An Amazon image gallery should behave like a six-step sales conversation. The main image identifies the exact product. The remaining images answer the next buyer questions: what is included, what it looks like in use, which details matter, how scale or fit works, and why this version is the right choice.
That is the missing layer between knowing Amazon's product-image rules and running one controlled image experiment. A seller does not need six unrelated “beautiful” images. They need six assigned jobs, six evidence sources, and six acceptance decisions.
Evidence boundary: the blue planter is a fictional controlled product reused from our Amazon image-experiment test. For this workflow, we made one new source-referenced lifestyle candidate and compared it with the approved white-background source. It was not uploaded to Amazon, tested with buyers, or approved by the product owner. The visual is a first-pass disposition, not an accuracy or sales claim.
Use six images as a working stack, not a platform promise
Amazon's public seller guidance says every detail page needs at least one image and recommends six. It assigns the first image as the MAIN image, while additional images can show use, environment, angles, and features. It also says images must accurately represent the product. Review Amazon's current product-image guidance.
That recommendation is a planning constraint, not permission to fill six slots with invented content. Category-specific rules can override general guidance, available gallery surfaces can change, and the live contribution in Seller Central remains authoritative.
Current merchant discussions expose the more useful operational pattern. Sellers describe assigning gallery jobs such as hero, benefits, features, included items, use case, comparison, and brand story. Other practitioners warn that several plain, question-answering images can outperform one cinematic asset because each visual has a reason to exist. Those discussions establish a recurring job, not a universal winning sequence. Read the seller role discussion and the image-job workflow discussion.
Assign each slot one buyer question
| Slot | Buyer question | Image job | Required authority | Release test |
|---|---|---|---|---|
| 1 | Is this the exact item? | MAIN identification | approved product photograph or permitted render | current main-image and category rules; exact SKU; deterministic white background and framing |
| 2 | What arrives in the box? | included-items proof | bill of materials, pack-out image, quantity record | every visible item is included; quantity and variant match |
| 3 | What does it look like in use? | restrained use context | approved product source plus documented scenario | product identity holds; props cannot be mistaken for included items |
| 4 | Which physical details matter? | angle or close-up | approved alternate photograph or product detail source | detail exists on the sold item and remains legible at gallery size |
| 5 | Will it fit? | scale, dimensions, or compatibility | approved dimensions and compatibility table | numbers and comparison objects are verified; no generated scale claim |
| 6 | Why choose this version? | approved comparison, benefit, or brand proof | signed claim matrix and competitor-safe evidence | wording is substantiated, current, and rendered deterministically |
The sequence is deliberately boring. It forces production to begin with buyer uncertainty and product authority. If reviews repeatedly ask what is included, slot 2 outranks a lifestyle scene. If returns cite size, slot 5 deserves the next production cycle. If no source authorizes a comparison claim, slot 6 remains HOLD.
Download the six-slot Amazon image manifest. It includes the buyer question, source authority, allowed generation scope, deterministic layer, acceptance test, owner, and release status for every slot.
Keep the MAIN image out of generative reconstruction
Amazon's current general guidance requires the MAIN image to show the product for sale against pure white, with the product filling at least 85% of the frame. Added text, logos, borders, color blocks, watermarks, and other graphics are prohibited. Amazon also says the image must accurately represent the product.
Those constraints make generative reconstruction a poor default for slot 1. A model can return a plausible planter, bottle, shoe, or appliance while changing a seam, cap, glaze, control, texture, label, or included part. The output can look cleaner and be the wrong SKU.
Use deterministic operations where they solve the job:
-
start from an approved photograph or permitted exact render
-
remove the existing background without regenerating the product
-
composite literal RGB 255, 255, 255
-
frame the product to the current role and category requirement
-
export the accepted format and filenam
e
-
inspect edges, color, quantity, included items, and every visible mark at full resolution
If the live main image is suppressed, use the main-image recovery workflow instead of treating the incident as a creative refresh.
Generate the environment, then judge the product
Slot 3 is the cleanest place to test a source-referenced generation workflow. The task is narrow: preserve the exact product and change the context so one documented use becomes easier to understand.
The candidate clears the article's first visual check: one planter and saucer remain visible, the pale blue glaze and terracotta-colored base are recognizable, the setting is restrained, and no text or unverified benefit was embedded. It still needs product-owner review. The model may have changed fine ceramic texture, rim thickness, lower-groove depth, or saucer spacing. The added plant and soil also need an explicit “not included” treatment wherever a buyer could confuse the configuration.
That distinction is the operating advantage of role-based production. The image can be useful enough to continue and still be rejected for publication. Record ACCEPT_FOR_OWNER_REVIEW, REVISE, or REJECT; do not collapse those states into “looks good.”
Add text and claims after the image is accepted
Supporting images often need dimensions, included-item labels, feature callouts, or comparison copy. Do not ask the image model to invent or spell those facts. Keep the generated or photographed visual separate from the factual layer.
A safe production order is:
-
approve the image pixels against the exact SKU
-
select wording from a signed product-information or claims source
-
render type, icons, lines, and measurements in a deterministic design template
-
compare every rendered string and number with the source
-
export a flat candidate and retain the editable sourc
e
-
review the current marketplace and category rules for that slot
This also makes localization and corrections cheaper. A changed dimension or translated label does not require a model to regenerate the product.
Release one stack with a manifest
Before upload, every row in the manifest should resolve the same questions:
- Which marketplace, parent, child ASIN, and sold variant does this asset belong to?
- What buyer question and gallery slot does it own?
- Which file or record is the product and claim authority?
- What was the model allowed to change?
- Which facts were rendered deterministically?
- Who accepted product truth, policy review, and final export?
- What live asset does this replace, and where is the rollback copy?
- What result will be read, on what date, before another slot changes?
Upload order is not evidence of exposure. After contribution, verify that the intended asset is selected and visible on the correct child ASIN and marketplace. Retain the previous published stack and record the observed publication time.
When the business question requires attribution, do not replace all six images and call the sales change an image result. Use Amazon Manage Your Experiments where eligible, change one declared role, and carry contribution, returns, support contacts, and product-truth incidents alongside the platform result. The Amazon A+ workflow handles the separate below-gallery module system.
The production rule
For one ASIN, the useful loop is:
- rank buyer questions from reviews, returns, support, search, and merchandising evidence;
- assign one question and one authority source to each image slot;
- keep the MAIN image source-based and deterministic;
- use AI only where the allowed change is explicit, usually a supporting environment;
- render approved claims and measurements after image acceptance;
- reject invented product details even when the output is attractive;
- publish with a versioned manifest and rollback asset;
- verify the exact live child ASIN and marketplace;
- measure one role at a time when the decision needs attribution.
The goal is not six images. It is six fewer unanswered reasons for the right buyer to hesitate.
