How a scan works

Four photos, taken the same way every time, so the AI always sees what it needs — and so you always know how much to trust the answer.

A single snapshot in bad light is the reason photo-based metal identification has a reputation. IsItGOLD? does not take single snapshots. The app walks you through four specific photos, coaches you while you frame each one, scores every shot before it is sent, and tells you afterwards exactly what moved the confidence figure. It never refuses a photo — but it never pretends a poor one is good.

The four shots

Each one answers a question the others cannot.
  1. Shot 1: Top-down

    The whole piece, flat, filling about two thirds of the frame. This is where the AI reads shape, colour and construction — and where it works out the size if you added a reference object.

    How: Plain, matte surface. Daylight from a window or a lamp from the side. Phone held level over the piece.

  2. Shot 2: Angled

    Tilted 30–45°, so the light rakes across the surface. Wear, solder seams, plating breaking through at the edges and the difference between cast and stamped work all show up here and nowhere else.

    How: Same spot, same light. Tilt the piece, not the phone. A small glint is fine; a white blob is glare — move a little.

  3. Shot 3: Hallmark close-up

    The stamp — 585, 750, 925, PT950, a maker or a sponsor mark — is the strongest evidence a photo can carry. Reading it moves the purity from an assumption to a fact, and the confidence rises with it.

    How: As close as your camera will focus, with the stamp centred in the small square. Tilt into the light until the numbers stand out. No stamp? Say so and continue — the score just will not get the bonus.

  4. Shot 4: Reverse

    The back of a pendant, the inside of a ring, the other face of a coin. Findings, clasps and posts are often a different metal from the front, and a coin is identified from both faces.

    How: Optional for most jewellery, required for coins. Same framing as the first shot.

Light

Daylight from a window, or one lamp from the side. Never the phone torch pointed straight down — it flattens everything and glares.

Framing

A square in the viewfinder shows where the piece should sit. Fill about two thirds of it. Live hints say when to move closer, back, or steadier.

Quality score

After every shot you see a score out of 100 and the reasons — soft focus, glare, busy background. Retake it, or use it anyway. Your call.

The reference object

Optional — but it turns size from a guess into a measurement.

A phone camera has no idea how big anything is. Place one everyday object of known size next to the piece in the first shot and the AI can measure the piece against it. The app asks which one you used:

  • Bank card85.6 mm long (ISO/IEC 7810 ID-1)
  • €2 coin25.75 mm across
  • €1 coin23.25 mm across
  • US quarter24.26 mm across
  • Printed calibration card85.6 mm long, with grey and white patches for colour

The printable calibration card

A bank-card-sized outline with a neutral grey patch and a white patch. Beyond scale it gives the AI a known reference for white balance, so a warm lamp does not make white gold look yellow. Print it at 100% — never “fit to page” — and check the ruler on the page before you use it.

Open the calibration card

What moves the confidence figure

The confidence you see is not the AI’s gut feeling. It is a score with a paper trail, and the app shows you the trail.

Raises it

  • A reference object in the first shot+5

    Size becomes a measurement instead of a guess.

  • A hallmark the AI could read+8

    Purity is read from the stamp rather than assumed.

Lowers it

  • Only one photo−10, and the score is capped at 79

    One angle cannot show wear, edges or the back.

  • No weight entered−10

    Without a weight the value has to be estimated from size.

  • No description, or a very short one−12 / −6

    Where it came from and what you already know narrows the search.

  • Photo quality below the bar−7 below 80, −15 below 60

    Each shot is scored for sharpness, exposure, glare, framing, background and colour cast before it is sent.

  • No hallmark close-up−8

    The strongest single piece of evidence is missing.

  • No reference object−6

    Scale has to be estimated from the photos alone.

  • Cannot tell solid from plated−5

    When the surface treatment is unclear, so is the value.

The AI’s own assessment sets the starting point. The adjustments above are applied on our servers, the same way every time, and the result never leaves the range of 1 to 100. Every scan result lists the ones that applied to it.

How the value is calculated

Weight, purity, and today’s price per gram — nothing hidden.

Weight. The weight you enter wins. Without one, the app estimates it from the size of the piece and the density of the metal — which is why the reference object matters, and why a kitchen scale is the single cheapest way to tighten the estimate.

Purity. Read from the hallmark when there is one (585 is 14 karat, 750 is 18 karat, 925 is sterling), otherwise assumed from what the piece looks like and flagged as an assumption.

Price. The live market price per gram of pure metal, fetched several times a day. Weight × purity × price is the metal value. The range around it widens as the confidence falls.

Plated and filled. A gold-plated piece is valued on its base metal, not on its colour. A gold-filled piece is valued on the small fraction of gold it actually carries. If the app cannot tell, it says so — and the confidence drops accordingly.

How accurate is it?

IsItGOLD? gives you an informed estimate, not a laboratory assay. No photograph can measure density or run an acid test, and no app should claim to. What the guided procedure does is make the estimate as good as a photograph allows, and make its confidence honest: the figure goes up only when the evidence does.

That is usually enough to decide whether something is worth a trip to a jeweller. For insurance, resale at scale, or anything legally binding, get a professional appraisal — the app will tell you the same thing on every result.

Read the full accuracy disclaimer in the terms of service.

Try the guided scan.

Four photos, three minutes, and a number you can trust.