How to use the Word and Token Counter
The token counter updates as you type, so there is no separate step to run it. The steps below describe the tool generically.
- Paste or type your text into the input box. It can be an article, a prompt, a page of documentation, a product description or an email.
- Read the live counts. You will see words, characters (with and without spaces), sentences and often paragraphs. These update immediately as you edit.
- Check the token estimate. This is an approximation of how many tokens a language model would use for the same text. See the section below for what that number can and cannot tell you.
- Check the reading time. It converts the word count into minutes at a typical adult reading speed.
- Edit and compare. Trim, rewrite or split the text and watch the numbers move. When you are within your target, copy the text back out.
TIP: To compare formats, count the same content twice: once as plain text and once as raw HTML or JSON. The difference in tokens shows how much markup costs when you give a page to a model.
What the counter measures
Each count answers a different practical question, so it is worth knowing what each one means and where counters disagree.
Words
A word is a run of characters separated by spaces or line breaks. That simple definition works well for English and most European languages. Hyphenated terms such as “well-known” count as one word in most counters, and numbers and URLs count as one word each. Languages written without spaces between words, such as Chinese and Japanese, cannot be counted this way, so word counts for those languages are unreliable and character counts are more useful.
Characters
Characters are counted with and without spaces. Character counts matter for fields with hard limits: meta descriptions, ad copy, social posts, SMS messages, form fields and database columns. Be careful with emoji, accented letters and some symbols, which can be stored as more than one unit in some systems even though they look like one character.
Sentences
Sentence counts are an estimate based on punctuation: a period, question mark or exclamation mark followed by a space or the end of the text. Abbreviations such as “e.g.” and “Dr.” and decimals such as “3.5” can confuse a simple counter, so treat the number as approximate. Average sentence length, which is words divided by sentences, is a better use of this number than the exact count.
Reading time
Reading time divides your word count by an assumed reading speed. Adults typically read silent English prose at roughly 200 to 250 words per minute, and technical or dense text is slower. The tool’s estimate is a guide for setting expectations (“a 4-minute read”), not a measurement of any individual reader. Speaking time is slower, around 130 to 160 words per minute, which is useful when you write scripts.
What is a token?
Language models do not read words. They read tokens, which are pieces of text produced by a tokenizer before the model sees anything. A token can be a whole common word, part of a longer word, a single character, a space attached to a word, or a piece of punctuation. The word “the” is usually one token. A rare word like “tokenization” may be split into two or three. A long string of digits is often split into several tokens.
The model’s limits and pricing are expressed in tokens. A context window, the maximum amount of text a model can consider at once, is measured in tokens and includes both your input and the model’s output. API providers charge per token, usually with different rates for input and output. So when you plan how much content to feed to a model, or how much a task will cost, tokens are the unit that counts.
Rules of thumb for English
For ordinary English text, a common rule of thumb is that one token is about four characters, or about three quarters of a word. Put the other way, 100 words is roughly 130 to 135 tokens, and 1,000 words is roughly 1,300 to 1,400 tokens. These are averages. The ratio shifts with the kind of text, and the real count can sit well above or below it.
This is how the counter produces its estimate. It applies those ratios to your text, so you get a useful ballpark in a split second without loading a model-specific tokenizer.
Why a counter can only estimate
Different models use different tokenizers. OpenAI models, Anthropic’s Claude models, Google’s Gemini models and open models such as Llama each have their own vocabulary and splitting rules, and vendors change them between model generations. The same sentence can come out as a different number of tokens depending on which model you use. The differences are often modest for plain English, but they can be larger for other languages, code and unusual text.
Accurate counts require the vendor’s own tokenizer or token-counting tool for the specific model you plan to use. Several providers publish tokenizers or offer a counting function in their API. If you are within about 10 or 20 percent of a hard limit or budget, use the real tokenizer. If you are well inside, an estimate is fine. Treat this tool’s token number as a planning aid.
What makes token counts go up
Several kinds of text use more tokens than the four-characters-per-token average suggests.
- Non-English text. Tokenizers are usually trained on text where English dominates, so many other languages break into more, smaller tokens. Text in languages such as Hindi, Arabic, Thai or Korean can use noticeably more tokens per word than English. Chinese and Japanese often use roughly one or more tokens per character.
- Code. Symbols, indentation, short variable names and punctuation tend to split into many small tokens.
- Numbers. Long numbers and identifiers are split into chunks of digits.
- URLs, hashes and IDs. Random-looking strings do not match common word pieces, so they split heavily.
- Markup. HTML tags, attributes, JSON braces and quotes all count. A page of HTML can use several times the tokens of the same content as plain text or Markdown.
- Unusual formatting. Repeated whitespace, emoji and special characters can add up.
Why words and tokens matter for AI search
It might seem odd to count words on an AI search site, but length plays a real part in how your content is used by answer engines and by the AI tools your team works with.
