One photograph, sixty-four bits.

Comparison sites match products by barcode, brand or title. Most things sold online are unbranded and carry no barcode, and every seller writes their own title. Two listings of one object can share almost no words.

• Seller AMountaineering Leisure And Tactical High-Capacity Backpack For Men And Women
• Seller BLarge-capacity Commuter Outdoor Travel Backpack

A reseller uploads the supplier’s media pack unchanged, so the same image file turns up on every listing of the same object. The titles lie; the photograph does not. Matching it is closer to comparing checksums than to judging similarity.

$6.61

$7.77

Five steps, in plain arithmetic, so the browser, the extension and the index builder all produce identical bits. Drag through them below on a real listing. It runs in your browser with the same code that built the index.

1. Greyscale. Colour is thrown away; brightness is what survives recompression.
   Y = 0.299·R + 0.587·G + 0.114·B
2. Trim. Marketplaces pad photos onto white squares by different amounts. A border is cropped only if it is genuinely uniform, so two encodings of one photo get the same crop.
3. Shrink to 32×32 by averaging the pixels in each cell. Deliberately the simplest resampler, because every platform must reproduce it exactly.
4. Transform. A 2-D discrete cosine transform turns the 32×32 picture into frequencies. The top-left 8×8 block is the coarse shape of the image; that is what we keep.
   F(u,v) = Σx Σy f(x,y) · cos(π(2x+1)u / 64) · cos(π(2y+1)v / 64) keep 0 ≤ u,v < 8, skip (0,0) → 63 coefficients

   Computed as two 1-D passes. The [0,0] coefficient is the average brightness, which says nothing about content, so it is skipped, leaving 63.

5. Threshold at the median. Each coefficient above the median becomes a 1, the rest 0. Half the bits are set by construction, which is why every distance in the index is even.
   bit(i) = 1 if F(i) > median(F), else 0

Count the bits that differ. That count, the Hamming distance, is the whole comparison. Two unrelated photographs differ by about 32 bits, because each bit is a coin flip against the other. The same photograph after a CDN resizes and recompresses it differs by a handful.

We measured it. Sixty real product photos were put through the resizing, recompression and padding a marketplace applies; those must match. Different products from the same search, the hardest fair test, must not.

Ten leaves six bits of room before the first false match. In the live index, every one of the 74,691 pairs lands like this:

The honest number: in the live index the closest two different products get is 12 bits, a narrower margin than the calibration set.

You describe what you want the way you would to a person. We pull out what can be checked, show it back to you as chips, and say plainly what we could not turn into a rule. No model and no API key: the parser is deterministic.

Every candidate is checked against every rule. One that fails is out, with the reason kept, rather than ranked lower. The answer is the cheapest that passes, and if none pass, the answer is none.

What passes the gate is ordered on price, then how well it matches what you named, then whether the listing is complete.

What is not on that list: anything we are paid for. Some outbound links earn a commission; it is added after the order is decided and cannot change it. What we don’t do is the full list.

• Breadth. One catalogue is connected today, so most products have one seller and the spreads are small. Adapters for more stores are written and waiting on accounts.
• Delivery dates. No source returns them, so we report them as not measured rather than guess.
• Pack size and revision. The same photo is strong evidence of the same product, not proof of the same quantity or version. Both photos are always shown.
• Price history. It cannot be backfilled, so it starts the day daily snapshots do.
• Heavy edits. A tiny thumbnail or a heavily padded photo can move a fingerprint past the line. Those are missed, not mismatched, which is the safe direction.