This page is a map of the selection problem, not a new theorem. It links peer-reviewed computational advertising work on how a dynamic creative system should choose the next combination of elements when impression budgets are thin and most combinations never get enough clicks for a stable CTR.

Primary computational source: Chen et al. (2021), Automated Creative Optimization for E-Commerce Advertising (WWW ’21, DOI 10.1145/3442381.3449909). AutoCO searches element interaction functions, carries uncertainty with variational inference, and explores with Thompson Sampling. Online A/B result: +7% CTR versus baseline.

The selection problem

A multi-element unit is a combination: template, scene, overlay, type treatment, and the other parts the assembler can swap. Each combination can be treated as an arm in a bandit. Pulling an arm means spending impressions on that combination. The reward is a click or conversion after delivery.

Two pressures compete. Exploration spends impressions on novel combinations so the system can learn whether they convert. Exploitation spends impressions on combinations that have already converted. Dynamic creative optimization has to do both inside one auction, with a finite budget of impressions.

Sparsity

Chen et al. state the binding constraint directly: combinatorial explosion (their example reaches 2,000 creatives from modest element sets) plus thin daily impression budgets and high creative churn. A policy that explores too widely never confirms a hook. A policy that exploits the first combination to show a click stops searching the scene space. AutoCO treats that tradeoff as an estimator problem — interaction search plus posterior sampling — not as a media-buying preference.

Related attention sources in this index

Those papers measure human attention and retention. AutoCO measures algorithmic selection among composed creatives. Field measurement of a specific hook contrast in live auctions remains a separate Darwin Labs record: No. 104.