Why anyone counts the words in a PDF
Three reasons, mostly. A translator quotes per word and needs to know what they are quoting on. A student or author has a limit to meet and only has the typeset version. An editor is estimating how long something will take to read, or to work through.
All three want a number they can rely on, and all three routinely find that the number they get does not match the number somebody else got from the same document. The reasons for that are worth understanding, because the disagreement is usually legitimate rather than a fault.
What a text layer is
Most PDFs carry their words as text. Underneath the visual page there are actual characters, which is why you can select a paragraph and copy it.
Scanned documents do not. A scan is a photograph of a page, and to software it contains no words at all — only an image that happens to look like words to a human. Nothing can count it without optical character recognition first, which is a separate and much heavier operation.
So the first thing any count tells you is which kind of document you have. A PDF that returns nothing is not empty; it is a picture. The [msm_tool slug="pdf-word-counter"] reads the text layer and says so plainly when there is not one, rather than returning a zero that looks like an answer.
Why the number is higher than you expected
A count taken from a PDF includes everything drawn on the page. That means the running header repeated on all forty pages, the footer, the page numbers, the figure captions, the table cells, the footnotes and the entire reference list.
Your word processor, counting the same document before export, was counting the body text. That is the usual source of a discrepancy of several hundred words, and neither number is wrong — they are counting different things.
It matters most where the count is contractual. A translation quote based on a PDF count includes boilerplate that repeats on every page and may not need translating forty times. A word limit measured against a PDF count includes the bibliography that the limit probably excludes. Say which count you mean, and where it came from.
What counts as a word
Splitting text on whitespace is the standard method and it is a convention rather than a truth. It counts a hyphenated compound as one word, an em dash between two words as part of neither, "e.g." as one, and a date as one. Every tool draws these lines slightly differently, which is why two honest counters disagree by a per cent or so on the same text.
PDF adds its own wrinkles. Text is stored as positioning instructions rather than as sentences, so words broken across a line by hyphenation may come back as two pieces. Ligatures — the joined "fi" and "fl" of a well-set font — sometimes extract as a single character and sometimes not. Multi-column layouts can interleave if the producer drew them in an unhelpful order.
None of this shifts a count by much. It does mean treating the last two digits as noise, and being suspicious of any tool that reports a word count to an implied precision it cannot have.
Characters, and when they are the real unit
Words are the wrong unit for several jobs. Character counts with and without spaces are what publishers use for casting off, what several translation markets price on, and what any field with a hard limit measures.
The two figures differ by roughly the word count, since a space follows most words. Which one is wanted is always worth asking rather than assuming, because the difference on a long document is substantial.
Words per page tells you something the total does not
Divide the words by the real page count and you learn about the document's density. Around 250 to 300 to a page is ordinary prose in a typical layout. Much below that means the pages carry tables, images or generous white space. Much above it means dense typesetting, small margins or a small point size.
That figure is a better predictor of how long a document takes to work through than the word count alone, and it is the number that tells a translator whether they are looking at flowing text or a mass of fiddly labelled diagrams, which price very differently.
Reading time, honestly
Reading time estimates divide by a constant, usually around 200 words a minute for careful reading of ordinary prose. It is a reasonable figure for an article and a poor one for anything technical, where the reader stops, re-reads and works through examples.
Take it as an order of magnitude. A document reporting an hour is an evening's work, not fifty-eight minutes, and treating a reading-time figure as a schedule is how a review deadline gets set badly.