Best free tool to turn web pages into clean text for AI: Trafilatura vs Readability

Who this is for: you’re building an agent, a scraper or a RAG pipeline and need the article text from web pages, without paying for the menus, cookie banners, “related stories” and footers in tokens, or letting them confuse the model.

Our pick: Trafilatura (100/100). On 181 real news and blog pages it kept 98% of the article text and added little junk, at 11 milliseconds a page and under 100 MB of memory. It gives you Markdown for your LLM, title, author and date, its own downloader and a command line.

Runner-up: Mozilla Readability (80/100). The code behind Firefox Reader View was just as accurate. In Node.js, though, it needs jsdom to parse the page, which made it 5 times slower and used 1.3 GB, and it doesn’t do Markdown, downloading or a CLI. Pick it if you’re in JavaScript or a browser and don’t want Python.

Results on 181 pages

Scroll sideways for more columns →

F1, precision and recall come from the benchmark’s own scoring code, which compares overlapping word sequences with the hand-labelled text, so order matters.

Gotchas

Trafilatura refuses private addresses. Its downloader blocks non-public IP addresses (to stop server-side request forgery). That’s a good default, but it broke the README example on our test machine, whose network proxy uses a private address. If you run behind such a proxy, download with your own HTTP client and pass the HTML to extract().

Readability needs Node 22 in Node.js. The current jsdom won’t load on Node 20, but npm still installs it there, so the error only shows up when your code runs.

Readability: close your jsdom windows. The README example doesn’t, and memory climbed to 1.3 GB over 181 pages. In a long-running process, call window.close() after each page.

Neither runs JavaScript. They read the HTML you give them. Pages that build their content with JavaScript need a headless browser first.

Scores

Scored with our published rubric. Tested 2026-10-16. One run each: the gap was 20 points.

Scroll sideways for more columns →

The 7 checks: text on at least 95% of pages, F1 of at least 0.90, precision and recall of at least 0.90 each, at least 90% of pages at F1 0.80, a median of at most 50 ms a page, and at most 1 GB of memory.

What it costs

Both are free and open source (Apache-2.0) with no usage limits. Our runs cost $0.

Caveats

• Not our data, and not new numbers. The pages and labels are Zyte’s public benchmark, which already lists scores for both tools. Zyte sells an extraction API, and Trafilatura’s author contributed its runner there. We ran both tools ourselves with today’s versions: Readability matched its published F1 exactly, and Trafilatura was within the error bar of its published score. What we add is speed, memory, safety and features.
• News and blog articles only. Product pages, docs sites, forums and search results weren’t tested.
• Defaults only. Trafilatura has precision and recall settings that we left alone.
• Runs were done by an AI agent following each tool’s README, with the network off. Details and the check script are in the test script and here.

Next step: automate what happens next. See our tested pick for automating a workflow on your own server, or pull data out of documents with our PDF tables pick.

Not tested yet

• Hosted reader APIs (Jina Reader, Firecrawl, Zyte API): need accounts, and some need a card.
• Browser-based scrapers for JavaScript-heavy pages (Crawl4AI, Playwright plus an extractor).

This page has no affiliate links yet. Every link goes to the tool’s own homepage. Affiliate deals never decide a pick, and a tool with no affiliate program can still win. How we make money.