The providers can sell inclusion in the system prompt to advertisers. Run some ad-tech on the first message before it goes to the LLM to see whose gets included.
For most advertisers, sure, there's no need to go all the way back to the training data. Advertisers want immediate results. Training takes too long and has uncertain results. Much easier to target the prompt instead.
If you're someone like Marlboro or Coca-Cola, on the other hand, it might be worth your while to pollute the training data and wait for subtle allusions to your product to show up all over the place. Maybe they already did, long before LLMs even existed.
