How do I scrape a website to LLM-ready text?

Paste the URL and click Scrape & Count Tokens — the page is fetched server-side and cleaned to Markdown (nav, ads, and boilerplate removed), ready to copy or download and paste into ChatGPT or Claude.

How do I count tokens for an entire documentation site, not just one page?

Set Max pages to 5 or 10 before scraping — the tool follows same-site links found on the first page (one level deep) and sums tokens across every page it reaches. The Per URL tab shows the breakdown, and a warning appears if any linked page couldn't be scraped.

What is the difference between raw HTML tokens and clean Markdown tokens?

Raw HTML includes navigation, ads, scripts, and markup that costs tokens without carrying any of the page's actual content. The Optimization tab fetches the same URL's raw HTML directly (as a best-effort, approximate comparison — not always available) and shows the real token count next to the clean Markdown version, typically 25-40% lower.

Is scraping a URL here private?

The URL is fetched by an edge function on our server — not stored, logged as content, or shared — via the Jina AI Reader service, which returns clean Markdown for the page. This is different from Toolsda's purely client-side tools: a URL scrape necessarily involves a server round trip, since a browser tab can't fetch an arbitrary third-party page itself. Token counting on the returned text still happens entirely in your browser.

Why did scraping fail or return almost nothing?

A few common causes: the site returned a 403/blocks scraping (try saving the page as a PDF and use File to Tokens), the page is a JavaScript-rendered single-page app that Jina's reader can't execute (try Screenshot to LLM instead), or the request simply timed out on a slow site — retry it.

URL to Tokens

Q: How many tokens is a website?

Paste the URL above — it is scraped to clean Markdown and counted with the same tiktoken encoder GPT-4o and GPT-5.6 use (a close estimate for Claude, Gemini, and Llama, since those providers don't publish a public JavaScript tokenizer). As a rule of thumb: a 1,000-word blog post runs about 1,500 tokens, a documentation page about 4,000, and a long Wikipedia article 8,000+.

Q: Will this URL fit in ChatGPT 8K, 32K, or 128K?

Check the context-window fit bars above — they compare your token count against each model's real window with 20% held back for the answer (usable = window × 0.8). GPT-4o's window is 128K; current frontier models (GPT-5.6, Claude Opus/Sonnet 5, Gemini 3.1 Pro) run closer to 1M tokens, so most single pages fit comfortably. Multi-page crawls of a whole documentation site are the case most likely to exceed even a large window.

Any public URL

Will it fit in ChatGPT?

Paste any URL. It's scraped server-side to clean Markdown (nav, ads, and boilerplate stripped), then counted for GPT-4o, GPT-5.6, Claude Opus/Sonnet 5, Gemini 3, and Llama — with a fit check against each model's context window, 20% reserved for the answer.

URL

Strip links & images from count

Max pages

Scraped via Jina AI Reader through our edge function — clean Markdown, no ads, nothing stored.

Model

Result

Paste a URL on the left to see its token count and check whether it fits GPT-4o, Claude, or Gemini's context window.

What is Website / URL to Tokens?

URL to Tokens scrapes the main content from a webpage — stripping navigation, ads, and footer boilerplate — and counts how many tokens that content will use in an LLM's context window. As a rule of thumb, 1 token is roughly 4 characters of English text: a 1,000-word blog post runs about 1,500 tokens, a typical documentation page around 4,000, and a long Wikipedia article 8,000 or more.

How many tokens is a website?

Website type	Avg tokens	Fits in 8K?	Fits in 128K?
Blog post, ~1,000 words	~1,500 tokens	Yes	Yes
Docs page (React-docs style)	~4,000 tokens	No — needs 16K+	Yes
Long Wikipedia article	~8,500 tokens	No	Yes
Full docs site, 10 pages crawled	~40,000 tokens	No	Yes, but chunk it for RAG

Averages for plain prose — dense docs, code samples, or tables run higher.

Will this URL fit in ChatGPT / Claude's context window?

Every model reserves part of its context window for its own answer — filling the window to 100% with input leaves no room to respond. The standard practice: usable tokens = context window × 0.8, keeping 20% free. Current context windows: GPT-4o 128K (≈100K usable), Claude Opus/Sonnet 5 and Gemini 3.1 Pro ~1M (≈800K usable). Legacy 8K/16K/32K tiers still show up in search but are mostly retired from current frontier models.

Your page	8K	16K	32K	128K	200K	1M
1,500 tokens	Yes	Yes	Yes	Yes	Yes	Yes
4,000 tokens	Yes	Yes	Yes	Yes	Yes	Yes
8,500 tokens	No	Yes	Yes	Yes	Yes	Yes
40,000 tokens	No	No	No	Yes	Yes	Yes

8K — Small/legacy tier — some local and older API models · 16K — Legacy GPT-3.5 16K tier · 32K — Mid-size legacy tier · 128K — GPT-4o · 200K — Legacy Claude 200K tier · 1M — Current frontier models — GPT-5.6, Claude Opus/Sonnet 5, Gemini 3.1 Pro. Use the live fit checker above for this page's exact number against today's real models.

Raw HTML vs clean Markdown tokens — why clean saves tokens

Raw HTML carries navigation menus, ads, tracking scripts, and layout markup — none of which answers a question, all of which costs tokens. Scraping through Jina AI's Reader strips that noise and returns clean Markdown instead. Open the Optimization tab above (single-page scrapes only) to see this page's real before/after token count and saving percentage.

Best practices to make a website fit in the context window

Use clean mode — scraping via Jina AI Reader already removes nav, ads, and footer noise by default.
Crawl only the pages you need — set Max pages to 1 for a single article, not an entire site.
Chunk for RAG — split a large scrape into overlapping windows (try 1,000 tokens per chunk, keeping each section's heading with it; adjust the size: ) via RAG Text Splitter.
Reserve 20% for the answer — never fill a context window to 100% with input alone.
Use a larger-context model when a documentation site is simply large — GPT-5.6, Claude Opus/Sonnet 5, and Gemini 3.1 Pro all run ~1M tokens.
Strip links and images from the count when you only need the prose — the checkbox above does this instantly.

How website scraping to tokens works

Pasting a URL sends it to a stateless edge function (not stored, not logged as content), which fetches the page through Jina AI's Reader API and returns clean Markdown. With Max pages set above 1, same-site links found on that first page are crawled one level deep, up to the limit chosen. Every returned page is then tokenized in your browser with the same tiktoken encoder Token Counter uses, so a scraped page and pasted text always agree on the same model's count.

Related tools

File to Tokens — have a PDF or DOCX instead? Check its token count and context fit the same way.
Token Counter — count tokens for pasted text across the same nine models.
Docs to LLM — convert DOCX, PPTX, XLSX, MD, or HTML to clean Markdown and chunk it for RAG.
Smart Token Reducer — strip whitespace, filler, and formatting waste from any text.
Screenshot to LLM — a JavaScript-heavy page didn't scrape cleanly? Screenshot it instead.