What a PDF really contains, and why the words come out in order
A PDF page is a program, not a paragraph: draw this character with that font at these coordinates, then that picture, then a line. The reading order a human sees is not stored anywhere - it is inferred from the positions. The PDF text reader used here walks the content stream and reports each piece of text in the order the page draws it, along with a flag on the last piece of each line.
That order was measured before this tool was written, on a six-page, two-column sample: 12,833 characters came out in 96 ms, the two columns arrived one after the other rather than interleaved (one switch between columns in 60 checks), and 85 text pieces carried line-end flags that rebuilt 65 lines. So the tool follows the content stream and uses the line-end flags to cut lines. Clustering text by coordinates - the other common approach - would reorder anything deliberately placed off the baseline, like a caption or a margin stamp, and the measurement shows it is not needed.
The alternative you will see elsewhere, rasterising each page and running OCR, is worse for a document that already has text: it guesses at characters it could have read exactly, it cannot tell a lowercase l from a one, and it turns a searchable file into a picture of one.
How the Word file is built
A .docx is a ZIP package holding XML. This tool writes it directly - three parts: the content-type map, the package relationships, and the document body - with a CRC-32 checksum and the browser's own compression. Nothing is embedded from a template, nothing is fetched, and nothing outside the page touches your text on the way through.
The document body is deliberately plain, because that is what a text layer supports: one paragraph per line, a page break where the PDF had one, US Letter pages with one-inch margins. Text is escaped for XML and characters XML forbids - a NUL or a form feed from a broken font encoding - are dropped rather than passed on, because Word rejects the whole document over one illegal character. Unicode travels through untouched, including accented letters, CJK, currency symbols and emoji.
The package is verified before it ships: it is read back the way a reader does (from the end of the file, through the central directory, checking every offset, size and CRC), and the finished download is opened and inspected in the browser. The same construction was also opened and checked with an independent office toolkit, and the files open in Word without a repair prompt.
What is refused, and what does not come across
A scan is refused rather than exported empty. That was measured too: a picture-only sample produced 56 characters of text for six pages - and those were a header that had been drawn on top, not the page content. Real scans produce zero. So the tool counts the text it found, and if there is none, it says the pages have no text layer, suggests OCR, and stops there instead of handing you a Word file that looks like a failure.
A PDF that needs a password is refused as well, because there is no way to ask for one in a page that never uploads anything; the message points at the unlock tool, which runs in the same tab and can remove the password you know.
Everything else is honest about its limits. Headings, lists, columns, tables and colors are not detected, because the text layer does not distinguish them: a table row arrives as one line of text with the spacing the PDF stored. Pages with no readable text are listed in the result, so a document that is half scans is obvious rather than silently incomplete. What you get is a clean editable draft of the words, which is what most people actually need - and it takes a second rather than a minute.
The page-break mode, and how to read the results
The page-break choice decides how the text lands in Word. One Word page per PDF page keeps the counts in step - page 4 of the Word file is page 4 of the PDF - which matters when numbering has to line up with a printed document. No page breaks lets the text flow as one continuous body, which is what you want when the job is searching, quoting or rewriting and page numbers stopped mattering the moment the file was accepted.
The plan below the form is drawn from the page count you type (up to 2000) before any file exists, then rebuilt from the real document once it lands. Pages that produce no text are reported individually in the results rather than vanishing, and the conversion takes a single PDF of up to 25 MB, read and written entirely in this tab.
Danfio