Circle the Keepers: How to Pick AI Ad Images to Test

Photographers used to judge a roll of film on a contact sheet: every frame printed small on one page, studied with a loupe and marked with a grease pencil. An X through the blinks and the soft focus, a circle around the frames worth printing, sometimes a star on the one that became the picture. The expensive part of the shoot was already over. The judgment happened on that sheet.

AI image generation has put ad teams back in front of a contact sheet. Once variants are quick to make, the bottleneck moves to choosing them. Our free guide to GPT-4o image generation for ad creative covers the production side, from prompt structure to the testing workflow. This post is about the grease pencil: the rules for deciding which frames earn a place in a live test.

First, know which model you are judging

OpenAI introduced GPT-4o image generation on March 25, 2025, and its launch post was candid about the limitations at the time, listing cropping, hallucinations, precise graphing, multilingual text rendering, editing precision and dense information with small text (Introducing 4o Image Generation, OpenAI). OpenAI's API documentation now lists newer image models (Image generation guide, OpenAI), so the tool your team uses may not be the one a given guide was written around. The picking method below works with any of them, because it judges the output and not the model.

Lay everything out at the same size

Put every variant from a batch on one board at the same thumbnail size before anyone comments. Seen one at a time, each image gets judged against the last one you looked at. Seen together, patterns jump out: the model's favorite lighting, the same face turning up again and again, a background that keeps drifting off-brand. A shared slide or a printed page both work. Nobody picks until everyone can see the whole roll.

The X marks: reasons to reject

Go through the sheet once looking only for reasons to reject. Write the reason next to each X in a word or two. Those words matter later.

• Product truth. The product in the image has to be the product you sell: right shape, color, label and packaging. A model will happily add a feature your product does not have, and an ad that shows it makes a promise you cannot keep.
• Every letter. Read all the text in the frame aloud, letter by letter, including small print on packaging and signs in the background. If a headline matters, I would rather set it in type afterward so the words come from a font and not from the model.
• Hands, faces and people. Check anatomy, and check that no face resembles a real person. A generated person should never be presented as a customer or a reviewer.
• Claims hidden in the picture. A before-and-after, a badge, a price or a stamp that says certified is a claim. If you could not print it in the ad copy, it cannot live in the image either.
• Message match. Put the frame next to the landing page it will send people to. If a visitor would not recognize the page from the ad, the click is wasted.

The circles: what makes a keeper

What survives the X pass gets a second look for strength. A keeper stops the eye at thumbnail size, reads clearly on a phone and carries one idea. Circle the ones that pass. Then look at the circled frames together and choose which to test. They should differ in one thing you want to learn about, such as the setting or the framing, while everything else stays the same. Two frames that differ in five ways can win or lose a test without teaching you anything.

Circle fewer than you think

Every variant in a live test needs enough spend to tell you something. Split a small budget across a dozen images and none of them gets there, so the test ends with a winner chosen by noise. The number of circles should follow the budget, not the number of frames the model produced. For a modest monthly budget, two or three circled frames against a control is usually plenty, and the rest of the keepers wait for the next round.

Run the session with two people. One marks the sheet, the other argues with every circle. Give it a fixed half hour and agree beforehand who has the final say. Picking by committee drifts toward the safest frame, and the safest frame rarely teaches you anything.

The star: keep a control

Every test needs a control, and your current best ad is usually the right one. If your team has human-designed creative that has been running for a while, keep it in the test. Comparing new AI variants only against each other tells you which AI image is best, which is a narrower question than whether any of them beats what you already have.

Keep the paperwork with the file

OpenAI's launch post says images generated by GPT-4o carry C2PA metadata identifying where they came from (OpenAI). Keep the original files with their metadata, note which prompt produced which frame, and check the AI-disclosure rules of each platform you advertise on before launch, since those rules belong to the platforms and change on their own schedule. The legal and platform checklist in the guide exists because this step is easy to skip when a launch is due.

Turn the X marks into rules

The reasons written next to each X are the most useful output of the session. If three frames failed on the same product detail, that detail belongs in your prompt library as a fixed instruction. If the model keeps placing text where a platform crops the image, that goes in too. Over a few rounds the reject reasons shrink and the sheet gets faster to mark.

I will concede that some teams will find this slower than launching everything and letting the ad platform sort it out. With a large budget and loose brand rules, that can work. With a small budget, every weak variant in a test takes spend away from a strong one, and an hour with a grease pencil costs less than a week of bad impressions.

Next step

The free GPT-4o Image Generation for Ad Creative: The Practitioner's Guide covers what comes before and after the contact sheet: a use-case comparison of GPT-4o, Midjourney and Adobe Firefly, the CSCM prompt framework (composition, style, copy, mood), building a brand-locked prompt library, a legal and compliance checklist for AI-generated creative, a way to benchmark AI creative against your human-designed baseline, and fitting it into Meta, Google and TikTok workflows. Bring your first marked-up sheet with you.

← Back to all articles