AI Product Images: What Actually Changed Conversion for 10 Sellers

I spent a few months redoing product images for 10 small sellers: Etsy, Shopify, a couple of Amazon FBA stores, one eBay shop. Most had decent products and weak photos. A few had good photos and still weren't converting. I went in assuming the AI part was the interesting bit. It wasn't.

One honest disclaimer up front: you will not find a conversion percentage in this article. Ten sellers, ten different traffic sources, different seasons, small samples, no controlled test. Anyone quoting a clean "+340%" from a setup like that is making it up. What you get instead is the workflow, the failure modes, and a measurement protocol that does not fool you.

What I actually did

For each seller I took one existing, accurate source photo, cut out the subject, and generated a fixed set of six image types from it:

1. Clean catalog shot on pure white, same framing as the original
2. One in-use / context shot
3. One scale reference shot (hand, coin, ruler, or a familiar object)
4. One close-up of the material, stitching, or texture
5. One "what's in the box" flat lay
6. One variant or comparison grid

Then we swapped only the images on a handful of listings and held title, price, description, shipping and ad spend frozen for two weeks. That constraint is worth more than any prompt in this article.

Generation was maybe 15% of the work. The rest was choosing a trustworthy source photo, throwing away bad outputs, and arguing about what not to change.

Seven changes that held up

1. Fix the framing before the lighting

Most weak hero shots aren't underexposed, they're badly cropped: the product takes 40% of the frame, sits off-center, or changes aspect ratio between the main image and the gallery. Standardizing subject size to roughly 75–85% of frame and centering consistently did more than any cinematic lighting prompt. On mobile the thumbnail is tiny. A product that fills it reads as a product; one floating in white space reads as a blank tile.

2. Add a scale reference

This was the single most consistent win. Hesitation and returns on small goods usually come from "I thought it was bigger". A hand holding the item, or a familiar object beside it, answers a question buyers won't bother to ask. It is not glamorous and it repeatedly beat the pretty lifestyle shots. Put it in gallery slots 2 or 3, never slot 7.

3. Shoot the objection

For each seller I asked what buyers zoom in on or complain about in reviews, then generated one close-up of exactly that: zipper teeth, strap stitching, wood grain, clasp, label, port. One bag seller's objection was the strap attachment. A single macro shot of the stitching did more than five lifestyle images.

4. Keep the catalog shot boring

Pure white, soft even shadow, no props, no gradient, no floating sparkles. The model wants to make it beautiful. Resist. This is the image marketplaces use for thumbnails and comparison grids, and it has to match the real product exactly.

5. Lock the product identity across every image

This is where AI quietly destroys trust. If the engraving moves, the logo changes, the handle shape drifts, or the color shifts half a shade between gallery images, buyers read it as a different product — or a fake one. I started writing an immutable identity block (counts, positions, exact color, explicit absences) and pasting it verbatim into every prompt, then doing a final side-by-side check against the original. Any output that changes the product gets deleted. I threw away a lot of good-looking images over this.

6. Put the objection-killer early

Most buyers never open image 6. Whatever answers the top objection — size, what's included, how it attaches, whether it fits — belongs in positions 1–3. Moving "what's in the box" from the last slot to the third stopped the same pre-purchase question on one store. Zero generation cost.

7. Check platform rules before generating anything

Amazon main images require a pure white background, no text, no watermark, no props, and the product filling most of the frame. Etsy and Shopify are far more permissive. I once spent an afternoon generating beautiful contextual hero images for an Amazon listing that could never use them. Check the rule first, then decide which of the six shot types are even legal for that SKU.

Three things that wasted time

• Lifestyle scenes as the hero. Great in a portfolio, weak as a thumbnail. Small subject, busy background, no clear read at 100px. Keep them in the gallery.
• Regenerating the whole gallery at once. Any change becomes uninterpretable and you cannot roll back cleanly. One or two slots at a time.
• "Enhancing" an already accurate photo. Two sellers had honest, slightly flat but correct photos. The AI versions were prettier and subtly wrong — smoother texture, half a shade off. Returns went up. If the photo is accurate, the highest-value work is adding scale and detail shots, not replacing the hero.

The prompt skeleton

Consistency beat cleverness. One fixed identity block, pasted verbatim into every prompt:

IDENTITY (never change):
[product], [exact color], [material], [count/position of distinctive
parts], [logo / engraving placement], [explicit absences].
Preserve the exact product from the reference image.
Do not add, remove, recolor, or reinterpret any part.
No generated text. No new logos. No watermark.

TASK:
[one of the six shot types], subject fills 80% of frame, centered,
[background], soft even neutral light from upper left,
50mm equivalent, eye level, [aspect ratio].

NEGATIVE:
extra parts, duplicated product, warped geometry, plastic-looking
reflections, oversaturated color, drop shadow, text, props.

REJECT IF:
identity drift, extra objects, deformed edges, unrealistically
perfect material, shadow that does not match the light direction.

Three habits mattered more than the wording: generate more than you need and reject aggressively (my keep rate was about 1 in 5), check at both thumbnail size and full zoom, and always keep the original file for comparison and rollback.

How to measure without fooling yourself

• Track product page → add to cart, not overall store conversion, which is polluted by traffic mix.
• Change one thing. Hold price, title, description, shipping, ads, and every other gallery slot frozen.
• Run at least 14 days, ideally one full purchase cycle.
• Record the sample size. Under a few hundred product-page views, a swing is usually noise.
• Watch returns and "not as described" complaints as a counter-metric. A prettier-but-wrong image can raise add-to-cart and still lose money.
• Keep the old images. Rollback should take one minute.

Platform main-image baselines

Compliance is not optional

Generated images can misrepresent the product. On several marketplaces that is a listing violation, not a marketing judgment call. Do not generate a competitor's product, another brand's logo, or a fake certification badge. Food, supplements, medical devices, children's products and branded goods have stricter category rules. Disclose AI use wherever a platform or client asks for it, and keep the source photo plus your prompt log — that record is your defense if a listing is challenged.

The unglamorous summary

The wins came from scale references, objection-specific close-ups, consistent framing, and not letting the model change the product. AI was just a cheap way to produce six shots instead of booking six setups. Every time I let it "improve" the product itself, it cost money.