A/B creative testing using predictive AI

A/B creative testing lets you compare versions of a creative before committing to one. Because the analysis runs without participants, you can test a full set of options quickly and narrow it down to the strongest one or two, then take those into a study with real participants. This article explains how to set the versions up so the comparison holds.

SETTING UP THE COMPARISON

Upload each version of the creative into the same study. Every version is analysed separately and keeps its own scores.

Three things need to match across the versions, or the difference you see will come from the setup rather than the creative. Assess every version against the same objective, since the objective controls what the recommendations optimise for.

Mark the same elements as areas of interest on each version, using the same AOI types, so the values line up element by element.

Upload every version at the same size and format, since a layout difference caused by a crop is not a creative difference.

HOW TO READ IT

Compare one measure at a time rather than judging the versions as wholes.

• The Decode Score tells you which version performs better overall against the benchmark. Where two versions score close together, the individual metrics tell you where they differ.
• Attention identifies which version is strongest on initial impact, and Engagement tells you which one holds the viewer after that. A version with high Attention and low Engagement is stopping people without keeping them.
• Focus and Clarity tell you which version gets attention to the right elements and communicates the message more easily.
• At element level, compare Time to Discover and Attention on the areas of interest you marked. On short exposure formats, Time to Discover on the CTA and the logo is usually what separates two versions that score similarly overall.

READ IT WITH A SYNTHETIC AUDIENCE

The scores tell you which version performs better for a general viewer. To understand how each version performs for a specific group, apply a synthetic audience to the study.

This gives you the scores per persona, so you can check whether a version that wins overall also wins for the audience you are targeting. Where the two disagree, that difference is the finding. See Using a Synthetic Audience in AI Creative Insights.

TAKING THE RESULT FORWARD

This is a predicted comparison, which is what makes it quick enough to run across a full set of options. Use it to shortlist, then take the shortlisted versions into a study with real participants for a measured result.