Write an article system that allows for multivariate A/B testing, performs multivariate regression and "decides" upon the version of the article that converts best. Then use that data for subsequent articles to suggest changes that convert better.

eg. {Wall Street Lays an Egg|Stock Market Crash 1928}

Today, {Chief Economist|Fed Chief|Money Czar} decided...

The system would randomly pick changes to the article and as it had more data it would more often pick the options that convert better. So lets say it found that "Wall Street lays an egg" and "Money Czar" converts best, then in the future when it sees Chief Economist the software could suggest that using Money Czar converts better and automatically add it as an option.

eg. Chief Economist -> {Chief Economist|Money Czar}

If you can get demographic information, lets say from facebook, then you're really golden as you can offer a service that will tailor an article to a demographic, or conversion goal. eg. articles that convert for car ads, or articles that convert for teenagers. With 5 million uniques you would probably have enough facebook connect users to get decent demographic data. Heck, if you had the facebook data you could tailor the article to the demographic profile of the user, so the system would use the general profile for unknown users and once demographic info was available would re-tailor the content. Put an API on the service and then charge a cent per milli "tailorized" articles.