PDF to text

How a PDF stores text, and why copying from a reader goes wrong

A PDF is built for printing, not for reflowing. A page does not contain paragraphs. It contains instructions to place short runs of characters at exact coordinates: “put these eleven characters here, then these six over there”. The order of those instructions is whatever order the program that made the file happened to write them. A page that looks like two tidy columns may be stored as one line from the left column, one line from the right, then the next.

Each character is stored as a code in a font, and the font says which shape to draw for that code. To turn codes back into text a program needs a mapping, usually a table called ToUnicode. Good generators include one. Some do not, or write a wrong one, and the extracted text then comes out as the wrong letters, or as symbols, even though the page looks perfect.

Extracting text is therefore partly a reconstruction. The tool reads every run with its position, sorts the runs into lines, joins runs that sit on the same line, groups lines into columns, and writes out what it finds in the order a person would read it. When that works, a two-column article gives you column one top to bottom, then column two. When the layout is unusual, such as text wrapped round a picture, a sidebar or a table, the result is a best effort.

What the output looks like, and what it leaves out

• Plain text only. No fonts, bold, italics, colours, images, links or shapes. Characters outside basic English, such as accented letters or Greek, are kept where the PDF holds them as real text, and the file is saved as UTF-8.
• Lines and paragraphs. Line breaks are kept where the page breaks a line, so text copied from a narrow column keeps its short lines. If you plan to paste into a word processor, expect to join lines.
• Hyphenated words. A word split at the end of a line with a hyphen stays split, because the hyphen is part of the stored text and is not distinguishable from a genuine one without reading the meaning.
• Ligatures. Joined letter pairs such as “fi” and “fl” can be stored as one character. Many fonts map them back to two letters, but some do not, and you may see a missing letter in words such as “office”.
• Tables. Cells come out as separate lines or runs of text in row order. The grid is lost; numbers keep their order, which is often enough to paste into a spreadsheet and tidy.
• Headers, footers and page numbers. They are included as text, in the position they appear, so they will interrupt the body text on each page. Page markers help you find and delete them.
• Pictures of text. Not read. Pages that are pictures return empty, and the tool says so.

Scanned PDFs: why they come back empty, and what to do

A scan is a photograph of a page stored inside a PDF. The letters on it are pixels, so there is nothing to extract. This tool detects the pages without text and tells you: if the whole file has none, a banner says “No text found” and points you to OCR PDF; if only some pages have none, it lists their numbers, with “…” after the first twelve.

OCR reads the pixels and adds a hidden text layer that matches them. Once a file has been through OCR PDF, extract the text from the result. Accuracy is good on clean, upright, printed pages and weaker on small type, skewed scans, colour backgrounds and handwriting, which is not supported. Always proof-read figures, names and dates against the page.

Some files mix both kinds: typed text with scanned appendices. In that case you will see pages of text and a few empty ones. Extract the typed pages, OCR the others, and combine the output.

How this differs from copying in a reader, opening in Word, or using Google Docs

If what you want is to keep the design, extract pictures instead with PDF to JPG, or stay with the PDF.

Practical uses

Checking a redaction

Extracting a file’s text shows you everything it holds in one view. If a name you removed with Redact PDF appears in the output, it is still in the file. This is the strongest check you can do without special software.

Quoting and note-taking

Pull one page or one range to cite a section of a report. A range such as 14-19 is quicker than selecting across six pages.

Counting words

The word and character counts give you the length of a document without opening it in a word processor, which matters for a dissertation limit or a submission field with a maximum. Counts are indicative: the tool counts runs of characters separated by spaces, which will differ a little from a word processor’s count on hyphenated words and numbers.

Feeding text to another tool

Turn off page markers, copy, and paste into a translator, a summariser, a spelling checker or a notes app.

Accessibility

People who use screen readers sometimes find a PDF easier to listen to as plain text. The plain text carries no headings or structure, so treat it as a fallback rather than an accessible version of the document.

Two worked examples with numbers

The figures are round numbers chosen to show how the counts behave. They are not benchmark results.

Length of a 12-page chapter

A chapter of 12 typed pages, with about 450 words on a full page, should come to roughly 5,400 words. The chips read, for example, 5,386 words, 31,200 characters and 12 pages read. Characters divided by words gives about 5.8, which is typical English prose including spaces. If the page count read is smaller than 12, some pages had no text and are listed in the banner.

A mixed file with scanned appendices

A 20-page file has 14 typed pages and 6 scanned appendix pages. Extraction returns text for 14 and a banner saying pages 15, 16, 17, 18, 19 and 20 have no text. Run the file through OCR, then extract again with the range 15-end to add the missing pages. The final text is 14 pages from the first run and 6 from the second, pasted in order.

Limits and edge cases

• Right-to-left scripts. Arabic and Hebrew are stored in many different ways, and the order of the output may need checking against the page.
• Text drawn as shapes. Logos and some design-heavy documents turn letters into outlines. Those words are not text and are not extracted.
• Rotated pages and vertical text. Positions are read as displayed, but heavily rotated or vertical layouts can come out in an odd order.
• Form fields. Typed answers in a fillable form live in the form, not the page text, and may not appear. Flatten with Flatten PDF first if you need them.
• Hidden or white text. Text drawn in white on a white background is still text and is normally extracted along with the rest. That makes this a useful check for hidden words you did not know were there.
• Copy-protected files. A PDF that restricts copying but opens without a password is read normally here. If you do not have the right to copy a document, do not.
• Very long documents. A thousand-page file takes longer and more memory. Use a range to do it in parts.

Privacy: your document’s words never leave the page you are on

Text extraction is a read-only job, which makes the privacy case easy to state: the file is opened in your browser, the engine reads it locally, and the text appears in a box on your screen. No server receives the file or the words. When you press Copy or Download, the text goes to your clipboard or your downloads folder and nowhere else.

This matters most for the documents people most often extract from: contracts, letters, medical reports and statements. The verification guide shows how to watch the Network tab while you extract, and how to confirm it with your connection off.