Which type of model to use and what's my target variable

I would say the variable 'spent' is your target variable. I'd suggest you to analyse the distribution of each feature and target to check if, eventually a linear regression could be appropriate. You might try different other methods however, like a decision tree regression, which does not require any assumptions as it doesn't estimate parameters. What I would suggest you to do, howeveer, is to concentrate on the distributions of parameters to check for eventual unbalances and make groups to make estimation robust.

If your model's output is an ad that people might like I would say ad_id is your target variable and you're building a recommendation system. For example, if Person A likes item Ad1, and Person B shows similar feature values (age, gender, etc.), then you could recommend Person B the item Ad1.

it depends on what's the question you want to get answered.

The target variable could be spent but if you want to optimize what you're spending I think that only make sense when you put it in relation with what you're trying to achieve which I guess are customer conversions.

I would approach this problem in a way that you try to predict what specific ads are bringing more conversions. The ratio between the money spent and the total_conversion or approved_conversion would be your target variable.

Assuming spent is time spent on the page, I'd say that approved_conversion is your target variable. Your customer does something (e.g. buy something, register for a newsletter...) on the site (1) or doesn't (0); and one simple classification algorithm to predict that value (0 or 1 with a certain probability) is Logistic Regression. See ISLR Book Page 131.