Why word counts disagree

Paste the same paragraph into four counters and you will get somewhere between two and four different answers. This is not a bug in any of them. It is four defensible answers to a question that has never had a single agreed definition, and knowing which decisions produce which gaps lets you predict the difference instead of being surprised by it.

A worked example

Take this sentence:

Dr. Alvarez's state-of-the-art rig logged 1,204 readings — see https://example.org/data — between 3.15 a.m. and dawn.

Reasonable tools report this as anywhere from 14 to 24 words, and as 1 or 3 sentences. Here is where each decision lands:

None of those choices is unreasonable in isolation. Together they produce a 70 per cent spread on one sentence.

The six decisions that cause almost every difference

1. Hyphenated compounds

This is the biggest one by volume. State-of-the-art, well-known, twenty-five, self-employed. We count each as one word, on the grounds that it is one lexical item. Microsoft Word splits on the hyphen. In a document that uses compounds freely — academic prose does, technical writing does — this alone accounts for one to three per cent of the total.

2. What ends a sentence

Abbreviations are the classic trap. A tool with no abbreviation list breaks after Dr., etc., e.g. and every initial in J. R. R. Tolkien. A tool with a list has to decide how long the list is. And the line-break question — is a heading a sentence? — splits tools straight down the middle. We say yes; Word says no. On a bulleted document that difference can halve or double the sentence count, which then corrupts every readability score derived from it.

3. What a character is

Three different things get called “a character”:

• Grapheme clusters — what a person sees. An emoji is 1. This is what we count.
• UTF-16 code units — what JavaScript's .length and Microsoft Word return. Most emoji are 2, composite emoji more.
• Bytes — what a database column limit often means. An accented Latin letter is 2 bytes in UTF-8, a CJK character 3, most emoji 4.

A 20-character string of emoji is 20, 40 or 80 depending on which of those you asked.

4. What is in the document at all

This causes the largest absolute surprises, and it is not a counting rule at all. Word can include or exclude footnotes, endnotes, headers, footers, text boxes and comments. A thesis with 400 footnotes counts very differently depending on one checkbox. When you paste into a browser counter, you get exactly what you selected — which is often not what your institution means by the word count of the document.

5. Numbers and symbols

Is 1,000,000 one word or two? Is a lone & a word? Is £50 one word or two? Tools differ, and documents full of figures — financial reports, lab writeups — are where it shows.

6. Weighted counting

Some platforms deliberately do not count characters. X weights CJK at 2 and fixes every URL at 23 units. Some SMS gateways count segments rather than characters. These are business rules wearing the costume of a character count.

Rough guide to the size of each gap

What to do about it

1. Find out what the enforcing system counts. If there is a submission portal, its number is the only one that exists. Paste into it early, not at the deadline.
2. Ask what is included. Most academic word counts exclude the bibliography, many exclude footnotes, some exclude block quotes, and almost none say so unless asked.
3. Leave a margin. Three to five per cent under a hard cap absorbs every difference described on this page.
4. Use one tool consistently while drafting. The absolute number matters less than being able to see whether you are going up or down.

Our rules are all published on How we count, which is the point: you can predict our number, and you cannot predict a number whose rules are hidden.

Common questions