What happens to your image after you press upload
Sellers optimise the file they upload. Shoppers never see it.
Between the two sits a processing pipeline that resizes, re-encodes, strips and caches. Knowing roughly what it does tells you which of your export decisions carry through and which are wasted effort.
The typical sequence
1. Validation. Dimensions, file size, format and often the background are checked. This is where a rejection happens, and it is the only stage where your exact file matters.
2. Metadata stripping. EXIF, IPTC and often the colour profile are removed. This is why orientation must be applied to the pixels rather than left as a tag, and why an untagged wide-gamut file is a problem you cannot fix later.
3. Colour conversion. Usually to sRGB. Good implementations convert; others assume the file was already sRGB and simply relabel it, which is where flat colours come from.
4. Derivative generation. Several sizes are produced from your upload: a grid thumbnail, a main view, a zoom view, and often mobile-specific sizes. This is a downscale, so how sharp your image is at the target size matters more than how many pixels beyond it you supplied.
5. Re-encoding. Into the platform's chosen format at its chosen quality. Often WebP for browsers that accept it.
6. CDN caching. The derivatives are pushed to edge servers under generated filenames.
What this means in practice
Your format choice barely matters. Upload JPEG. It will be re-encoded regardless, and JPEG is the format every uploader definitely accepts. The comparison articles on WebP and AVIF are about serving your own site, not about uploads.
Your quality setting matters less than you think, but not zero. It is re-encoded, so a very high quality upload is mostly wasted. But quality low enough to introduce artefacts gets those artefacts baked in and then re-encoded on top, which compounds them.
Sharpness matters a lot. Downscaling softens. An image that is sharp at 2000 pixels produces a better 400-pixel thumbnail than one that is soft at 4000.
Metadata is wasted effort. Copyright fields, keywords and captions are generally stripped.
Dimensions matter for thresholds, not beyond them. Above the zoom threshold you are supplying pixels that get discarded.
Caching is why fixes seem not to work
This causes more support tickets than anything else on the list.
You notice a bad image, fix it, re-upload, and the old one is still there. Sometimes for hours.
The derivatives are cached at the CDN under keys the platform controls. Your new upload generates new derivatives, but the old ones may still be served until they expire or the platform invalidates them. Uploading a third time does not help and may reset a queue.
The reliable trick, where the platform allows it, is to change the filename. A new filename usually produces a new cache key. Beyond that, the answer is to wait, and to check on a device that has never seen the listing before. Your own browser cache is a second layer of the same problem.
Why validation sometimes seems inconsistent
Two things explain most of it.
Not every check runs at upload. Some platforms accept an image immediately and run a compliance sweep later, in batches. A listing can go live and be suppressed weeks afterwards, which is why a background that measures 250 rather than 255 can seem fine for a month and then not be.
Some checks are automated and some are human. Prop and text rules in particular are often reviewed by people, which makes them slower and less consistent than dimension checks.
The practical consequence: passing upload is not the same as passing. Check the things you can measure yourself rather than treating acceptance as confirmation.
Finding out what your platform actually does
Right-click a live product image on your own listing, open it in a new tab and look at the URL and extension. Then open the network inspector and reload: the response headers give the content type and often the cache policy.
Do this once per marketplace and you will know exactly which parts of your export survive. It takes two minutes and it is more reliable than any article, including this one, because it is measuring your platform today rather than someone else's last year.