Why PDF to Word is never perfect

What a PDF actually stores

A PDF page is a sequence of drawing operations. Set a font, move to a position, show a string of glyphs, move again. There is no object representing "a paragraph" and no object representing "a table" — a table is some horizontal and vertical lines drawn on the page, with text positioned in the gaps.

Reading order is not stored either. The order glyphs appear in the file is the order the generator emitted them, which for a two-column layout may be down the left column then down the right, or may alternate between them, or may be scattered by an optimiser. Nothing in the file says which is meant to be read first.

Even words are an inference. Many PDFs contain no space characters at all — the gap between words is a positioning instruction, and the converter must decide from the gap width whether two glyph runs are one word or two. Get the threshold wrong and you get "th e docum ent".

What the converter has to guess

Paragraphs from vertical gaps and line-start positions. Headings from font size and weight relative to surrounding text. Lists from repeated leading characters and indentation. Tables from drawn rules and from text alignment where no rules exist. Columns from horizontal whitespace running down the page. Reading order from all of the above combined.

Each of these is a heuristic, and each fails in a predictable way. Borderless tables become tab-separated text or a jumble. Merged cells break grid detection. A table split across two pages becomes two unrelated tables. A pull quote in the middle of a column gets spliced into the sentence it interrupts. A footer repeated on every page becomes a paragraph in the body.

Why some PDFs convert beautifully

The best case is a PDF exported from a word processor. The layout is simple and regular, the fonts are embedded with proper character mappings, and the generator emitted content in reading order. The converter's guesses are all easy and the result can be close to the original document.

Better still is a tagged PDF. Tagging adds a structure tree describing headings, paragraphs, lists, table rows and reading order explicitly — the very information the format otherwise lacks. It exists for accessibility, so screen readers can navigate a document, and it is a requirement of PDF/UA and of accessibility rules in many jurisdictions. A properly tagged PDF converts well because the converter no longer has to guess.

You can check: in most readers, document properties will indicate whether the file is tagged. Untagged, complex, multi-column, heavily designed layouts are where conversion goes wrong, and design-tool output — magazine layouts, brochures, annual reports — is the hardest case, because visual position carries meaning that was never written down.

The scanned document case

If you cannot select text in the PDF, there is no text in it. Every page is an image, and a converter will faithfully produce a Word document containing pictures of pages — technically correct and completely useless.

The fix is OCR: optical character recognition reads the pixels and produces text, adding a searchable layer. Convert after that and you get real editable text.

The caveat is that OCR introduces its own errors, and they are different from conversion errors. Character confusions in poor scans, mangled table structure, and lost formatting are all normal. OCR output should be proofread rather than trusted — particularly for numbers, where a misread digit will not look wrong the way a misread word does.

Getting the best result

Check whether the source document still exists before converting anything. A conversion is always worse than the original, and the person who sent the PDF often still has the DOCX.

Ask what you actually need. If you need the text, converting to plain text or Markdown is far more reliable than DOCX, because it discards the layout the converter would otherwise guess at badly. If you need a specific table, extracting to Excel targets that problem directly. If you need to change three words on a page, editing the PDF avoids the round trip entirely — a full conversion to make a small edit is usually the wrong move.

And check the output against the original, especially tables and numbers. The failure mode of conversion is not a file that obviously breaks; it is a file that looks right and has a column silently shifted.