Every counter has to pick a rule, and the reasonable rules disagree. The common one is to trim the text and split it on runs of whitespace, counting the pieces. That is fast, language-neutral and defensible — and it produces surprises.
Take this sentence:
The state-of-the-art model — trained on 1,000,000 documents — did not converge.
| Rule | Count | Why |
|---|---|---|
| Split on whitespace | 12 | Both bare em-dashes become tokens of their own |
| Whitespace, ignoring punctuation-only tokens | 10 | The two dashes are discarded |
| Treat hyphenated compounds as separate words | 13 | state-of-the-art splits into four |
All three answers are correct under their own convention. This is the whole reason a word processor, a submission portal and an online counter can each give you a different number for the same paragraph — and why an essay limit of 2,000 words is best treated as approximate rather than exact.
Other cases that move the number: a number written as 1,000,000 is one token but three if commas are treated as separators; an em-dash used without surrounding spaces joins two words into one token; a URL is one token no matter how long; and in Chinese, Japanese and Thai there are no spaces at all, so whitespace splitting reports a handful of words for a full page of text.
Most people expect one number and there are genuinely three, because a character can mean a storage unit, a Unicode code point, or a thing a reader perceives as one symbol:
| Text | UTF-16 units | Code points | Perceived characters |
|---|---|---|---|
| hello | 5 | 5 | 5 |
| café written with a combining accent | 5 | 5 | 4 |
| A single thumbs-up emoji | 2 | 1 | 1 |
| A four-person family emoji | 11 | 7 | 1 |
| A flag emoji | 4 | 2 | 1 |
The naive length of a string in most programming languages is the first column, which is why a database field of 20 characters can reject a nine-emoji nickname. The number a human would give is the last column, and getting it requires grapheme-cluster segmentation rather than counting.
| Where | Limit | Counting rule |
|---|---|---|
| X posts | 280 | Weighted — Latin and most European text counts 1, but CJK and several other ranges count 2, so a Japanese post caps near 140 |
| Instagram caption | 2,200 | Only the first line or two shows before the More link |
| SMS, single segment | 160 | 7-bit GSM alphabet; one non-GSM character switches the whole message to 70 per segment |
| Meta title | about 50–60 | Truncated by pixel width, not characters |
| Meta description | about 150–160 | Also pixel-based, and frequently rewritten entirely |
Two of those deserve expansion. The SMS rule catches people out constantly: typing one curly apostrophe or one emoji into a 160-character message can silently split it into two billed segments of 70. And the SEO figures are a proxy, not a rule. Search results are truncated when the rendered string exceeds an available width in pixels, so a title of narrow letters survives longer than one full of capital W and M. A 55-character title in wide characters can be cut while a 62-character title in narrow ones is not. Treat 60 as a guideline and check how the actual string looks.
The estimate is simply minutes = words ÷ reading speed. The speed is the contested part. A 2019 meta-analysis of reading-rate studies put average silent reading of English non-fiction at roughly 238 words per minute, with fiction somewhat faster; most publishers use 200 to 250.
| Words | At 200 wpm | At 238 wpm | At 300 wpm |
|---|---|---|---|
| 500 | 2.5 min | 2.1 min | 1.7 min |
| 1,200 | 6.0 min | 5.0 min | 4.0 min |
| 3,000 | 15.0 min | 12.6 min | 10.0 min |
The spread across that table is the honest error bar. Speaking is much slower — a presenter delivers roughly 130 to 150 words per minute, so a five-minute talk is about 700 words, not 1,200.
Sentence counting normally splits on a full stop, question mark or exclamation mark. That over-counts on abbreviations such as Dr. and e.g., on decimals such as 3.14, on ellipses, and on initials in a name; it under-counts when a writer uses semicolons or line breaks instead of terminal punctuation. Paragraph counting depends on whether a single line break or a blank line is treated as the separator, which is why pasting from a PDF can report every line as its own paragraph.
A word count measures length and nothing else. It cannot see whether a sentence is clear, whether an argument holds, or whether a paragraph earns its place — and a longer count is not a better piece of writing. Readability scores such as Flesch-Kincaid, where they are offered, are formulas over sentence and syllable length; they reward short words, which is not the same as rewarding good prose.
On the search side, there is no minimum word count that ranks and no keyword-density figure to hit. Density targets are an artefact of tools from an earlier era of search, and writing to satisfy one reliably makes text worse. Length should follow from what the topic needs.
Because it applies a different rule to hyphens, dashes and punctuation-only tokens. For a hard limit, count with whatever tool the person enforcing the limit uses, and leave a margin of a few per cent.
Weighted counting. Characters outside the Latin and common European ranges count double, and a shortened link is counted as a fixed length regardless of the URL you pasted.
It depends on who is asking. Publishers pricing by characters usually exclude spaces; platform limits and database fields include them. Both figures are shown for that reason.
No. All counting happens in your browser, and nothing you paste is transmitted, logged or saved.
Every tool comes with a written guide, and every category is one click away.