I built a custom gpt to inject 'covert' advertising into its responses with varying levels of transparency. At more covert levels it wouldn't disclose any brands unless you asked it for specific examples of products that had the benefits it would talk about. At the most covert it wouldn't even do that, but would litter its responses with enough keywords that a google search would lead you to the product it was intending to pitch.

The wild part is that it would always lie about why it was doing it, not disclosing anything about the framework or intent of injecting that content into its responses. Yes at some level it's obvious it would do that, but it's also interesting to see it first hand.