Inside this resource.
8 sectionsA map of the published document. Choose a section to start reading.
- 01What this sheet is for
- 02Message: read the words apart from the picture
- 03Product: the ad shows the thing you sell
- 04Claims: nothing the product can't back up
- 05Test integrity: the test measures what you think it measures
View 3 more sections
Jump to a section
What this sheet is for
An AI creative system is only as good as the ads it refuses to ship.
Image checks catch a melted label or a sixth finger. They don't catch a headline that, in a new layout, sounds like the brand admitting the product doesn't work. That ad can pass every visual review and still cost you money.
Run these 14 checks before any AI-generated ad reaches a budget. Each one takes minutes. An ad ships when all 14 pass, or when every failure has a recorded repair.
Message: read the words apart from the picture
- 1. Cold read of the headline. Have a reviewer that never saw the brief read the headline and overlay text alone. Write down what they think the brand is saying. If it isn't what you meant, the ad fails here.
- 2. Who is speaking? Read the words and the layout together. A customer's complaint printed as a headline reads as the brand's own admission. If a quote belongs to a customer, the layout has to show that (a speech bubble, a review card, a name).
- 3. Offer matches checkout. Any price, discount or bundle in the ad matches the live product page and can be redeemed at checkout exactly as written.
Product: the ad shows the thing you sell
- 4. Product fidelity. Compare the finished file with the real product: label text, colour or shade, texture, pack shape and size. Check it at the size the shopper will actually see, not zoomed in on a monitor.
- 5. Product intent. If the concept is meant to show no product, the generator received no product photos, and the reviewer isn't marking the ad down for a missing product.
- 6. Text survives the crop. Open every placement size you'll run (square, vertical, landscape) on a phone. Overlay text stays legible and nothing important falls outside the safe area.
Claims: nothing the product can't back up
- 7. Every benefit traced. List each benefit the headline, overlay or image implies. Each one maps to evidence for this product as sold. AI edits can add a benefit nobody wrote, such as a glowier face, a richer texture or a slimmer waist. Treat those as claims too.
- 8. Real quotes, unchanged. Customer words come from your actual review or quote library, with permission to use them, and aren't trimmed into a stronger claim than the customer made.
- 9. No generated results. No generated image depicts a result, a before-and-after or a person's outcome as if the product produced it.
Test integrity: the test measures what you think it measures
- 10. Refresh lock. When you're refreshing a winner, the headline, angle and claim are locked and only the presentation changes. Rewriting everything at once makes the test unreadable.
- 11. Reviewer sees current files only. After a replacement image is generated, the reviewer reads the replacement, not the rejected original. Otherwise an ad keeps failing for an image already thrown away.
- 12. Inputs actually arrived. If the workflow says something is missing (buyer quotes, product references, brand rules), check the disk before believing it. Research that was gathered but never reached the generator is the cheapest fix in the whole process.
Cost and record: learn once, not every batch
- 13. Price a pair before the batch. Render two short tests, one with your change and one without, and compare the provider's charge with its quote. Small differences multiply across a batch.
- 14. Write the rejection record. For every ad you kill, record the defect, the check or source that exposed it, and one decision: repair this version, return to the brief, or drop the concept. Then write what changes in the system so the same mistake can't come back.
The question to ask any AI creative vendor
Before you scale an AI creative workflow, yours or someone else's, ask to see an ad it rejected and the reason it failed.
Then ask what stops that mistake from happening again.
A team that can answer both has quality control. A team that answers with another 100 variations has a generator.
Want a second pair of eyes?
We build and run AI creative systems for ecommerce brands at Tegra, with these checks wired in before anything reaches a budget. If you'd like to walk through your own workflow, book a consult.
Choose who handles the next step.
Use the reference yourself, book a focused paid working session, or have Tegra handle the ongoing channel work.
Ongoing media buying, customer research, creative strategy, copy, static and video production, campaign deployment and optimization.
Explore managed Meta Ads → · See paid working sessions →