Models read in limited chunks
Retrieval systems break pages into chunks before storing them, and they pick a handful of chunks to show the model when answering. Chunks tend to be a few hundred tokens long. A section that makes one clear point in a few short paragraphs fits cleanly in a chunk. A section that rambles across 1,500 words gets cut into pieces that may not make sense alone. Counting words per section helps you keep each answer self-contained.
Content has to fit when you feed it to a model
When you use an AI assistant to summarize, rewrite, translate or audit a page, the text must fit in the model’s context window along with your instructions and the answer. A long page plus a long set of instructions can exceed the limit, and the model may truncate or refuse the input. Counting tokens before you paste saves a failed attempt, and trimming boilerplate gives the model more room for the work you want.
Cost scales with tokens
If you run content through an API at scale, for example to classify thousands of pages, cost is tokens times price. A reliable estimate of the average page length lets you forecast the bill before you start. Converting HTML to Markdown first usually cuts the cost sharply. The HTML to Markdown converter is made for that step.
Length is not quality
Longer pages are not better for AI visibility, and shorter pages are not either. Engines quote passages, not pages. A 600-word page with direct answers can be cited where a 4,000-word page that never states the answer is skipped. Use word counts to check that a page is long enough to be thorough and short enough to be clear, not to hit an arbitrary length target. The AI answer readiness checker looks at the features that matter more for being quoted.
Useful length guidelines
These are practical ranges, not rules. Always let the content decide.
Page titles and descriptions
Titles do best at about 50 to 60 characters and meta descriptions at about 120 to 155, though search engines truncate by pixel width, not by characters. The title and meta description checker shows pixel estimates.
Direct answers
A direct answer to a question, placed right under its heading, works well at about 40 to 60 words, roughly two to four sentences. This is long enough to be complete and short enough to quote whole.
Section length
A section under a heading is easiest to retrieve and quote at around 100 to 300 words. If a section grows beyond that, consider splitting it with sub-headings.
Articles
Guides that cover a topic thoroughly often run 1,500 to 3,000 words. Quick answers can be far shorter. There is no magic number, and padding to reach a target hurts readers.
Prompts
Good prompts are as long as they need to be. For recurring tasks, short clear instructions with examples usually beat long rambling ones. If you are building a prompt that is sent thousands of times, token count is a direct cost, so trim it.
FAQ answers
Answers in FAQ sections and FAQ schema work best at one short paragraph, often 40 to 80 words, so they can be lifted without editing.
Best practices
Use these habits to get more from a counter.
Count before you paste into a model
If a page, a PDF extract or a transcript might be long, check the count first. If it is near the limit of the model you are using, trim, split into parts or summarize in stages.
Count what the model will actually see
Count the real input, including your instructions, examples and any system prompt, not just the article. Remember that the output also uses the context window. Leave room for the answer.
Compare formats
Paste a piece of content as HTML, then as plain text, then as Markdown. The differences make the case for cleaning content before sending it to a model.
Use a margin
Because the token number is an estimate, keep a margin of at least 10 to 20 percent below any hard limit, and use the vendor’s tokenizer when you are close.
Watch non-English content
If you work in other languages, remember that the same meaning can take more tokens than in English. Budget accordingly, and prefer a model-specific tokenizer for non-English text.
Edit for the reader first
Do not cut useful content just to hit a number. Use counts to find rambling and repetition, not to remove substance. A fact-rich page earns citations; a short empty one does not.
Keep a record of targets
Write down your length targets for each content type, such as the lengths for titles, descriptions, intros and answers, so everyone on the team writes to the same standard.
Examples
These examples show how the numbers relate. The token figures follow the rule of thumb and are estimates, not outputs of any specific tokenizer.
Text: "Answer engines quote short, direct answers."
Words: 6
Characters (with spaces): 43
Estimated tokens: about 9 to 11
A longer sample shows the scaling at paragraph length.
Plain-English article section
Words: 300
Characters (with spaces): about 1,800
Estimated tokens: about 400 (300 words / 0.75)
Reading time: about 1.5 minutes at 200 to 250 words per minute
Markup changes the picture. Here is the same short sentence as plain text and wrapped in HTML.
Plain text:
Our store ships free over $50.
HTML:
<p class="shipping-note"><span class="strong">Our store ships free over $50.</span></p>
The visible words are identical, but the HTML version has several times as many characters and uses noticeably more tokens, with no extra meaning. At the scale of a full page with navigation, scripts and styles, the gap is much larger, which is why feeding clean text or Markdown to a model is cheaper and often gives better results.
Estimating a bulk job
Here is how you can use the numbers to plan an API job on a fixed list of pages.
Pages to process: 500
Average words per page: 900
Estimated tokens/page: 900 / 0.75 = 1,200
Instruction tokens: 300 per request
Input tokens total: (1,200 + 300) x 500 = 750,000
Output tokens/page: 200 -> 100,000 total
Multiply those token totals by your provider’s per-token rates for the model to estimate cost. Add a safety margin, since your estimate will not be exact, and run a small batch first to compare against the provider’s usage report.
