Methodology: how our numbers are built

The short answer

• Three kinds of number appear on this site: published measurements, agency guidance, and our own derived estimates. Each is labeled.
• Preference scores are a five-point condensation of published species accounts and feeding research, scored conservatively where sources disagree.
• Kernel yield is the softest input in our cost analysis and the one that moves the results most.
• Everything is versioned and published as JSON so the analysis can be checked rather than taken on trust.

This page exists so that any number on this site can be traced back to where it came from. It covers the three datasets that most of our analysis rests on, the derived figures we calculate from them, and the places where our confidence is lower than the tidiness of a table suggests.

Three kinds of number#

Published measurements. Energy, fat and protein values for seeds and kernels come from standard food composition references. These are stable and we do not adjust them.

Agency guidance. Cleaning intervals, disinfectant dilutions, nectar ratios and disease responses follow current guidance from wildlife agencies and long-running citizen science programs. Where guidance differs between agencies, we say so on the page rather than picking one silently.

Our derived estimates. Kernel yield, cost per pound eaten, calories per dollar, preference scores, feeder attribute scores and the cleaning-interval model are ours. They are computed from the inputs above using the formulas below, and they are labeled as our figures wherever they appear.

The preference matrix#

Thirty-four species, thirteen foods, scored 0 to 4.

How scores were assigned. For each species we worked from published species accounts (diet, foraging guild, bill morphology), from feeding-preference research where it exists for that food, and from documented feeder behavior. Foods with strong, well-documented associations (nyjer and goldfinches, safflower and cardinals) score 4 with high confidence. Middle scores carry more uncertainty than the crispness of a single digit implies.

Where we scored conservatively. When sources disagreed, or when a preference appeared to be regional or individual rather than general, we assigned the lower score. A 2 in this matrix should be read as "do not buy this food for this bird", not as an endorsement.

Known weaknesses. The matrix does not encode consumption volume, seasonal shifts in preference, regional learning, or competitive displacement. A species that would rate a food highly in isolation may never reach it at a crowded feeder. We consider adding a volume dimension the single largest improvement available to this dataset.

Kernel yield, and why it matters most#

Kernel yield is the fraction of purchased weight that a bird can actually swallow. It is the input that moves our cost analysis more than any other, and it is the one with the widest real variation.

These are midpoints of published ranges, not measurements of a specific lot. Real lots vary with cultivar, growing season, grade and how the lot was cleaned, and a poor lot can fall several points below these figures. Safflower is the least certain: hull share across cultivars spans roughly 18 to 59%, so a safflower figure should be read as an order of magnitude rather than a precise number.

These values were revised upward in version 2026.2 of the dataset. An earlier version used lower kernel yields drawn from bird-feeding convention (60% for black oil sunflower) rather than from the agronomic literature, which overstated hull waste and therefore overstated the corrected cost of every hulled seed. The change is logged on the corrections page.

Sensitivity. At benchmark prices, black oil sunflower at 75% kernel costs about $1.80 per pound eaten. Take it to the pessimistic end of the published range, 70%, and it becomes $1.93. Take it to the optimistic end, 80%, and it becomes $1.69. None of those change its position against the other foods, which is the useful test: the conclusion is robust even though the input is not precise. We flag this because a single tidy percentage in a table can look more certain than it is.

Derived formulas#

All of these are computed, not looked up. You can reproduce every published figure from the dataset with them.

cost per pound eaten   = bag price per pound ÷ kernel yield
kcal per pound of food = kcal per 100 g of kernel × 4.536 × kernel yield
kcal per dollar        = kcal per pound of food ÷ bag price per pound

Benchmark prices are 2026 United States large-bag (roughly 20 lb) retail rates, stated on the cost per calorie page and used only as defaults. Every tool accepts your own prices.

The cleaning interval model#

The cleaning schedule builder starts from a per-design baseline drawn from agency guidance, then applies three multipliers:

• Temperature: 0.5× at 85 °F and above, 0.75× from 70 to 84 °F, 1.5× below 40 °F
• Moisture: 0.6× in humid or rainy conditions, 1.2× in dry conditions
• Traffic: 0.75× for heavy use, 1.25× for light use

This is a model, not a measurement, and it is deliberately conservative. It exists because a single published interval cannot be right for a nectar feeder in Phoenix in July and a suet cage in Minnesota in January.

Squirrel geometry figures#

The 4 ft vertical and 8 to 10 ft horizontal jump distances used throughout the squirrel and baffle pages are the widely used working figures for eastern gray squirrels. They describe typical capability rather than a maximum: individual animals, downhill launches and hard frozen ground all shift the numbers, and our recommended clearances include margin for that.

Versioning and corrections#

The dataset carries a version string (currently 2026.3) and is published at /datasets/bird-food-matrix.json under a Creative Commons Attribution 4.0 license. Substantive changes to a page update its displayed date. Factual errors are logged and dated on the corrections page rather than edited away silently. If something here is wrong, tell us and it will be fixed and recorded.

Frequently asked questions#

Are the preference scores measured or judged?

Judged, from published sources, and we say so plainly. There is no single dataset that scores every common feeder bird against every common food on one scale, so building the matrix required condensing species accounts, feeding-preference research and agency guidance into a five-point scale. The judgement is in the condensation. Where sources disagreed we scored one point lower rather than higher.

Why do you publish prices that will go out of date?

Because a cost analysis with no prices in it is useless, and because the conclusions do not depend on the prices being current. The corrections that matter, kernel yield and energy density, are properties of the seed. The prices are benchmarks that make the arithmetic concrete, they are dated, and every tool lets you substitute your own.

How often is the data updated?

The matrix carries a version number and is revised whenever a correction is accepted or a source changes its guidance. Page-level update dates reflect substantive revisions to the text, not cosmetic edits. Corrections are logged publicly.

Sources and further reading

1. FoodData Central US Department of Agriculture Energy, fat and protein reference values
2. Project FeederWatch Cornell Lab of Ornithology Species accounts, feeder types and food preferences
3. All About Birds Cornell Lab of Ornithology Diet, range and behavior for species in the matrix
4. Sunflower hulls and sunflower screenings Feedipedia (INRAE, CIRAD, AFZ and FAO) Hull share of whole sunflower seed
5. Safflower seeds and oil meal Feedipedia (INRAE, CIRAD, AFZ and FAO) Hull share and its variation across safflower cultivars
6. Niger (Guizotia abyssinica) Feedipedia (INRAE, CIRAD, AFZ and FAO) Hull share of niger (nyjer) seed