If you paste webpages into ChatGPT or Claude regularly, you are probably wasting a third of your context window on nav menus, cookie banners, and "related articles" widgets.
I measured it. A typical article page is ~24,000 characters as rendered HTML-to-text; the actual article is ~8,000. Everything else is website furniture — and the model has to spend attention (and you have to spend tokens) figuring out which is which.
The usual fixes, and where they break
Copy-paste. Fast, but you drag in everything: sidebars, footers, comment CTAs. Tables turn to mush, code blocks lose their fences.
r.jina.ai URL prefix. Genuinely clever — prefix any public URL and get Markdown back, no install. But it's a server-side fetch, so it breaks exactly where you need it most: Reddit threads (blocked), X posts (login wall), paywalled articles, anything behind your session.
Readability-based scripts. Work fine on clean blog posts, fall over on structured sites (docs with tabs, forums, video pages).
What actually solved it for me: convert in the browser
The insight is boring but important: your browser already has the fully rendered, logged-in page. Server-side readers have to fight for access; an extension just reads what you can see.
My daily workflow now:
- Open any page — Reddit thread, X post, YouTube video, docs.
- One click → clean Markdown (headings, lists, tables, code fences preserved).
- Send straight to ChatGPT/Claude, or save as
.mdfor Obsidian.
On token count: clean Markdown of the same article runs ~65% fewer tokens than raw copy-paste, because none of the furniture comes along. Over a month of heavy research use, that's real money on API workflows.
Site-specific extractors matter more than I expected. Generic readability guesses; a Reddit extractor knows what a comment thread is (score, nesting, OP), a YouTube extractor grabs the transcript. Web2MD ships 16 of these.
When each tool wins
| Scenario | Best tool |
|---|---|
| Public article, one-off, no install |
r.jina.ai prefix |
| Automation/pipelines on public URLs | Jina Reader API / Firecrawl |
| Reddit, X, paywalls, logged-in pages | Browser extension |
| Batch-convert 50 URLs for an agent | Extension + MCP |
I build Web2MD (free tier, no signup for the web tools), so the browser-extension enthusiasm is earned bias — but the comparison table holds regardless of whose extension you use. If your AI workflow chokes on exactly the pages that matter, move the conversion into the browser.
Happy to answer questions about the extraction internals — the Reddit one was the most annoying to get right.
Top comments (0)