The Word file was empty and the PDF was not
Fourteen pages, a scanned contract, a perfect thumbnail of every clause. PDF to Word handed me a document with headings and no body. Nothing was “broken” in the romantic sense. getTextContent returned no strings because the pages were images. The tool writes those strings into an HTML document and downloads that as a Word-friendly file. No strings, empty file. I should have looked at the PDF with the cursor first. If I cannot highlight a word, this export will not invent one.
The second attempt was a born-digital policy, three columns, footnotes, a sidebar of definitions. The Word file had all the words and none of the geography. Column one ran to the bottom, then column two, then the sidebar glued itself to a footnote. That is reading order from the PDF content stream, not a layout reconstruction. I stopped calling it a conversion and started calling it an extraction. The disappointment went away. The cleanup did not, but it was the cleanup I expected.
What the extractor will and will not carry
| In the PDF | In the download |
|---|---|
| Real text, simple paragraphs | Editable paragraphs, usually in a sensible order |
| Two or three columns | Linear text. You rebuild the columns if you still want them |
| A simple table | Often tab-separated lines or stacked cells, not a Word table |
| A scanned page | Nothing, until AI OCR has produced text you paste in |
| Headers and footers | Repeated on every page, in the flow, where they annoy you |
| A text box in the margin | Wherever the content stream emitted it, which may be mid-sentence |
I delete running headers in one pass with find-and-replace before I touch tone. “Page 3 of 14” forty times will poison any later rewrite. Same for “CONFIDENTIAL” banners that were drawn as text. They are text. The extractor is correct to include them and you are correct to remove them.
A cleanup order that wastes less of the afternoon
First, confirm the PDF highlights. Second, extract. Third, skim the start of each old page — I drop a temporary marker in Word when I see a heading jump — and fix only the joins that became nonsense. Fourth, rebuild tables that hold numbers someone will calculate. Fifth, only then send a paragraph to AI Rewriter if the prose is the problem. Rewriting a scrambled column produces confident garbage that is harder to untangle than the raw extract.
Numbers move. I have seen a line break split “1,200” across a column boundary so the extract reads “1,” at the end of one paragraph and “200” at the start of the next. I check every figure I might quote against the PDF, not against the Word file. The Word file is a draft surface. The PDF is the witness.
Fonts and styling do not survive as design. You get text. Bold sometimes comes through as structure, sometimes as nothing. Do not spend an hour matching the corporate typeface on an extract you are about to rewrite anyway. Match it at the end, on the paragraphs you kept.
Scans, in the only order that works
Rotate upright if the scan is sideways. OCR. Read the OCR for names and for the word “not,” which vision models drop more often than they drop adjectives. Paste the corrected text into Word, or extract again only if you have a PDF that now contains a real text layer. This Word tool will not OCR for you. Chaining them in the other order — Word first, panic at the empty file, OCR second — is what the empty download is for. It is a signal, not a failure of the scanner.
Locked PDFs do not yield text until you open them with a password you are allowed to use. Extract from the unlocked copy. Do not paste the unlocked text into a shared doc if the lock was there for a reason. Extraction removes the inconvenience of the PDF. It does not change who is allowed to have the words.
Very long files make a huge HTML download and a Word document that is miserable to scroll. I split the chapter I need, extract that, and leave the rest. A 4-page extract I can proof. A 90-page extract I will skim, and the error will be on page 70.
Hyphenation at line endings is the quiet mess. The PDF stored “docu-” and “ment” as the line broke. Word now contains “docu- ment” or “document” depending on how the producer encoded it. I search for hyphen-plus-space and I fix real compound hyphens by hand. A global replace will turn “state-of-the-art” into “state of the art” and also fix the line-break hyphens. I accept the collateral damage only when I am going to edit every paragraph anyway.
If the goal is a filing that must look like the PDF, stop. You do not want this tool. You want the PDF. If the goal is a paragraph you can edit, quote, or translate, the linear text is a feature. I treat layout loss as the price of getting a cursor in the sentence. Paying that price on a scan, without OCR, buys an empty file and a false belief that the contract had no text. It had text. It had it as pixels. Different door.
Footnotes that arrived as a second article
A journal PDF put footnotes after each page’s body in the content stream, which is sane, and also repeated the short title in the header as text, which is also sane for a PDF and poisonous in Word. After extraction I had the article, then a pile of footnotes in one block because I had been too aggressive with a “remove headers” pass and I had deleted the markers. I re-extracted and did the header removal with a search for the exact running title, not with a paragraph-size heuristic. The markers [12] stayed. I could jump again. The first extract went in the bin. It looked cleaner and it had detached every citation from its sentence.
Tables of contents come through as a list with dot leaders. I delete the leaders and I do not ask a rewriter to “clean the TOC.” A rewriter will turn the page numbers into prose. I have seen “see also page” appear where a dotted line used to be. The TOC is navigation. If I need navigation in Word, I rebuild it from the headings I kept, after the headings are real headings and not just bold lines. That is a Word job at the end, not an extraction feature.
I spot-check one middle page, not only page 1. Page 1 of a report is often a cover with a simple text frame, and it extracts beautifully, which is how you declare victory while page 18’s sidebar is welded to a table. Pick the ugliest page in the thumbnail strip. If that page’s extract is merely linear and complete, the rest is probably tolerable. If that page’s extract is missing a column, the whole export is a maybe, and I extract that section alone so I can watch it.
Line endings from the PDF sometimes arrive as hard returns on every visual line. Word then looks like a poem. I join those lines when the paragraph is obvious, and I do not join them inside a table I have not rebuilt. A global “unwrap” has merged two table cells into a sentence that was never written. I unwrap one section, look, then the next. The poem is annoying. The merged cells are a false document.