Firecrawl: our review
Firecrawl turns a site into Markdown pages ready to feed a language model, plus an AI-driven structured-extraction layer. The content cleanup is mechanical; it's mostly this second layer that earns the word "AI."
What we take away from it
What counts in its favor
- Built for RAG: the clean Markdown output skips the cleanup work a raw HTML scraper forces before loading content into a model.
- Credit-based billing, one unit per page, which makes a site's cost predictable before launching the crawl.
- An explicit ban in the terms on using the service for purposes contrary to the GDPR or the CCPA — the clause exists, which isn't systematic among its competitors.
What counts against it
- Servers located in the United States only: the privacy policy itself warns European residents that their data travels outside their country.
- The privacy policy says nothing about using data to train a model: neither a written ban nor a written authorization — a silence to note, not to fill in by assumption.
- The terms don't explicitly require respecting robots.txt or a targeted site's terms: the obligation is inferred from a general law-compliance clause, not from a clause dedicated to scraping.
Useful for preparing clean data for a legitimate RAG use on content you have the right to process. Rule it out if hosting must stay in Europe, or if you're looking for an explicit policy on training models.
The facts
- First paid tier
- $16/month — Hobby, billed annually
- Data hosting
- Data in the United States · source
- Categories
- Data collection,RAG and knowledge bases
- Data verified on
- 30 July 2026
What happens to your data
- Is your data used to train their models?
- Not verified
- Subprocessors
- Four named in the privacy policy: Stripe (payment), PostHog (product analytics), Crisp (customer support) and Vercel Analytics (audience measurement). No exhaustive list published beyond that.
- Retention
- Personal data is kept "until a written deletion request": no default duration is set in the text reviewed.
Firecrawl in detail
Firecrawl starts from a specific problem: a classic scraper returns HTML full of menus, banners, and scripts, unusable as is to feed a language model. Firecrawl converts a page or an entire site into clean Markdown, and adds an AI-driven structured-extraction layer on top.
What it actually does
Three uses coexist. Basic scrape and crawl walk a site and render each page as cleaned-up Markdown: a mechanical conversion operation, not AI work. Structured extraction takes a natural-language instruction — "the price, the release date, the rating" — and a language model takes on finding that information on the page, even when its location changes from one site to another. The most recent feature, the agent, goes further: describe what you want without even giving a URL, and the service searches, crawls, and returns JSON. This third component weighs most in the price: it's billed per token, not per page.
Why the issue is legal first, technical second
Feeding a model with web content raises the same question as any extraction: who has the right to collect what. Firecrawl's terms require use in compliance with applicable laws and explicitly forbid any use contrary to the GDPR or the CCPA — a clause not all its competitors display as clearly. They contain, however, no specific obligation to respect the targeted site's robots.txt: that obligation is inferred, not stated. The robots.txt and the terms of the collected site remain the law in fact, whatever the tool technically allows.
Firecrawl's own point of attention is hosting: servers are located in the United States, and the privacy policy says so plainly, including a warning for European users. For a firm or an administration that wants to keep its data in Europe, this point settles the debate before even looking at price. And on whether Firecrawl itself trains a model on collected content, the policy's silence is notable: neither a written ban nor a written authorization, which should push you to ask for a contractual guarantee before any sensitive use rather than assume an answer.
Where it falls short
The public documentation leaves two gray areas: no clause dedicated to robots.txt, and no commitment on model training. Add exclusively American hosting, and the choice narrows fast for any organization bound by a strictly European data policy.
Our recommendation
Firecrawl suits a team building an internal assistant or a knowledge base in RAG from public content it has the right to process, and that accepts American hosting. It doesn't suit an organization bound by a European hosting requirement, nor anyone looking for a written guarantee on model training — on that point, only a direct agreement with the vendor would settle it, not the public documentation.
Who it's for, and who it isn't
Suited for
- A team building an assistant or an internal knowledge base (RAG) that needs cleaned-up web content rather than raw HTML
- A developer who wants a predictable per-page price rather than a fixed-volume subscription
Not a fit
- An organization bound by a strictly European hosting requirement: Firecrawl hosts in the United States, full stop
- Anyone who needs a written guarantee on the absence of model training: the current policy doesn't settle the question
Frequently asked questions
- Does Firecrawl use AI to extract data, or is it just a scraper?
- Both, for two different uses. The base functions (scrape, crawl) convert a page to Markdown mechanically, with no language model. Structured extraction, and the agent that responds to a natural-language instruction, rely on a model and are billed differently, per token. Don't confuse the two when assessing price or reliability.
- Does the robots.txt of the site I collect apply?
- Yes, whatever Firecrawl's technical ability to ignore it. The terms of use require compliance with applicable laws and contractual commitments, which includes the terms published by the targeted site. The documentation reviewed contains no clause dedicated to respecting robots.txt: that's a blind spot to check yourself before large-scale collection.
- Does my data stay in Europe?
- No. The privacy policy explicitly states servers are located in the United States and warns European users that their information is transferred outside their home country. If European hosting is a requirement for your activity, this point rules out the tool from the start.
- Can I collect email addresses or profiles with Firecrawl?
- Technically yes, but the terms explicitly forbid using the service in violation of the GDPR or the CCPA. A named professional email address remains personal data: collecting it in bulk requires a legal basis on your side, which the tool neither provides nor replaces.
- How much does Firecrawl cost to extract 10,000 pages a month?
- The Standard tier, at $83 a month billed annually, includes 100,000 credits, one page per credit for a simple scrape or crawl. AI-driven structured extraction or a search costs more credits per call: page volume alone isn't enough to estimate the bill if you use extraction.
Alternatives to Firecrawl
Data collection
Apify
Marketplace of ready-made web scrapers, controllable via API.
Data collection
Diffbot
Web data extraction platform that automatically structures pages and sites into usable information for AI applications.
Data collection
Zyte
Web data extraction platform at scale, with a scraping API and a managed service for complex targets.
Our sources
- Pricing and credits pageaccessed on 30 July 2026
- Terms of use (indemnification, GDPR/CCPA)accessed on 30 July 2026
- Privacy policy (hosting, subprocessors, retention)accessed on 30 July 2026
Prices and hosting terms change often. We recheck them periodically, but the publisher's page remains the only up-to-date reference.