How to check word count, and why two tools disagree

There is no single definition of a word

Paste the same text into three tools and you can get three numbers. None of them is broken. They are applying different rules to cases that genuinely are ambiguous.

The disagreements are almost always these:

Hyphenated compounds. Is state-of-the-art one word or four? Most tools say one. Some say four. On a technical document this alone can shift the count by a percent or two.

Numbers and dates. Is 2026 a word? Is £1,500? Is 3.14? Counting anything separated by whitespace says yes to all three. A dictionary-based count says no.

Contractions. Don't is one word everywhere sensible, but a naive apostrophe split makes it two.

Em dashes without spaces. word—word is two words to a human and one to anything splitting purely on whitespace.

URLs and email addresses. One long token, or several?

The most common convention (and the one to assume unless told otherwise) is split on whitespace: a word is any run of non-space characters. It is what Word approximates, it is simple to explain, and it is what most submission systems use.

Which number does your editor use?

• Microsoft Word splits on whitespace with some special handling, and

counts numbers as words. It excludes headers, footers, footnotes and endnotes from the main figure by default, which is the single biggest source of surprise on an academic manuscript.

• Google Docs is close to Word and differs in the margins.
• Most web tools, including this one, split on whitespace.

If a count has to match somebody's threshold exactly, check in the tool they will use. Everything else is an estimate accurate to within a percent, which is fine for a blog post and not fine for a submission with a hard limit.

Characters are the unambiguous measure

Where word count is fuzzy, character count is exact, which is why every hard limit in the wild is in characters.

The one thing to know is with or without spaces. Both are legitimate, they differ by roughly 15%, and they are frequently confused. When a limit says "280 characters", it means with spaces.

Then there is the awkward part: a character is not always a character. Platforms that count in UTF-16 code units treat most emoji as two, and some composed emoji as many more. A family emoji can cost eleven. This is why a tweet that looks short can be rejected.

The limits worth knowing

The search ones are approximations on purpose: Google truncates on pixel width, not character count, so a title in wide capitals is cut sooner than one in narrow lowercase. Treat 60 as a target rather than a boundary.

Reading time and speaking time

Both are word count divided by a rate, and both are worth sanity-checking rather than trusting.

Reading: 200–250 words per minute for ordinary prose. Technical material is slower, 150 is more honest for anything with code or equations in it.

Speaking: 130–150 words per minute for clear delivery. Audiobook narration sits near 150. Conference talks run slower than people expect, and nerves make them faster than they rehearsed.

A useful check: a 20-minute talk is about 2,600 words. If your script is 4,000 you are going to overrun, and no amount of speaking quickly will fix it.

Counting as you type, without uploading

The entire operation is arithmetic on a string, which means there is no reason for it to happen on a server. A tool that uploads your draft in order to count its words has your draft.

That matters more than it sounds for the things people actually count: cover letters, dissertations, legal submissions, articles under embargo.