Where calorie data actually comes from (and why two apps disagree about your sandwich)

You log a turkey sandwich. One app says 320 calories, another says 410, a third says 580. Same sandwich, same words. The temptation is to assume one app is "right" and the others are broken — but that's not what's happening. Each number is sourced from a different place, built by different people for different reasons, with different levels of scientific rigor behind it. Understanding those sources is the difference between trusting a number blindly and knowing when to trust it.

There are essentially three origins for any calorie number you'll ever see in a tracking app: a government laboratory, a global crowdsourced barcode database, and an AI estimate. They have very different DNA.

Source 1: USDA FoodData Central — the laboratory

The closest thing to ground truth for whole foods is USDA FoodData Central (FDC), the U.S. government's food composition system. When you look up "raw banana" or "chicken breast, roasted," the underlying numbers often trace back to actual chemical analysis of food samples in a lab.

FDC isn't one database — it's five data types of varying rigor, and the distinction matters enormously:

For an unpackaged whole food — an egg, an apple, a sweet potato — Foundation Foods and SR Legacy are about as good as nutrition data gets. They reflect real measurement, they report units consistently, and they even capture natural variability between samples. This is why a serious tracker prefers USDA whole-food rows over a random branded entry when it can match one: the underlying analysis is more trustworthy than a manufacturer's rounded label or a crowd guess.

One subtlety hides inside even the lab data: calories themselves are usually calculated, not directly burned in a calorimeter. FDC applies Atwater factors — roughly 4 calories per gram of protein, 9 per gram of fat, 4 per gram of carbohydrate — to the measured macronutrients (with food-specific refinements where available). So even the gold standard is a model. A well-validated, decades-tested model, but a model — which is part of why no two sources land on exactly the same number.

Source 2: Open Food Facts — the crowdsourced barcode

USDA is excellent for whole foods and weaker for the packaged, branded, international universe of products people actually scan. That's where Open Food Facts (OFF) comes in: a free, collaborative database of over 2.8 million packaged products from 150+ countries, built Wikipedia-style by volunteers photographing and entering labels.

OFF's strength is coverage. That obscure German muesli or the regional energy drink that USDA has never heard of probably has an OFF entry, because someone in that country scanned it. For barcode-driven logging of branded products, nothing matches its breadth.

Its weakness is the flip side of the same coin: it's crowdsourced, so quality is uneven. By OFF's own account the data is community-contributed and may contain errors or be incomplete — some nutrients carry 10–40% missing values, entries can be outdated, and a typo in a serving size can throw a number wildly off. The project runs moderation and increasingly uses AI to extract and cross-check label values, but the floor is lower and more variable than a government lab. An OFF entry can be perfect or it can be a fat-fingered guess, and from the outside they look identical.

So already you can see why two apps disagree: one matched your sandwich to a curated USDA whole-food composite, the other matched a branded OFF entry someone uploaded with a misread serving size.

Source 3: AI estimates — the educated guess

The third source is the newest: a language or vision model estimating calories from your text ("turkey sandwich on sourdough") or a photo. There's no database lookup at all here — the model is producing a number from patterns in its training data.

AI estimates are extraordinary for friction. They handle the messy, unbarcoded, homemade reality of real meals — "leftover stir-fry, about two cups" — that neither USDA nor OFF can match cleanly. But as we covered in depth in how accurate photo calorie counters really are, they inherit a specific weakness: identifying what the food is has gotten very good, while estimating how much of it there is remains hard. A model can nail "this is carbonara" and still be 30% off on the portion, because volume is a 3D problem captured from a 2D image, and hidden fats and sauces are functionally invisible. The number is plausible, not measured — and plausible is exactly the failure mode that feels trustworthy while being wrong.

Why they disagree — the four sources of divergence

Put the three origins together and the disagreements between apps fall into predictable buckets:

1. Different underlying source. A lab composite, a manufacturer's label, and a model's guess will rarely converge on the same number for the same food. This is the biggest driver.
2. Portion and serving-size mismatch. Most divergence isn't about calories-per-gram at all — it's about how many grams the app assumed. "One sandwich" is not a unit; the gram weight behind it varies hugely between entries.
3. Recipe and preparation ambiguity. Was the chicken roasted dry or pan-fried in oil? Is the mashed potato 30% butter? The same named dish spans a wide caloric range, and each source picks a different point on it.
4. The legal margin baked into labels. Even a "perfect" packaged number isn't exact: the FDA permits roughly a 20% margin of error on declared values. The label on the box is itself a regulated estimate.

