How to convert a web page to Markdown for ChatGPT

The problem with pasting a link

You find an article you want summarised. You paste the link into ChatGPT. Sometimes it works. Often you get an apology instead — the model couldn't open the page.

There are three separate reasons this happens, and they matter:

• The site refuses automated visitors. Reddit, X, and a great many news publishers sit behind bot protection like Cloudflare or DataDome. These systems are designed to tell a browser apart from a script, and they are good at it.
• The page has no content until JavaScript runs. Single-page apps send an almost-empty HTML shell and build the article in your browser afterwards. A fetcher that only reads the initial response sees nothing.
• The content is behind a login or a paywall. The server returns a subscription prompt, which is exactly what it is supposed to do.

Why copy-paste isn't the answer either

The obvious workaround is to select the page and paste it in. It works, but it drags along everything else: the navigation, the cookie banner, the newsletter prompt, the related-articles rail, the footer. Headings arrive as ordinary sentences. Tables collapse into unreadable runs of text. Code loses its indentation.

All of that costs tokens, and tokens are finite. Worse, it costs clarity: a model reading a flattened table has to guess at which number belonged to which column.

What Markdown actually fixes

Markdown keeps the structure and throws away the presentation. A heading stays a heading. A table stays a table. Code stays fenced, with its language attached. Everything that was decoration — the styling, the scripts, the layout scaffolding — disappears.

The saving is substantial. Cloudflare published a measurement of one of its own announcement pages: 16,180 tokens as raw HTML, 3,150 tokens as Markdown. That is roughly an 80% reduction for identical content. Documentation and blog pages typically land somewhere in the 60–80% range.

The practical effect is that a page which wouldn't fit now fits, and a model that was skimming can now read the whole thing.

The five-step version

1. Copy the address of the page you want the AI to read.
2. Paste it into the converter and press Convert.
3. Check the token count against your model's context window.
4. Press Copy Markdown.
5. Paste it into your AI chat, above your actual question.

One detail worth getting right: put the Markdown above your question, not below it. Models attend more reliably to an instruction that comes after the material it applies to.

Using the toggles

Three switches sit under the input box, and the defaults are right most of the time.

• Title adds the page title as a top-level heading. Leave it on — it gives the model immediate context about what it is reading.
• Links keeps inline links as real Markdown links. Turn it off when feeding an AI: link URLs cost tokens and rarely add anything the model can use. Turn it on when you are saving the page as a note you will come back to.
• Clean removes images, figures and embeds. Useful for image-heavy articles where you only want the argument.

When a page won't convert

Some pages will fail, and the tool will tell you which kind of failure it was rather than showing a generic error. The useful responses differ:

• Blocked. The site refused the request outright. For Reddit, try the old.reddit.com version of the URL — the older interface serves plain HTML and usually works. Otherwise, open the page yourself and copy the text: you are logged in, and a server never will be.
• Paywalled. Nothing to be done from here, by design. If you subscribe, copy from your own browser tab.
• No readable content. The page is built in the browser. Look for a permalink, a print view, or an AMP version, all of which tend to be server-rendered.

A general trick: many sites publish a plain-text or Markdown version at a predictable address. Appending .md to a docs URL works surprisingly often, and it is faster than converting.

Getting better answers once it's pasted

Converted text is only half of it. A few things reliably improve what comes back:

• Say what the text isbefore you ask about it — “this is a Reddit thread about X” frames everything that follows.
• Ask for something specific. “Summarise this” produces a summary. “What does this disagree with in the mainstream view?” produces something worth reading.
• Convert two sources on the same subject and paste both. Models are considerably better at comparison than at recall.