What Is Firecrawl in 2026?
Firecrawl is a hosted service, and also an open source project, that turns a web page into clean text a language model can read. You give it a URL and it returns markdown, JSON, HTML, a screenshot or page metadata, with the navigation, scripts and layout noise stripped away. In 2026 it is billed in credits rather than per seat, and it exposes its work through a small set of named endpoints.
The endpoints matter more than the marketing copy, because each one costs a different number of credits and answers a different question. The table below reproduces the full published list as it stood in August 2026.
| Endpoint | What it does | Published credit cost |
|---|---|---|
| Scrape | Turns one page into clean data: markdown, JSON, screenshot and more | 1 credit per page |
| Crawl | Follows links from a starting URL across a whole site or section | 1 credit per page |
| Map | Returns a list of links found across a site, for choosing what to scrape | Sources disagree, see below |
| Search | Searches the web and returns full page content from the results | 2 credits per 10 results |
| Extract | Pulls structured data out of pages against a JSON schema you define | Additional credits for advanced features |
| Interact | Scrapes a page then acts on it: click, scroll, type, wait, press | 2 credits per browser minute |
| Monitor | Watches pages and reports when they change | 1 credit per page |
| Agent | Autonomous data gathering, in preview at time of writing | Dynamic pricing, 5 free runs daily |
| Parse | Document parsing for PDFs, DOCX and HTML files | Not published on the pricing page |
Map deserves a note because the sources are not consistent. The documentation describes it as the fastest way to go from one URL to a map of an entire website, assembled primarily from the sitemap and topped up with search results and previously crawled pages, returning each link with a title and description where available. That documentation states each Map request costs 1 credit regardless of how many URLs come back; the homepage credit table lists Map at 1 credit per page. Check your usage dashboard rather than trusting either.
Output formats in 2026 go well beyond the markdown default. The scrape documentation lists markdown, summary, cleaned HTML, raw HTML, screenshot, links, JSON, images, branding, product, audio, video and query, though an audit needs only the first four. Official SDKs cover Python, Node.js, Go, Rust, Java and Elixir, alongside a CLI and a REST API, with MCP integrations for tools including Cursor, Claude Code and Windsurf.
Why Does Extractability Decide AI Visibility in 2026?
Because an AI answer engine cannot cite a fact it cannot parse. When a model answers a question about your category in 2026, it retrieves pages, converts them to text, and quotes from that text. A price rendered inside a product image, a specification list drawn by a JavaScript widget, a doctor's qualification that only exists as a styled badge: all of these are visible to a human and absent from the text a model receives.
This is the gap that separates traditional SEO from AEO and GEO work. Ranking decides whether a page is retrieved. Extractability decides whether, once retrieved, it can supply the sentence the model needs. A page can win position one and still lose the AI answer, because a competitor stated the number in plain text and you did not.
Most teams treat this as theoretical, discussed in the abstract and never tested. Our pet e-commerce marketing guide for 2026 poses it as an AI search test brands ought to run on themselves. Firecrawl turns that test into a five-minute check with an output you can read.
If a fact about your business does not appear as plain text in the markdown output, treat it as a fact your website does not communicate to AI systems in 2026, however prominent it looks on screen.
How Can You Audit Your Own Site's AI Extractability in 2026?
By scraping your own pages and reading what comes back. The workflow is short enough to run in an afternoon and cheap enough to sit inside the free tier. You do not need to write software: the hosted playground accepts a URL and shows you the markdown. The skill involved is editorial judgement about what should be there, not engineering.
- Map the site. Run Map on your domain for the URL list. It tells you what exists, which for most sites in 2026 is not what the team believes exists.
- Choose the pages that carry commercial weight. Ten to thirty for a first pass: homepage, service or category pages, top product pages, pricing, key location pages, About, and any practitioner profiles.
- Scrape each one to markdown. One credit each. Save the output where you can read it beside the live page.
- Read the markdown against a checklist of the facts that page is supposed to communicate. Work down the table below.
- Log every miss as a page defect, not an SEO task. A missing price is a content bug with a specific fix.
- Fix, re-scrape, confirm. This is the step teams skip. One credit tells you whether the fix worked.
- Put priority pages on Monitor so a template change does not quietly remove a fact you fixed.