The honest takeaway is that no single source is right for everything. USDA wins on whole foods. OFF wins on branded coverage. AI wins on messy real-world meals nothing else can match. A tracker that uses only one of them is strong in one column and weak in the other two.

How CalBurndown grounds the number: label, then USDA, then AI

The accuracy strategy that follows from all this isn't "pick the best database" — it's "use the most trustworthy source available for this particular food, and fall back gracefully." CalBurndown resolves every logged item through an explicit precedence:

Package label → USDA database → AI estimate.

In practice that means:

• If there's a real package label, that wins. A manufacturer's measured Nutrition Facts panel for the specific product you ate beats any database lookup or model guess. Label-sourced values are treated as ground truth and aren't overridden.
• If there's no label but the food cleanly matches a USDA whole-food row, the calories and macros are sourced from USDA × your portion rather than from the AI's direct guess. The model still does the job it's good at — recognizing the food and estimating grams — and then the code scales the trustworthy USDA composition to that portion. The AI estimates the size; the lab data sets the density. This override is deliberately gated to USDA-source matches, because the noisier branded and crowdsourced entries don't earn the right to overrule the calorie number.
• If neither exists — a homemade or unmatched dish — the AI estimate stands, because a plausible model guess beats no number at all.

That ordering is the whole philosophy in one line: prefer measurement to lookup, prefer a curated lookup to a guess, and only guess when there's nothing better. It's why grounding an estimate in USDA improves accuracy without sacrificing the speed of just snapping a photo — you get the AI's frictionless capture and the lab's density, instead of having to choose.

What this means for you

A few practical consequences fall out of knowing where the numbers come from:

• Don't treat any single number as exact. Even the best source is a model with a margin. Consistency over time matters more than the precision of any one entry — a small, steady bias still reveals your trend.
• Lean on labels when precision counts. If you're breaking a plateau or in a high-stakes cut, scanning the actual package beats any estimate, because it sits at the top of the precedence for a reason.
• Add detail to AI logs. "Two tablespoons of olive oil" or "80/20 ground beef" removes the single biggest blind spot in a model's guess — the invisible fats it can't see. The meal logging guide covers how to capture that context as you go.
• Expect whole foods and branded foods to behave differently. The banana you matched to USDA is on firmer ground than the artisanal granola you matched to a crowdsourced barcode. Knowing which source you're standing on tells you how much to trust the step.

The bottom line

The calorie number on your screen isn't handed down from a single authority — it's assembled from a government lab, a global crowd, or an AI's pattern-matching, each with its own rigor and its own blind spots. Two apps disagree about your sandwich because they reached into different sources and made different portion assumptions, not because one is broken.

The fix isn't to find the one true database; it's to use the right source for each food and fall back sensibly when the best one isn't available. Label first, USDA second, AI last — measurement over lookup, lookup over guess. That precedence won't make any single number perfect, because nothing can. It just makes the number you log the most defensible one available for the food actually on your plate.

Citations

• USDA FoodData Central — About Us. The integrated U.S. food composition data system.
• USDA FoodData Central — Data Documentation and Data Type Comparison. Foundation Foods, SR Legacy, FNDDS, Branded, and Experimental data types, including Atwater energy calculation.
• Open Food Facts and the Open Food Facts database. Crowdsourced, multi-country packaged-product database; coverage and data-quality characteristics.
• Zheng, J., Wang, J., Shen, J., & An, R. (2024). "Artificial Intelligence Applications to Measure Food and Nutrient Intakes: Scoping Review." J Med Internet Res 26:e54557. AI identification vs. portion-estimation accuracy.
• U.S. Food and Drug Administration, 21 CFR 101.9. Nutrition labeling of food, including the permitted compliance margin on declared values.