How accurate is NutriShot AI?

We measured it instead of guessing. We logged 412 real meals, weighed every ingredient on a kitchen scale, and compared NutriShot AI's numbers to ground truth from USDA FoodData Central, across 5 food categories and 7 cuisines.

The results

• 90% average calorie accuracy once you confirm the portion size: within 10% of the scale, roughly 60 calories on a 600-calorie meal
• 71% of meals landed within 10% of the true calories; 91% within 20%
• From the photo alone, without confirming the portion, estimates landed within 11% on average

That gap between 11% and 10% is why the portion confirmation question exists: answering it improves accuracy. Macros hold up too: protein within 6.7%, carbs within 5.9%, and fat within 9.5% (oils and dressings are hard to see in a photo).

It varies by food type

Packaged and branded foods score best at within 4.8%, simple home cooking lands within 7.8%, and drinks within 8.1%. Restaurant and takeout plates land closer to 15%, with sauce-heavy cuisines like Thai (16.4%) and Indian (17.6%) the toughest.

How to read this honestly

We ran and funded this test ourselves, so it isn't an independent study. That's why we publish the raw data and the analysis code, so you can check our math instead of taking our word.

• See the full accuracy results and data