When does it go in?

garlic sits at a mean position of 0.28 across 1,302 recipes — 56% of its mentions fall in the first third of a recipe and 6% in the last. onion sits at 0.35.

Pooling hides things. Cinnamon has a mean of 0.28 in mains and 0.48 in desserts — early in a savory pot, where it is bloomed in fat, and later in a sweet one. Butter runs 0.25 in desserts, where it is creamed at the very start, and 0.47 in mains, where it is often the finish. Egg has the widest spread of any ingredient here, 0.21 to 0.80 depending on what you are making.

One honest caution about the garlic-and-onion argument, since that is the one people wanted settled. This tool shows when each ingredient is mentioned, pooled over recipes that may contain only one of them. That is a different question from which goes in first when a recipe uses both — for that, see the order of operations, where of 785 recipes naming both, 38.5% add garlic after onion, 39.4% add them together and 22.2% start with garlic.

Everything here is a statement about the recipes in this corpus, grouped by the cuisine written on the recipe — not a claim about a cuisine itself. With that said, the split surfaces real structure:

Onion is the clearest, and the sturdiest: a mean of 0.28 across 138 Mediterranean recipes — the base of the pot — against 0.52 across 77 Southeast Asian ones, where 41% of its mentions land in the final third. That gap holds whichever slice of the corpus you look at on its own, which is the test that matters and which most cuisine differences here do not pass.

Soy sauce is a null, and a sturdy one: it goes in early everywhere it clears the floor — 0.27 East Asian (n=122), 0.29 Southeast Asian (n=53), 0.31 American (n=74) — and that holds slice by slice too. We expected the East Asian curve to sit earlier than the rest. It doesn’t.

“Herbs go in later in Mediterranean cooking” looks backwards here, though more weakly than it first appeared: basil averages 0.66 in Mediterranean recipes (n=33) against 0.89 in American ones (n=52), and parsley 0.65 in Middle East & North Africa (n=57) against 0.85 in American recipes (n=123). Compare only recipes written in the same batch and the basil gap narrows from 0.23 to 0.13 — real, but half the size the pooled number suggests.

And one we had to withdraw. An earlier version of this page reported that the “bloom your spices early” pattern was not strongest in South Asian recipes. It is not a finding — 49 of those 56 recipes came from a single authoring batch weighted toward restaurant-style curries, which build a sauce rather than temper whole spice in hot fat at step one. The number measured that batch, not South Asian cooking. We can’t answer the question with this corpus yet, and saying so is the honest version.

• A mention is not an addition. “Remove the bay leaf” counts as a bay-leaf mention, and a step that says “season to taste” without naming salt does not count as salt. Step text is a proxy for practice.
• Matching is by word, not by understanding. Terms are drawn from the corpus’s own ingredient names and matched with a plural-tolerant word-boundary search, so “pepper” catches both the spice and the vegetable, and “oil” catches every kind at once. Where a more specific term exists (“olive oil”), prefer it.
• Floors are enforced, not faded. An ingredient needs 25 recipes to appear at all, and a kind of dish or a cuisine needs 15 to get its own curve — counted on the exact ingredient-and-cuisine cell, never on either side alone. Anything below is omitted rather than drawn faint — a curve from nine recipes is noise wearing a shape.
• Cuisine and authorship are tangled together here, and that limits what a cuisine curve can mean. This corpus was assembled in batches, and the batches did not cover cuisines evenly — some cuisines are almost entirely from one batch, others from another. So a difference between two cuisine curves is partly a difference between the people and processes that wrote those recipes. Where we state a cuisine finding above, we checked it survives inside a single batch; one earlier claim did not, and was withdrawn rather than reworded. Treat any cuisine comparison you make with the tool yourself as a hypothesis, not a result.
• Cuisine labels describe the recipe as written, not the tradition. A “Thai” recipe in an American cookbook is labelled Thai; that is a fact about this corpus, not about Thailand. Raw labels are grouped through a hand-curated taxonomy (95 labels → 50 cuisines → 11 families) because the raw field has regional labels sitting beside their parents. And a difference between groups is a description of these recipes, not a cause — “garlic goes in earlier in Latin American recipes” means earlier in the 104 such recipes here, nothing more.
• Position is by step count, not by the clock. Step 3 of 4 might be “simmer for three hours”. Under half of steps state a duration, so the count is what there is.
• ~40% of the corpus was written for this project and may state sequence more explicitly than recipes in the wild. Its cuisine distribution is not the world’s, either — American-family recipes are the largest group.

Every number on this page is a deterministic measurement over the same recipe corpus — nothing here was estimated or embellished. Want to check the work, or run your own analysis? A 250-dish sample of the corpus — ingredients, 18-note flavor profiles, timing, macros — is free to download.