Search the same food twice and you may get different numbers. That is not a bug, and understanding why is the difference between a food diary you can act on and one you quietly stop believing.
Nothing in this app is a measurement. Every figure is an estimate, and the app is deliberate about telling you where each one came from and how sure it is.
Where the numbers come from
A lookup checks several sources in turn and shows you which ones it used. Each is strong somewhere and weak somewhere else.
- Your saved foods. Things you have logged or corrected before. The most trustworthy figures in the app, because you checked them.
- Open Food Facts. A public database of packaged products, contributed by volunteers. Strong on branded goods with a barcode. Weak on anything loose, fresh or home-cooked, and occasionally out of date when a manufacturer reformulates.
- AI estimates. A model reasoning from what it knows about the food. Reasonable on composition, for example that chicken breast is mostly protein, and least reliable on quantity.
- Your own history and patterns. What you have eaten before, used to rank results and drive suggestions.
The confidence badges
Confidence describes how well the sources agreed with each other, not how correct the answer is. Two sources can agree and both be wrong. Disagreement, though, is a reliable signal that something needs your eye.
- High confidence: cross-checked. More than one source, and they matched. Log it and move on.
- Moderate confidence: usable, worth a glance at the serving.
- Low confidence: sources disagree. Check it against the package if you have one, and correct it if you can.
Two specific warnings go further. "Sources disagree on this food's serving size" means the quantity is the uncertain part. Switch to grams and enter a real amount. "AI and the web returned different figures for this food" means treat the whole result as a rough guide.
How AI gets it wrong
AI models fill gaps with confident guesses, known as hallucinations. A model asked about a product it has never encountered will rarely say so. It will produce a plausible nutrition panel instead, formatted exactly like a correct one. Nothing in the presentation separates a figure the model knows from one it invented, which is why the app names its sources and shows its confidence rather than handing you a bare number.
The onboarding asks you to confirm that you will check what the AI returns. That is not boilerplate, it is the actual arrangement.
The biggest source of error is not the app
It is portion size. A database entry that is 10% out matters far less than a serving that is 50% out, and portions are easy to misjudge by that much. Weighing the things you eat most often will improve your diary more than any amount of checking sources.
How to read your totals
Use them for comparison, not for accounting. A day logged at 2,100 kcal is an estimate rather than a measured 2,100 kcal. If you log consistently, though, this week against last week is a real signal. Chasing a daily number to the calorie is chasing a level of detail the data does not have.
What genuinely improves your diary over time is correcting foods as you go. A corrected food is saved as yours and comes back first next time, so your regular meals get closer to what you actually eat while everything else stays an estimate.