1. Web Science & Technologies University of Koblenz ▪ Landau, Germany Online Advertising Steffen Staab

2. Introduction to Web ScienceSlide 2 of 71 http://west.uni-koblenz.de Topics  Introduction to online advertisement  Understanding the participants and their roles.  Targeted advertising.  Privacy Issues  Solutions  User based solutions  Collaborative solutions  Conclusions

3. Introduction to Web ScienceSlide 3 of 71 http://west.uni-koblenz.de Introduction  Online Advertising plays a critically important role in the Internet world.  advertising is the main way of profiting from the Internet, the history of Internet advertising developed alongside the growth of the medium itself

4. Introduction to Web ScienceSlide 4 of 71 http://west.uni-koblenz.de Facts and short history  First internet banner, 1994, AT&T.  Also in 1994, the first commercial spam, a "Green Card Lottery".  The first ad server was developed by FocaLink Media Services and introduced on 1995.  In March 2008, Google acquired DoubleClick for US$3.1 billion in cash.

5. Introduction to Web ScienceSlide 5 of 71 http://west.uni-koblenz.de Parties  Advertiser  Got money, wants publicity  e.g., Coca-Cola  Publisher  Got content, wants money  Cnn.com  Ad-network  Got advertising infrastructure, wants money  e.g., Google AdSense, Yahoo  Consumer  Wants free content

6. Introduction to Web ScienceSlide 6 of 71 http://west.uni-koblenz.de Ad embedding

7. Introduction to Web ScienceSlide 7 of 71 http://west.uni-koblenz.de Business Model  CPM = Cost Per thousand impressions  Impression: user just sees the ad.  Rates vary from $0.25 to $100  CPC = Cost Per Click  This is the cost charged to an advertiser every time their ad is "clicked" on  Rates around 0.3$ per click

8. Introduction to Web ScienceSlide 8 of 71 http://west.uni-koblenz.de Click fraud  clicking on an ad for the purpose of generating a charge per click without having actual interest.  Might be:  The publisher  Advertiser’s competitor  The publisher’s competitor  Ad-networks deal with it by trying to identify who clicks on the ads.

9. Introduction to Web ScienceSlide 9 of 71 http://west.uni-koblenz.de Online Advertising and Ad Auctions at Google Vahab Mirrokni Google Research, New York

10. Introduction to Web ScienceSlide 10 of 71 http://west.uni-koblenz.de At the beginning: Traditional Ads  Posters, Magazines, Newspapers, Billboards. What is being Sold:  Pay-per-Impression: Price depends on how many people your ad is shown to (whether or not they look at it) Pricing:  Complicated Negotiations (with high monthly premiums...)  Form a barrier to entry for small advertisers Traditional Advertising

11. Introduction to Web ScienceSlide 11 of 71 http://west.uni-koblenz.de Online Ads:  Banner Ads, Sponsored Search Ads, Pay-per-Sale ads. Targeting:  Show to particular set of viewers. Measurement:  Accurate Metrics: Clicks, Tracked Purchases. What is being Sold:  Pay-per-Click, Pay-per-Action, Pay-per-Impression Pricing:  Auctions Advertising on the Web

12. Introduction to Web ScienceSlide 12 of 71 http://west.uni-koblenz.de 1994: Banner ads, pay-per-impression Banner ads for Zima and AT&T appear on hotwired.com. 1998: Sponsored search, pay-per-click 1 st -price auction GoTo.com develops keyword- based advertising with pay-per- click sales. 2002: Sponsored search, pay-per-click 2 nd -price auction Google introduces AdWords, a second-price keyword auction with a number of innovations. 1996: Affiliate marketing, pay-per-acquisition Amazon/EPage/CDNow pay hosts for sales generated through ads on their sites. History of Online Advertising

