The Dual-Representation Web

Markdown for Agents

HSD Mastermind Power Session

Guerin Green & Holly | February 26, 2026

What Cloudflare Just Did

~20%

of the internet runs through Cloudflare

80%

token reduction: HTML to Markdown

Markdown for Agents — any site on Cloudflare (Pro+) can now serve clean markdown to AI agents automatically. No code changes. Toggle it on.

Claude Code and OpenCode already send Accept: text/markdown headers.

How Content Negotiation Works

AI Agent

Sends request with Accept: text/markdown

Cloudflare Edge

Fetches HTML from origin, converts to markdown on the fly

Origin Server

Sees the Accept header, knows it's an AI agent, can respond differently

Humans get HTML. Agents get markdown. Same URL. Different content.

This is standard HTTP content negotiation — not a separate URL.

The Crisis for SEOs

What the markdown version strips out:

GONE

• CSS grid / div-based comparison tables
• JSON-LD schema markup
• JavaScript-rendered content (React, Vue)
• FAQ accordions that need JS to open
• Navigation, sidebars, footers
• Ads, tracking scripts, popups

SURVIVES

• Semantic HTML headings (H1-H6)
• Body text paragraphs
• HTML <table> elements
• Bullet and numbered lists
• Image references (alt text)
• Hyperlinks

The markdown version of your client's page may look nothing like what you optimized.

Front-Loading is Now Critical

Markdown strips everything to headings + body text in order.

No sidebars. No visual hierarchy. No tabs. Just sequence. Whatever is at the top is what the model processes first and most reliably.

Connects to our January research:

Lost-in-the-middle: 20-60% accuracy loss for mid-document content
Every additional word before your core claim reduces citation probability.
The first 500 words of the markdown = highest-citation-probability zone.

The Cloaking Vector

The Accept: text/markdown header is forwarded to your origin server.

Your server knows an AI agent is asking

Standard content negotiation, not user-agent sniffing

You can serve different HTML

Entity-enriched, citation-optimized content to agents

Cloudflare converts whatever you send

It doesn't validate your HTML matches the human version

Google hasn't addressed this

Mueller criticized separate .md pages, NOT content negotiation

That silence is either a warning or a window. Probably both.

Content Signals: Know What You're Opting Into

Content-Signal: ai-train=yes, search=yes, ai-input=yes

ai-train=yes

You grant permission for AI training on this content

search=yes

Content can appear in search results and AI overviews

ai-input=yes

Agents can use content as input for tasks and responses

These are defaults. Custom policies are coming but aren't available yet. For client sites, this is a documented decision — not a casual toggle.

Free Competitive Intelligence

Current AI bot content type distribution (Cloudflare Radar):

What you can track:

Which crawlers request markdown
How often, from which properties
Per-bot breakdown
Global vs. your zone trends

radar.cloudflare.com/ai-insights
radar.cloudflare.com/bots/directory/gptbot
            

That's the Landscape.

Now We Build.

Members-only from here.

Four paths for non-Cloudflare sites. Implementation steps.

How to audit what agents actually see on your pages right now.

The cloaking play. The agent-side integration.