|Photo: Benedict Evans|
Hence, most of the experiments that Google has launched over the years are best seen as tests to see if they fit this model. Can you apply a vast expertise in understanding data, large numbers of computer scientists and data scientists, lots of infrastructure and a model of total automation and get something interesting and useful - can you get massive amounts of new data in, can you do something unique with it, and can you surface it back out? And, for all of these, are you solving hard, important problems with global scale?
That is, Google tests new opportunities to see if they fit in the same way that a shark bites a surfer to see if they're a seal. If not, you don't change Google to fit the opportunity - you spit out the surfer (or what's left of him).
Naturally, sometimes it turns out that you need other capabilities - e.g. radio advertising. Sometimes it ought to be a good opportunity but the friction in actually unlocking the data is too great - e.g. Google Health, where there were too many different and reluctant parties involved. Sometimes Google's skills are just a condition of entry and other skills are more important (Google Plus in social), and sometimes the opportunity is just too small - e.g. Google Reader. But equally, there are projects for which Google's core skills and needs have fit very well. Maps had little obvious to do with web search and nothing to do with PageRank, but was a big problem that Google's assets could be applied to (and of course, a decade later, Maps turned out to have a huge strategic leverage in mobile). The same may be true of self-driving cars - this is not a search question, but it is a data and machine intelligence problem where Google is uniquely placed to do things (or at any rate, that's what Google believes).