AI-Generated Product Imagery Governance at Haven & Oak
Haven & Oak sells rugs and accent furniture: 1,500 SKUs, 380 AI-generated lifestyle scenes produced in Product Studio, and 90 virtual try-on images placing rugs in room settings. Legal reviewed a sample and flagged 22 scenes where the rug appears roughly 30% larger relative to the sofa than the listed dimensions allow, plus 15% of generated scenes still show products discontinued last quarter. The images run in two Performance Max asset groups and one Demand Gen campaign. Marketing wants to scale generation to 2,000 scenes next quarter. It also claims the images lifted CTR by 18% in a two-week test where generated creatives ran in one campaign and originals in a different campaign, with different budgets and bidding targets. Returns on rugs already run at 14% against a 9% category benchmark.
1. What are the governance and policy risks here, and what must be in place before scaling to 2,000 scenes?
Read the worked answer
The core risk is misrepresentation, not aesthetics. A scene showing a rug 30% larger than its listed dimensions is a factual misstatement that drives wrong-size purchases, and in a category where returns already sit at 14% against a 9% benchmark, it converts directly into refund cost and margin loss. Google Ads policies on misrepresentation apply to AI-generated creative the same way they apply to photography, so scaling a known-inaccurate asset set amplifies compliance exposure, not just returns.
Discontinued products are a separate, cheaper problem. Roughly 15% of scenes wasting spend on unavailable SKUs is a feed-and-asset hygiene failure; the creative should be tied to product status so it drops out automatically.
Prerequisites before scale:
- Dimension-accurate compositing: render at true scale against a reference object, and review any scene where scale is ambiguous.
- A human review queue with a documented checklist before an asset goes live.
- Asset-to-SKU mapping so discontinued items trigger automatic removal.
- A version log so legal can trace any live image back to its source and reviewer.
- A color-accuracy check, since rug color is the top return driver after size.
2. Critique the 18% CTR lift test and design a test that would actually support the scaling decision.
Read the worked answer
The test is confounded on every axis. Generated creatives ran in one campaign and originals in another, with different budgets and different bidding targets. Any difference could come from budget, bid strategy, audience, or seasonality rather than imagery, and a two-week window is short for apparel-style consideration cycles. Worse, CTR is the wrong headline metric: a more attractive scene can raise clicks while lowering conversion if it misrepresents size, and the 14% return rate is the signal that this may already be happening.
A better design:
- Run a campaign experiment or asset-level test inside the same campaign, holding budget, bid strategy, and targeting constant, with a 50/50 split.
- Randomize at the asset level where possible, and run at least four weeks to cover a full purchase cycle.
- Pre-register the metrics: primary is cost per conversion or ROAS; guardrails are return rate, size-related return reason codes, and average order value.
- Report CTR as a diagnostic only.
- Require the size-accuracy issue to be fixed before the test starts, otherwise you are measuring the appeal of an inaccurate image.
Scale only if conversion efficiency holds and returns do not rise; otherwise the CTR gain is borrowed against future refunds.