Common mistakes
- Treating the token count as exact. It is an estimate based on an average ratio. Different models count differently, and the real number can be higher or lower.
- Assuming one token equals one word. On average a word is more than one token, about 1.3 for English. For other languages and code, the ratio is higher.
- Forgetting output tokens. The context window must hold your input and the model’s reply. If you fill it with input, the answer gets cut off.
- Counting only the article and not the instructions. System prompts, examples and attached files all count toward the limit.
- Sending raw HTML to a model. Tags, scripts and styles waste tokens. Convert to Markdown or plain text first.
- Using word count for Chinese or Japanese. Without spaces between words, word counters are unreliable. Use characters and a model-specific tokenizer.
- Optimizing for length alone. A page is not better because it is longer. Answer quality, clarity and specificity matter more.
- Trusting reading time as a promise. It is an average. Dense technical material takes longer, and skimmers take less.
- Ignoring character limits that count differently. Some systems count emoji or accented letters as more than one unit, so a string that looks short may exceed a limit.
- Pasting confidential text into tools you do not trust. This counter runs in your browser, but check any other tool before pasting private material.
How the estimate is calculated
The counter takes the character count and the word count and uses both rules of thumb: about one token per four characters, and about 0.75 words per token. For typical English prose those two methods land close to each other, and the tool shows a single estimate derived from them.
It does not run a real tokenizer. Real tokenizers, such as byte-pair encoding implementations, need a large vocabulary file for each model, and different vendors do not share a vocabulary. Bundling them would make the page heavy and would still cover only some models. The estimate is therefore a deliberate trade: you get a fast, private ballpark that works for any text, and you should reach for the vendor’s own tool when exactness matters.
You can improve your own estimate by calibrating. Take a sample of your typical content, count its tokens with the vendor’s tokenizer, and compare that with this tool’s number. If your content runs consistently 15 percent higher, apply that factor. Code-heavy documentation, for instance, will land higher than a plain blog post.
Privacy
The tool counts everything locally in your browser. Your text is not sent to aeoora or to any model, and no AI service is called to produce the numbers. You can safely paste drafts and internal documents. As with any web page, do not paste passwords, keys or personal data you would not want in your browser’s clipboard history.
Frequently asked questions
How accurate is the token count?
It is an estimate. The tool applies the common English averages of about four characters or three quarters of a word per token. For ordinary English prose it is usually close. For code, numbers, non-English text and markup, real counts are often higher. If you are near a hard limit, check with the tokenizer from the model’s vendor.
Why does my text use more tokens in one model than another?
Each model family has its own tokenizer with a different vocabulary and different splitting rules. A phrase that one tokenizer keeps as a single piece, another may split in two. Vendors also update tokenizers between model generations, so even two versions of one model family can disagree. That is why a general counter can only approximate.
How many words is 1,000 tokens?
For English, about 750 words. The reverse also holds: 1,000 words is roughly 1,300 to 1,400 tokens. These ratios drop for text with lots of code, numbers or non-English words, where 1,000 tokens will cover fewer words.
Does the context window include the answer?
Yes. The context window covers the input you send and the output the model generates. If your input uses nearly the whole window, there is little room left for the response. Plan so that input plus the expected output stays comfortably inside the limit.
How is reading time calculated?
The tool divides the word count by an assumed reading speed in words per minute, typically in the range of 200 to 250 for adults reading English silently. It is a rough guide. Complex or technical material takes longer, and readers who skim finish faster. For spoken delivery, use a slower rate of about 130 to 160 words per minute.
Is there an ideal word count for ranking in AI answers?
No. Engines quote passages that answer the question, so the content of a section matters more than the length of the page. Aim for pages that are as long as the topic needs, with sections that stand on their own and direct answers near the top of each. Padding adds nothing and can dilute the quotable parts.
Why do words and characters differ from my word processor?
Different tools make different choices about hyphens, numbers, URLs, line breaks and special characters. Small differences are normal. For hard limits, such as an ad platform’s character cap, always check against the platform’s own counter.
Is my text stored or sent anywhere?
No. Counting happens in your browser. The text is not uploaded, stored or sent to a model.
Go further with aeoora
Word counts and token estimates help you write content that is easy for models to read and quote. To see whether that content is actually showing up in AI answers, aeoora Cloud tracks your visibility across engines, the citations you earn and the visits AI assistants send to your site. Its content editor applies writing rules and quality checks as you draft. The content editor documentation explains how, and the prompts and topics guide shows how to choose the questions worth tracking.
See plans and pricing or start a 2-day free trial; a card is required and you can cancel before day 3. If you run WordPress, the free aeoora plugin scores each post with a 24-check AI readiness check and builds Markdown, llms.txt and robots rules for you. For more small utilities like this one, browse the free tools page.