They can easily LoRA-finetune each model based on user preferences expressed in the past conversations. That would improve accuracy compared to Google's ad targeting by orders of magnitude.