1. Social Networks 101 P ROF. J ASON H ARTLINE AND P ROF. N ICOLE I MMORLICA

2. Lecture Twenty-Two: Online advertising.

3. In the beginning, …

4. Posters

5. Newspapers

6. Magazines

7. Billboards

8. Television

9. What is being sold?

10. Pay-per-impression Price depends on how many people your advertisement is shown to. (whether or not they look at it)

11. How is the price determined?

12. Complicated negotiations with high monthly premiums, forms a barrier to entry for small advertisers.

13. And on the web, …

14. Banner Ads

15. Sponsored Search Ads

16. How are these ads different than the ads from the offline media?

17. Online ads have very accurate metrics. Was the ad clicked on? Did the click result in a purchase? can be shown to very particular viewers. (Targeting)

18. What is being sold?

19. Pay-per-click Price depends on how many people saw your advertisement and then clicked on it. (we know they read it) Except for banner ads, which are still sold largely on a pay-per-impression basis.

20. Past: Pay-per-impression Present: Pay-per-click (Typically very low price) (Typically much higher) 3 cents per impression for 2009 superbowl $50 per click for “car insurance”

21. How is the price determined?

22. Matching (per keyword) Slot One Slot Two Ad A Ad B Ad C

23. Matching market 1 2 3 A B C SlotsAdvertisers Click-through rates (# of clicks/hour) 10 4 0 Value per click 7 6 1

24. Value of advertiser j for slot i is: v ij = (value per click) x (click-through rate)

25. Market-clearing matching 1 2 3 A B C 10 4 0 7 x (10, 4, 0) = (70, 28, 0) 6 x (10, 4, 0) = (60, 24, 0) 1 x (10, 4, 0) = (10, 4, 0)

26. Market-clearing prices 1 2 3 A B C (= 10 * 4) 7 x (10, 4, 0) = (70, 28, 0) 6 x (10, 4, 0) = (60, 24, 0) 1 x (10, 4, 0) = (10, 4, 0) 40 (= 4 * 1) 4 (= 0 * 0) 0

27. What if we don’t know values? Run an auction.

28. Generalized Second Price Auction Each advertiser j announces a bid b j Slot i is assigned to the i th highest bidder at a price per click equal to the (i+1) st highest bidder’s bid

29. 1 2 3 A B C Click-through rates 10 4 0 Value 7 6 1 Slot i is assigned to the i th highest bidder at a price per click equal to the (i+1) st highest bidder’s bid Bid 7 6 1 (per click) Payments: A pays 6*10 = 60B pays 1*4 = 4C pays 0*0 = 0

30. How should you bid? Experiment: Click-through rates Write your name and your bid on your card. Point total = payoff / 10. 1 2 3 A B C 10 4 0 Value 7 6 1 Bid ? ? 1

31. Truthful Bidding is Not Necessarily an Equilibrium! (and therefore also not a dominant strategy) If each bidder bids their true valuation, then A gets Slot 1 and her payoff is 7*10-6*10=10 1 2 3 A B C Click-through rates 10 4 0 Value 7 6 1 Bid 7 6 5

32. Truthful Bidding is Not Necessarily an Equilibrium! (and therefore also not a dominant strategy) If A were to bid 5, then she gets Slot 2 and her payoff is 7*4-1*4=24 (which is higher than 10!). 1 2 3 A B C Click-through rates 10 4 0 Value 7 6 1 Bid 5 6 5

33. Bidder A bids 5, B bids 4 and C bids 1 is an equilibrium Bidder A bids 3, B bids 5 and C bids 1 is also an equilibrium (and it’s not socially optimal, since it assigns B the highest slot 1 2 3 A B C Click-through rates 10 4 0 Value per click 7 6 1 5 4 1 3 5 1

34. What are the “nice” equilibria?

35. “Market-clearing equilibrium” 1 2 3 A B C (= 10 * 4) 7 x (10, 4, 0) = (70, 28, 0) 6 x (10, 4, 0) = (60, 24, 0) 1 x (10, 4, 0) = (10, 4, 0) 40 (= 4 * 1) 4 (= 0 * 0) 0 ValueBid > 4 4 1 Market-clearing prices/matching  players maximize payoff  no one wants to raise or lower bid

36. Is There a Way to Encourage Truthful Bidding?

37. Second Price Sealed Bid Auctions Revisited If bidders values in decreasing order were v 1, v 2, v 3, …, v n Then bidder 1 would win If bidder 1 were not present, the object would go to bidder 2, who values it at v 2 Bidders 2, 3, …, n collectively experience a harm of v 2 because bidder 1 is there

38. Vickrey-Clarcke-Groves Mechanism Each individual is charged the harm they cause to the rest of the world

39. First assign items to buyers so as to maximize total valuation What is the harm caused by bidder A’s existence? If bidder A was not there, B would make 60 and C would make 4, which improves their combined valuation by 24. So A has to pay 40. 1 2 3 A B C Click-through rates 10 4 0 Value per click 7 6 1 (70, 28, 0) (60, 24, 0) (10, 4, 0)

40. What is the harm caused by bidder B’s existence? (70 + 4) – (70 + 0) = 4 What is the harm caused by bidder C’s existence? (70 + 24) – (70 + 24) = 0 1 2 3 A B C Click-through rates 10 4 0 Value per click 7 6 1 (70, 28, 0) (60, 24, 0) (10, 4, 0)

41. If items are assigned and prices computed according to the VCG procedure, then truthfully announcing valuations is a dominant strategy for each buyer, and the resulting assignment maximizes the total valuation of any perfect matching of slots and advertisers

42. The economist as an engineer: Ad quality: ad-dependent click-through rates Click fraud: incentive-based machine learning Bidding language: budgets, slot-dependent bids, enhanced targeting Competing platforms: prevalence of Google’s “mistake”

43. Next time Advertising auctions.