A word cloud is a frequency chart that has given up precision in exchange for being glanceable. Size maps to how often a word appears, position and colour carry no meaning at all, and the whole thing is drawn on an eight hundred by six hundred canvas you can export as an image.
Two things happen to your text before anything is drawn. Common function words — the, and, is and their relatives — are removed, because otherwise they would dominate every cloud ever made. Anything of two characters or fewer goes too. What survives is the vocabulary that actually distinguishes your text from any other text.
There is also a weighted mode, which is the more interesting one for planning. Instead of pasting prose and letting the tool count, you supply words with their own weights in a word and value format, so the picture reflects the emphasis you intend rather than the emphasis you happened to type.
A built-in list of common English function words, plus anything two characters or shorter. Without that filter every cloud would be dominated by the same handful of words and would tell you nothing about the text. It does mean the tool is tuned for English; another language keeps its function words and they will swamp the result.
No. Only size carries information. Placement is decided by what fits inside the chosen shape without overlapping, and colour is assigned from the scheme you picked rather than from the data. Two words sitting next to each other are neighbours by accident, not because they are related.
Supplying your own numbers instead of having the tool count. You give each word a value and it is drawn at that size, which is useful for showing intended emphasis — a planned chapter balance, a set of themes, a survey result you already have totals for — rather than measuring text you have written.