13. Introduction to Web ScienceSlide 13 of 71 http://west.uni-koblenz.de Banner Ads

14. Introduction to Web ScienceSlide 14 of 71 http://west.uni-koblenz.de Pay-per-1000 impressions (PPM): advertiser pays each time ad is displayed  Models existing standards from magazine, radio, television  Main business model for banner ads to date  Corresponds to inventory host sells Exposes advertiser to risk of fluctuations in market  Banner blindness: effectiveness drops with user experience Barrier to entry for small advertisers  Contracts negotiated on a case-by-case basis with large minimums (typically, a few thousand dollars per month) Pay-Per-Impression

15. Introduction to Web ScienceSlide 15 of 71 http://west.uni-koblenz.de Pay-per-click (PPC): advertiser pays only when user clicks on ad  Common in search advertising  Middle ground between PPM and PPA Does not require host to trust advertiser Provides incentives for host to improve ad displays Pay-PerClick

16. Introduction to Web ScienceSlide 16 of 71 http://west.uni-koblenz.de Advertisements sold automatically through auctions: advertisers submit bids indicating value for clicks on particular keywords  Low barrier-to-entry  Increased transparency of mechanism Keyword bidding allowed increased targeting opportunities Auction Mechanism

17. Introduction to Web ScienceSlide 17 of 71 http://west.uni-koblenz.de Initial GoTo model: first-price auction  Advertisers displayed in order of decreasing bids  Upon a click, advertiser is charged a price equal to his bid  Used first by Overture/Yahoo! Google model: stylized second-price auction  Advertisers ranked according to bid and click-through- rate (CTR), or probability user clicks on ad  Upon a click, advertiser is charged minimum amount required to maintain position in ranking Auction Mechanism

18. Introduction to Web ScienceSlide 18 of 71 http://west.uni-koblenz.de Graph from [Zhang 2006] Bidding history in Yahoo! First-Price Auction: Bidding Patterns

19. Introduction to Web ScienceSlide 19 of 71 http://west.uni-koblenz.de Graph from [Zhang 2006] Bidding Patterns

20. Introduction to Web ScienceSlide 20 of 71 http://west.uni-koblenz.de 4 Targeting Populations Advert Creation Keyword Selection Bids and Budget 3 2 1 “You don’t get it, Daddy, because they’re not targeting you.” Bidding Process

21. Introduction to Web ScienceSlide 21 of 71 http://west.uni-koblenz.de

22. Introduction to Web ScienceSlide 22 of 71 http://west.uni-koblenz.de

23. Introduction to Web ScienceSlide 23 of 71 http://west.uni-koblenz.de

24. Introduction to Web ScienceSlide 24 of 71 http://west.uni-koblenz.de

25. Introduction to Web ScienceSlide 25 of 71 http://west.uni-koblenz.de 4 Targeting Populations Advert Creation Keyword Selection Bids and Budget “Here it is – the plain unvarnished truth. Varnish it.” 3 2 1 Bidding Process

26. Introduction to Web ScienceSlide 26 of 71 http://west.uni-koblenz.de Ad title Ad text Display url

27. Introduction to Web ScienceSlide 27 of 71 http://west.uni-koblenz.de 4 Targeting Populations Advert Creation Keyword Selection Bids and Budget “Now, that’s product placement!” 3 2 1 Bidding Process

28. Introduction to Web ScienceSlide 28 of 71 http://west.uni-koblenz.de

29. Introduction to Web ScienceSlide 29 of 71 http://west.uni-koblenz.de

30. Introduction to Web ScienceSlide 30 of 71 http://west.uni-koblenz.de

31. Introduction to Web ScienceSlide 31 of 71 http://west.uni-koblenz.de 4 Targeting Populations Advert Creation Keyword Selection Bids and Budget 3 2 1 Bidding Process

32. Introduction to Web ScienceSlide 32 of 71 http://west.uni-koblenz.de Daily Budget

33. Introduction to Web ScienceSlide 33 of 71 http://west.uni-koblenz.de

34. Introduction to Web ScienceSlide 34 of 71 http://west.uni-koblenz.de A repeated mechanism! Upon each search,  Interested advertisers are selected from database using keyword matching algorithm  Budget allocation algorithm retains interested advertisers with sufficient budget  Advertisers compete for ad slots in allocation mechanism  Upon click, advertiser charged with pricing scheme CTR updated according to CTR learning algorithm for future auctions Auction Mechanism

