Markdown for agents: serve your pages to AI

Because a product page carries a lot besides the product. Navigation, scripts, styles and tracking sit around the description, and an agent has to read past them to find it. Cloudflare's converter drops "Non-content elements (such as headers, footers, navigation, scripts, and styles)" and keeps the page's JSON-LD at the end as "a fenced json code block". Shopify's Help Center, which serves its own pages this way, says the result "drops site chrome and navigation, leaving just the page body, which is the preferred representation for LLM ingestion". In our words, the agent gets the shop window without the scaffolding.

Markdown is not a replacement for structured data. Your prices and stock should still be in the Product markup that the structured data guide covers, and Cloudflare carries that markup across. Markdown helps an agent read the prose around it: descriptions, sizes, delivery and returns.

With the Accept header, the ordinary HTTP way of asking for one format over another. Cloudflare says a client "needs to add the Accept negotiation header with text/markdown as one of the options". The server answers with Content-Type: text/markdown, and because the same address now has two answers, the response says so in Vary. Cloudflare's page puts it this way: "Vary includes Accept (any Vary dimensions your origin already declared are preserved) so that caches store separate variants for Markdown and HTML." The text/markdown type comes from RFC 7763, an Informational RFC of March 2016, which says "charset is REQUIRED".

Servers differ in how strictly they read the header. Try a few:

HTTP/2 200
content-type: text/markdown; charset=utf-8
vary: Accept

---
title: Linen duvet cover | Northwind Linen
---

# Linen duvet cover

Stonewashed linen, £120…

Shopify's Help Center is the strict one, and its rules are a good model: "The media type must be named explicitly", */* returns HTML, "Markdown wins an equal-weight tie", and "The only media-type parameter honoured is charset=utf-8". Those are the rules of Shopify's documentation site. A Shopify storefront is another matter: when we tested three Shopify stores on 24 September 2026, each answered Accept: text/markdown with a text/markdown label on the full HTML page. We have not re-run that test since.

In the dashboard. Cloudflare's steps are to select your zone ("you need a Pro or Business plan"), visit the AI Crawl Control section and enable Markdown for Agents. You can also switch it on for one subdomain or path with a configuration rule, or through the API with a PATCH to /client/v4/zones/{zone_tag}/settings/content_converter and {"value": "on"}. Cloudflare says it "is available to Pro, Business and Enterprise plans, and SSL for SaaS customers at no cost", so the Free plan is not on the list.

Cloudflare's page also sets limits and a default:

• The size limit: "Conversion is limited to 6 MiB (6,291,456 bytes) of decompressed HTML", and "The limit applies after decompression, not to the compressed response size."
• Only HTML: "We only convert from HTML, other types of documents may be included in the future."
• No rendering: it converts the HTML your origin sends. For pages built in the browser, Cloudflare points to its Browser Run /markdown endpoint, "if you need to render a dynamic page or application in a real browser before converting it".
• A default policy: if your origin sends no content signal, Cloudflare "adds a default Content-Signal: ai-train=yes, search=yes, ai-input=yes". If you want a different answer for training, send your own; the robots.txt guide covers content signals.

Build it into your site. Vercel describes the pattern for Next.js: "The implementation has two parts, a rewrite rule in next.config.ts that detects the header and a route handler that returns markdown." Any framework can do the same: look at Accept, and when it names text/markdown, return a Markdown version of the page with the right headers.

Shopify's Help Center adds a second route and prefers it: "Append .md to the URL." Its reasons are practical. "The .md URL is distinct from the HTML URL, so responses are cache-safe", and "A .md request for a page that doesn't exist returns 404", where a negotiated request may redirect instead. Our own site does both: this guide answers Accept: text/markdown at its own address and also lives at /guides/markdown-for-ai-agents.md. Link the .md versions from your llms.txt file so agents can find them.

Paste a page's HTML to see roughly what a converter keeps:

---
title: Linen duvet cover | Northwind Linen
description: Stonewashed linen duvet cover in four sizes.
---

# Linen duvet cover

Stonewashed European linen, **£120.00**. In stock.

## Sizes

- Single
- Double
- King
- Super king

## Delivery and returns

Free UK delivery in 3 to 5 working days. Returns within 30 days, see [our refund policy](/policies/refund-policy).

![Oatmeal linen duvet cover on a bed](/img/duvet.jpg)

```json
{"@context":"https://schema.org","@type":"Product","name":"Linen duvet cover","offers":{"@type":"Offer","price":"120.00","priceCurrency":"GBP","availability":"https://schema.org/InStock"}}
```

If the Markdown comes out empty, your page is probably built by JavaScript in the browser, and so is what an agent that does not run scripts sees: nothing much.

What the HTML page says, minus the furniture. This part is our advice rather than anyone's rule. For a product page that means:

• the product's name as the first heading, and its price written with the currency, as in "£120.00", not a bare number;
• the options a shopper chooses between (sizes, colours) and whether each is in stock;
• a line on delivery and returns, with links to the policy pages, ideally to their own .md versions;
• the page's canonical address, so an agent that quotes you links to the real page.

Keep the Markdown to what the HTML shows. A Markdown copy that promises a faster delivery or a lower price than the page invites an agent to repeat a claim your checkout will not honour, and the shipping and returns guide covers keeping those terms the same in every place you state them. Generating both from the same product data keeps them in step.

Ask for it and read the answer. Run curl -si -H 'Accept: text/markdown' https://your-store/ and paste the output here:

Our DISC-009 sends one such request to your home page. It passes a text/markdown answer whose body is Markdown, warns when the label sits on the HTML page or on an empty body, fails when the answer is HTML, and skips when a bot wall or a server error answers instead. It reads the label and the body, and leaves the charset and Vary to you, which is why the fourth example passes our check and still misses two things RFC 7763 and caching ask for.

Some say they do, and nobody publishes a count we could check. Vercel writes that "Many agents already send Accept: text/markdown when fetching pages". The crawler pages of OpenAI and Anthropic that we read describe their bots and robots.txt, and say nothing about the Accept header. So treat Markdown as a cheap courtesy: it costs a setting on Cloudflare or a route handler on your own stack, and an agent that asks gets a cleaner page.

What is Cloudflare Markdown for Agents?

A Cloudflare setting that converts your HTML to Markdown at the edge when a request asks for text/markdown. It "is available to Pro, Business and Enterprise plans, and SSL for SaaS customers at no cost".

What is the size limit for Markdown for Agents?

"Conversion is limited to 6 MiB (6,291,456 bytes) of decompressed HTML", measured after decompression.

Which headers should a Markdown response have?

Content-Type: text/markdown; charset=utf-8, since RFC 7763 says "charset is REQUIRED", and Vary: Accept, so caches keep the HTML and Markdown versions apart.

Does Shopify serve Markdown for product pages?

Shopify's own Help Center does, at .md URLs and through Accept. When we tested three storefronts on 24 September 2026, they put a text/markdown label on the HTML page, which our DISC-009 reports as a warning.

Do I still need structured data if I serve Markdown?

Yes. Markdown carries the prose, and Product markup carries the price, currency and stock in fields an agent can trust. Cloudflare keeps your JSON-LD at the end of its Markdown, so the two travel together. Our agent readiness scanner checks both.

The sources below were last checked on the date shown.