How to convert a PDF to Word

A PDF page contains text-showing operators: set this font at this size, move to this coordinate, draw this string of glyphs. That is the whole model. There is no paragraph object, no sentence, no reading order, no notion that two lines belong to the same block of prose. The information was discarded when the document was written, because a PDF only needs to know where to put ink.

A converter therefore reverses an inference. It groups glyph runs into lines by their vertical position, groups lines into paragraphs by their spacing and left edges, and guesses at headings from font size. Word then needs all of that, because a .docx describes paragraphs that reflow.

Where the guesses are right the output is genuinely editable. Where they are wrong you get the familiar symptoms: a paragraph broken at every visual line, a two-column page interleaved line by line, a header repeated as body text, a table rendered as rows of tab-separated text. None of these are defects in the reading of the PDF — the reading was accurate — they are the cost of having to invent structure.

Single-column prose in one consistent font converts well: a letter, a memo, a report body, a policy document. The inference has an easy job because spacing is regular and there is one obvious reading order.

Documents converted from Word in the first place are the best case of all, because their layout came from the same paragraph-flow model you are trying to recover and the geometry still reflects it.

What converts badly: anything multi-column, anything heavily designed, forms, invoices and statements laid out as grids, and pages where text is positioned as labels around artwork. For these the honest assessment is that the output needs more repair than it saves, and reproducing the layout in Word from scratch while copying the text across is often faster.

Plain text extraction takes the same glyph runs and writes them out as characters, without pretending to know about paragraphs. It has options for approximating the page layout with whitespace and for inserting page markers, and it makes no claim about structure.

Choose it whenever what you want is the words rather than the document: quoting a passage, feeding content into another system, searching a stack of files, or pulling figures out of a report. It never fabricates a structure that has to be unpicked later, and for those uses the formatting a Word conversion works so hard to produce is something you would have stripped anyway.

The decision comes down to a simple test. If you need to edit the document as a document — track changes, restyle it, send it for review — convert to Word and accept the repair work. If you need the content, extract text and skip the middle step.

If the pages are scans, there are no glyph runs at all. A page holding one large image has nothing that either conversion path can read, and the output is an empty or near-empty document. This is the most common reason a conversion appears to fail silently.

The fix is OCR: recognise the text, produce a searchable PDF with a real text layer, then convert that. Be aware you are now stacking two inferences, since OCR guesses the characters and the converter guesses the structure, so proofread the result against the scan rather than trusting it.

The selectable-text test at the start is worth repeating because it takes two seconds and it tells you which of these two workflows you are actually in.