Images and text now explore and exploit as one treatment

Spectare uses a multi-armed bandit algorithm to select images. Each image in your library can be tagged with audience affinities and buying-stage affinities, and the bandit learns over time which images produce conversions for which patterns.

Previously, image selection and content atom selection each rolled independently to decide whether to explore. That meant a visitor could receive an experimental image alongside a proven text atom, or the reverse, making it impossible to attribute a conversion cleanly to a single treatment.

Image selection now follows the arm assigned to the request as a whole. If the signal pattern for this visitor has been assigned the explore arm for the week, both the content assembly and the image selection explore together. If the pattern is on the exploit arm, both select the best known option. The image for a cached serve is always on the exploit path by definition, since cached pages are the result of a previously resolved assembly.

The arm itself is determined once per signal pattern per ISO week by hashing the pattern against the week boundary, so every visitor sharing a pattern shares an arm for the week and re-randomises together at the boundary. This makes each visitor's full page a single coherent experiment, and it makes the explore arm's conversion signal attributable to a real treatment rather than a random mixture of two independent decisions.