What happens to a scanned PDF?

It is refused, deliberately. The conversion counts the visible characters in the text layer first, and when that total is zero the run stops with a message saying every page is an image, which is what a scan or a phone photograph produces, and that there is nothing to copy into a Word file. The refusal carries a machine-readable reason, which is what lets the page offer you a button to open the same file in the character-recognition tool instead. A file that is not a PDF at all is refused separately, as unreadable.

How does it decide what a heading is?

By size, against the whole document. The median character size across every page is taken as the body size, and any line set at least 1.18 times that is treated as a heading: it is written bold and at its own measured size. A document whose body text is already large and whose headings are only slightly larger will therefore have none detected, and a document with a lot of large display text will have several. Headings are written as direct formatting rather than as named Word styles, because the file carries no styles part, so they will not appear in a navigation pane.

Why is my two-column page mixed up?

Because lines are grouped by their baseline. Two characters at the same height on a two-column page are treated as belonging to the same line whichever column they sit in, so a two-column layout reads across the page rather than down each column. There is no column detection in this tool. If the source is in columns, expect to reorder the paragraphs after the conversion, or to split the pages first.

Do I have to give an email address to get the file?

No. The .docx is written by this page and saved straight to your disk by your own browser. There is no account, no email step, no queue and no watermark, and nothing about the file is metered — the conversion happens on your computer, so there is no per-run cost to recover.

PDFTo Word

PDF to Word

Pull the text out of a PDF into an editable .docx file. The PDF is read by this page and never sent to a server.

On-device prototype

What this does, and what it does not

It recovers the text: reading order, paragraphs, page breaks, and headings where the PDF sets them in larger type. It does not rebuild the page layout — columns, tables as real tables, images, and fonts are not carried across. If your PDF is a scan or a photo of paper it holds no text at all, and this page will tell you so rather than hand you an empty document.

Choose a PDF from this device to convert.

No file-count limit, no daily limit, no queue — the work happens on this machine. Choose one file for the usual single-file flow, or select many for batch results and one ZIP.

About this PDF to Word converter

Choose a PDF of up to 150 MB and convert it. The text layer is read in the page, regrouped into lines and paragraphs by the coordinates of the characters, and written to a .docx named after your PDF. This recovers the words, not the page: it is a text extraction, and the page says so above the button.

A PDF stores glyphs at coordinates, not paragraphs, so reading order, line grouping and paragraph boundaries all have to be reconstructed from geometry — and that reconstruction is what you get. Reading order, paragraphs, page breaks and headings set in larger type do come across. Layout, columns, tables as tables, images and fonts do not, and calling the result a conversion rather than an extraction would be overstating it. A PDF with no text at all — a scan, or a photograph of paper — is refused by name rather than handed back as an empty document, with a button that passes the same file to the character-recognition tool on this site. A file over 150 MB is refused, and in a multi-file run it is skipped with a message naming the limit.

Runs in this tab — no uploadFree, no account, no watermark

Step by step

1Choose the PDFUp to 150 MB. The file is read into the page; it is not sent anywhere and the copy on your disk is not modified.
2The text layer is counted firstBefore any conversion runs, the visible characters in the text layer are counted. A total of zero means every page is an image — a scan or a phone photograph — and the run stops there rather than producing an empty Word file.
3Lines and paragraphs are rebuilt from coordinatesCharacters are grouped into lines by their baselines and lines into paragraphs, a page break is written between every page, and any line set at least 1.18 times the document’s median character size is treated as a heading and written bold at its measured size.
4Download the .docx and read the countsThe file takes your PDF’s name. The result panel reports the page count, the paragraph count, the character count and how many pages carried no text, so a part-scanned document cannot be silently half-converted.

Why nothing can convert a PDF to Word perfectly

It is worth being plain about this, because every tool that claims otherwise is selling something. A Word file describes a document as structure: paragraphs, styles, tables, sections, a flow that reflows when you change the margins. A PDF describes a document as marks on a fixed page: this glyph at this coordinate in this font at this size. Going from Word to PDF throws the structure away deliberately, because fixing the appearance is the whole purpose of the format. Going back means guessing the structure from the marks.

Some of those guesses are safe. Two characters sharing a baseline are on the same line; a gap larger than the leading is a paragraph break; a run of text well above the document’s median size is a heading. Some are not safe at all, which is why this tool does not attempt them: reconstructing a table from the positions of its numbers, or deciding which of two columns a line belongs to, is where a converter starts producing plausible, wrong documents.

