A/B Testing Product Images: A Seller's Playbook
Image tests move more revenue than copy tests — if you run them properly. What to test first, the three ways to run a test, and how to avoid fooling yourself.
Sellers rewrite bullet points for weeks and never touch the image that decides whether anyone reads them. A/B testing product images is the highest-leverage experiment on a listing: the main image is the only asset that affects both click-through rate in search and conversion on the page.
This playbook covers what to test first, the three practical ways to run a test, how much traffic you actually need, and how to read the result without fooling yourself.
Why image tests beat copy tests
A shopper scanning search results makes a decision in well under a second, on a thumbnail. Nothing in your copy is even legible at that moment. That is why the biggest measured wins on marketplaces come from the main image — and why it is worth being systematic instead of swapping photos on a hunch.
Two things are being measured, and they can move in opposite directions:
- Click-through rate (CTR) — how many people tap your thumbnail in search
- Conversion rate (CVR) — how many of those buy once they land
A high-contrast, zoomed-in main image often lifts CTR and dents CVR. Only the combination — units or revenue per session, not per click — tells you whether you made money. Our guide to product images that convert covers the underlying principles this playbook tests.
What to test, in priority order
| Priority | Test | Typical question |
|---|---|---|
| 1 | Main image | Product-only on white vs. product with scale or context cue |
| 2 | Second image | Lifestyle vs. infographic in the first swipe |
| 3 | Gallery order | Does moving the scale shot to slot 3 reduce returns? |
| 4 | Infographic style | Text-heavy callouts vs. one big benefit |
| 5 | Lifestyle casting | Model vs. product-only; which demographic |
| 6 | Background | Pure white vs. very light grey or gradient |
Test one variable at a time. If you change the crop, the background, and the shadow at once, a win tells you nothing you can reuse on the next SKU. The point of testing is a repeatable lesson, not a single lucky image.
Good main-image hypotheses to start with:
- Fill the frame more tightly (85% of the frame vs. a smaller product floating in white)
- Change the angle: straight-on vs. three-quarter
- Add a legitimate scale cue where the category allows it
- Show the product in its packaging vs. unboxed
- Lead with the best-selling colour instead of the "hero" colour
Three ways to actually run the test
1. Amazon Manage Your Experiments
Amazon's built-in A/B tool splits shoppers between two versions of a listing asset and reports the difference. It supports main images, titles, bullet points, product description and A+ content.
- Eligibility: brand-registered sellers with enough traffic on the ASIN
- Duration: typically eight to ten weeks
- Strength: real marketplace traffic, real purchases, no attribution guesswork
- Limit: not every ASIN qualifies, and you can only run one experiment per ASIN at a time
This is the cleanest option when you qualify. Check eligibility in Seller Central before planning a schedule around it.
2. A Shopify split test
On your own store you control the page, so you have more options and more responsibility. Use a dedicated A/B testing app that splits traffic at the theme or product-page level and tracks orders, not a homemade redirect. Two rules:
- Split users, not sessions, and keep each user on one variant
- Track revenue per visitor as the primary metric; add-to-cart is a proxy, not the goal
If your store traffic is thin, do not run store-side tests — you will spend two months measuring noise. Test on ads instead.
3. The ad-creative proxy test (fastest)
The cheapest, fastest signal is to run both images as creative in a paid social or shopping campaign, with everything else identical. You can get a directional read in days instead of months, for the cost of a small budget.
- Measures: CTR and cost per click reliably; conversion only if volume allows
- Watch out: ad audiences are not marketplace search audiences, so treat it as a strong hint, not proof
- Best use: shortlisting three candidate images down to the one you then test properly on the listing
See how to make product images for ads for building the creative side.
How much traffic you need
| Baseline conversion rate | Effect you want to detect | Sessions per variant (rough) |
|---|---|---|
| 10% | +20% relative | ~3,000 |
| 10% | +10% relative | ~11,000 |
| 5% | +20% relative | ~6,500 |
| 2% | +20% relative | ~17,000 |
Rough figures for a standard two-variant test — the exact number depends on your variance, but the shape of the table is the lesson: small effects need a lot of traffic. If your listing gets 200 sessions a week, you cannot detect a 10% lift, so test big swings or use the ad proxy.
Reading the result without fooling yourself
- Don't peek and stop. Checking daily and stopping the moment one variant leads is the most common way to ship a losing image. Decide the end date before you start.
- Compare like periods. Prime Day, Black Friday, a viral video, or a stock-out will dominate any image effect. Note anything unusual in the test window.
- Judge on revenue per session, then check returns a month later. An image that oversells causes returns that never show up in the test.
- A flat result is a result. It means the image is not your bottleneck — move on to price, reviews, or the offer.
- Write it down. One line per test: hypothesis, variants, dates, metric, outcome. After ten tests you will have category-specific rules worth more than any blog post.
Producing variants fast enough to test
The reason most sellers do not test images is production cost. A studio reshoot per variant means every test costs hundreds of dollars and weeks of lead time, so nobody runs the second test.
With HedaAI you upload one or more real photos and get a full set of 12 e-commerce images — 8 main and gallery shots plus 4 A+ banners — in minutes, with your product's true shape, colour and labels preserved. That gives you multiple legitimate main-image candidates from the same source photo, so a variant costs a couple of minutes instead of a shoot day. New accounts get $2 in free credits (about two products free), then $1.00 per product; the first run is a watermarked preview, and your first payment removes watermarks and unlocks 2K HD downloads. See the examples page and pricing.
If a test tells you the current image is the problem, our image audit checklist is a good place to find the next hypothesis.
The takeaway
Test the main image first, change one variable at a time, run for at least two full weeks, and judge on revenue per session rather than clicks. Use Amazon's experiment tool where you qualify, an app-based split test on Shopify with real traffic, and ad creative as a fast pre-filter everywhere else. The compounding value is not any single winning photo — it is the list of rules you build about what your buyers respond to.