Optimising 40 AI-Generated Ad Creatives
A direct-to-consumer cosmetics brand uses AI to generate 40 ad variants every week (image, headline, call to action). A classic A/B test with 40 variants would take weeks and spend most of the budget on poor creatives. In addition, creatives 'burn out' in 10-14 days. The CMO wants to learn fast and not waste budget.
Where the money comes from
What to do now
- 1Define the bandit's reward as the margin of the sale, not the click.
- 2Set brand guardrails and pre-review of the creatives.
- 3Launch with 10% exploration and compare with a classic A/B test.
For your committee
One page with the decision, the impact, the risks, the steps and the KPI to track.
The print dialog opens: choose "Save as PDF".
Watch out: more volume is not always more profit
The creative that converts the most may be that of a low-margin product or an aggressive offer. Optimise the bandit by margin per impression, not by clicks or sales.
KPI for the committee
Margin per 1,000 impressions.
Detail for your data team
Proposed technique: Multi-Armed Bandit
Why: The goal is not to predict, but to decide while learning. The bandit balances exploration (trying variants) and exploitation (showing the best one), shifting traffic towards the winners in hours instead of weeks. Ideal when there are many variants and a short life.
Cost of getting it wrong: Exploring too much spends budget on poor creatives; exploiting too early locks in a false winner and fails to discover the best one.
Data needed: Impressions, clicks and purchases per creative in real time (ad platform APIs).; Cost per impression by channel and time slot.; Metadata for each creative: format, tone, product, colours.