PDF to Word test results: what our converter gets right and wrong

Last reviewed: 2026-10-25

Every converter says it “keeps your formatting”. We would rather show you. These are the results of running thirteen test PDFs through the converter on this site, including the files it does badly on.

How we tested

We made thirteen one- and two-page A4 PDFs ourselves, so we know exactly what text each one contains: a plain report, an invoice with tables, a two-column article, a Greek application form, scanned copies of these at different qualities, a “searchable” scan of the kind phone scanner apps make (page picture with invisible text underneath), a file mixing a normal page with a scanned page, and two password-protected files. Each was converted on our own server with the same software the site uses, one file at a time.

For each result we counted the words in the Word file and compared them, in order, with the words of the original. Word match below is the share of words that came out the same and in the same order, so a value under 100% can mean a wrong word or a word that moved. Time is the conversion step on the server only; upload and download depend on your connection.

Test file	Engine used	Words kept	Word match	Time
Report: headings, paragraphs, bullet list	pdf2docx	304 of 304	100%	1.1 s
Invoice: ruled table, borderless address block, borderless schedule	pdf2docx	142 of 142	90.8%	1.7 s
Two-column article	pdf2docx	307 of 307	98.7%	0.8 s
Greek form with a table	pdf2docx	116 of 116	100%	0.9 s
Scanned report, 200 dpi	OCR, then text rebuild	304 of 304	100%	5.2 s
Scanned report, 100 dpi	OCR, then text rebuild	304 of 304	99.7%	3.6 s
Scanned report, tilted 1.5° with specks	OCR, then text rebuild	304 of 304	100%	4.9 s
Scanned Greek form	OCR, then text rebuild	117 (original 116)	93.6%	4.2 s
Scanned invoice	OCR, then text rebuild	143 (original 142)	91.9%	3.6 s
Searchable scan (picture + invisible text, as phone apps save)	Existing text reused	304 of 304	100%	0.2 s
Normal page + scanned page in one file	pdf2docx	Page 1 only	see below	2.1 s
Needs a password to open	refused	–	–	instant
Opens freely, copying/printing restricted	pdf2docx	304 of 304	100%	1.0 s

What came out well

Ordinary text PDFs (anything exported from Word, Google Docs or a “print to PDF”) came through word for word. Headings kept their size and weight, and the bullet list stayed a bullet list.
Greek text in a normal PDF was perfect, including accents and diaeresis (προϋπόθεση, Μαΐου), and its bordered table stayed a real Word table.
Ruled tables, meaning tables with visible lines between cells, became real Word tables you can edit cell by cell. The invoice’s line-item table kept all 9 rows and 6 columns.
Clean scans were read almost perfectly by OCR, even at 100 dpi (one wrong word) and when the page was tilted and speckled.

What needed fixing by hand

Two columns. Reading order was right (left column, then right), but every paragraph kept a manual line break at the end of each original line, and a large right indent kept the text in the left half of the page. Both are quick to undo: our formatting guide shows how.
Tables without lines. The invoice’s borderless delivery schedule became a table without borders, but its bold header row was left outside the table as a single line, so the headings no longer sat above their columns. The borderless address block also became a small table. See keeping tables.
Column widths. In the line-item table every column came out the same width, so the description column wrapped onto several lines. Drag the column borders in Word to fix it.

Where it falls short

Scans lose their layout. A scanned page comes back as editable paragraphs of text. Tables become rows of text separated by tabs, sometimes with a stray “|” where a cell border was, and pictures on the page are not included.
Scanned Greek is good, not perfect. The large bold title came out wrong, and letters with diaeresis were sometimes read with the wrong accent. Body text was mostly right. More in the Greek guide.
Mixed files. When a normal page came first and a scanned page second, the scanned page arrived as a picture with no editable text. The converter only runs OCR when the first pages of a file have almost no text (fewer than 20 characters across the first five pages). If your file mixes the two, convert the scanned pages as a separate PDF.
Password to open. Files that need a password to open are refused with a message saying so. See protected PDFs.

Convert a PDF now → Free, no sign-up, files deleted as soon as you download the result

What changed because of this test

Running these files showed three real problems, and all three are fixed. Scanned pages used to come back as dozens of floating text boxes laid over the original page picture, so the page looked printed twice. They now come back as normal paragraphs, in about half the time. And scanned Greek used to be read as if it were English: only about 3% of words were right. It is now about 94%. And “searchable” scans from phone apps, which already carry invisible text, were treated as normal PDFs and ended up with 307 floating text boxes over the page picture; their existing text is now turned into normal paragraphs, in 0.2 seconds instead of 7. All three are listed on the updates page.

What this test does not tell you

Thirteen files we made ourselves are not a guarantee for yours. Real PDFs have unusual fonts, broken text encodings, photographed pages with shadows and complicated tables. Treat the numbers as a fair picture of what to expect, and always read through an important document after converting it.

Frequently asked questions

How accurate is the OCR on scanned PDFs?

On our clean English test scans, 99.7% to 100% of words came out right, even at 100 dpi and tilted 1.5 degrees. Our scanned Greek form reached 93.6%, with most mistakes in a large bold title and in letters with diaeresis.

Does the converter keep tables?

Tables drawn with visible lines in a normal PDF became real Word tables in our test. Tables without lines may lose their header row alignment, and tables in scanned PDFs come back as tab-separated text, not Word tables.

How long does a conversion take?

The one-page text PDFs in our test took about 1 to 2 seconds on the server, and the scanned pages about 4 to 5 seconds, plus upload and download time.