Skip to main content
Language models read text, not markup. Most of a page’s HTML is scripts, styles and layout that a model pays for in tokens and never uses. Pass markdown=true to /getPageSource and you get the rendered page as clean Markdown instead: headings, paragraphs, lists, tables, links and images with alt text - without scripts, styles, navigation or footer. The page is fetched exactly like an HTML call: JavaScript-rendered, behind Cloudflare, DataDome, PerimeterX or Akamai, from the country you choose. Only the output changes, and it costs the same as an HTML call.
A Wikipedia article that is 230 KB of HTML comes back as about 50 KB of Markdown. A GitHub repository page shrinks from 317 KB to 10 KB.

Request

The SDKs have a dedicated method: get_markdown() in Python and Ruby, getMarkdown() in Node.js and PHP.

Response

The body is Markdown, Content-Type: text/markdown; charset=utf-8. For https://books.toscrape.com/catalogue/a-light-in-the-attic_1000/index.html:
What you get:
  • Text structure: headings, paragraphs, bulleted and numbered lists (nested lists kept), block quotes and code blocks with their language.
  • Tables as Markdown tables. Key/value lists (<dl>, common on spec sheets) become - **Weight**: 300 g lines.
  • Absolute links: every link and image URL is resolved against the page, so an agent can follow it. mailto: and tel: links are kept.
  • Images with alt text as ![alt](url). Images without alt text are decoration or tracking pixels and are left out.
What is removed: scripts, styles, <svg>, iframes and embeds, buttons and form controls, dialogs, and hidden elements (hidden, aria-hidden="true").

Main content only

By default (main_content_only=true) you get the page’s main content:
  • If the page has a <main> element holding at least 30% of the page’s text, that element is used, without the menus inside it.
  • Otherwise the whole page is used, without navigation, the site header, the footer and sidebars (<nav>, <header>, <footer>, <aside> and their ARIA roles). An article’s own <header> is kept.
Cookie consent banners (Cookiebot, OneTrust, Didomi, Usercentrics, Quantcast and other consent managers, and any dialog labelled as a cookie or consent dialog) are removed too, so their cookie lists never come before the content. If that leaves nothing, the whole page is returned rather than an empty answer. Set main_content_only=false to keep the whole page, for example when you need the footer’s contact details or the site’s navigation links.

Combining with other parameters

  • proxy_country works as usual: the page is fetched from that country, then converted.
  • steps run first: click “load more”, type into a search box or wait for a selector, then the resulting page is converted. See browser steps.
  • parsed_data, get_cookies and list_elements each return their own format, so they cannot be combined with markdown=true. Such a request answers 422 and is not billed.

Errors and billing

A Markdown call is billed like an HTML call: only when the page is delivered. Errors are the same as for HTML (see page source) and keep their plain-text body. When the target itself answers 404, 410 or a 5xx error page, that page comes back as Markdown with the target’s status in the X-Origin-Status header, as described in handling failures.

Markdown or parsed data?

Using an AI assistant? The MCP server has a fetch_markdown tool, and search_and_read searches Google and returns the top results as Markdown in one call.