How AI meal planning works

Two separate machines run inside an AI meal plan, and confusing them is the commonest misunderstanding in the category: the target is arithmetic, the food is generated.Your daily calorie and macro target is the output of a published equation, so the same inputs always give the same number. The meals that fill that target come from a generative model, and the same inputs can give a different answer. How much you should trust any given number depends on which side of that line it sits.

"AI" means three different things here

Apps use one word for very different jobs, and marketing language blurs it:

• Generating a plan: deriving a target from your profile, then building a week of meals that hits it. Cookrange and Eat This Much sit here.
• Back-calculating the target from your logs: estimating real metabolic rate from logged intake and weight trend. That is MacroFactor's approach.
• Recognising food from an image: working out what is on the plate from a photo. Not the same job as plan generation, and far more variable in accuracy.

Seeing all three under one "AI-powered" label makes it harder to find the app that does the job you actually want. The category view is in theAI nutrition apps roundup.

Step 1: the target number - there is no AI in this part

Your daily calorie target is a calculation, not a guess. Height, weight, age and sex give a basal metabolic rate; an activity multiplier turns that into total daily energy expenditure; a deficit or surplus is applied according to your goal. The equation is usually Mifflin-St Jeor, published in 1990, and it returns the same number in every app given the same inputs.

The macro split is arithmetic too: protein, carbohydrate and fat in grams convert to calories with the Atwater factors (protein and carbohydrate 4 kcal/g, fat 9 kcal/g). You can check it yourself - the site's calorie calculator andBMR calculator run the same equation and show you the numbers going in.

So: at this step no app's "AI" is smarter than another's. The difference is which version of the equation and which activity multipliers it uses - a choice, not an intelligence gap.

Step 2: the meals - this is where generation happens

Here is the hard part. Around 28 meals for a week, each with calories and macros that land on the daily target, without repeating themselves, and cookable by an actual person. On top of that sit allergens, ingredients you dislike, and kitchen reality.

If a language model does this step, one consequence follows naturally:the same inputs produce different plans on two separate runs. That is not a bug, it is what generation means. But it has a cost worth stating: you lose reproducibility. Ask for the same plan again and you get something similar, not the same thing. If you like a plan, saving it is safer than expecting to regenerate it.

Step 3: where the numbers come from - and the weakest link

A meal's calorie figure can arrive by three routes, and they are not equally reliable:

• A database record. A packaged product's barcode is looked up. Cookrange uses Open Food Facts for this;coverage and its gaps are on their own page. Where a record exists this is the most reliable route.
• The label. The manufacturer's declaration. Where a label and a database disagree, the label wins.
• A model estimate. If a generated recipe's macros were not computed from its ingredients, they were estimated. This is the weakest link in the chain and the least-discussed thing in the category. When a plate says "about 520 kcal", that "about" sometimes hides a wide range.

Practical consequence: you can work with the daily total against a weight goal, but you should not read a single generated meal's calorie figure as a measurement.

The same uncertainty exists in photo analysis, with one difference: there the estimate carries a confidence score, and that score is shown to you.Where photo calorie estimation fails is on its own page.

What a plan cannot know

Nothing that was not given as an input can reach the plan. That is a boundary rather than a flaw, but the boundary is worth knowing: your blood panel, your medication, pregnancy, a diagnosed condition, your budget, your cooking skill, what was in the fridge that week, and whether you actually had the energy to cook that evening. For some of those the right counterparty is not an app;who should not use these tools at all is a separate page.

Five things to check before trusting one

1. How was the target derived? Which equation, which activity multiplier. If it will not say, ask.
2. Where does the nutrition data come from? A named database, or an estimate.
3. Does the allergen filter run before generation or after? A filter that runs afterwards only removes a meal that was already created.
4. Does the target move on its own during the day? If it does, knowing what you are progressing against gets harder - see net calories.
5. Does it state in writing that it is not medical advice? If not, treat that as a warning sign.

How these steps work in Cookrange

As a worked example: the target is arithmetic, allergen and preference filters applyat generation time (a meal that breaks them is never created), the weekly plan does not regenerate itself, and you can change a single meal or a single snack on its own. The step-by-step account is on theplan generation page.

One more limit that ought to be stated plainly: generation is capped at2 runs a day on the free tier (plan, recipe generation and chat share the same quota). The app is at v0.9.6 internal alpha and is not in the stores.