Dynamic Creative QA: A Modular E-commerce Asset Checklist

A product image can be accurate. A video clip can be accurate. A headline can be approved. Yet when an ad system joins them to a different SKU, a different offer state, or a different disclaimer at impression time, the finished ad can still make the wrong promise.

This is no longer a resizing problem. On September 30, 2026, Amazon Ads announced a reimagined Responsive eCommerce Creative (REC) with in-banner video, multi-image carousels, conversational prompts powered by Alexa for Shopping, and rich-media layouts. Amazon says the system can assemble and optimize ads in real time using audience signals, campaign objectives, and placement context. Its specifications also describe REC as a component-based system that can rotate as many as 20 ASINs and draw on product detail pages, brand assets, uploaded media, and generative tools.

That announcement proves what Amazon has launched; it does not prove that every platform or advertiser has completed the same transition. But it changes a practical procurement question for brands: the deliverable can no longer be a folder of individually approved files. It must include rules for what may be combined, what must stay together, and what must never appear together.

The finding: platforms optimize combinations while brands approve parts

Three details in Amazon's REC documentation matter to production and marketing teams.

The system can continuously test combinations of images, headlines, product titles, and layouts rather than choosing only among finished ads.
Live prices, deal badges, ratings, coupons, and Prime eligibility can enter from the product listing, so a visual asset meets commerce data at delivery time.
Preview and Advanced Preview filters can inspect results by ASIN, on- or off-Amazon placement, size, language, and specific asset, but capabilities vary by market. Brand controls are currently US-only; in-banner video and multi-image carousel are available across a broader list of markets; conversational prompts remain US-only.

Amazon reports that AI-optimized REC ads saw, on average, a 14% higher click-through rate and a 7.2% higher return on ad spend. Those figures are Amazon internal data from 2025, not an industry benchmark or a performance promise. The more durable signal is the operating model: the platform is moving from choosing finished ads to choosing and assembling components.

Our earlier analysis of Amazon DVA+ focused on the information capacity required across video, display, and audio. This article addresses a different failure point: every channel can carry enough information, and the chosen components can still create a wrong combination.

Four combination errors that file-level approval misses

Product mismatch

A lifestyle image shows a black headphone while the commerce card calls a silver variant. A video demonstrates active noise cancellation while the selected ASIN only offers passive isolation. Each asset may depict a real product, but their combination changes what the viewer believes about the SKU being sold.

Claim mismatch

A headline about comfort on long trips can safely accompany a wearability scene. Pair it with an extreme-sports shot and the same words may imply impact resistance or waterproofing. The copy has not changed; the visual context has added a claim that was never approved.

Market and rights mismatch

A model, music track, or creator clip may be licensed only for the United States, social channels, or six months. An asset can be fully licensed and still be ineligible for European off-site display, Fire TV, or a longer automated campaign.

Freshness mismatch

Promotion voiceovers, old packaging, seasonal imagery, and bundled accessories all expire. A live price can update automatically; the gift mentioned in a video cannot remove itself. More automation therefore requires explicit expiry data, not a warning buried in a chat thread.

The ONCE Combination-Safe Pack

For multi-SKU, multi-market, or promotion-heavy brands, we recommend adding a Combination-Safe Pack to the production brief. This is not another ad platform. It is a six-layer instruction set delivered with the assets.

Delivery layer	What the brand and production team must lock
Invariant truth	SKU, color, features, accessories, packaging version, mandatory qualifiers
Swappable modules	Hook, product detail, use scene, proof shot, background, pre-CTA frame
Relationship rules	Must-appear-together, never-appear-together, SKU-only and claim-only elements
Rights scope	Talent, music, fonts, market, media, term, and generative-edit permission
Freshness fields	Start and expiry dates for offers, inventory, packaging, and seasonal assets
Preview sampling	High-risk combinations across SKU, market, language, placement, and size

The point is not richer file tags. It is to let both people and systems answer three questions: Which product does this module describe? What must accompany it? Under what conditions must it stop running?

Production should also shift from “shoot enough for one film” to “capture a complete set of combinable actions.” For a headphone, separate the exterior, wearability, controls, charging, storage, and environmental-use shots. Do not permanently burn the product, price, and CTA into one frame. For localization, preserve a clean master, separate captions, replaceable voice tracks, and modular end cards.

This builds on the traceability principles in our creative asset chain-of-custody framework: provenance tells you whether a component may be used; combination rules tell you what it may be used with. ONCE's cross-border production approach brings these rules into the shot list and delivery manifest before the edit is locked.

Stop pricing the job by the number of exports

Real-time assembly creates a tempting assumption: if the platform can generate many versions, production should become almost free. Automation does reduce repetitive layout and export work. It increases product-fact preparation, relationship tagging, rights verification, preview sampling, and update maintenance.

A more useful procurement equation is:

Cost per approved combination = (module production + rule annotation + legal and market review + preview sampling + maintenance) / approved SKU × market × placement combinations

The denominator is not the theoretical number of permutations. Five video clips, six images, and four headlines do not automatically create 120 usable ads. Only combinations that respect rules, carry the necessary rights, and pass risk-based review count as production capacity.

When budget is constrained, prioritize risk instead of attempting an exhaustive review:

Start with high-value products, regulated claims, and visually similar SKUs.
Test promotion, packaging, bundle, and inventory changes next.
Then test cross-market, cross-language, and off-site placements.
Finally sample routine combinations for cropping, muted comprehension, and small-format legibility.

Adobe's GenStudio approval documentation brings assets and experiences into a traceable multi-stakeholder review cycle. Celtra's explanation of dynamic creative optimization emphasizes locked templates, approval workflows, and channel-native output. These are useful signs that review infrastructure is moving toward system-level creative, but neither replaces live platform previews or the brand's responsibility for product truth.

When a single finished ad is still the right buy

If a campaign has one SKU, one market, stable pricing, a fixed placement, and no dynamic product data, one fully approved finished ad is usually more direct. The Combination-Safe Pack earns its keep through complexity; a simple job should not be turned into a database project.

Move to modular procurement when at least two of these conditions apply:

One campaign contains visually similar SKUs with different capabilities.
Price, inventory, offer, or bundle status changes frequently.
Assets must travel across markets, languages, or on- and off-site placements.
Talent, music, or creator rights have market or term limits.
The platform can rewrite copy, generate a background, or source assets from another campaign.
The team cannot see all plausible finished outputs in one default preview.

The next step: run a wrong-combination drill

Do not begin by asking how many versions the platform can generate. Take one live SKU, three clips, three images, two headlines, and one offer state. Ask the team to find five cases in which every input is individually accurate but the assembled ad is wrong. Those counterexamples will reveal the fields missing from the next production contract.

For brands preparing to supply dynamic ads, shopping assistants, or AI agents, the immediate decision is not “AI or no AI.” It is whether your content system can preserve product truth when another system chooses the final combination.