PDF to Word Online
Extract text from a PDF into a Word-compatible document. Complex layouts and scanned PDFs may not preserve perfectly.
Drop a PDF here
or choose a file
Strings from the content stream
PDF to Word opens the file with PDF.js and walks getTextContent. The strings are written into an HTML document and downloaded as something Word will open. It is an extraction, not a layout conversion. A three-column policy comes back as column one, then column two, then the sidebar glued to a footnote. That is reading order, and it is useful once you expect it.
Scans produce an empty body. The thumbnails can look perfect. The cursor test is the spec: if you cannot highlight a word, there is no text layer. Locked files fail until they open. There is no upload and no server size cap. A very long file makes a huge HTML download. Split first. HTTPS loads the page. The PDF never becomes a request body.
| In the PDF | In the download |
|---|---|
| Real paragraphs | Editable text, usually in a usable order |
| Columns, captions, sidebars | Linear. You rebuild geography if you still need it |
| Scan | Empty until OCR |
| Line-end hyphens | Often kept as hyphens. Search before a global delete |
Drafts, not filings that must look identical
A policy you have to edit
Extract, delete the running header with an exact search, fix the joins that became nonsense, rebuild any table whose numbers will be calculated. Then, if the prose is the problem, rewrite a paragraph. Rewriting the scrambled extract first produces fluent garbage.
A journal with footnotes
Keep the markers. A “clean” pass that deletes them turns the notes into a second article at the end. Re-extract rather than repair that.
A quote for counsel
Copy the sentence from the PDF if the extract’s line break looks suspicious. Do not quote the Word file as the record.
A scan someone called a Word job
Stop. OCR the upright pages. Proof every digit. Paste. The empty download was a signal, not a broken converter.
Highlight, then extract, then the ugly page
- Drag across a sentence in the PDF. If nothing selects, switch to OCR.
- Drop the text PDF and download the document.
- Open the ugliest page in the thumbnail strip, not only the cover. The cover often extracts beautifully while page 18 welds a sidebar to a table.
- Search for the running header and for hyphen-space. Join poem-like line breaks one section at a time. Do not unwrap a table you have not rebuilt.
Extract PDF text into Word includes the footnote marker mistake.
Proof the page that looks worst
Covers extract cleanly. The failure is a page with a sidebar, a table, and a footer. Open that page in Word first. If a caption has been glued to a cell, stop global find-and-replace. Rebuild the table from the PDF and leave the rest of the extract as paragraphs.
Hyphens at line ends are real characters in the content stream. Deleting every hyphen will also delete the ones that belong in compound words. Search, look, then join. The download is local. Nothing in this step calls /api/ai.
Frequently asked questions
Will tables and fonts match the PDF?
No. getTextContent returns strings in content-stream order. A simple paragraph often lands in a sensible order. Two columns become one sequence. A simple table often arrives as stacked lines, not a Word table. Fonts are not reconstructed as a design.
The Word file is empty and the PDF looks fine.
The pages are images. If you cannot highlight a word in the PDF, this export has nothing to write. Run OCR, proof the numbers, then paste the text. This page will not OCR for you.
Headers appear in the middle of a sentence.
Running headers and footers are text, so they are extracted. Search for the exact running title and remove those lines. Do not delete by paragraph size or you will detach footnote markers.
A number split across a column.
Line breaks can separate 1,200 into “1,” and “200”. Check every figure you might quote against the PDF. The Word file is a draft surface. The PDF is the witness.
Can I extract one chapter?
Split that chapter first. A 90-page extract is miserable to proof, and the error will be on page 70. A 4-page extract you can actually read.
Is the text uploaded?
Not by this page. Extraction stays in the browser. If you later paste the result into Rewrite or Translate, that paste goes to /api/ai. The extract itself does not.