Why Doesn’t Google Use VCG?
After learning about VCG and GSP auctions in class, I was interested to find out why some of the big tech companies, like Facebook, use VCG auctions to sell ad slots, while others, like Google, use GSP.
In the paper linked below, The VCG Auction in Theory and Practice, the Chief Economist at Google, Hal Varian, compares the GSP and VCG auctions and then discusses Google’s decision to keep the GSP as their algorithm of choice for auctioning off ad slots. One key, and important, scenario where VCG and GSP differ is that of broad matches. The advertisers who participate in the auction are only those whose key word matches the users query. However there are queries that are very similar but not exact matches; these are called “broad matches”. For example, if the advertiser’s key word is “vegan restaurants”, a potential broad match is “plant-based restaurants”. The advertiser may see a value in advertising for this broad match, but it is probably different than their value for a direct match to their keyword. Similarly, different broad matches may have different values to the same advertiser. The paper explains that for VCG auctions, advertisers can state their average value for broad matches and the algorithm works out pretty well. However, for GSP auctions, this is not the case. Advertisers can appear in different orderings for different auctions, thus complicating the procedure.Thus the only scenario where GSP is good for broad matches is when the impact of a broad match on the value change is very small and therefore insignificant. The article then also follows with an explanation for why the VCG auction is still implementable even without knowing click through rates.
So, if VCG auctions work better than GSP auctions in these cases, why does Google still use GSP?
One of the main answers is that it is hard to change systems. The original GSP auction was designed and implemented by Google in 2001. They realized there is a problem with not getting truthful bids, and the solution one member came up with was akin to the VCG auction. However, Google did not change their system at the time because it was hard to engineer the new system while maintaining the original, which required a lot of attention. Also, they believed it would be harder to explain VCG to advertisers, and it would require advertisers to greatly increase their bids. Google reconsidered changing to a variant of VCG a decade later in 2012, but by then there were other systems in place, such as dynamic resizing, that worked better for Google’s specific use cases. Finally, the paper also mentions that VCG systems are a lot harder to implement in reality because the engineer must consider a lot more edge case, increasing complexity.
Overall this paper gave a great insight into how the concepts we use in class are used in the real world and into how companies make important decisions like their ad pricing scheme.
Soure: https://courses.cs.washington.edu/courses/cse490z/20sp/slides/Varian.pdf