Why conversion is hard — the one thing to understand
Once you accept that, the results stop being mysterious. A clean single-column memo converts almost perfectly, because the inference is easy. A two-column academic paper with a sidebar and a borderless table is genuinely ambiguous — even a human would have to decide where one column ends.
Three kinds of PDF, three different outcomes
Exported from Word, LaTeX or a design tool. You can select the text in a viewer.
PDFnoted: Structure is reconstructed from coordinates: lines into paragraphs, blocks into columns, ruled grids into Word tables. Headings are styled where the font signals them.
A photocopy, a phone photo, a fax. Selecting text selects nothing.
PDFnoted: PDFnoted detects the missing layer, runs OCR to build one, and then converts — so you get editable text instead of a document full of pictures.
A digital document with scanned signature pages, stamped appendices or pasted screenshots.
PDFnoted: Each page is treated on its own terms: text pages reconstructed, image pages OCR'd. The seam between the two is where you should look first when checking the result.
What breaks, and why
| Naive extractor | Typical online converter | PDFnoted | |
|---|---|---|---|
| Multi-column text | Runs together | Sometimes runs together | Detected and separated |
| Tables with ruling lines | Lost | Sometimes | Rebuilt as Word tables |
| Borderless tables | Lost | Lost | Inferred from alignment — verify |
| Scanned pages (no text layer) | Empty document | Empty or image only | OCR first, then text |
| Headers and footers | Mixed into body | Mixed into body | Detected and separated |
| Where the file goes | Local | Uploaded to a server | Your browser |
Five checks on any converted document
Page count matches
The fastest check there is. A dropped page is invisible until someone signs the wrong version.
Headings are headings
If every paragraph came back as body text, apply Word styles before you start editing — the table of contents depends on it.
Tables: rows, columns and merged cells
Compare one complex table against the original. Financial figures that have shifted a column are the classic silent error.
Columns did not merge
In two-column layouts, check that the left column ends where it should and the right one does not begin mid-sentence.
Numbers, dates and names
OCR is good but not perfect on digits. Any figure that will be relied on is worth reading against the source page.
Frequently asked questions
Why does my converted document look nothing like the original?
Because a PDF does not record "this is a paragraph" or "this is a table" — it records each glyph and where to draw it. The converter has to infer the structure from positions and spacing. Inference is good on simple documents and imperfect on dense ones: multi-column layouts, tables without ruling lines, and text wrapped around images are the three classic failure points.
Why did I get a document with no text in it?
The PDF is a scan — a picture of a page with no text layer behind it. There is nothing to extract. Run OCR first: PDFnoted detects this case and builds the text layer before converting, so the output has real editable text instead of an image.
Will my tables survive?
Tables with visible ruling lines convert well — the lines are detected and rebuilt as a Word table. Tables that are only implied by alignment are harder: the converter has to guess column boundaries from whitespace. Check any financial table cell by cell against the original, especially merged cells.
What about headers, footers and page numbers?
They are detected as repeating content across pages and separated out rather than dropped into the body text mid-paragraph. Where a footer is unusual — a different one on the last page, say — it is worth a look, because the detection is statistical.
Can I convert a password-protected PDF?
Yes, with the owner password. Without it there is no conversion — a file you cannot open cannot be read.
Does conversion keep images and formatting?
Images are carried over. Formatting is reconstructed, not copied: bold, italics and font sizes come from the embedded fonts, and Word styles are applied where the structure is clear. Expect to restyle headings rather than having them arrive as perfect Word heading styles.
How long does conversion take?
Text-based documents are near-instant because nothing needs recognising. Scanned pages need OCR first, which runs at roughly a second or two per page depending on resolution and language.
What should I always check in the output?
Five things: that headings are headings and not body text, that tables have the right number of rows and did not split, that multi-column text did not run together, that headers and footers are not sitting in the middle of a paragraph, and that no page is missing. Numbers in financial documents deserve a spot-check against the original.
Try it on your own file
Upload a document and see what comes back. Free to start, nothing to install.
Convert a PDF now — free to startKeep going
Sources and further reading: ISO 32000 (PDF specification); Adobe PDF Reference; Schema.org structured data; Aggarwal et al., GEO: Generative Engine Optimization (arXiv:2311.09735). This page is a technical guide, not legal advice.