Word Clouds, Explained

This guide explains what a word cloud actually measures, how the layout gets built without overlaps, how word size maps to frequency, and where the picture can mislead you. After reading you will be able to make one that tells the truth.

What a word cloud is

A word cloud is a picture of a frequency count. You take a block of text, count how often each word appears, throw away the common function words, and draw what is left so that more frequent words are bigger. The result is a fast visual summary: the words that dominate a text grow large, the rest stay small.

Take a short example. The sentence "the cat sat on the mat and the cat was happy" has 11 words. Three of them are the. Once you remove function words like the, on, and, was, you are left with cat (2), sat (1), mat (1), happy (1). In the cloud, cat is drawn at roughly twice the area of the others. That is the whole idea: size encodes count.

Word clouds are decoration as much as analysis. They read at a glance, they look good on a slide, and they invite people to lean in. They are also easy to misread, which is why the rest of this article is about the numbers underneath.

When to use one, and when not to

Reach for a word cloud when the goal is a quick, informal impression: the recurring themes in open-ended survey answers, the vocabulary of a speech, the words a brand keeps repeating in its reviews. It works best when a handful of content words genuinely dominate and you want that fact to jump out.

Skip it when you need to compare counts precisely. Human eyes cannot tell a word drawn at 48 pixels from one at 52 pixels, so a cloud hides small differences. It also cannot show change over time, correlation, or anything about word order and grammar. If your question is "how much more often did X appear than Y," a bar chart or a plain frequency table answers it, and honestly so.

A word cloud shows relative frequency of single words after stop-word removal. It says nothing about sentiment, importance, or truth. A word appearing large means only that it appears often. In a set of one-star reviews the biggest word might be the product name, not a complaint.

The math: counting, then mapping to size

Two steps produce the picture. First, count. For each distinct word w you get a raw frequency f_w, the number of times it occurs after filtering. Second, map that count to a font size. The map is where clouds quietly lie or tell the truth, so it deserves attention.

The honest choice is a linear map. Let f_{\max} be the largest count, and let s_{\min} and s_{\max} be the smallest and largest font sizes you allow in pixels. Then a word's size is:

s_w = s_{\min} + (s_{\max} - s_{\min}) \cdot \frac{f_w}{f_{\max}}

Here s_w is the font size for word w, f_w its count, and f_{\max} the top count in the text. A word with half the top count gets a size halfway between s_{\min} and s_{\max}. That is the "linear" frequency mapping. (The generator on this page anchors the rarest kept word at s_{\min} by using (f_w - f_{\min})/(f_{\max} - f_{\min}) in place of f_w/f_{\max}; the shape of the argument is the same, and its default map is the square root, not linear.)

Two things complicate the reading. First, area grows as the square of the font size, so a word drawn at twice the height covers about four times the ink. Doubling the font from 24 to 48 pixels multiplies the area by 2^2 = 4. Second, some tools apply a square-root or logarithmic map to compress a long tail. A log map turns counts of 100 and 10 into sizes that differ by a fixed step, not a factor of ten. That makes the cloud legible but breaks the "twice as big means twice as often" reading you probably assumed.

How the words get placed without overlapping

Sizing is only half the job. The words still have to fit together. The classic approach, from Jonathan Feinberg's 2009 Wordle, places words one at a time from largest to smallest. Each word starts at a preferred spot, then spirals outward along an Archimedean spiral until it lands on empty space. Empty is tracked on a bit-packed occupancy map, so the collision test is fast and no two words ever touch.

Shapes work the opposite way. A heart, star, or your own silhouette is simply a region the spiral is forbidden to enter, or forbidden to leave. The search treats the outline as another obstacle, so words pack up to the edge and stop. Because placement is greedy and order-dependent, the same words can land differently on two runs. A seed number fixes the random starts so a layout is reproducible, and a share link carries that seed.

Placement position carries no meaning. Center is not "most important" and edges are not "least important." Only size (and sometimes color, if you deliberately map it to a value) encodes data. Everything else is packing.

