that move real citation and ranking numbers are more foundational: clear content structure, technical SEO health, solid internal linking, and a steady stream of genuine mentions across the web that AI models can trust.

The hard part for most brands is not knowing that these fundamentals matter, it is not having a clear way to see where they currently stand across every AI answer engine, or the hands-on support to close the gaps once they are found. That is the specific problem a platform built for this moment needs to solve.

Where This Fits into AI Search Strategy

This is exactly the kind of nuance that trips brands up. Chasing every new file format is not a strategy. What actually moves the needle for AI visibility is structured, clear, well-linked content that both traditional search engines and AI answer engines can trust and cite, backed by the technical basics like fast pages, clean schema, and solid crawlability.

This is where Quattr comes in.

Quattr is a platform built to unify SEO, AEO, and GEO in one place, so brands do not have to guess which emerging standards actually matter for their traffic. It tracks where a brand shows up across Google AI Overviews, ChatGPT, Claude, Perplexity, and AI Mode, pulling results from real consumer-facing AI responses rather than APIs, so the data reflects what users actually see, including citation tracking, brand mention monitoring, share of voice, and competitor comparisons.

Beyond monitoring, Quattr’s AI agent, GIGA, goes a step further by actually executing fixes, from optimizing content and internal links to closing technical gaps that block AI crawlers, rather than just flagging problems in a dashboard. Every account also gets support from a dedicated SEO and GEO expert with more than a decade of experience in enterprise search, useful when you are trying to figure out whether something like llms.txt is worth your engineering team’s time.

The results show up in practice. Kiteworks saw a 79 percent expansion in AI Overview presence after working with Quattr, with its content citation rate climbing 20 percent above baseline within just one week. The company also reversed a site-wide indexation decline, growing indexed pages by 30 percent and increasing top-3 rankings by 22 percent.

The AI search landscape will keep producing new file formats and acronyms. The brands that win will separate genuinely useful signals from passing trends and put their attention on the fundamentals that consistently earn citations and traffic.

Want to see where your brand currently stands in AI search? Book a demo with Quattr to see your AI citation share and get a plan to grow it.

FAQs