Not really a satisfying response, but what you want to do here is very close to what CausalPy does. I mean that in terms of estimating impacts of treatments (advertising) by comparing actual sales to estimates of what would have happened with no advertising.
Depending how many cities you have, it feels like a synthetic control / geolift type problem. Maybe with a hint of difference in differences with staggered treatment effects. I don’t think there’s anything in CausalPy right now would flat out solve the problem - but I think it could/should in the future.