What should appear in the markdown output
The checklist below is the part worth keeping. Each row is a fact type AI systems commonly need in order to cite a business, paired with what its absence from the markdown means. Not every row applies to every page, but a page failing four or five applicable rows is not competing in AI answers in 2026 regardless of its rankings.
| Look for in the markdown | If it is missing, it usually means |
|---|---|
| What the business does, in plain text, within the first 200 words | The value proposition lives in a hero image or a slider, so the model has no definition to work from |
| Prices or price ranges as text | Pricing is rendered by a script or sits inside an image, and you are absent from every cost comparison |
| Product specifications as a list or table | Specs are in a graphic or a tabbed widget the scraper never opened |
| Credentials, qualifications and years of experience beside the person's name | Trust signals are styled badges rather than text, so the entity has no verifiable attributes |
| FAQ questions and their answers as readable text | The accordion is collapsed in the DOM or built client-side, so your best AEO asset is invisible |
| City, service area and contact details | Local relevance is only in a map embed or the footer image, which no model reads |
| Review counts and ratings as numbers in text | Social proof is inside a third-party widget that loads after the page and never reaches the output |
| Heading hierarchy showing up as markdown levels | Headings are styled divs, so the page has no structure the model can navigate |
| Dates: published, updated, or the date a claim applies to | Nothing tells a model whether your figure is current, and it will prefer a dated source |
| Tables that survive as tables | Comparison data was an image, and comparison queries are where buying decisions get made |
Some page types fail far more than others. Practitioner profiles are the worst offenders, which is why the doctor-profile entity block in our pediatric hospital marketing guide for 2026 is written to be extractable rather than merely attractive. Scrape any three profile pages in your sector and you will usually find qualifications, registration numbers and languages spoken all missing from the text.
Extractability is one half of a technical review. Pair this output with a conventional crawl audit of indexation, canonicals and speed, along the lines of the technical audit table in our guide to marketing for IEPF and share recovery firms in 2026. One tells you whether the page can be found, the other whether it can be quoted.
What Else Can a Marketing Team Use It For in 2026?
Three secondary uses justify the subscription once the audit habit is established. All three follow one pattern: they replace a person manually reading and copying from web pages with a scheduled job that returns clean text. None needs a developer beyond initial setup, and all carry the diligence considerations set out further down this page.
Competitor content and messaging monitoring
Crawl a competitor's blog, pricing page or product pages on a schedule and diff the markdown between runs. Monitor is built for exactly this. What you get in 2026 is a factual log of when positioning changed, when a price moved, when a feature was quietly dropped, and which topics a rival is publishing into, rather than an impression formed from occasional browsing.
A knowledge base your AI tools can actually use
Crawl your own site plus a set of trusted external sources into markdown, and you have the corpus for a retrieval system that answers questions from your own material instead of guessing. This is the practical route to grounded internal AI in 2026, and it pairs with the model cost planning in our Gemini 3.7 Flash guide for marketing teams. Clean input is most of the work in any retrieval build.
Structured market research
Extract takes a JSON schema and returns pages as filled fields rather than prose. Point it at a set of category pages with a schema of product name, price, pack size and claim, and you get a spreadsheet instead of a fortnight of copy and paste. It is the right tool for pricing surveys, assortment tracking and competitive feature grids in 2026.
How Does Firecrawl Pricing Work in 2026?
You buy a monthly allowance of credits and each operation consumes some. A page scraped is one credit, a page crawled is one credit, ten search results are two credits, and browser interaction is two credits per minute. Plans also cap concurrency, which is how many requests can run at once, and that limit affects how fast a large crawl finishes rather than what it costs.
| Plan | Price per month, billed yearly | Credits per month | Concurrent requests |
|---|---|---|---|
| Free | 0 USD | 1,000 | 2 |
| Hobby | 16 USD | 5,000 | 5 |
| Standard | 83 USD | 100,000 | 50 |
| Growth | 333 USD | 500,000 | 100 |
| Scale | 599 USD | 1,000,000 | 150 |
| Enterprise | Custom | Custom, unlimited pages | Custom |
Three published conditions matter to a budget. Scale lists extra credits at 397 US dollars per additional 350,000. Firecrawl states it does not currently offer a pay-per-use plan. And credits do not roll over on self-serve plans, so an allowance sized for an occasional audit is wasted in months you do not run one. Enterprise adds dedicated support, an SLA, bulk discounts, zero-data retention and SSO. The Free tier includes five daily runs of the Agent preview.
An extractability audit of a 400 page site: 1 credit to map it, 400 to crawl every page, 40 targeted re-scrapes after fixes, and 25 priority pages monitored weekly for a month at 100 credits. That comes to roughly 541 credits, inside the 1,000 credit Free tier. Paid tiers exist for continuous competitor crawling and research volume, not for auditing your own site. Your figures will differ with page count and re-check frequency.
Pricing figures on this page were checked in August 2026 against the published pricing page. Software pricing moves, and credit definitions move with it, so confirm the current rates and the current per-endpoint credit costs before you commit a budget.
What Does the AGPL Licence Mean for Agencies in 2026?
It means the phrase "it's open source" needs a second look before an agency builds a client-facing service on a self-hosted copy. The repository states that the project is primarily licensed under the GNU Affero General Public License v3.0, and that the SDKs and some UI components are licensed under the MIT License. The repository carries roughly 167.6 thousand stars and publishes a self-hosting guide.
