The conversion that looks worse and is better

A PDF can be reproduced in Word almost exactly, and the result is impossible to edit. The useful conversion throws the layout away and infers paragraphs instead.

Two conversions, and only one of them is useful

There are two ways to turn a PDF into a Word file, and the difference decides whether you get something you can work with.

The first reproduces the page. Every line of text becomes an absolutely positioned text box at exactly the coordinates it occupied in the PDF. Open it and it looks perfect — indistinguishable from the original. Then try to change a sentence. Nothing reflows, because nothing was ever a paragraph; the text boxes stay where they are and the words spill out of them. Delete a line and a hole appears. It is a picture of a document with a Word extension.

The second throws the layout away. It works out where the paragraphs were, which lines were headings, which were list items, and rebuilds the document from that — then lets Word lay it out again, in Word’s own way. The result looks less like the PDF. It is the only version anybody can edit.

If you wanted the appearance preserved exactly, you already have that: it is the PDF. The reason to convert is to change the words, and that argues for the second conversion every time.

What has to be inferred

A PDF holds glyphs at coordinates. There are no paragraphs in it, no headings, no lists, and no reading order. Every piece of structure in the Word file is a guess made from position and size — usually a good one, and worth understanding so you know where to look when it is wrong.

Headings, from size relative to the body

Never from absolute points. A document set in 9pt and one set in 14pt have completely different bodies and exactly the same structure, so a threshold like “anything over 13pt is a heading” works on one and fails on the other. What matters is the ratio to whatever most of the words are set in.

Short and bold counts as a heading too. Bold on its own deliberately does not: plenty of documents bold an entire paragraph for emphasis, and promoting that produces a document outline full of paragraphs — which then poisons the navigation pane, the table of contents and anything else built from the heading structure.

Paragraphs, from three weak signals

A new paragraph starts where the line spacing grows, where the previous line stopped well short of the right margin, or where the kind of line changes. Each of those is unreliable alone. A paragraph’s last line is short, but so is a line before a page break. Spacing grows before a heading as well as between paragraphs. Together they are usually right.

Words broken across lines

Justified text hyphenates, and a PDF preserves the hyphen with no record of whether it was inserted by the typesetter or typed by the author.

“informa-” followed by “tion” is one word that must be rejoined. “well-” followed by “Known” is a real hyphen that must survive. The rule that works: lower-case on both sides is a break; anything capitalised, numeric or punctuated keeps its hyphen. That gets “COVID-positive” and “PDF-based” right, and when it does get one wrong the mistake is visible in the text rather than hidden in the formatting.

Columns, and the failure that reads perfectly

This is the one worth understanding properly, because it does not look like a failure.

A PDF stores text in the order it was drawn, which has nothing to do with the order it should be read. On a two-column page the two columns usually share their vertical positions — the first line of the left column sits at the same height as the first line of the right. Read the file naively and you get alternate clauses from two different arguments, spliced into sentences that are grammatical, fluent, and meaningless.

That is far worse than obvious rubbish. Obvious rubbish gets noticed. A page of confident, well-formed sentences that quietly say nothing can survive a proof-read.

The fix is to find the vertical band the text never crosses and split the page there — and to do it before the lines are assembled. Grouping text into lines first merges the two columns into single lines, and at that point the columns are gone; no amount of reordering afterwards can recover them.

Bold and italic, where the font is willing to say

Emphasis is read from the font’s name, because that is all a PDF reliably offers. A document using Helvetica-Bold announces itself. A document that embeds its fonts as g_d0_f1 says nothing at all, and emphasis on those pages is simply lost. A tool that stays quiet about that is letting you find out later; the honest thing is to say which pages were affected.

What no browser tool can do

Scans. A scanned page holds an image of text, not text. There are no glyph positions because there are no glyphs, and no amount of cleverness in this kind of tool will help. The test takes a second: open the PDF and try to select a line. If you cannot, you need optical character recognition — a different job with a much larger error rate and a real need for proof-reading afterwards.

Tables. They come through as text. A table in a PDF is not a table either — it is glyphs in tidy positions — and reconstructing one properly is a different problem from reconstructing prose. A dedicated table extractor does far better and gives you a spreadsheet.

Images. Not carried across.

Why this should happen on your own machine

The documents people convert are contracts, reports, statements, letters and drafts. Uploading one to a website means handing it to somebody whose retention policy you have not read, in exchange for arithmetic your browser can do on its own.

Browsers have had a capable PDF engine for years, and a Word file is a zip of XML that JavaScript can assemble without difficulty. There is no technical reason for the document to leave your device, so it should not.

Frequently asked questions

Why does the Word file not look exactly like the PDF?

Because it is built to be edited. A PDF can be reproduced almost exactly by wrapping every line in an absolutely positioned text box, and the result is unusable — change one sentence and nothing reflows, because nothing was ever a paragraph. Inferring real paragraphs produces a file that looks less like the original and is the only version anybody can work with. If you wanted the appearance preserved exactly, the PDF already does that.

Can I get a Google Doc?

Download the .docx and open it from Google Drive. Google Docs imports Word files directly, so a separate format would gain nothing.

How are headings identified?

By size relative to the body text, never by absolute points — a document set in 9pt and one set in 14pt have different bodies and the same structure. Short and bold counts too. Bold alone deliberately does not: plenty of documents bold a whole paragraph for emphasis, and promoting that to a heading wrecks the outline and the navigation pane built from it.

Why do two-column PDFs come out scrambled?

Because a PDF stores text in drawing order and the two columns usually share their vertical positions. Read naively you get alternate clauses from two different arguments spliced into sentences that parse perfectly and mean nothing. The columns have to be separated before lines are formed — clustering first merges them and nothing afterwards can undo it.

How is a word broken across two lines handled?

Rejoined when it should be. “informa-” plus “tion” is one word; “well-” plus “Known” is two. Lower-case on both sides is treated as a break, and anything capitalised, numeric or punctuated keeps its hyphen.

Why is bold and italic sometimes missing?

Emphasis is read from the font's name, which is all a PDF reliably offers. A document that embeds its fonts as g_d0_f1 gives nothing away. A good tool reports that rather than leaving you to discover it.

Will it convert a scan?

No, and no tool of this kind can. A scan holds an image of text, so there are no glyph positions to work from. Try selecting a line in your PDF viewer: if you cannot, you need optical character recognition, which is a different job with a much larger error rate.

What happens to images and tables?

Images are not carried across, and complex tables arrive as text rather than as Word tables. For a table specifically, a dedicated table extractor does a far better job and gives you a spreadsheet.