Problem
“Better thumbnail, more clicks” is easy to assert and hard to prove. Views are confounded by the channel, the topic, the algorithm, and timing, so a naive comparison measures all of that at once. This study asks what the thumbnail itself causes.
Approach
[PLACEHOLDER: the identification strategy. What is the treatment, what is the counterfactual, and how do you rule out the obvious confounders.] The methodology writeup goes deeper than the results summary, on purpose, because the method is the contribution.
What I built
[PLACEHOLDER: the dataset, the model, and the tooling, as they come together.]
Status
In progress. This page is a placeholder that fills in as the analysis and writeup land. Follow the updates below.
Why it is here
It shows how I think when the honest answer is “it depends on identification,” which is most of the interesting questions in product and experimentation work.