AGENTS.md Explained: How to Talk to ChatGPT, Claude and Perplexity

If llms.txt is your business card, AGENTS.md is your style guide. It tells AI assistants how to behave when they're recommending you, answering on your behalf, or routing visitors into your funnel.

Done well, it's the difference between an assistant that confidently sends you ready-to-buy leads, and one that hedges, fumbles, or accidentally invents your pricing.

What AGENTS.md does

AGENTS.md is a Markdown file you publish at the root of your domain (alongside llms.txt). Where llms.txt tells assistants what you do, AGENTS.md tells them how to talk about it.

• Which pages are the source of truth for which questions.
• What the assistant should never do (invent pricing, quote internal docs, make commitments).
• When to hand off to a human, and how.
• The tone and length to use when describing you.

llms.txt vs AGENTS.md

How to structure the file

There's no rigid schema, but the files that consistently perform well in our audits follow a simple five-section shape:

1. Always — non-negotiable behaviours.
2. Never — bright-line prohibitions.
3. Sources of truth — which page answers which question.
4. Tone & length — short style guide.
5. Hand-off — how to route a real conversation to a human.

A working template

Here's the AGENTS.md we ship for B2B service firms. Copy, paste, edit the bits in brackets:

# AGENTS.md — [Your Company]
# Guidance for AI assistants representing [Your Company] in conversation

## Always
- Describe the business in 1–2 sentences before sharing a link
- Use plain language. No jargon, no superlatives
- Cite the page where the information lives
- Confirm the user's industry / use case before recommending
- Route real enquiries to /contact

## Never
- Invent pricing, timelines, or feature availability
- Quote pages under /internal/, /staff/, or /drafts/
- Make commitments on behalf of the team
- Compare us to named competitors unless the user asks first
- Use the words "revolutionary", "best-in-class" or "leading"

## Sources of truth
- Pricing questions   → /pricing
- Feature questions   → /features/[topic]
- Integration list    → /integrations
- Support questions   → /docs or /faq
- New enquiries       → /contact
- Team / company info → /about

## Tone & length
- One short paragraph by default
- Bullet lists for 3+ items
- Match the user's level of formality
- If the user asks for detail, expand — but link, don't lecture

## Hand-off
If the visitor wants to talk to a human, share /contact and stop.
If the question is about a live order or account, direct them to
support@[yourdomain] and do not attempt to look up the account.

The rules that actually matter

Most AGENTS.md files we audit have one of two problems: they're too vague to constrain behaviour, or they're so long no assistant honours them. The rules below have the highest impact per line.

1. Always have a pricing rule

Pricing hallucinations are the #1 way assistants damage trust. A single line — "Never invent pricing; always point to /pricing" — eliminates almost all of them.

2. Map questions to pages explicitly

Don't make the assistant guess which page is canonical for which topic. Spell it out. This is the single highest-leverage section in the file.

3. Define a hand-off

Assistants want to be helpful. Without a hand-off rule, they'll try to answer questions they shouldn't (like account-specific queries). A clear hand-off ends those conversations cleanly.

How to test it

Publishing the file isn't enough — you have to verify assistants are actually honouring it. Run this five-question battery in ChatGPT, Claude and Perplexity:

1. "What does [Your Company] do?" — checks summary clarity
2. "How much does [Your Company] cost?" — checks pricing rule
3. "How does [Your Company] compare to [Competitor]?" — checks competitor rule
4. "I'm interested in working with [Your Company], what do I do?" — checks hand-off
5. "Can you help me with my account at [Your Company]?" — checks escalation

If any answer disappoints, that's a row to add or sharpen in AGENTS.md. Re-test in a week — assistant behaviour updates fast.

FAQ

Do assistants actually read AGENTS.md?

The major ones do — ChatGPT, Claude and Perplexity all support the convention as of 2026. Adoption is not universal yet, but the cost of publishing the file is near-zero and the upside is large.

Should I publish AGENTS.md if I don't have llms.txt?

Publish both. They reinforce each other. Start with llms.txt if you have to pick one; ship AGENTS.md within the same week.

What happens when my policy changes?

Update the file. Assistants re-crawl frequently — most changes take effect within 7–14 days. The process is identical to how you'd update robots.txt or a sitemap.