Recent Twitter outages have caused much consternation. These seem to be symptoms of recent rapid growth. With a platform business like Twitter that does not charge its users, it gets no marginal revenue but incurs marginal costs as its user base grows, which leads to cashflow problems.
Obviously the strategy is to raise revenue from other sources such as advertising or making some other use of Twitter data. The problem is that it’s hard to increase these revenues in direct sync with the user base. Scaling problems become more acute in this type of business model where an additional users brings additional costs immediately, but brings marginal revenue later.
Interestingly, some Twitter users are so upset by the outages that they are organising a ‘charity’ drive to raise money for the company to help it buy servers. Another new revenue model?
Facebook’s “walled garden” strategy is sometimes annoying. I was thinking it would be nice to have a little social aspect to my blog. I can create a group related to my blog on Facebook, but to join the group or see who’s a member etc, you actually have to go to the Facebook site. I doubt that many people will be bothered to do that. Instead I’d like to have a little widget on my own site where people can sign up to the group and easily view the members and any other data related to the group. Unfortunately, this doesn’t seem to be possible, as Facebook wants to keep people coming back to facebook.com.
I agree with The Economist that this is going to backfire for Facebook. There’s nothing special about social networking that means it must be embedded only in facebook.com. The same idea could be set free, “widgetized” and embedded anywhere. A more open and “sociable” platform could easily become more attractive than Facebook’s garden. Such a platform would probably be a lot harder to monetise than a garden that generates a lot of traffic, but it would certainly be a lot more useful.
Google has announced a new display advertising platform for small and medium sized websites. Unlike Adsense, websites hosting these ads get paid by the number of pageviews, rather than clicks.
Google’s service is going to be free to web publishers that host the ads. There are a number of other companies that offer display ad platforms like Google’s, and they typically charge web publishers for their service. So one might wonder if this is predatory pricing by Google. Although the marginal cost of serving an ad will be very small, it will not be zero, so Google is pricing its service below marginal cost to web publishers. Isn’t that anti-competitive?
The answer is that this is a two-sided market and things are more complicated. Google has two groups of customers for its advertising platform — advertisers and web publishers. It can raise revenue from both, and costs are jointly generated by both sides. Google will set prices on both sides of the market jointly, and we cannot look at the price on one side in isolation. So just looking at the price charged to web publishers and observing that it is below marginal cost does not tell us that Google is behaving anti-competitively. We also need to consider prices charged on the advertiser side.
This paper (pdf) has a very good explanation of some of the mistakes you can make if you apply one-sided thinking in two-sided markets.
This article in Wired has a great account of the contest for the Netflix Prize. In late 2006, Netflix offered a US$1m prize for anyone who could come up with a movie recommendation algorithm that performed 10% better than its own. So far, the prize is unclaimed, although the leaders are close on aroud 8.75%.
The Wired article is mainly a profile of one of the contestants, Gavin Potter, who was a solo late entrant and quickly beat some of the bigger teams. He’s currently ranked eighth with an 8.14% improvement. He makes an insightful comment:
“The 20th century was about sorting out supply,” Potter says. “The 21st is going to be about sorting out demand.” The Internet makes everything available, but mere availability is meaningless if the products remain unknown to potential buyers.
What Potter is saying is that in the past few decades, quite a lot of business progress was made on the supply side of things — just in time, outsourcing, supply chain management etc. Now the Internet has released shelf-space constraints and opened up the fabled long tail of demand. However it’s not enough just to have a million products in your catalogue, you also have to have a way to match consumers with products as effectively as possible. This matching is a crucial aspect of a business like Netflix, and is the reason why they’re willing to pay a million dollars for a ‘mere’ 10% improvement in matching quality.
Following on from my previous post about the value of content filters, here’s some more thoughts on how to do better filtering …
Like I said, the big problem is finding the gems amongst all the crap. I find myself continually frustrated with the lack of sophistication of the filters that YouTube and similar sites offer. Basically, you can sort items according to popularity or other people’s star-ratings over various periods of time. But the good or unique stuff doesn’t always make the top-10 lists.
One idea for a better filter is just to have more flexibility in the way that ranked results are presented. Check out this video for example:
It has about 375,000 views, which is a decent number, but the most popular videos have 30 million or more views, so if you just rank all videos by views you’ll have to scroll through dozens of pages of results until you get down to the 300,000 views range. So why not let users define the lower and upper limits of the rankings in their search results? For example if you graph all the videos on YouTube according to their number of views, you’ll probably get the fabled “long tail” distribution. This could be presented to users and they could select a range within the ranking to view:
Another thing that would be nice is to be able to identify videos that are likely to become hits but haven’t reached that status yet. I suspect that the path of total views over time for hit videos follows the typical S shape. If YouTube can track downloads over time, they could identify videos that may be in the acceleration stage of this process:
A more sophisticated approach to this would be to use time-series models to try to forecast future popularity. Unlike, say, stock prices or exchange rates which behave like random walks, I suspect there is a lot of forecastable serial correlation in daily views for any given video. It’d be interesting to see if some simple models can reliably predict which videos are likely to become popular once they’ve been on the site for a little while and accumulated some views.
Another alternative approach would be to try to identify users who are good at identifying popular content early on. Are there some users who consistently gave high ratings to videos before they became highly popular? If so, these users’ ratings could be tracked (secretly, so as not to induce them to manipulate the process) to generate video recommendations.
These are just a few random ideas that I had. It seems that with so much data available, a few simple tools could go a long way to helping people to sort the good from the bad, which is obviously a valuable thing for sites like YouTube.
