Methodology

How The Panel is built.

The Panel maps captured US dog-food ingredients by how it's actually used — not what it is. Two ingredients sit close together when they show up in the same kinds of recipes, with the same companions. Distance summarizes similarity in usage; it does not establish replacement suitability.

2,776
Recipes
408
Brands
1,212
Ingredients
14
Clusters
5
Formats
01The corpus

What's captured

Ingredient panels captured from brand sites and public product pages for 408 US dog-food brands — 2,776 distinct recipes in all. Coverage is uneven: 211 of those brands are a single recipe, the median brand is one panel, and the top 41 brands hold 60% of everything captured. Each panel is parsed into its individual ingredients, which are then de-duplicated and canonicalized (e.g. "Chicken Meal" and "chicken meal, dehydrated" collapse to one ingredient id) down to 1,212 distinct mapped ingredients.

Coverage spans five formats: Kibble, Canned / Wet, Freeze/Air-Dried, Raw, Fresh.

02The method

How the map is built

Every recipe becomes a co-occurrence signal: which ingredients appear together. Ingredients are embedded from that co-occurrence pattern and projected into two dimensions (UMAP), then grouped into 14 clusters. Position is what an ingredient is used with. Color is what it is (its role — protein, mineral, oil…) or, in cluster view, which of the 14 groups it belongs to.

Position is what it's used with.
Color is what it is.

We publish a trustworthiness metric with every regeneration — currently 0.99 — a measure of how faithfully the 2D layout preserves the original high-dimensional neighborhoods.

03The caveats

Honest limits

  • Coverage is distinct recipes, not every SKU a brand sells — a brand with one mapped recipe may have a dozen more we haven't captured yet.
  • Coverage is concentrated. 211 of 408 brands are captured as a single recipe, and 195 have two or more. A thin footprint on the map usually means thin capture, not a thin product line.
  • An ingredient absent from the map may be uncaptured, too rare, or filtered from the model. Absence does not mean no brand uses it.
  • Clusters and neighbor lists are a usage-similarity shortlist, not a guarantee two ingredients are functionally swappable.
  • An ingredient is only offered as a usage neighbor once we've seen it in 5+ recipes. Below that there isn't enough co-use to place it against anything. Rarer ingredients still appear on the map and still have their own neighbor list.
The live version

See it live.

The interactive version — searchable, filterable, every brand's footprint on demand — is part of The Underbite Pro.

Access the Panel
The Panel
Methodology · Free report · Feedback