35. Introduction to Web ScienceSlide 35 of 71 http://west.uni-koblenz.de Click-through rate (CTR): a parameter estimating the probability that a user clicks on an ad A separate parameter for each ad/keyword pair Assumption: CTR of an ad in a slot is equal to the CTR of the ad in slot 1 times a scaling parameter which depends only on the slot and not the ad CTR learning algorithm uses a weighted averaging of past performance of ad to estimate CTR Click-Through Rates

36. Introduction to Web ScienceSlide 36 of 71 http://west.uni-koblenz.de Advertiser A B C BidAllocationPrice $102$5 $50 X 1 $0 $10 per click! Ad slot 1 Ad slot 2 Keywor d Algorithmic search results (Old) Yahoo! 2 nd -Price Auction

37. Introduction to Web ScienceSlide 37 of 71 http://west.uni-koblenz.de Advertiser A B C BidCTRBid x CTRAllocationPrice $100.101.02$5 $50 0.50 0.01 2.5 0.5 1 X $2 $0 (expected bid per impression) per click! Ad slot 1 Ad slot 2 Keywor d Algorithmic search results Google Single-Shot Auction

38. Introduction to Web ScienceSlide 38 of 71 http://west.uni-koblenz.de Exact match: keyword phrase equals search phrase Phrase match: keyword phrase appears in search (“red roses” matches to “red roses for valentines”) Broad match: each word of keyword phrase appears in search (“red roses” matches to “red and white roses”) Issues:  Tradeoff between relevance and competition  How to handle spelling mistakes Keyword Matching

39. Introduction to Web ScienceSlide 39 of 71 http://west.uni-koblenz.de Basic algorithm  Spread monthly budget evenly over each day  If budget leftover at end of day, allocate to next day  When advertiser runs out of budget, eliminate from auctions Issues:  Need to smooth allocation through-out day  Allocation of budget across keywords Budget Allocation

40. Introduction to Web ScienceSlide 40 of 71 http://west.uni-koblenz.de Keyword Price in 3 rd slot# of Keywords $20-$502 $10.00 - $19.9922 $5.00 - $9.99206 $3.00 - $4.99635 $1.00 - $2.993,566 $0.50 - $0.994,946 $0.25 - $0.495,501 $0.11 - $0.245,269 PPC of most popular searches in Google, 4/06 Typical Parameters

41. Introduction to Web ScienceSlide 41 of 71 http://west.uni-koblenz.de KeywordTop Bid2 nd Bid mesothelioma$100 structured settlement$100$52 vioxx attorney$38 student loan consolidation$29$9 Bids on some valuable keywords CTRs are typically around 1% Typical Parameters

42. Introduction to Web ScienceSlide 42 of 71 http://west.uni-koblenz.de Avoiding click fraud Bidding with budget constraints Externalities between advertisers User search models Typical Parameters

43. Introduction to Web ScienceSlide 43 of 71 http://west.uni-koblenz.de Adwords FrontEnd: Bid Simulations  Clicks and Cost for other bids. Google Analytics  Traffic Patterns, Site visitors. Search insights:  Search Patterns Interest-Based Advertising  Indicate your interests so that you get more relevant ads Measurement: Information

44. Introduction to Web ScienceSlide 44 of 71 http://west.uni-koblenz.de AdWords FrontEnd

45. Introduction to Web ScienceSlide 45 of 71 http://west.uni-koblenz.de

46. Introduction to Web ScienceSlide 46 of 71 http://west.uni-koblenz.de Web Analytics

47. Introduction to Web ScienceSlide 47 of 71 http://west.uni-koblenz.de 47 Distinguish Causality and Correlation. Experimentation:  Ad Rotation: 3 different creatives  Website Optimizer  E.g. 6000 search quality experiments, 500 of which were launched. Repeated experimentation:  Continuous Improvement (Multi-armed bandit) Re-acting to Metrics

