>This incentive is inevitably going to lead to ad supported LLM's not answering questions properly the first time so more ads served.

The extra inference cost is going to be greater than whatever fraction of a penny they would get from doing this.