The AGPL is not the same as the permissive MIT or Apache licences most marketers have half-heard about. Its Section 13 is headed "Remote Network Interaction; Use with the GNU General Public License" and opens with this text:
Notwithstanding any other provision of this License, if you modify the Program, your modified version must prominently offer all users interacting with it remotely through a computer network (if your version supports such interaction) an opportunity to receive the Corresponding Source of your version by providing access to the Corresponding Source from a network server at no charge, through some standard or customary means of facilitating copying of software.
Read plainly, that clause turns on modification combined with remote network interaction, which is the shape of an agency that forks the tool, wires it into a client dashboard and charges for access. Whether any specific deployment triggers it is a legal question about your architecture and your modifications. Nothing here is a legal conclusion. Take advice before you build.
Two facts reduce the practical exposure for most readers in 2026. The hosted plans do not raise the question, because you are a customer of a service rather than a distributor of software. And the repository notes that the cloud version at firecrawl.dev includes additional features beyond the open source version, so self-hosting is not a free copy of the hosted product in any case.
What Are the Lawful Use Considerations in 2026?
Readers will point this at competitor websites, so the diligence question deserves a straight treatment. There is no single answer, because what matters is the combination of what you collect, whose data it is, where the subjects are located and what the source site's terms say. The table below separates the uses that are generally low-risk from those that warrant advice before you start.
| Generally lower risk | Get legal review first |
|---|---|
| Your own websites and properties | Anything containing personal data about identifiable people |
| Publicly available pages containing no personal data | Content behind a login, a paywall or an account agreement |
| Respecting robots.txt and any stated crawl directives | Platforms whose terms of service restrict automated access |
| Conservative rate limits that do not burden the source server | Republishing or reselling substantial copied content |
| Reading facts to inform your own original work | Bulk collection intended to reconstruct a competitor's database |
Two regimes are worth naming for Indian and international teams in 2026. Collecting personal data about people in India engages the Digital Personal Data Protection Act 2023, and collecting personal data about people in the European Union engages the GDPR, both of which apply to scraped personal data as much as to data a person typed into your form. Contact names, email addresses and reviewer profiles pulled off competitor sites are personal data.
The practical rule is narrow and easy to hold. Collect facts about businesses, products and prices; do not collect facts about people. Read the current terms of any site you intend to crawl at scale, because terms change. Where a project sits near personal data, put it in front of counsel before the first request. None of this is legal advice.
What Are the Common Mistakes in 2026?
The failures cluster in three places: treating the tool as developer-only when the work is editorial, running the audit once and assuming the result holds, and skipping the licence and lawful use questions until a client asks. Each is cheap to avoid in 2026 and expensive to discover late. The list below covers the eight we see most often.
- Assuming it needs an engineer. The audit needs a URL and someone who knows what the page is supposed to say. The judgement is editorial.
- Auditing once and never again. A theme update or a new widget can remove an extractable fact overnight. That is what Monitor is for.
- Only reading the pretty output. If your fix works in a screenshot but not in the markdown, it did not work.
- Crawling the whole site at full depth. Map first, choose thirty pages, then crawl.
- Reading the credit table once. Interact bills per browser minute and Extract can cost more than a basic scrape, so an unattended job spends an allowance quickly.
- Buying a tier for a job the Free tier covers. Credits do not roll over on self-serve plans in 2026.
- Saying "it's open source" without reading which licence. The core is AGPL-3.0 and the SDKs are MIT. Those are materially different positions.
- Pointing it at logged-in or personal data without advice. The DPDP Act 2023 and the GDPR do not care that the collection was automated.
Key Takeaways for 2026
The short version for anyone deciding whether this belongs in a 2026 marketing stack. The tool is genuinely useful for building AI applications, but its value to a marketing team is diagnostic: it is the cheapest honest answer to the question of what AI systems can actually read on your own website.
- Firecrawl converts pages into clean markdown, JSON, HTML, screenshots and metadata, through Scrape, Crawl, Map, Search, Extract, Interact, Monitor, Agent and Parse endpoints, billed in credits.
- The highest-value marketing use in 2026 is auditing your own site: scrape your commercially important pages and check whether prices, specifications, credentials, locations, review counts and FAQ answers survive into the text.
- A fact absent from the markdown cannot be cited by an AI answer engine, regardless of how the page ranks or how it looks in a browser.
- An audit of a mid-size site fits inside the Free tier of 1,000 credits per month. Paid tiers are for continuous monitoring and research volume.
- Published pricing checked in August 2026 runs from Free at 0 dollars to Scale at 599 dollars per month billed yearly, with no pay-per-use option and no credit rollover on self-serve plans.
- The core project is AGPL-3.0 with MIT-licensed SDKs. Agencies that modify and self-host it behind a client-facing service should take legal advice on Section 13 first.
- Pointing it at other people's sites is a diligence question. Business facts sit at the low-risk end, personal data engages the DPDP Act 2023 and the GDPR.
The transferable habit has nothing to do with any one tool. In 2026, the only reliable way to know what AI systems can say about your business is to read your website the way they read it: plain text, design removed. That takes an afternoon and a free account, and most teams have never once done it.