48. Introduction to Web ScienceSlide 48 of 71 http://west.uni-koblenz.de 48 Google Ad Systems:  Sponsored Search: AdWord Auctions.  Contextual Ads (AdSense) & Display Ads (DoubleClick)  Ad Exchange  Social Ads, YouTube, TV ads. Bid Management & Campaign Optimization for Advertisers  Short-term vs. Long-term effect of ads. Planning: Ad Auctions & Ad Reservations.  Stochastic/Dynamic Inventory Planning  Pricing: Auctions vs Contracts Ad Serving  Online Stochastic Assignment Problems Other Online Advertising Aspects

49. Introduction to Web ScienceSlide 49 of 71 http://west.uni-koblenz.de 49 Efficiency, Fairness, Smoothness. Sponsored Search: Repeated Auctions, Budget Constraints, Throttling, Dynamics(?) Display Ads: Online Stochastic Allocation  Impressions arrive online, and should be assigned to Advertisers (with established contracts) Online Primal-Dual Algorithms. Offline Optimization for Online Stochastic Optimization: Power of Two Choices.  Learning+Optimization: Exploration vs Exploitation?? Ad Exchange Ad Serving: Bandwidth Constraints. Social Ads: Ad Serving over Social Networks Ad Serving

50. Introduction to Web ScienceSlide 50 of 71 http://west.uni-koblenz.de Itay Gonshorovitz Foundation of privacy TARGETED ONLINE ADVERTISING

51. Introduction to Web ScienceSlide 51 of 71 http://west.uni-koblenz.de Online behavioral advertising  Online behavioral advertising refers to the practice of ad- networks tracking users across web sites in order to learn user interests and preferences.  Benefits  Advertisers targets a more focused audience which increases the effectively.  Consumer is “bothered” by more relevant and interesting ads.

52. Introduction to Web ScienceSlide 52 of 71 http://west.uni-koblenz.de How ad-networks match ads  Most behavioral targeting systems work by categorizing users into one or more audience segments.  Profiling users based on collected data  Search history – analyzing search keywords  Browse history - analyzing content of visited pages  Purchase history  Social networks  Geography

53. Introduction to Web ScienceSlide 53 of 71 http://west.uni-koblenz.de How Ad-Networks track users  Cookies  3 rd Party cookies  Flash cookies  Web bug  IP address  User-agent Headers  Browser + OS  More than 24,000 signatures

54. Introduction to Web ScienceSlide 54 of 71 http://west.uni-koblenz.de Levis.com case study

55. Introduction to Web ScienceSlide 55 of 71 http://west.uni-koblenz.de Levis.com case study

56. Introduction to Web ScienceSlide 56 of 71 http://west.uni-koblenz.de Privacy  Tracking and categorizing users by the ad-networks tend to violate user’s privacy.  The gathered information, linked with the users real identity, form a violation of privacy in its most basic form.  For example, if a person is searching the web for information on a serious genetic disease, that information can be collected and stored along with that consumer's other information - including information that can uniquely identify the consumer.

57. Introduction to Web ScienceSlide 57 of 71 http://west.uni-koblenz.de So… What we have so far?  User - Preserve his privacy  Ad-Network & Publisher –  Maintain targeting and preserve their effectiveness and income  Still want to be able to fight click fraud  Questions:  Do the two goals necessarily conflict?  Or can they be both achieved?