Worked example: five counts, one map

Take five words that survive stop-word removal with these counts, and size them with the linear map described above:

Top word counts and the resulting font sizes with a linear map
WordCount f_wRatio f_w/f_{\max}Size s_w (px)Relative area
data401.00096.01.000
cloud300.750075.00.6104
text200.500054.00.3164
word100.250033.00.1182
shape40.100020.40.04516
  1. Set s_{\min} = 12 and s_{\max} = 96 pixels, a range of 84.
  2. For cloud: 12 + 84 \cdot 0.75 = 12 + 63 = 75 px.
  3. For text: 12 + 84 \cdot 0.5 = 12 + 42 = 54 px.
  4. Relative area is the size squared, normalized to the largest: (75/96)^2 = 0.6104 for cloud.

Notice the trap. cloud has 75% of the count of data, but only 61% of the ink. shape has 10% of the count and just 4.5% of the ink. The eye reads area, so a linear font map makes small words look even smaller than they are. This is a fundamental limit of the format, not a bug in any one tool.

For every word the drawn area (right bar) sits below its true count share (left bar). The gap widens for smaller words, which is why clouds understate the tail.

See how the size map changes the picture

The single most instructive control is the frequency-to-size mapping. Move it between linear, square root, and logarithmic and watch the same counts produce very different clouds.

With counts 40, 30, 20, 10, 4 and a size range of 12 to 96 px, a linear map gives sizes 96, 75, 54, 33, 20; a square-root map gives 96, 84, 71, 54, 39; a logarithmic map gives 96, 89, 80, 64, 44. Square root and log lift small words so the tail stays readable, at the cost of the "twice as big means twice as often" reading.

Reading a cloud without fooling yourself

Read the biggest three or four words as the genuine signal. They dominate the count and the difference between them is usually large enough to see. Treat the small words as a list of "also present," not as ranked data. Do not compare two medium words and conclude one beat the other by a nose; the sizing and packing hide differences that small.

Watch stop words closely. The tool removes the most frequent function words in eight languages, and it guesses the language from how many of those words appear. Mixed-language text can fool the guess, so if your German quotes leave und and der looming large, pick the language by hand. Domain words act like stop words too: in a set of pizza reviews, pizza will swamp everything. Knock it out with a click and the real vocabulary surfaces.

Finally, remember lemmatization. If you do not merge plurals, review (18) and reviews (14) are counted separately, and neither reaches the size their combined 32 deserves. Merging them changes which word is largest.

Common mistakes

Reading position as importance
Center placement is an artifact of the spiral, not a rank. Only size carries data.
Comparing two clouds by shape
Different texts, seeds, or shapes rearrange everything. Compare the underlying counts, not the pictures.
Forgetting the area effect
A word at double the font height has quadruple the ink. Your eye reads the ink, so linear font maps understate small words.
Leaving domain words in
The product name or topic word drowns the cloud. Remove it to see the rest.
Trusting a tiny text
With 30 words, one repeat swings the whole picture. Clouds are stable only when counts are large enough that a single occurrence does not dominate.

Related tools

A word cloud is the visual layer over a frequency count, so the analytical tools pair naturally with it. The Word Frequency & N-gram Analyzer gives you the exact ranked counts, plus bigrams and trigrams the cloud cannot show. For plain totals and reading time, use the Word & Character Counter, and for grade-level scores the Readability Scorer. If you suspect you overuse certain phrases in a draft, the Repeated Phrase Finder is more precise than a cloud. To compare two texts for a shared author, try the Stylometry & Authorship Comparator. And if odd invisible characters are throwing off your counts, inspect them with the Unicode Inspector.

Frequently asked questions

Why is a small word bigger than a more frequent one?

It should not be, under a single map. If it happens, check whether two-word phrases are enabled (a phrase counts as one item), whether plurals are merged, or whether you supplied the weights yourself. With a "word, weight" list you set the numbers directly, so the counts are whatever you typed.

Does color mean anything?

By default no. Palettes assign colors for looks, not data. Some layouts let you tie color to frequency deliberately, but unless you set that up, treat color as decoration and read only size.

Why does the cloud look different every time I generate it?

Placement uses random starting points, so the packing shifts run to run even with identical counts. Fix the seed number to reproduce a layout exactly, and use the share link to send that same layout to someone else.

Can I trust a cloud made from 50 words of text?

Not much. With so few words, a single repeated word can become the largest, and removing one occurrence reshapes the picture. Clouds are reliable summaries only when the top counts are well separated from the rest, which usually needs a few hundred words at least.

Why did my French text keep the word "les"?

The language guess failed, probably because the text mixed languages or was short. Choose French manually so the French stop-word list applies, and les, des, and une drop out.