Every AI calorie tracker is really two systems. One reads your sentence and works out what you ate. The other looks up what that weighs and what it contains.

The second part is mostly solved. The first part is entirely up to you.

"Dal" could be 90 calories or 400 depending on the dal, the katori and the tadka. The model is not guessing badly — it is guessing from a sentence that genuinely does not contain the answer. Add nine words and the same estimate lands within about 10%, which is all the accuracy anyone actually needs.

This is the skill nobody teaches, so here it is.

The five details that move the number

Ranked by how much difference each one makes. If you only ever add one, add the first.

1. Cooking fat

The single biggest variable in home cooking, and the one people leave out most.

A teaspoon of oil or ghee is roughly 40 calories. A dal cooked with a restaurant-style tadka can carry three or four of them. Two identical-looking sabzis, one cooked dry and one cooked properly, differ by 150 calories or more.

Vague — "aloo sabzi"

Better — "aloo sabzi, cooked in about a tablespoon of oil"

You do not need to measure it. "Lightly cooked", "normal home tadka" and "restaurant style, quite oily" are three different numbers and the model treats them as such.

2. A countable anchor

Give it one thing it can count. Rotis, idlis, eggs, slices, katoris, cups.

Counted objects are the most reliable input you can provide, because they collapse the portion question entirely. "Rice" is a guess. "One cup of cooked rice" is nearly a measurement.

Vague — "dal and rice"

Better — "one katori dal, one cup of cooked rice"

If you are unsure what a katori or a cup comes to in practice, tracking calories without weighing food has the conversions worth memorising.

3. Cooked or raw

This one quietly ruins more logs than anything else. Rice roughly triples in weight when cooked. Dal more than doubles. Chicken loses about a quarter.

100 g of raw rice and 100 g of cooked rice differ by a factor of three, and both are things a person might reasonably type. Say which you mean whenever you give a weight.

Vague — "150 g rice"

Better — "150 g rice, cooked weight"

4. The additions after cooking

Butter on the roti. Malai on the paratha. Sugar in the chai. Cheese on the pasta. Dressing on the salad.

These get left out because they feel like garnish rather than food. Four cups of sweetened chai is 200–280 calories a day, every day, and it is invisible in a log that just says "chai".

Vague — "two rotis"

Better — "two rotis with ghee on top"

5. Restaurant versus home

A single word that reprices the whole meal. Restaurant kitchens use far more fat, more cream and larger portions than home cooking, and the model adjusts substantially when you say so.

Vague — "butter chicken and naan"

Better — "restaurant butter chicken and one butter naan"

What to put in a sentence, and what to skip

Worth addingNot worth adding
Cooking fat, roughlyExact gram weights of everything
Counts — rotis, eggs, katorisBrand names of basic ingredients
Cooked or raw, when giving a weightThe spices
Butter, ghee, sugar, cream added afterWhether the onions were browned
Restaurant, home, canteen, streetPrecise cut of vegetable
Rough size — small, medium, largeCooking method, if the fat is stated

The right-hand column is the trap. It feels like precision and it changes the estimate by single-digit calories, while people who spend their effort there routinely omit the tablespoon of ghee that changes it by a hundred.

Spend your words on fat, portions and additions. Everything else is decoration.

A worked example

The same dinner, described three ways.

One word. "Dinner." Useless — this is a diary entry, not a log.

A normal sentence. "Dal, rice, sabzi and two rotis." Now it can estimate, but every quantity is assumed. Expect ±35%.

Nine more words. "One katori dal with a normal tadka, one cup cooked rice, bhindi sabzi cooked in about a tablespoon of oil, two rotis with a little ghee."

That took six seconds to say and it removed almost all the ambiguity. Nothing was weighed and nothing was looked up.

Log a meal the way you would write it down. BetterCal turns a plain sentence — spoken or typed — into calories and macros. Free to start.

Be consistent before you try to be accurate

Here is the part that matters more than any of the above.

If you always describe your rotis the same way, and your estimate is 15% low, your log is still completely usable. Your day-to-day comparisons hold, your weekly average moves in the right direction when you eat less, and adjusting your target by 150 calories fixes the offset entirely.

A log that is consistently wrong is a working instrument. A log that is sometimes precise and sometimes vague is not, because you cannot tell which weeks the number means anything.

So pick a way of describing your regular meals and reuse it. This is also why saved meals matter so much — a repeat is not just faster, it is identical, which removes the variance you would otherwise introduce by describing the same dinner differently on Tuesday.

Correcting an estimate

When a number looks wrong, correct the input rather than overriding the output.

Say what was actually different — "less oil than that", "a much smaller katori", "no ghee" — instead of typing a calorie figure you also made up. The corrected estimate is usually closer than your override, and more importantly it teaches the system what your normal portions look like.

The exception is packaged food. If you have a label in your hand, the label wins. Nothing beats a number the manufacturer had to print.

When the sentence cannot save you

Some meals genuinely resist description, and it is worth knowing which ones so you stop trying.

  • Mixed restaurant curries. You cannot see the cream or the ghee, and there is usually a lot of both. Assume the high end of any range.
  • Deep-fried anything. Oil absorption varies with temperature and batter in ways nobody can eyeball. Street food and snacks sit in a wide band by nature.
  • Food someone else cooked for you. You did not see the tadka go in.
  • Shared dishes. The portion question is unanswerable, not the calorie question.

In all four cases, log the honest wide range and move on. Estimates are always ranges anyway — these are just wider ones.

The short version

  1. Say the cooking fat. It is the biggest number you are leaving out.
  2. Give one thing that can be counted.
  3. Say cooked or raw whenever you give a weight.
  4. Include what went on top after cooking.
  5. Say restaurant, home or street.
  6. Describe your regulars the same way every time.
  7. Correct the description, not the number.

Six seconds of extra sentence buys most of the accuracy that a food scale would, without the scale — and unlike the scale, it still works at a restaurant.

Log a meal the way you would write it down. BetterCal turns a plain sentence — spoken or typed — into calories and macros. Free to start.

Frequently asked questions

How do I make AI calorie estimates more accurate?

Say the cooking fat, give one countable anchor such as a katori or a number of rotis, state cooked or raw whenever you give a weight, and include anything added after cooking. Those four details move the estimate more than everything else combined.

What details do not matter when logging a meal?

Spices, exact vegetable sizes, brand names of basic ingredients, and cooking method once the fat is stated. They feel like precision and change the estimate by single-digit calories.

Why do calorie estimates vary so much for the same dish?

Because the sentence rarely contains the answer. Dal can be 90 calories or 400 depending on the dal, the katori and the tadka — the model is guessing from genuinely insufficient information.

Should I correct the calorie number or the description?

The description. Say what was different — less oil, a smaller katori, no ghee — rather than typing a figure you also made up. The corrected estimate is usually closer, and it teaches the system your normal portions.

Does it matter if my estimates are consistently a bit low?

Much less than you would think. A log that is reliably 10% low still shows real changes correctly, and adjusting your target cancels the offset. Inconsistency is the real problem.

Which meals are hardest to describe accurately?

Mixed restaurant curries, anything deep-fried, food someone else cooked, and shared dishes. In all four, log an honest wide range and move on rather than inventing precision.


Related: how accurate are AI calorie counters · text vs photo calorie tracking · tracking calories without weighing food