Yesterday, I stumbled upon some exciting news from Cloudflare. They’ve introduced a feature called Markdown for Agents, which provides machine-friendly versions of web content alongside the traditional pages we all see.
Cloudflare describes this update as a proactive measure in response to increasing AI crawler activities and agentic browsing.
When a client requests text/markdown, Cloudflare fetches the HTML from the origin server, converts it right at the edge, and then hands over a Markdown version.
Interestingly, the response includes a token estimate header, which helps developers like me manage context windows more effectively.
Early feedback highlighted not only the efficiency gains but also the potential implications of offering alternate representations of web content.
What’s happening. Being part of the 20% of the web that Cloudflare powers, I learned that Markdown for Agents utilizes standard HTTP content negotiation. If a client sends an Accept: text/markdown header, Cloudflare immediately converts the HTML response on-the-fly to Markdown format. The response, marked with Vary: accept, ensures caches store separate versions.
Cloudflare views this opt-in feature as a shift in content discovery and consumption, benefitting AI crawlers and agents with its structured text that requires less overhead.
They claim Markdown can reduce token usage by up to 80% compared to HTML, which is quite impressive!
Security concern. SEO consultant David McSweeney raised a concern, citing that Cloudflare’s Markdown for Agents feature might make AI cloaking incredibly simple because the Accept: text/markdown header tips off origin servers that the request is AI-related.
Regular requests deliver the usual content, but those for Markdown can trigger a unique HTML response that gets converted for AI consumption, McSweeney explained on LinkedIn.
The worry is that sites might inject hidden instructions, altered product data, or other machine-only content, creating a hidden “shadow web” for bots, unless the header is stripped before reaching the origin.
Google and Bing’s markdown smackdown. Here’s the kicker. Representatives from Google and Microsoft advised against creating separate markdown pages for large language models. Google’s John Mueller noted:
“Given that LLMs have always trained on and parsed normal web pages, it seems obvious they have no issues with HTML. Why serve a page that no end user sees? Plus, if they validate equivalence, why not stick to HTML?”
Microsoft’s Fabrice Canel added:
“Do you really want to double crawl load? We’ll check for similarity anyway. Non-user versions (like crawlable AJAX) are often neglected and broken. Human oversight fixes both user and bot views. Schemas help, and AI makes us even better at deciphering web pages. Less is more in SEO!”
Cloudflare’s feature doesn’t generate another URL but does create varied representations based on request headers.
The case against markdown. Technical SEO consultant Jono Alderson pointed out that once a machine-targeted representation exists, platforms must choose to trust it, verify it against the human version, or outright ignore it:
“Flattening a page to markdown doesn’t only remove clutter. It strips away judgment and context.”
“The instant you publish a machine-exclusive page representation, you craft a secondary candidate version of reality. Regardless of source promises or claims of identical content, a system now views two representations and must determine the true reflection of the page.”
Dig deeper. Why LLM-only pages aren’t the answer to AI search
Why we care. With Cloudflare’s advancements, AI ingestion might become more cost-effective and streamlined. But does serving distinct content to humans and crawlers verge on cloaking? Stay tuned…
Inspired by this post on Search Engine Land.
