Generate-Critique-Revise Ad Loop
Improve AI ads through specific critiques, variants, scoring, and testing
- Difficulty
- Moderate
- Time to result
- ~weeks to results
- Steps
- 8
- Confidence
- 99%
Assume the first AI-generated advertisement is a draft, then run a controlled improvement loop. Review each asset for concrete defects in its claim, copy, imagery, brand treatment, hierarchy, legibility, and platform fit. Select the affected ad or component and provide explicit feedback describing both the defect and the desired replacement. Generate multiple revisions rather than relying on a single repair, then compare or score the viable headlines and concepts. Once an asset clears the quality threshold, run it with real traffic and collect performance data. Those results become evidence for the next creative batch, reducing uncertainty and making later iterations faster. The mechanism depends on precise feedback, sufficient variation, and repeated learning rather than one-shot generation.
Origin
Extracted from Marketing Against The Grain during a live critique and revision of Replit-generated LinkedIn, Instagram, and Google advertisements.
Core principles
- 01Treat the first generation as a draft
- 02Critique observable defects rather than giving vague reactions
- 03Generate enough variants to discover strong creative
- 04Use scoring as guidance, not final proof
- 05Feed performance data into the next production cycle
How to run it
- 1
Generate a baseline batch
Produce several initial ads from the researched creative specification.
Pro tip Request meaningfully different concepts rather than minor styling changes.
Watch out Do not mistake a complete batch for a finished campaign.
- 2
Audit each asset
Examine copy, imagery, logo use, color, layout, legibility, proof points, and platform suitability.
Pro tip Review ads individually at a readable scale.
Watch out A strong headline can conceal a visual execution that makes the ad unusable.
- 3
Describe the defect
State exactly what is wrong and why it prevents the ad from working.
Pro tip Name issues such as an incorrect logo, irrelevant image, overlapping text, or weak message.
Watch out Feedback such as 'make it better' gives the system little usable direction.
- 4
Specify the revision
Give a positive replacement direction for the selected element or entire advertisement.
Pro tip Retain components that already work while changing the failed mechanism.
Watch out A total regeneration may discard good copy or proof points unnecessarily.
- 5
Create enough variants
Generate multiple alternative executions until several credible candidates emerge.
Pro tip Budget for ten or more revisions when visual quality matters.
Watch out Iteration can consume substantial credits even in an economy mode.
- 6
Score and compare
Use model-assisted grading and human judgment to compare headlines, messages, and layouts.
Pro tip Ask how each candidate can reach a higher grade instead of accepting the ranking blindly.
Watch out An AI score is a prioritization aid, not market validation.
- 7
Test with real traffic
Run the strongest candidates and measure their response against the campaign objective.
Pro tip Match search copy to high-intent queries and social creative to feed behavior.
Watch out Do not infer actual performance solely from visual appeal.
- 8
Evolve the next batch
Carry observed winners, failures, and audience data into the next creative cycle.
Pro tip Preserve successful constraints and references in the revised prompt.
Watch out Failing to record learnings forces every batch to restart from zero.
In the wild
A generated ad uses an unrelated image, places an image over the word 'support,' and leaves value propositions in disconnected blue text. The marketer selects that ad, names each defect, requests a full revision, and then reviews the improved message and remaining logo and image problems.
→ The revision is materially better, while the next required edits remain explicit.
A Google search-ad output provides several headline options and grades them. The marketer compares the candidates, retains the strongest intent-matched copy, and asks for further iterations aimed at improving the grade.
→ The campaign develops a testable set of stronger search headlines instead of relying on one option.
Common mistakes
Expecting one-shot quality
AI creative commonly needs many revisions, especially for imagery, logos, and layouts.
Giving nonspecific criticism
A broad negative judgment does not identify what the model should preserve or replace.
Stopping before market feedback
Internal grading can improve candidates, but real advertising data is needed to evolve the campaign reliably.
Is it for you?
Best for
Marketers willing to spend several hours refining AI-generated campaign assets before purchasing traffic.
Not ideal for
Anyone expecting a polished, brand-accurate advertisement from one prompt and one generation.
From the transcript
“And AI is really good at scoring and grading and helping you iterate.”
“I probably need to do 10 to 20 versions of this ad and rev through it before I get something really good.”
“It will get faster and better every time.”
From the episode
Can AI Actually Make Good Ads? Replit Ad Maker Review