Why word counts differ between tools
Two counters can both be right and still disagree. The difference comes down to a handful of rules, and knowing which one applies tells you which number to trust.
Paste the same document into two word counters and you will often get two numbers. In a short piece the gap is a handful of words. In a dissertation it can be a few hundred. It is never random: every counter follows a rule, and two tools that disagree are disagreeing about a specific kind of text.
Four kinds of text account for most of it — hyphenated compounds, links and email addresses, numbers written with separators, and contractions. A second source of disagreement is larger and has nothing to do with where words begin and end: whether footnotes, captions, headers and text boxes are part of the document being counted at all.
What follows is where each disagreement comes from, the rule this site applies so that its number is predictable, and what to do when you are close to a limit somebody will enforce.
The rules that split the count
Every counter has to decide where one word ends and the next begins. One approach splits the text only at whitespace. The other splits at anything that is not a letter, which promotes the hyphen, the apostrophe, the full stop and the slash into word boundaries.
That single design choice produces every row below.
| Text | Counted here | Where tools commonly disagree |
|---|---|---|
| well-known | 1 word | A tool that treats the hyphen as a boundary reports 2. "State-of-the-art" is 1 word here and 4 there. |
| don't | 1 word | Usually 1 everywhere. A tool that reports 2 is splitting on punctuation, which will also inflate your links and your numbers. |
| 3.14 | 1 word | Mostly 1. Splitting on the full stop makes it 2, and turns "U.S.A." into 3. |
| 1,300 | 1 word | Mostly 1 in English. The same construction written 1.300, with a full stop as the thousands separator, is the one that trips tools tuned for a single locale. |
| https://example.com/report | 1 word | The widest spread of all. Splitting on dots and slashes can turn one link into five or six. |
| ana@example.com | 1 word | Usually 1, sometimes 2 or 3. A contact list or an appendix full of addresses makes it visible. |
| — (a dash on its own) | 0 words | Counters that keep every non-space run count it as 1, quietly adding a word per item in a dashed list. |
| 12 March 2026 | 3 words | Agreed almost everywhere. Dates only cause trouble written as 12/03/2026, which is 1 word here and up to 3 elsewhere. |
How big the gap gets
The arithmetic is worth doing once, because it decides whether the difference is trivia or a problem.
Take a 2,000-word literature review with 40 hyphenated compounds in the prose and 15 web addresses printed in the body text. Under a whitespace rule that document is 2,000 words. Under a punctuation-splitting rule each compound gains at least one word and each link gains four or five, so the same file reports somewhere near 2,115. That is nearly six percent, on a document that has not changed by a single character.
Now take 2,000 words of plain narrative prose: no links, no technical compounds, no figures. Two tools will usually land within five words of each other, and which one you used does not matter.
The size of the gap is therefore a property of your writing rather than of the tools. If your text is dense with hyphenated terminology and URLs, its count is fragile and no single number deserves much confidence. If it is not, the disagreement is noise and you can ignore it.
The rule this site uses
Stated as a procedure: cut the text at every space, tab and line break; keep each resulting piece that contains at least one letter or digit; the number of pieces you kept is the word count.
Stated formally, a word is a maximal run of non-whitespace characters containing at least one letter or digit. Nothing about meaning enters into it. There is no dictionary, no list of compounds and no language detection, which is why the behaviour does not change when you switch between English and Portuguese, or when you paste something the rule has never seen before.
It is worth being clear about what that rule gets wrong rather than only what it gets right. It has no way to know that "state-of-the-art" reads as four words to a person. It counts a fifty-character link as one word even though the link takes up half a line on the page. Both outcomes are defensible. Neither is the only defensible answer, and a marker counting by hand would probably disagree with both.
The reason to publish the rule this plainly is that a predictable rule is more useful than one tuned to approximate the average of other tools. You can look at a sentence and know in advance what this page will say about it.
What the tool decides is part of the document
Where words begin and end is the visible disagreement. The larger one is about scope.
An online counter measures exactly what you pasted into it and nothing else. A word processor measures what it considers to be the document, and applications draw that boundary in different places. Footnotes and endnotes may be folded into the main total or reported separately. Headers and footers are often left out, on the reasoning that a running title is not body text. Words inside a shape, a text box or a diagram are frequently treated as a separate object and skipped. Captions and the contents of tables can go either way. Comments and tracked-change balloons are normally excluded, though a deletion you have not yet accepted is still sitting in the file.
Little of this is documented anywhere a writer will find it, and it varies between applications and between versions of the same application. The reliable move is to stop asking what your software includes and start controlling it: select the text you actually want measured, and count the selection. If your word processor reports a figure for a selection, that figure and a count of the same selection pasted here are describing the same words, which reduces any remaining difference to a pure question of where words are split.
For coursework this matters most with footnotes. An essay with substantive footnotes can carry a tenth of its words down there. Whether those words count is a question about your brief rather than about your software.
Near a hard limit, only one count matters
If a limit is enforced by software, the number that decides is the one that software produces. Every other count is an estimate of it.
That sounds obvious and is ignored constantly, because the figure you watch all day is the one in the application you write in. A submission form that refuses anything over 3,000 words is not consulting your word processor. It applies its own rule to the file or the text it receives, and it will not explain that rule to you before you submit.
The same holds one level up. If a person enforces the limit, their count is the one that matters, and a human count follows the brief rather than any algorithm. A brief that excludes quotations means somebody is subtracting quoted words by hand, and no tool — this one included — will reproduce that number for you.
The practical consequence: "my counter said 2,998" is not a defence. It is evidence about a rule nobody agreed to use.
What to do when you are close to the limit
This takes about five minutes and removes the entire category of problem.
- Read the brief for what is excluded. References, quotations, footnotes, appendices and the abstract are all commonly outside the limit, and getting that wrong dwarfs every difference described above.
- Count in the system that will judge you, if it shows a number at all. Upload a draft early, read the figure it reports back, and treat that as the real one.
- If nothing will show you a number before the deadline, count in at least two tools and write both down. The gap between them tells you how sensitive your document is.
- Leave a margin. Two to three percent under a hard ceiling absorbs every disagreement on this page. On a 3,000-word limit that is 60 to 90 words, which is roughly the paragraph you were going to cut anyway.
- Recount after any edit that touches footnotes, captions or the bibliography, because those are precisely the regions tools disagree about including.
- Do not let a single number decide anything on the final day. If two tools differ by more than your margin, assume you are over until you have checked in the place that counts.
What a word counter cannot tell you
Being candid about this is more useful than claiming a precision that does not exist.
This site counts the text you give it, exactly, under one stated rule. It does not know what your institution counts. It does not know whether your marker will subtract the block quotations. It cannot see the footnotes you did not paste. When its figure differs from a submission system's, the submission system is not making a mistake — the two are answering slightly different questions, and only one of those questions is being asked by the person marking your work.
It also does not touch your writing. It measures what you have already written and reports numbers about it. It will not generate, expand, shorten or rewrite a sentence, because deciding what stays and what goes requires knowing what the piece is for, and a counter has no access to that.
Used as an instrument it is a good one: fast, private, consistent, and honest about its own rule. Trusted as an oracle it is a way to be confidently three percent wrong on the day it matters.