58. Introduction to Web ScienceSlide 58 of 71 http://west.uni-koblenz.de Naive (paranoid) solution  Surf only across anonymizing proxies.  TOR  Surf in private mode  Advantages  Effective from the user’s perspective.  Disadvantages  Are proxies really anonymizing?  Very awkward  Slower  Damages targeted advertising

59. Introduction to Web ScienceSlide 59 of 71 http://west.uni-koblenz.de TrackMeNot (Howe, Nissenbaum, 2005)  Implemented as a Firefox plugin.  Achieves privacy through obfuscation.  Generates noisy queries.  Starts with fixed a seed query list and evolve queries base on previous results.  Mimics user behavior so fake queries be indistinguishable:  Query timing  Click through behavior

60. Introduction to Web ScienceSlide 60 of 71 http://west.uni-koblenz.de TrackMeNot  Advantages  Simple  Disadvantages  Still the real queries can be connected to real identity.  Might have problems with offensive contents.  Again, damages targeted advertising

61. Introduction to Web ScienceSlide 61 of 71 http://west.uni-koblenz.de Privad (Guha, Reznichenko, Tang, et al., 2009)  Require client software:  saves locally database of ads (served by the ad-network)  Learn user interests in order to match ads.  Match add from the local database according to the User interests.

62. Introduction to Web ScienceSlide 62 of 71 http://west.uni-koblenz.de Privad  Introduce new party – Dealer:  Proxies anonymously all communication between the user and the ad-network.  might be government regulatory agency.  hides user’s identity from the ad-network, but itself does not learn any profile information about the user since all messages between the user and ad-network are encrypted.

63. Introduction to Web ScienceSlide 63 of 71 http://west.uni-koblenz.de

64. Introduction to Web ScienceSlide 64 of 71 http://west.uni-koblenz.de Privad  Advantages  Ad-Networks can still target ads without violates user privacy.  Disadvantages  Complicated to add the new party.  Ad-Network has to trust the dealer in order to fight click- fraud which might unmotivated them to cooperate.

65. Introduction to Web ScienceSlide 65 of 71 http://west.uni-koblenz.de Adnostic (Toubina, Narayanan, Boneh, et al., 2009)  Two party solution:  Client side: Implemented as a Firefox plugin.  Server side: requires Ad-Network support  User’s preferences and interests are stored locally by the plugin, instead of at the Ad-network.  The targeted ad is selected by the plugin locally at the users computer, instead of at the Ad-Network servers.

66. Introduction to Web ScienceSlide 66 of 71 http://west.uni-koblenz.de Adnostic - Accounting  “charge per click” model remains unchanged.  “charge per impression” is harder.  It uses homomorphic encryption scheme.  given the public key and ciphertexts, anyone can calculate  given the public key and ciphertexts, and scalar c, can be calculated.

67. Introduction to Web ScienceSlide 67 of 71 http://west.uni-koblenz.de Adnostic - charge per impression protocol  Client: Track user activity and maintains the data locally.  Visits an Ad supported website.  Server: Sends a list of n ads ids along with public key  The browser chooses an ad to display to the user. Then creates that matches the selected ad, then send, Along with zero-knowledge proof that and each is 0 or 1.

68. Introduction to Web ScienceSlide 68 of 71 http://west.uni-koblenz.de Adnostic - charge per impression protocol  Validates the proof. If the proof is valid then using homomorphic encryption calculates when c is the price of viewing the ad.  The server save encrypted counter for each ad and add to it the previous values. Only one counter’s real value change.  At the end of the billing period, say a month, each counter is decrypted (should be done by trusted authority) and the advertisers pays for the ad- network.

69. Introduction to Web ScienceSlide 69 of 71 http://west.uni-koblenz.de Adnostic  Advantages  Ad-networks can still target ads without violates user privacy.  Ad-networks can still detect click fraud though it will be difficult without gathering information on IP even for a short time.  Disadvantages  Ad-networks become weaker.  Ad-networks can still track user if they are willing to, and the protocol is built on trust.

70. Introduction to Web ScienceSlide 70 of 71 http://west.uni-koblenz.de Measurements Pricing Experimentation Other form of Advertising:  TV Ads  Ad Exchanges  Social Ads Future of Online Advertising

71. Introduction to Web ScienceSlide 71 of 71 http://west.uni-koblenz.de Conclusions  In my opinion, It is hard to believe that ad-networks will give up the power of tracking users without legislation.  Nevertheless, There are reasonable solutions that still support targeted advertising without violating users privacy.