PDF to text converter

What comes out, and in what shape

A PDF does not store its text as sentences. It stores a drawing program: choose this font at this size, move to these coordinates, draw these glyph codes, move again. To get the words back, this page runs that program — tracking the text matrix, the font size, character and word spacing and every glyph's width — until it knows which character each code stands for and where on the page it lands. Then it puts the pieces back together into lines.

By default you get the lines as they sit on the page, with a blank line wherever the gap between two lines is clearly bigger than the page's normal line spacing. That keeps addresses, tables and headings readable. Tick Join wrapped lines into paragraphs when you want flowing prose instead: lines inside a paragraph are joined with a space, and a word that was split across two lines with a hyphen (infor- / mation) is glued back together, as long as the second half starts with a lower-case letter.

Tabs between table columns turns a wide horizontal gap into a tab character instead of a space. It is on by default because the most common reason to pull text out of a PDF is a statement or a report with columns in it, and tab-separated lines paste straight into a spreadsheet as separate cells. Switch it off if you only want prose.

Why a PDF can have no text at all

If a document was scanned, photographed, or “printed” through an image-only driver, each page is a single picture. There are no characters in the file to find, only pixels that happen to look like letters. This tool tells you exactly which pages are in that state rather than handing you a blank result without explanation. Getting text out of them needs optical character recognition, which is a different job and is not done here.

Scanners and phone apps often run OCR themselves and hide the recognised text behind the image, drawn in invisible ink. That hidden layer is real text as far as the file is concerned, and it is extracted like any other. Its quality is whatever the scanner's OCR managed, so a scan of a creased page can produce text with odd substitutions that were already in the file.

Gibberish, and how it is avoided

You may have copied text out of a PDF and pasted something like pUHYLRXV\HDU¶V instead of “previous year’s”. That is not corruption. The font in the file numbers its glyphs in its own order, and the file forgot to include the table (a ToUnicode map) that says which glyph is which letter. A naive extractor reads the glyph numbers as if they were character codes and every letter comes out shifted.

When that table is missing or empty, this page goes looking inside the embedded font file itself: its own character map, read backwards, and failing that its glyph names. For the Microsoft core fonts — Arial, Times New Roman, Courier New, Verdana, Tahoma and Georgia, which all share one glyph order — it can fall back on that order when iPhones and Macs strip everything else out, which they regularly do. On a set of iPhone-generated forms where a well-known PDF library returned shifted nonsense, this recovers the actual sentences.

When none of that works, the characters are counted and left out, and the report under the text tells you how many. A gap you know about is more useful than a paragraph of plausible-looking garbage that you only notice after you have pasted it into an email.

Reading order

Text comes out in the order the file draws it, which for almost every word processor, report generator and browser is the reading order: columns are drawn one after another, not interleaved line by line. That is the right default, because sorting everything top to bottom would splice the left and right columns of a two-column page into nonsense.

The cost is that a few generators draw a page in an odd order — the footer first, or a table cell by cell down each column — and the extracted text follows them. When that happens the text is all there, just in a sequence you need to rearrange. Rotated text, such as a sideways label on a chart, comes out as its own line.

Two things that look like text are not included. What someone typed into a fillable form's fields lives in the form layer, not in the page's drawing, and comments and sticky notes are annotations laid on top of the page. Both are left out, so a filled-in form gives you its printed labels without the answers.

Fake bold and repeated text

Some programs make text look bold by drawing it several times, each copy nudged a fraction of a point sideways. Bank statement systems do this constantly. A simple extractor faithfully reports every copy, so “Deposit $31,746.61” appears four times in a row. Here, a run of text that repeats the previous one at almost exactly the same position is recognised as the same text and kept once. The tolerance is deliberately smaller than the narrowest letter, so two genuine letter l’s in “William” are never mistaken for a copy of each other.

Letter-spaced headings get the reverse treatment. When a heading is drawn one glyph at a time with extra tracking, every letter has a small gap after it; the tool measures the usual gap on that line and only calls it a space when a gap is clearly bigger. Otherwise “YOUR PRIVATE LINK” would come out as a row of single letters.

How it was checked

Before this page went up, its output was compared against PyMuPDF, a widely used open-source PDF library, on 288 real documents from more than a hundred different producers: Word, Excel and PowerPoint exports, Chrome and Safari print-to-PDF, LaTeX, bank statement systems, iText, ReportLab, scanners with OCR, and more. For the median document the characters matched exactly. Where they did not, the difference was usually on the reference side — the repeated fake-bold copies described above, or the letter-shifted text from forms whose fonts had lost their mapping.