Titles are split into words, with two-word names kept whole ("small batch", "cream cheese", "brown butter") and spellings folded together ("crisp" and "crispy"). The dish's own name, words like "the" and "recipe", and any word in half the titles or more are left out. A word needs at least 10 recipes, and there must be 10 without it. English titles only, so a word cannot stand in for a language.
A difference is shown when the medians sit at least a third of the middle half of all these recipes apart (about half a standard deviation, a medium effect) and a rank test puts it at 3 standard errors or more. For an ingredient many recipes skip, the share using none counts too: 25 points apart at the same 3 standard errors. With dozens of words and every ingredient tested, about one chance difference in 370 comparisons would pass, so a lab may show one that is luck; a finding that repeats across labs is unlikely to be.
Shares are rounded down. Each card links two recipes that use the word, nearest its own medians.