Amazon listing images in the age of AI search: what actually changed

On 13 May 2026 Amazon retired the Rufus brand and folded its technology into Alexa for Shopping, which now sits in the main search bar and answers shoppers before they reach a results page. Almost every guide written since has been about copy — titles, bullets, attributes. Almost none has been about the images. That gap matters, because the way an AI assistant reads a listing is not the way a shopper looks at one, and the difference changes how a gallery should be planned.

What changed in May 2026 — and what did not

The rename is not cosmetic. Alexa for Shopping rolled out across the Amazon shopping app, Amazon.com and Echo Show, and it generates an answer layer above the ordinary results. What did not change is the underlying logic: the assistant still assembles its answer from your product data, your reviews and your Q&A. If your listing already answered real buyer questions in plain language, you are mostly fine. If it relied on visual polish to do the persuading, you have a gap you cannot see.

The uncomfortable part: the assistant may not see your best work

Here we have to be honest about the state of the evidence, because a lot of confident advice is circulating. Amazon has not published documentation stating that the shopping assistant performs OCR on gallery images or interprets them visually. Some third-party guides claim it reads text overlays and infers materials and scale from photographs; others describe the system as text-driven and do not mention images at all. Those two positions cannot both be safe to plan around.

So we plan for the conservative case: assume anything that exists only as pixels may not reach the assistant. That assumption costs nothing if the optimistic reading turns out to be true, and it protects the listing if it does not.

The rule we design to: nothing important lives only in an image

Most listings we audit fail this test in the same place. The dimension that answers "will it fit?" is set in small type inside infographic four. The material is shown, beautifully, and named nowhere. The compatibility list is a graphic. To a human scrolling the gallery it all works. To a system reading the listing as text, that product has no dimensions, no material and no compatibility.

Mirror every visual claim into text

The fix is not to make the images uglier or more literal. It is to run a simple audit after the creative is signed off: list every claim the gallery makes visually, then confirm each one also exists as real text somewhere in the listing — a bullet, the description, a structured attribute or A+ body copy. Anything that appears only inside a JPEG gets written down. This is a thirty-minute pass, and it is the highest-return thing most brands can do to their existing listings this quarter. We made the same point for A+ specifically in choosing A+ Content modules: text baked into an image is not a substitute for real text.

Attributes stopped being admin work

Structured attribute fields used to be something an operations person filled in once. They are now one of the cleanest sources an assistant has for facts about your product, because they are unambiguous. An empty attribute is a question the assistant cannot answer about you — so it answers it about a competitor instead. Fill them from the same fact sheet that briefs the shoot, so the photograph, the infographic and the attribute field all agree.

What did not change: the gallery still closes the sale

None of this demotes photography. The assistant may surface your product, but a person still decides. On mobile the gallery is effectively the whole listing, and a shopper arriving through an AI answer arrives warmer and further along — which makes the first image more decisive, not less. The technical floor has not moved either: pure white background on the main image, roughly 85% frame fill, at least 1,000px on the longest side for zoom. Those rules are unchanged, and we set them out in Amazon product image requirements.

The right conclusion is not "write for the robot". It is that a listing now has two readers with different abilities, and a good listing serves both: designed for the human, transcribed for the machine.

A workflow that covers both readers

• Start from the buyer question list, not the shot list. Every question becomes one frame and one line of text.
• Design the gallery for the human: one message per image, legible at thumbnail size, no slot repeating another.
• After sign-off, transcribe. Every visual claim gets a text home in a bullet, the description, an attribute or A+ copy.
• Fill every structured attribute field that applies, using the same numbers shown in the images.
• Ask the assistant your own buyer questions and read which product it describes. If the answer is vague about you, the gap is in your text, not your photography.

That last step takes five minutes and is the closest thing to a free diagnostic available right now. We use it as a standing check on client listings — including the gallery system we built for POM Industries — and it is part of why we plan Amazon listing and product photography as one brief rather than two deliverables.