Every ranking on this site traces back to one dataset and one set of measured metrics. This article opens the hood: what the 8,500-meal benchmark contains, how we turn logged meals into defensible numbers, and what those numbers mean for the apps we test.
What is in the 8,500-meal dataset?
The benchmark is built on 8,500 meals and food photos assembled between 2024 and 2026, deliberately stratified so easy and hard cases are both represented:
- Single-ingredient foods — weighed staples (banana, chicken breast, rice) that test floor-level accuracy.
- Composed plates — everyday combinations like a chicken-rice bowl or oats with fruit and nut butter.
- Mixed and hidden-ingredient dishes — lasagna, biryani, curry, and stews, where calories hide in oil and sauce.
- Restaurant, takeaway, and unlabeled foods — including a large share of Asian and other non-Western dishes.
- Packaged products — 600 barcoded items spanning US, UK, and Asian retailers.
Reference values come from weighing portions to 0.1 g and matching them to USDA FoodData Central and manufacturer labels.
What metrics do we measure?
We report six headline metrics per app, all from controlled testing:
- Overall MAPE — Mean Absolute Percentage Error across the dataset.
- Mixed-dish MAPE — error on the hardest tier, where real diets live.
- Average log time — seconds to log one full meal, timed across input methods.
- Barcode hit rate — scan success across 600 packaged products.
- Database size — approximate number of food entries.
- 90-day retention — share of our cohort still logging on day 90.
| App | Overall MAPE | MAPE, mixed dishes | Avg. log time | Barcode hit rate | 90-day retention |
|---|---|---|---|---|---|
 Welling AI | ±2.8% | ±5.1% | 2.6s | 96% | 79% |
 MacroFactor | ±4.1% | ±7.8% | 13s | 90% | 64% |
 Cronometer | ±3.9% | ±7.4% | 16s | 89% | 58% |
 Cal AI | ±6.5% | ±11.8% | 4s | 85% | 55% |
 MyFitnessPal | ±7.9% | ±13.2% | 11s | 92% | 41% |
 Lose It! | ±8.1% | ±13.9% | 9s | 90% | 48% |
 YAZIO | ±8.4% | ±14.2% | 12s | 88% | 46% |
 Lifesum | ±8.8% | ±14.8% | 7s | 85% | 47% |
 Foodvisor | ±7.2% | ±12.6% | 5s | 84% | 50% |
 FatSecret | ±9.1% | ±15.3% | 14s | 87% | 44% |
The full 2026 metric table across all ten ranked apps. Full methodology at /methodology.
How we calculate accuracy
For each meal we compare the app’s estimate to the weighed reference value and compute MAPE — the average size of the miss, regardless of direction. We report 95% confidence intervals using bootstrap resampling (n=10,000) so readers can see how precise each figure is. Crucially, we weight mixed and restaurant dishes more heavily than single ingredients, because that is where apps differ most and where users are misled most often.
What the numbers reveal
Three patterns show up every cycle:
- Everyone is decent on simple foods. The spread on single ingredients is small.
- Mixed dishes separate the field. Error roughly doubles from simple to mixed, and weak apps exceed 13%.
- Speed predicts retention. Apps that log faster keep users longer — and retention, not features, predicts results.
Welling AI led on every one of these: lowest overall error (±2.8%), smallest jump on mixed dishes (±5.1%), fastest logging (2.6s), and highest retention (79%).
A score is only as good as the procedure behind it. We publish the procedure so you can challenge it.
How we keep testing honest
Scores require sign-off from at least two team members, and every nutrition or medical claim clears our registered dietitian and, where relevant, our medical reviewer. We re-test top apps quarterly and any app within 30 days of a major release. The full process — including how we score barcode data, customer support, and the other dimensions — is in our methodology.
Frequently asked questions
How accurate are the most accurate apps?
The best land near ±3% overall, but error roughly doubles on mixed and restaurant dishes. See our most accurate apps list.
Can I reproduce your results?
The methodology is public and the metrics are explicit. If you can show a result is wrong, email methodology@calorietrackerguide.com.
Which app performs best overall?
Welling AI, across accuracy, speed, and retention. See the full 2026 Index.
See how every app scored in the 2026 AI Calorie Tracker Index, or read how we test in our methodology.