How accurate is Olivka?

Where the numbers come from

Olivka doesn't guess calories out of thin air. Every estimate is built from reference nutrition databases: USDA FoodData Central, the U.S. Department of Agriculture's public database of foods and their laboratory-measured nutrients, and Open Food Facts, the world's largest open database of packaged products — including tens of thousands of products sold in Poland and across Europe. For restaurant meals and branded items missing from both, Olivka can additionally search the web for published nutrition information. When you log "grilled chicken breast with rice", each ingredient is matched to a database entry, and your meal's calories, protein, carbohydrates and fat are calculated from those entries and the estimated amounts.

You can watch this work in every meal card: Olivka lists the ingredients it recognized and the totals it calculated, and the card has Confirm and Edit buttons. Nothing is logged as final until you've had the chance to correct it.

How a photo becomes a calorie count

When you send a photo, a vision AI model first describes what is actually on the plate: the dish, the visible ingredients, and an estimate of the portion sizes. That structured description is what the calorie calculation works from. Each recognized ingredient is matched against the nutrition databases above, and the portion estimates convert per-100-gram values into your meal's totals. Voice notes are transcribed first (Polish and English both work well), and plain text goes straight to the same matching step. The whole loop usually takes about ten seconds.

When a meal is ambiguous — a stew where the fat content could swing the total, a smoothie that might or might not have sugar — Olivka asks a short follow-up question instead of silently guessing.

What accuracy should you expect?

Calorie counting has an uncertainty floor that no app can remove. In the United States, printed nutrition labels are legally allowed to deviate from the true value by up to 20 percent, and European rules allow similar tolerances. Two reputable databases can list the same food with noticeably different values, because brands, recipes, and preparation differ. A tablespoon of oil that stays in the pan versus lands on your plate changes a meal by over 100 kcal. Any tool that promises exact calories from a photo is overpromising.

Within that floor, how you log matters more than which tool you use. Text with amounts ("200 g of cooked rice, one egg") is the most accurate — essentially database-level. A clear photo of a simple meal identifies the composition well; portion size is then the main source of error. Mixed dishes, soups, and sauces carry the widest error bars, because fat and sugar hide in them.

The practical rule: treat any single meal as an estimate, and base your decisions on daily and weekly totals logged the same way every day. Classic studies of hand-kept food diaries show people routinely under-report what they eat — a consistent, low-friction logger you actually use every day beats a precise one you abandon in week two.

How to get the most accurate counts

Add amounts when you know them: "200 g", "half the pack", "two slices". Weights beat photos; even rough amounts beat none.

Name what's invisible. Oil used for frying, butter on the pan, sugar or milk in your coffee, dressing on a salad — these are the biggest silent calorie sources, and no photo can see them. One extra word ("fried in oil") fixes it.

For packaged products, send the brand name or a photo of the label — packaged foods resolve to exact database entries, so they're the most precise thing you can log.

Photograph at an angle rather than straight down, with the whole plate in the frame, before you start eating. A top-down shot records the outline of the food but not its height, and portion size is where most of the error lives — an angled view keeps some of that depth. Leaving cutlery or your hand in the frame gives the estimate a known size to scale against.

Correct the card. If the estimate is off, tap Edit and say what's wrong ("that was 300 g, not 150") — the correction takes seconds. Your confirmed history also makes repeat logging faster: "my usual breakfast" resolves against meals you've already confirmed.

One meal, three ways

The same dinner logged three different ways shows where the uncertainty actually enters. Take grilled chicken with rice and roasted vegetables. Sent as text with amounts — "180 g chicken breast, 150 g cooked rice, roasted courgette and pepper, one tablespoon of olive oil" — every component resolves directly to a database entry and the only uncertainty left is the database's own: the label tolerance described above. This is the most accurate input Olivka accepts, and it is more accurate than a photograph of the same plate.

Sent as a photo, the vision model identifies the three components reliably and estimates the portions from what it can see. The identification is rarely the problem; the portion is. A 150-gram serving of rice and a 220-gram serving look similar in a picture, and that difference alone is around 90 kcal. The olive oil the vegetables were roasted in is invisible — that is another 120 kcal that no photograph contains. This is why Olivka asks a follow-up when a dish looks like it is hiding fat, and why one added word ("roasted in oil") closes most of the gap.

Sent as a voice note, the audio is transcribed and then treated exactly like text, so its accuracy depends on what you said rather than on the recording. "Chicken, rice and veg" carries the same ambiguity spoken as typed. "About 180 grams of chicken, a cup of rice, veg roasted in a tablespoon of oil" is as good as the typed version, and takes four seconds to say while walking.

What Olivka does not do

A methodology page is only useful if it also states the limits. Olivka does not weigh your food, and no estimate here is a measurement — if you need laboratory precision for a medical reason, a kitchen scale and manufacturer data are the correct tools, not this or any comparable product. Olivka does not track exercise, body weight, or data from a wearable, so it cannot tell you your energy balance, only your intake. It has no barcode scanner: packaged food is best logged by sending the brand name or a photo of the label, which resolves to the same database entry a scan would.

Olivka also does not prescribe. It will not set you a calorie target, design a diet, or tell you what to eat next — it counts what you actually ate and keeps the daily totals honest. Working out a target is a separate calculation you can do yourself, and the calculators on this site run the standard Mifflin-St Jeor equation for it without an account. Finally, Olivka is not a medical device and nothing it reports is medical advice; for a clinical question about your diet, ask a doctor or a registered dietitian.

How we check Olivka's work

Accuracy isn't a set-and-forget feature. We routinely re-run samples of logged meals through an independent evaluation — comparing Olivka's estimates against reference models and the underlying databases — and investigate significant discrepancies. When a class of meals is systematically off, we fix the pipeline, not the individual answer. And because every estimate ships as an editable card, mistakes are visible and correctable by the person who knows the meal best: you.

Sources

• USDA FoodData Central
• Open Food Facts
• US FDA — nutrition label tolerances (21 CFR 101.9(g))
• European Commission — guidance on tolerances for nutrient values (Reg. 1169/2011)