Bulk Article Extractor: Clean Text & Markdown for LLMs

Turn article URLs into clean Markdown with author, date and language, split into RAG chunks. $0.70 per 1,000, failed pages are free.

Give it a list of article or page URLs. For each one you get:

• the main content as plain text, Markdown, or both, without menus, ads, sidebars, cookie banners and footers
• metadata: title, author, publication date, site name, description, main image, the language the page declares
• the language detected from the text (ISO 639-1 code)
• word count and an approximate token count
• optionally, the content split into overlapping chunks ready for a vector database (RAG)

Only the URLs you give are fetched. Links on the pages are not followed, so you always know what you pay for.

Input

{
    "urls": [
        "https://en.wikipedia.org/wiki/Web_scraping",
        "https://www.bbc.com/mundo/articles/cwvgr021p725o",
        "https://docs.python.org/3/tutorial/controlflow.html"
    ],
    "outputFormat": "both",
    "includeLinks": false,
    "includeImages": false,
    "chunking": true,
    "chunkSize": 500,
    "chunkOverlap": 50,
    "detectLanguage": true,
    "respectRobots": true,
    "timeoutSecs": 20
}

• urls (required): pages to extract. A URL without scheme gets https://. Duplicates are fetched once, invalid lines are skipped with a warning.
• outputFormat (default both): text, markdown or both. Markdown keeps headings, lists, tables, code blocks, bold and italics.
• includeLinks, includeImages (default off): keep links and images of the article in the Markdown. The plain text never has markup.
• chunking (default off): add a chunks array. Chunks are made from the Markdown when outputFormat is markdown, otherwise from the plain text.
• chunkSize (default 500) and chunkOverlap (default 50): in approximate tokens, counted as 4 characters per token. Chunks break at line ends where possible, then at sentence ends, then at spaces. Each chunk starts with the last chunkOverlap tokens of the previous one.
• detectLanguage (default on): fill detectedLanguage.
• respectRobots (default on): see below.
• timeoutSecs (default 20): limit for each request.

Output

One dataset item per URL. A successful one, shortened:

{
    "input": "https://www.bbc.com/mundo/articles/cwvgr021p725o",
    "url": "https://www.bbc.com/mundo/articles/cwvgr021p725o",
    "finalUrl": "https://www.bbc.com/mundo/articles/cwvgr021p725o",
    "status": "ok",
    "httpStatus": 200,
    "fetchedAt": "2026-10-30T04:26:26+00:00",
    "error": null,
    "title": "\"No hay palabras para describir tanto dolor\": Urda ...",
    "author": "Lorena Motos",
    "date": "2026-10-29",
    "sitename": "BBC News Mundo",
    "description": "...",
    "language": "es",
    "detectedLanguage": "es",
    "image": "https://ichef.bbci.co.uk/...",
    "extractor": "trafilatura",
    "text": "...",
    "markdown": "...",
    "wordCount": 1712,
    "approxTokens": 2491,
    "paywallFlag": false,
    "chunks": [{ "index": 0, "text": "...", "approxTokens": 498 }]
}

status tells you what happened:

status	meaning	charged
ok	content extracted	yes
blocked	the site answered 401, 402, 403, 405, 407, 429 or 451, or showed a bot check page (Cloudflare and similar)	no
paywalled	the page marks itself as paid content (schema.org isAccessibleForFree: false) and fewer than 150 words were readable. The teaser text is still returned.	no
robots_disallowed	robots.txt does not allow this URL, the page was not requested	no
no_content	fewer than 25 words of main content, for example a login page or a menu	no
unsupported	not an HTML page (PDF, image, JSON)	no
error	DNS failure, timeout, TLS error, 404, 5xx. The reason is in error.	no

A failing URL never stops the run.

How pages are requested

• One GET per URL, redirects followed, 5 MB limit per page. Plus one request for /robots.txt per host when respectRobots is on.
• At most 10 URLs at a time overall and 2 at a time per host, so a list from one site does not hammer it.
• No proxies, no browser, no JavaScript. Pages that only render their text with JavaScript come back as no_content.
• User-Agent: Mozilla/5.0 (Windows NT 10.0; Win64; x64) AppleWebKit/537.36 (KHTML, like Gecko) Chrome/130.0 Safari/537.36 (compatible; BulkArticleExtractor/1.0; +https://apify.com/isnard/article-extractor). It looks like a browser to avoid naive filters, and says who it is.

robots.txt

With respectRobots on, the Actor reads /robots.txt of each host once and follows RFC 9309 for the user-agent BulkArticleExtractor: the group that names it if there is one, otherwise the * group, longest matching rule wins. A 4xx answer means no restrictions. A 5xx or 429 answer means "do not crawl", so all URLs of that host are reported as robots_disallowed. If robots.txt does not answer at all, the page request goes ahead and reports the real problem.

How extraction works

1. trafilatura finds the main content and the metadata.
2. If trafilatura finds no text, readability-lxml picks the main block and trafilatura converts it. extractor says which one was used.
3. py3langid detects the language of the first 5,000 characters.

Known limits: the author field is sometimes wrong or empty (we drop values longer than 8 words, which are usually a block of text, not a name). Dates on documentation pages are often the last update, not the publication. Token counts are approximate: real counts depend on the model and are usually higher for Spanish than for English.

Licenses

• trafilatura: Apache-2.0 (since version 1.8.0, before it was GPL-3.0). The Actor requires 2.x.
• readability-lxml: Apache-2.0
• py3langid: BSD-3-Clause
• lxml: BSD-3-Clause

You are responsible for how you use the extracted content. Respect copyright and the terms of the sites you extract from.

Pricing

Pay per event: article-extracted, once per URL with status ok. Everything else is free.

FAQ

What is charged? Only URLs with status ok. Blocked, paywalled, robots-disallowed, empty and failed pages are free.

Does it run JavaScript? No. Pages that only show their text after JavaScript runs come back as no_content.

Does it follow links? No. Only the URLs you give are fetched, so you always know what you pay for.

How accurate is the token count? It is an estimate of 4 characters per token. Real counts depend on the model and are usually higher for Spanish.

Related tools

• RSS & Atom Feed Change Detector: find new articles to extract.
• PDF to Text, Markdown & Tables (CSV): extract text and tables from PDFs.
• Website Screenshot API (Full Page, Mobile, No Cookie Banners): keep a screenshot of each page.