What is kept, and what is lost

Kept: the words, the order they are read in, paragraph breaks, a real page break between every page, bold where the font name says bold, and headings detected by size. Lost: columns, tables as real tables, images, page furniture, colours, margins and the original fonts.

A two-column page is the case to watch. Lines are grouped by baseline, so two characters at the same height on a two-column page are read as one line whichever column they sit in — the text comes out reading across the page rather than down each column. There is no column detection here. If your source is in columns, expect to reorder paragraphs afterwards, or split the pages first.

A table comes through as its text in reading order, which is usually not the shape you want. The PDF to Excel tool on this site is the one that reconstructs a table as a grid, and it shows you the column boundaries it found so you can correct them.

What the Word file itself contains

A minimal three-part .docx: paragraphs, runs carrying bold and a size, and page breaks. Headings are written as direct formatting rather than as named Word styles, because the file carries no styles part, so they will not show up in Word’s navigation pane.

The page size written into it is fixed at A4 — 11906 by 16838 twips — whatever size the PDF pages were, so a US Letter original will reflow when you open it. Control characters that XML forbids are dropped during the write, because leaving one in makes the whole document refuse to open, which is a worse loss than one glyph.

Why the document is not uploaded

Every response from this site is served with a Content Security Policy whose connect-src directive is set to none. That is not a promise in marketing copy; it is an instruction to your browser, and the browser is the one enforcing it. While this page is open it cannot open a connection to anywhere — not to this site, not to anyone else — so there is no path by which your document could be uploaded, not by a bug, not by a future change, not by a script that should not be here. e2e/egress-proof.spec.ts asserts the served header and then tries to send data out by every route a page has and requires each attempt to be refused, and it runs on every build.

The documents people convert to Word are drafts, contracts, reports and letters — material that is usually confidential for exactly as long as it is useful. A conventional online converter has your file on its disk for as long as its retention policy says, and you have no way to check the policy is kept.

The second half of the guarantee is that the code has nowhere to send anything from. lib/tools/local-source-policy.test.ts reads every source file under lib/tools, workers, components and app on every test run and fails the build if any of them contains fetch, XMLHttpRequest, WebSocket, EventSource, sendBeacon or a peer connection, or even a remote address written down in a comment. A tool here cannot acquire an upload by accident, because the file that would have to contain one cannot be committed.

Questions people ask

What happens to a scanned PDF?: It is refused, deliberately. The conversion counts the visible characters in the text layer first, and when that total is zero the run stops with a message saying every page is an image, which is what a scan or a phone photograph produces, and that there is nothing to copy into a Word file. The refusal carries a machine-readable reason, which is what lets the page offer you a button to open the same file in the character-recognition tool instead. A file that is not a PDF at all is refused separately, as unreadable.
What is kept, and what is lost?: Kept: the words, the order they are read in, paragraph breaks, a real page break between every page, bold where the font name says bold, and headings. Lost: columns, tables as real tables, images, page furniture, colours, margins and the original fonts. A table will come through as its text in reading order, which is usually not the shape you want — the PDF to Excel tool on this site is the one that reconstructs a table as a grid.
How does it decide what a heading is?: By size, against the whole document. The median character size across every page is taken as the body size, and any line set at least 1.18 times that is treated as a heading: it is written bold and at its own measured size. A document whose body text is already large and whose headings are only slightly larger will therefore have none detected, and a document with a lot of large display text will have several. Headings are written as direct formatting rather than as named Word styles, because the file carries no styles part, so they will not appear in a navigation pane.
Why is my two-column page mixed up?: Because lines are grouped by their baseline. Two characters at the same height on a two-column page are treated as belonging to the same line whichever column they sit in, so a two-column layout reads across the page rather than down each column. There is no column detection in this tool. If the source is in columns, expect to reorder the paragraphs after the conversion, or to split the pages first.
What does the Word file itself look like?: A minimal three-part .docx: paragraphs, runs carrying bold and a size, and page breaks. The page size written into it is fixed at A4, 11906 by 16838 twips, whatever size the PDF pages were, so a Letter-size original will reflow when you open it. Control characters that XML forbids are dropped during the write, because leaving one in makes the whole document refuse to open, which is a worse loss than one glyph. The result panel reports the page count, the paragraph count, the character count and how many pages carried no text, so a part-scanned document is not silently half-converted.
Do I have to give an email address to get the file?: No. The .docx is written by this page and saved straight to your disk by your own browser. There is no account, no email step, no queue and no watermark, and nothing about the file is metered — the conversion happens on your computer, so there is no per-run cost to recover.