1. FLOWER AUCTIONS IN AMSTERDAM

2. Ad Auctions March 16, 2007

4. Overture slide

5. Motivation Market inherently interesting –98% of Google’s and ~50% of Yahoo’s revenues –“Future of advertising” Unusual auction rules –Multiple units, but only one bid. Continuous time. Structured market –Rules. Almost like a lab. Good data. Purely electronic market –No goods ever shipped anywhere. Flexibility to change auction rules from time to time

6. Market history & evolution early banner ads (circa 1994) Overture (Goto.com) (1997) per-impression pricingper-click pricing limited targetingkeyword targeting person-to-person negotiations automated acceptance of revised bids generalized first-price auction rules

7. Generalized first price auctions Problem: Generalized first price auctions are unstable. No pure strategy equilibrium, and bids can be adjusted dynamically. Bidders want to revise their bids as often as possible.

8. Edelman and Ostrovsky, 2007 Yahoo data from June 15, 2002 to June 14, 2003 1000 top markets 10,475 bidders 18,634,347 bids Observe bids at the quarter-hour

9. Cycling TimeMarketBidderBid 6/17/2002 6:30 AM2413$5.91 6/17/2002 6:30 AM24810$5.92 6/17/2002 6:30 AM2414$5.93 6/17/2002 6:30 AM2413$5.94 6/17/2002 6:30 AM2460$5.95 6/17/2002 6:30 AM2414$5.96 6/17/2002 6:45 AM24810$5.97 6/17/2002 6:45 AM2413$5.97 … 6/17/2002 11:30 PM2413$9.98 6/17/2002 11:30 PM2414$9.98 6/17/2002 11:45 PM2414$10.00 6/17/2002 11:45 PM2460$10.00 6/17/2002 11:45 PM2413$10.00 6/17/2002 11:45 PM24810$10.01 6/17/2002 11:45 PM2414$10.02 6/17/2002 11:45 PM2413$5.12 6/17/2002 11:45 PM2414$5.13

10. Cycling

11. Alternative mechanisms Generalized first-price Generalized second-price –Pay the bid of the next-highest bidder –First implemented by Google (2002), later adopted by Yahoo VCG –Each bidder pays the externality he imposes on others

12. Generalized second-price auctions PositionBidderBid 1A$7 2B$6 3 C$5 Payment $6.01 $5.01 $0.10

13. Computing VCG payments: example Position# clicks 1100 280 BidderValuation A $8 B $5 C $10 C’s payment: C pushes A from 1 to 2 C pushes B out completely So C should pay $160+$400=$560  loss of surplus (100-80)*$8=$160  loss of surplus 80*$5=$400

14. GSP in use Adv BidPayment A$3.01 $3.01 B$3.00 $2.81 C$2.80 $1.11 D$1.10

15. GSP versus Vickrey and VCG With only one slot, GSP is identical to standard second price auctions (Vickrey, VCG) With multiple slots, the mechanisms differ –GSP charges bidder i the bid of bidder i+1 –VCG charges bidder i for his externality “[Google’s] unique auction model uses Nobel Prize-winning economic theory to eliminate … that feeling that you’ve paid too much.” - Google marketing materials

16. Truth-telling is a dominant strategy in a single-unit second-price auction

17. Truth-telling is not a dominant strategy under GSP Intuition: Sometimes, bid below your true valuation. You may get less traffic, but you’ll earn greater profits. bidderbid A $8 B $5 C’s valuation: $10 C bids $10, pays $8 → payoff ($10-$8)*100=$200 C bids $6, pays $5 → payoff ($10-$5)*80=$400 Suppose there are 3 bidders but 2 positions. Positions have click-through rates 100 and 80. $400>$200. So C should place a bid below its valuation.

18. Bidders’ actual strategies

19. Optimal reserve prices What reserve price maximizes search engine revenue? How do outcomes differ from optimal reserve price? From the reserve price that maximizes advertiser surplus? Method: Simulate a set of vectors of valuations. Use equilibrium bid formula to compute bids. Compute outcomes under each minimum bid rule.

20. SE revenues and adv. surplus s i ~ lognormal ( 1, 0.1 2 ) maximum search engine revenue maximum total surplus and advertiser surplus

21. Number of bidders remaining

22. Individual bidders’ per-click payments bidder with lowest valuation bidder with highest valuation (in each simulation)

23. Optimal reserve prices: results set minimum bid to maximize difference SE RevAdv & Ttl Surp Min Bid0.8400 SE Rev1.0291.0130.016 Adv. Surplus0.0730.090-0.017 Total Surplus1.1021.103<0.001 p1p1 1.0751.0700.005 pKpK 0.8400 α1p1α1p1 1.0751.0700.005 αKpKαKpK 0.0030 s i ~ lognormal ( 1, 0.1 2 )

24. With more variation in valuations s i ~ lognormal ( 1, 0.5 2 ) Max SE revenue Max total surplus and advertiser surplus

25. With more variation in valuations set minimum bid to maximize difference SE RevAdv & Ttl Surp Min Bid0.7400.0000.740 SE Rev1.1741.1590.015 Adv. Surplus0.5540.581-0.027 Total Surplus1.7281.740-0.012 s i ~ lognormal ( 1, 0.5 2 )

26. With fewer bidders K=5 s i ~ lognormal ( 1, 0.5 2 ) Max SE revenue Max total surplus and advertiser surplus loss in total surplus if search engine chooses reserve price gain in SE revenue if search engine chooses reserve price loss in adv surplus if search engine chooses reserve price