[This falls into the category “Old News”, btw.]
End of last year I was allowed to observe the same struggle at two different tier 1 carriers between the advertisement department of the carrier’s portal business and the carrier’s operational “content delivery guys” (both carriers are differently structured, but for the sake of simplicity, there are some dudes that are responsible for the CDN structure and congestion). Now there’s some research why prediction of popularity of online content behaves differently from the prediction of its advertisement revenue. But here’s the story:
The ad guys are using the usual platforms to determine and fill ad space pricing. The content delivery guys have to operate their platform as cost efficient as possible and would like to see revenues at least covering their costs, but also have to deal with different campaigns and different portal businesses for different target groups. Here’s their dilemma:
The ad guys were discrediting the content guys that their prediction methods must be not up to par, as for them ad revenue prediction worked out pretty nicely. The content guys were complaining about the ad guys how they could charge always similar amounts for completely different content management lifecycles and efforts. They also pointed out the unfair fact that the ad guys treated “popularity” of written content or user-submitted links the same way as video content — they were interested in click-through rates and ad-prices — while for the content guys video and text have very different cost mechanics.
Fortunately for me, I could present some research from 
November 2008 of Gabor Szabo and Bernardo A. Huberman at HP Laboratories, titled “Predicting the popularity of online content”. They offer a method to accurately predict the long time popularity of online content from early measurements of user’s access on the examples of Digg and YouTube. While the research presents many great findings how to improve or establish prediction models on different user and presentation behavior, here is the part that helped me understand the discrepancies between the ad guys and the content guys:
If the popularity count is tied to advertising revenue such as what results from advertisement impressions shown beside a video, the revenue may be fairly accurately estimated, since the uncertainty of the relative errors stays acceptable. However, when the popularities of different content are compared to each other as commonly done in ranking and presenting the most popular content to users, it is expected that the precise forecast of the ordering of the top items will be more difficult due to the large dispersion of the popularity count errors.
and earlier:
The mechanism that gives rise to these two markedly different behaviors is a consequence of the different ways of how users find content on the two portals: on Digg, articles be- come obsolete fairly quickly, since they often most refer to breaking news, fleeting Internet fads, or technology-related stories that naturally have a limited time period while they interest people. Videos on YouTube, however, are mostly found through search, since due to the sheer amount of videos uploaded constantly it is not possible to match Digg’s way of giving exposure to each promoted story on a front page (except for featured videos, but here we did not consider those separately).
At the end, the research gives some exciting pointer to related areas of interest, which I recommend to explore:
In related works video on demand systems and properties of media files on the Web have been studied in detail, statistically characterizing video content in terms of length, rank, and comments [6, 1, 19]. Video characteristics and user access frequencies are studied together when streaming media workload is estimated [11, 7, 13, 24]. User participation and content rating is also modeled in Digg, with particular emphasis on the social network and the upcoming phase of stories [18]. Activity fluctuations, user commenting behavior prediction, the ensuing social network, and community moderation structure is the focus of studies on Slashdot [15, 12, 17], a portal that is similar in spirit to Digg. The prediction of user click-through rates as a function of document and search engine result ranking order has over- laps with this paper [4, 2]. While the display ordering of submissions plays a less important role for the predictions presented here, Dupret et al. studied the effect of document position in a list on its selection probability with a Bayesian network model that becomes important when static content is predicted [10]; a related area is online ad click-through rate prediction also [20].
