Chaos Monkeys: Obscene Fortune and Random Failure in Silicon Valley

by Antonio García Martínez

According to Facebook Research, the inclusion of social context (i.e., the addition of “your friend Joe liked” to a text ad) lifted clickthrough rates something like 40 to 60 percent versus similar ads without your friend’s smiling face. That sounds like a lot, and certainly it was better than the alternative of no effect. In practice, though, this was terrible news. The clickthrough rates on Facebook ads were abysmal, mostly due to their crappy targeting. A 0.05 percent clickthrough rate was average; achieve a rate of 0.11 percent and you were ordering the truffled steak at Alexander’s in SoMa (on the client’s dime, of course). Compared with regular display advertising, Facebook ads didn’t even register; even the worst-targeted ads (e.g., those moronic “punch the monkey” pseudogame banners) got clickthrough rates of at least 0.1 percent, and well-targeted ads using first-party data and rich, dynamic creative got click rates as high as 1 percent or more. It meant Facebook ads underperformed by as much as twenty times, if you were actually doing the bookkeeping (which many FB advertisers were not). Thus, a 60 percent bump in performance was nice, but it was piddly-boo compared with the lift generated by real, piping-hot targeting data. Facebook had bet its entire monetization future on a scheme that modestly boosted clickthrough, and even then it wasn’t clear why; the Facebook Research paper mentioned a host of possible experimental confounds. I had seen an early version of...