Checked against the current Windows build. Where an article names a menu, a button or a message, that is the wording in the application, not a paraphrase.
There is no layout in the file to preserve
A PDF does not store paragraphs. It stores glyphs at coordinates, and the exporter rebuilds paragraphs from geometry: characters are grouped into lines by how far their boxes overlap vertically, lines are grouped into paragraphs by the gap between them, and each paragraph carries one size and one weight. The output page is A4.
Every consequence people report follows from that one sentence. A table has no rows in the file, only text at positions, so it arrives as lines. Two columns are two sets of coordinates with nothing recording which to read first, so they interleave. An image is not text, and a text-and-paragraphs model has nowhere to put one. A heading is only a heading because it is larger; a heading set at body size has nothing left to distinguish it.
Which documents come across well enough to use
Single-column prose converts acceptably: a report, a letter, a judgement, a set of minutes, a policy document. You will re-apply the headings and repair a few line breaks, and the words and the paragraph structure will be there waiting.
Documents where the layout is the content do not: a brochure, a form, a certificate, a statement with ruled columns, anything with a sidebar or a caption sitting beside a picture. For those, retyping the part you need is faster than repairing the conversion, and it is a much shorter job than it looks from the outside. Work out which of the two you are holding before you convert. That decision costs a minute in advance and an afternoon afterwards.
If nothing came out at all
The result names any page it could not read: page 4 has no extractable text, and if it is a scan, run OCR first. Confirm it by dragging the cursor across a word in the viewer. If nothing highlights, the page is a photograph of words and there is nothing for an exporter to recover from it.
Local OCR puts a text layer under the image, and that layer can then be exported. It is a machine's reading of a picture, so check the figures and the names against the scan before you reuse any of it — a recognised amount can be wrong by one digit and look entirely calm about it.
An encrypted document is refused unless you supply the password. An owner-restricted file opens without one and the export is still refused, because taking the text out against a restriction the document carries is a decision for the person holding it rather than for the tool.
When to stop, and when not to start
For a wrong date or a misspelt name, do not go to Word at all. Editing the words where they sit keeps the design and finishes in one step, and the text-editing guide covers the two places a page will fight back. Going out to Word and rebuilding the layout costs you the design and buys nothing.
For a folder, Batch has a job that exports the text of every PDF in it, and the command is creasepoint todocx file.pdf --out out.docx, with --pages 1,3 when only part of the document matters. Running the conversion a hundred times does not change what the conversion keeps, so read a sample of the outputs rather than the count of files that finished.
The two places worth looking next.
Read the matching guide if you want the reasoning rather than the fix, and check limitations before spending any longer on it — some of what people look for here is absent on purpose.