Our own accuracy test · updated June 2026
How Accurate Is NutriShot AI?
We weighed every ingredient of 412 real meals on a kitchen scale, looked up the true calories in the U.S. government's food database, and compared NutriShot AI's estimates against them. In June 2026 we upgraded the AI model and re-ran all 412 meals. Every number below is from that re-test. The raw data and the script that computes it are linked below, so you can check every number yourself.
What this is, and what it isn't
We designed, ran, and paid for this test ourselves. We're the company behind NutriShot AI (Questopia LLC). No outside lab was involved, no independent scientists have reviewed the results, and we didn't lock in our test plan publicly beforehand the way formal studies do. That's exactly why everything is published below: the full meal-by-meal data and the code, so anyone can check our work.
The complete writeup of our original test, including the method, is published at The Pantry Notes: NutriShot AI accuracy test (this page uses the same method and shows the latest results).
The headline numbers
Why we show two numbers
Snap a photo and NutriShot AI estimates your meal. From the photo alone, estimates land within 11% of the scale on average. The app then asks one quick question: does the portion size look right? That few-second answer reduces the average gap to 10%. So whenever we say 10%, we mean the estimate after you've answered that question.
| Calorie accuracy | Photo alone | After confirming portion |
|---|---|---|
| Average distance from the scale | 11.0% | 10% |
| Meals within 10% of the true calories | 63% | 71% |
| Meals within 20% | 88% | 91% |
Protein, carbs, and fat
We also checked protein, carbs, and fat. Oils and dressings are hard to see in a photo, which can affect fat estimates. After confirming the portion size:
| Nutrient | Average distance from the scale | Within 10% of the truth |
|---|---|---|
| Protein | 6.7% | 87 out of 100 meals |
| Carbs | 5.9% | 85 out of 100 |
| Fat | 9.5% | 79 out of 100 |
Where it's strong, and where it isn't
Easiest for the AI
- Packaged and branded foods4.8% from the scale
- Simple home-cooked dishes7.8%
- Drinks8.1%
Hardest for the AI
- Restaurant and takeout meals15.0% from the scale
- Thai dishes16.4%
- Indian dishes17.6%
All figures show the average distance between the calorie estimate and the kitchen scale. Restaurant dishes often contain oil, butter, and sugar in sauces that a photo doesn't show, and the numbers reflect that. If most of your meals are takeout or come with a lot of sauce, expect an average gap around 15-18%.
How we ran the test
- We cooked or bought 412 meals and weighed every ingredient on a gram scale.
- For each weighed ingredient, we looked up the true calories and nutrients in USDA FoodData Central, the U.S. government's public food database.
- The meals covered packaged foods, home cooking, drinks, mixed plates, and restaurant takeout, across 7 cuisines.
- We logged each meal with NutriShot AI and wrote down two numbers: the estimate from the photo alone, and the estimate after confirming the portion size.
- The accuracy figures on this page are the average percentage gap between NutriShot AI's numbers and the scale's.
What to keep in mind
- We ran and paid for this test ourselves. Nobody independent has checked it.
- We picked the meals, and we didn't publicly lock in the test plan beforehand the way formal studies do.
- The "true" values are only as good as the government food database they come from.
- Results reflect the app as tested in June 2026. Updates can move accuracy in either direction.
Check our math
One script recalculates every number on this page from the raw data. Both files are free to download:
Put both files in the same folder and run:
python3 analyze-validation.pyThe complete writeup, with the full method and every table: NutriShot AI accuracy test, the full writeup (The Pantry Notes)
What this is good for, and what it isn't
Landing within 10% on average is plenty for everyday tracking: spotting trends, staying near a calorie target, keeping your protein up. But NutriShot AI is not a medical tool. If you need exact numbers for medical reasons (insulin dosing, or a diet supervised by a doctor), use the methods your doctor recommends, not photo estimates.
Common questions
How accurate is NutriShot AI's calorie count?
We weighed 412 real meals and compared NutriShot AI's estimates to the scale. Once the portion size was confirmed in the app, calorie estimates landed within 10% of the scale on average (about 90% accuracy), and within 11% from the photo alone. 71 out of 100 meals came within 10% of the true calories, and 91 out of 100 came within 20%.
What does "within 10%" mean in practice?
On a 600-calorie meal, 10% is about 60 calories. Over a full 2,000-calorie day, it adds up to roughly 200 calories. That's small enough to spot trends and stay near a goal, too big for medical decisions that need exact numbers.
When was this test run?
These results are from June 2026, when we re-ran the same 412 meals after a model update. The average gap was 10%. The downloadable data below is from that test, and the original test's full writeup stays linked above.
Was this an independent study?
No. We designed, ran, and paid for it ourselves, and no outside scientists have reviewed it. That's exactly why we publish the raw data and the code, so you can check the numbers without taking our word for it.
Where is NutriShot AI least accurate?
Restaurant and takeout meals (15.0% from the scale on average) and dishes where most of the fat is in the sauce. Indian (17.6%) and Thai (16.4%) dishes were the hardest in our test. Packaged foods (4.8%) and simple home cooking (7.8%) were the easiest.
Is NutriShot AI more accurate than other calorie apps?
We don't know. We only tested NutriShot AI, so this test says nothing about other apps and we make no comparison claims. If another app publishes its own test data, you can line the numbers up yourself.
Try it on your own plate
Free forever: daily AI meal scans and a coaching insight. Weigh a meal yourself and put us to the test.
