A causal method traces ad engagement to chosen labels
In plain words
Marketers use labels such as product benefits and audience cues to guide AI-written ads. But the same labels can produce different copy, and hidden factors can affect both the labels chosen and how people respond. D-VAE is a method for estimating how labels affect engagement despite those complications.
It learns a representation of label meaning in the ad copy and predicts engagement after changing labels. In a small experiment with 30 foundation ads and 120 RED users, D-VAE’s label recommendations produced 32.7% more combined likes, comments, and collects per post than the original labels.
The research
When generated ads change from one run to another, how can marketers tell whether engagement came from the labels they chose or from copy variation and hidden factors?
The paper addresses this attribution problem for label choices that steer generated marketing copy, focusing on first-stage engagement in comparable creative contexts.
D-VAE combines shared-confounder estimation with a label-aligned representation to predict how label changes affect engagement through generated copy.
To estimate the effect of a marketing label, researchers must account for generated copy that varies even under the same labels, as well as hidden factors that may affect label choices and engagement.
The authors introduce D-VAE, which estimates shared confounding and learns a label-aligned representation for predicting engagement under alternative label configurations.
In a controlled experiment, users’ combined likes, comments, and collects per post were 32.7% higher for D-VAE-modified labels than for original labels. The comparison with C-VAE did not show a significant improvement over the original control, while D-VAE’s result was statistically indistinguishable from the expert condition.
This supports the framework’s potential in the tested setting, alongside its theoretical treatment of remaining confounding.
The user experiment covered 30 foundation-related ads and 120 recruited beauty-focused RED users.
The framework estimates first-stage engagement effects in comparable creative contexts; it does not establish results across other marketing domains or later budget, channel, or influencer-allocation decisions.
Analysis
Deep
- Original paper authors
- Danhong You, Wei Jiang, Zhiqiang, Zheng
- Submitted to arXiv
- License
- arXiv non-exclusive