Zapier watched Claude Code read its site. The agent pulled the Markdown version, and a second model summarized it first

Agents are reading your clients' websites right now. Zapier tested what actually reaches them.

What happened

Zapier engineers built a tool called Agent Arm Runner. It runs Claude Code in a Docker container behind a proxy, so the same prompt and the same URL can be pointed at either the live page or a preview version. Each run starts fresh, and each version is repeated because agent output varies.

Their first test emptied almost all content from the Slack integration page. The agent still answered with detailed Slack triggers and actions. The proxy logs showed why: WebFetch asked for Markdown before HTML, and Zapier's server returned its Markdown mirror of the page, which they had not changed. Zapier says Markdown mirrors are already in the live request path for agents.

The second finding: the fetched page went to a separate model call, Haiku 4.5 with no tools, under its own instructions, including a 125 character limit on quotes. Claude Code's documentation confirms that WebFetch converts HTML to Markdown and, for most fetches, runs the prompt against the content in a separate model call, so Claude gets that result rather than the raw page. The docs call it lossy by design. Zapier notes it has not assumed other agents behave the same way.

My take

For a client's service pages, this changes what "publishing" means. There may be two versions of every page now, HTML for people and Markdown for agents, and if they drift apart, the agent answers from the stale one.

There is also a summarizer between your page and the model answering the question. Long, wordy pages lose detail in that step. Clear headings, short facts near the top and plain lists give the summary something solid to keep.

Practical steps:

1. If a site serves Markdown or llms style files, put them in the same publishing pipeline as the HTML so both update together.
2. Put prices, service areas and contact steps near the top in plain text.
3. Test the page by asking an agent about it, not just by viewing it in a browser.

I am adding a "Markdown and HTML in sync" check to the content automations I build.