Just for clicks: the Google ad model
Traditionally, advertising is sold by salespeople who quote prices for ads. The tech company decided to auction advertising space instead
When Hal Varian arrived at Google as a part-time economic adviser in 2002, he asked the then chief executive, Eric Schmidt, how he might make himself useful. Schmidt suggested that Varian might “take a look” at the way Google sold its advertising because “it might make us a little money”. That was an understatement: in 2011, Google’s ad revenues were more than $36bn.
Nice as it would be to give Varian all the credit for this – his textbook was my microeconomics bible – the foundation stones of Google’s advertising success had been laid before he arrived. Traditionally, advertising is sold by salespeople who quote prices for advertisements. Google decided to auction advertising space instead. And when Varian, who is now Google’s chief economist, took “a look” at the auctions that Google’s computer scientists had designed, he found that they were pretty much perfect.
If you search on Google, one set of Google computers tries to deliver the best possible search results; a second set is running an auction with the aim of delivering the most effective ads to be displayed alongside them, in 11 different positions of varying prominence. An auction is run every time a user searches – billions a day.
Beyond the sheer computational demands, there are two reasons why these auctions are tricky to run well. First, these advertising spaces are substitutes for each other. If I sell flights to Reykjavik and you Google “flights to Iceland”, I want one of those ad spaces. I probably don’t want all of them, and that might also irritate users, which is in the long-run interests of nobody (except possibly Yahoo and Microsoft, Google’s rivals in this business).
Google doesn’t want to sell slots in parallel because advertisers fear winning multiple redundant slots. The solution is something called a “generalised second price auction”: the winning bidder gets the top slot and pays whatever the second bidder offered; the second bidder gets slot two and pays the third-highest bid. (This is a slight oversimplification, as we shall see.) Google’s willingness to accept less than each bidder actually offers might seem odd, but it encourages higher bids and may well raise more money overall.
The second problem is what the metric of bidding should actually be. Google could charge per “impression” – that is, for each time an advertisement is displayed. Or it could charge per “click” – each time a user clicks on the ad and travels to the advertiser’s website. The difficulty here is that Google’s costs – such as the forgone opportunity to sell space to someone else – are based on impressions, whereas the advertiser chiefly cares about clicks. I typed “Picasso prints” into Google and was offered the chance to buy some posters, but also to bid at an auction at Christie’s. I’m sure Christie’s gets far fewer clicks but is willing to pay much more for each of them.
Google’s solution is to create a “quality” metric, largely based on expected clicks, that serves as an exchange rate between impressions and clicks. If Christie’s is willing to pay $1,000 a click, and Google expects one such click, Art.co.uk will beat them with a bid of 10 cents a click, as long as Google expects more than 10,000 clicks – rightly so, since Google’s expected revenue from Art.co.uk is higher. Art.co.uk will pay a sum related to Christie’s bid and to the “quality” of both adverts.
Despite the wrinkles it is a simple idea, executed well. The biggest surprise for me is that many Google searches are “undersold”, with a few advertisers getting a bargain-basement rate – or no advertisers at all. Type “Hal Varian Google ad auctions” into Google and you’ll see no ads. Type “flowers” into Google and, I assure you, all 11 advertising spaces will be filled. It is on such searches that Google earns that $36bn.
Also published at ft.com.