What temperature and top_p do you use for ad copy with the API?

if your model still accepts it, a small presence penalty helps too, it stops every headline in the list starting with the same word. and put the character limits in the prompt but still check them in code afterwards. it goes over more often than you'd expect.

change one, not both. that's the general advice and it matches what i've seen. but the variety problem is mostly the prompt, not the sampling. asking for ten headlines in one call, each with a different angle (price, speed, pain point, social proof, question), gets way more spread at temperature ~0.8 than ten separate calls at 1.2. separate calls don't know what the others said, so they converge.