In my last post, I discussed that clicks are a bad metric to optimise against in advertising systems. However, optimising against metrics that are closer to conversion do have their limits. Basically, if feedback to AI systems (e.g. Ad optimisation sy...

R-bloggers

AI ad optimization systems built for large markets like the US fail in smaller markets like Sweden or Germany due to insufficient feedback data. These systems require a minimum number of conversion events to converge, but smaller populations generate far fewer events. A 9-arm personalization test requiring 34 million impressions takes 7 weeks in the US but nearly 2 years in Sweden. Practical solutions include simplifying test designs to 2-arm A/B tests, using higher-frequency proxy metrics (site visits, add-to-cart) instead of conversions, pooling data across markets, and accepting that manual human targeting often outperforms AI stuck in perpetual cold-start mode.

The Limits of AI: Why Automated Optimization (often) Fails in Smaller Markets

Short practical advice on ad optimization in smaller markets:

The fundamental constraint: events per model