Amazon's white background rule is a number, not a colour

Amazon's main-image requirement is usually described as "a pure white background". That phrasing costs sellers a lot of time, because the actual rule is not about appearance. It is a number: RGB(255, 255, 255).

A photograph shot on a well-lit white sweep will look white to you and to anyone else. Sampled, its corners are more likely to read 250, 248, 251, near-white, clean, professional, and not what the rule says.

Why "looks white" is not white

Three things push a background off 255 without you noticing.

Light falls off. A sweep lit from one side is brighter on that side. The difference across the frame can easily be ten or fifteen levels, which is invisible next to a product and obvious to a sampler.

Cameras protect highlights. Most metering aims to keep the brightest part of the frame just below clipping, because clipping is usually a mistake. On a white-background product shot it is the goal, and the camera is working against you.

Compression moves pixels. JPEG is lossy. A subject close to the frame edge, at a low quality setting, can drag nearby background pixels off pure white through chroma artefacts, even when the pixel was 255 before the file was written.

That last one is the sneakiest, because it happens after everything you did was correct.

What we found testing this

We measured it rather than guessing. Encoding a white frame with a hard-edged subject through mozjpeg, then decoding and reading the actual corner pixels:

So pure white survives almost everywhere, and fails in one specific combination: a subject close to the frame edge, a low quality setting, and chroma subsampling switched on.

That combination is not exotic. Walmart asks for 2000 x 2000 images under 1MB, which pushes any quality-reducing pipeline into exactly that range on a detailed photograph.

How to check yours

Open the image in any editor with an eyedropper and sample the four corners. You want 255, 255, 255 in all three channels, at all four corners. Not 254. Not "white" in the swatch name.

If you are checking a lot of images, sampling by hand is not practical, which is roughly why we built PrepShot, it verifies the corners of the encoded file, not the buffer before encoding, because the encoded file is what Amazon receives.

How the corners get measured

Worth knowing what an automated check is actually doing, because it explains which parts of your image matter and which do not.

The check samples a small region at each of the four corners, typically a few pixels in from the edge, sometimes a small block averaged rather than a single pixel. It compares against 255 in all three channels. Some platforms allow a tolerance of one or two levels; the strictest do not.

Two consequences follow.

Only the corners are sampled, not the whole background. A background that is 255 at the corners and 251 in the middle behind the product will usually pass. This is not an invitation to be sloppy, a visibly uneven background looks bad in a search grid regardless, but it explains why a shot with slight falloff can pass while one with a bright corner reflection fails.

A single stray pixel can fail it. Dust on the sweep, a sensor hot pixel, or a compression artefact in exactly the wrong place. This is why checking programmatically beats eyeballing: a pixel reading 254 is invisible to you and decisive to the check.

Why the rule exists at all

It is worth understanding, because it explains why the marketplaces are inflexible about a number that seems pedantic.

Search result grids are assembled from many sellers' images on a white page. If each image carries its own slightly-different near-white, the grid looks patchy: one product sitting on a faintly warm background, its neighbour on a faintly cool one. Exact white makes every image's background disappear into the page.

It also enables the zoom and lightbox behaviour. When the background is known to be exactly 255, the interface can composite the image onto its own white panel without a visible seam.

So the requirement is not arbitrary strictness. It is the condition that makes a page of hundreds of sellers' photographs look like one catalogue.

Three fixes

Leave a margin. Keeping the product away from the frame edge removes the chroma-artefact case entirely. Amazon's own guidance asks for the product to fill about 85% of the frame, which conveniently leaves enough room.

Do not let quality drop too far. If you are squeezing a file to meet a size limit, watch what quality you ended up at. Below about 60 with subsampling on is where the problem lives.

Replace the background rather than correcting it. Brightening a near-white background risks clipping product detail along with it. Cutting the product out and compositing onto true white is both safer and exactly reproducible.

The order of operations matters

Here is a sequence that sounds reasonable and produces failures.

1. Resize the photograph to 2000 x 2000.
2. Remove the background and composite onto white.
3. Verify the corners. They read 255. Good.
4. Encode as JPEG at whatever quality fits the size limit.
5. Upload.

The check in step 3 proved something true about a buffer in memory. Step 4 then changed the pixels, and nobody looked again.

The correct sequence puts verification last, after encoding, and, if you want to be thorough, decodes the encoded file to check it rather than trusting the encoder. That sounds paranoid until you see the table above: the encoder demonstrably changes corner values under specific, reachable conditions.

The same applies to watermarking. A mark placed near a corner, at low opacity, can leave that corner at 254. Verification before the watermark step proves nothing.

Shooting so this never comes up

The most reliable fix is upstream of all the software.

Do not clip the background in camera. Aim for the background reading around 240 to 250 rather than a blown-out 255. A background clipped in camera has already lost the transition where the product edge meets it, and that transition is what any background removal tool uses to find the edge. Blow it out and the fine detail is gone unrecoverably: fabric edges, wire handles, stray hairs.

Light the background separately from the product. A background lit only by spill will be darker at the corners, which is where it gets measured. Angling each light so its edge falls on the sweep behind is the single most useful adjustment in a two-light setup.

Keep the product away from the frame edge. It removes the compression case entirely, and marketplaces ask for a margin anyway.

Checking a whole batch

Sampling four corners by hand takes about ten seconds per image. At four hundred images that is over an hour, every time you produce a batch, and it is exactly the kind of task people stop doing after the second week.

Two things make an automated check worth having rather than a formality.

It must read the encoded file, not the buffer. A tool that composites onto white and then declares success has checked its own arithmetic, not the file you are about to upload. The encoder sits between those two things, and as the table above shows, the encoder can change the answer.

It must check after every transformation, in order. Resize, composite, watermark, encode, each step can move a pixel. Checking in the middle proves nothing about the end.

What to do when a corner fails

In order of how often it works:

1. Move the product further from the frame edge. Removes the chroma-artefact case entirely.
2. Raise the quality setting, if a file size limit is not forcing it down.
3. Turn off chroma subsampling for that preset. Costs file size, fixes coloured-edge artefacts outright.
4. Check the source photograph is not clipped. If the product's own highlights are at 255, a background removal tool has nothing to separate them from the background, and the cutout may have punched through the product.

The thing nobody tells you

A listing with a nearly-white background usually goes live. It is not always caught immediately. What happens instead is that it gets flagged later, sometimes weeks later, in a bulk sweep, and by then you have a suppressed listing and no idea which of your four hundred images was the problem.

Checking the number takes a second. Finding it afterwards does not.