Measuring and Improving Persuasiveness of Large Language Models