A likely match is an honest answer, not a dodge.

A likely match, not a verdict

A food photo shows what was visible: the dish's surface, its colours, its plating, at one moment and one angle. What it cannot show is the recipe — the stock it simmered in, the nuts inside the sauce, the oil it was fried in. A lookup reads the visible evidence and names the most likely dish; that is a starting point, not a determination.

The status line on a result exists to say this out loud. "A likely match" means the visible evidence points one way; "A few possibilities" means more than one dish fits what the photo showed; "A little more context would help" and "This may not be food" cover the rest. None of them is a claim about ingredients, allergens or safety — those are never readable from a picture, which is why the result says so alongside every estimate.

Research on image-based dietary assessment reaches the same conclusion at scale: a 2020 review in the International Journal of Obesity found smartphone photo methods promising but noted the technology for automatic food recognition and portion estimation was still in its infancy, and that no single method was adequate in all settings. Hedged answers are what an honest system looks like. Source: 2020 review of image-assisted dietary assessment (opens in a new tab).

Dishes that genuinely look alike

Some uncertainty is baked into food itself. A chicken tikka masala and a red Thai curry can share a bowl, a hue and a sheen; a carbonara and a mac and cheese can be the same pale swirl; a cheesecake and a flan can differ only in the crumb you cannot see from above. Two dishes with different names, ingredients and origins can legitimately share an outline.

Researchers built a benchmark for exactly this problem: Food-101, from ETH Zurich's Computer Vision Lab, collects 101 food categories and 101,000 images precisely because telling pictured dishes apart is hard enough to measure. The lab is candid that its own training images were left noisy — intense colours and sometimes wrong labels — because real food photography is like that. Source: Food-101 dataset description (opens in a new tab).

When a lookup hedges, pairs like these are often why — dishes whose differences live in the recipe, not the silhouette. A hedged status and a more cautious summary are information, not indecision: they mark the specific ambiguity your photo could not resolve.

The same name, different recipe

Names travel loosely. "Stew", "dumplings", "curry" and "salad" each cover families of dishes that vary by region, household and occasion — and even a specific name like bibimbap describes a dish whose toppings and sauces genuinely vary from bowl to bowl. Matching the name does not pin down the recipe.

This is why a correct answer can still describe a different dish than yours: the lookup names the dish your photo most resembles, while your plate follows its cook's version. The field notes make the same point from the other side — they describe common context for a dish and leave room for variation, because variation is what real food does.

Treat the match as a map to the right neighbourhood, then let the menu, the person who cooked it, or the label do the precision work.

Mixed plates and partial views

Composition limits what any photo can carry. A crowded plate presents several dishes at once; a shot from table level hides the surface; a covered pot, a dark restaurant or a heavy garnish conceals the very details that separate lookalikes. The model reads what is visible — if the deciding detail is hidden, the honest result hedges.

The fix is mostly photographic: the photo guide covers the habits that help — whole dish in frame, daylight over filters, one dish per lookup. A photo taken for the record and a photo taken for identification are different exercises.

What 'unsure' looks like in a result

The clearest cue is the status chip itself — "A few possibilities" instead of "A likely match". Below it, the hedged summary and the "what was visible" observations show the actual evidence: the textures, colours and accompaniments the reading relied on. Reading those tells you more than the headline.

Portion and nutrition carry their own honesty labels: figures appear only when serving details support an estimate, and they are marked approximate with the assumption printed beside them — the reasoning is in the portion guide.

One boundary never softens with confidence: uncertainty about identity never licenses certainty about allergens. Even a confident-looking match cannot confirm ingredients, allergens or dietary suitability — a sauce or garnish may be invisible in the frame. When an allergen matters, the route is the questions in the eating-out guide, aimed at whoever made the food.

How to nudge a better answer

Three things reliably sharpen a lookup. Better light and a straighter angle — the whole dish from above beats a flattering side shot. One dish at a time — a second lookup for the second plate beats a crowded frame. And context — if you know the cuisine or saw the menu description, the name-entry mode and the note field let you say so.

If the first result hedges, the photo is usually the cheapest variable to change: retake with the tips in the photo guide, and the same lookup often resolves. The photo and portion guide covers what even the sharpest photo cannot establish.

And when the answer still comes back unsure, that is the system working — an honest "a few possibilities" is more useful than a confident guess. Try the lookup yourself at the scanner, or browse the field notes to see how a confident match gets written up.