Corsi di Laurea Specialistica in Ingegneria Informatica/Gestionale Corso di Sistemi Informativi per il Web A.A. 2005 2006 Sponsored Search Game Theory Azzurra Ragone
Sponsored Search Results in Google
Sponsored Search in common web sites
Banner ads
Search engines Search engines index billions of Web pages [Feng et al. 2007]: 800 million pages the publicly indexable Web containing 6 terabyte of text data on 2.8 million servers (Lawrence and Giles 1999); today: billions of pages over 40 million Web servers. Google gets over 250 million search request each day: over 13% of traffic to commercial sites generated by search engines over 40% of product searches on the Web initiated via search engines
Sponsored Search Sponsored links (SL) are a new marketing instrument SL are only shown to users who have expressed an interest in a search term that is related to the product that the advertiser seeks to sell. SL are a major component of the total revenue of companies that run search engine advertising. [Borgers et al. 2007]
Sponsored Search: some numbers Google: $4.65 billion (of $4.71 of total revenue) originated in sponsored search incomes (Jan-Jun 2006) Yahoo: $2.77 billion (of $3.14 of total revenue) originated in marketing services (Jan-Jun 2006) [Borgers et al. 2007]
Characteristics of Sponsored Search Pay per impression (Banner ads, es $1 for 1000 display) Pay per click Real-time auctions capturing the advertisers' true willingness to pay (WTP) Separate auctions for different search terms (ex.: 'plasma television', 'investment advice') Hundreds of thousands of advertisers compete for positions alongside several million search queries every day (e.g. Expedia and edreams for 'London travel') Auction are dynamic (change bids at any time, new auction for each search query, eg 'Olympic Games')
Sponsored Search: challenges Revenue is the product of (click price*click volume) Maximize the search engine revenue: display ads of advertisers with the greater WTP or the greater likelihood of ads of being clicked Maximize users' utility: refuse irrelevant or offensive ads Maximize advertisers revenue How much to rely on automatic filtering or how many resources to allocate to human editorial review? Finding the optimal trade-off between sponsorship and user retention
Sponsored Search: challenges (2) Management of sponsored search is ultimately a game-theoretic balance among users, advertisers and the search engine [Feng et al. 2007] The broker is a forth party to consider in the balance (e.g. Overture(now Yahoo!) with MSN, Google with AOL, LookSmart with Lycos) Challenge: maximize revenue, keeping both advertisers and users from defecting to other search engines Challenge2: detecting and ignoring robot clicks and fraudulent clicks by people with malicious intents
Sponsored search: challenges (3) Sponsored slots are scarce resource that need to be allocated carefully High demand for popular and commercially relevant search terms
Types of Ads Local advertising: targeted to users in specific geographic regions Contextual advertising: diplayed on particular web sites containing specific content (sports web sites, high-tech, etc.)
Sponsored search: Revenue Model s companies k paid slots v j advertiser j's willingness to pay (WTP) a j relevance score of listing ad j (users' utility) Hp.: 0<v j <1, 0<a j <1 f(a j,v j ) joint density function
Sponsored search: Revenue Model (cont'd) Expected revenue (ER) depends on price of each slot and the clickthroughs generated at each slot Exponentially decaying attention model δ>1 Average clickthrough α j /δ i-1 Ranking function r: I J allocate position i to company j Payment for position i: P i =p i α j /δ i-1 (p i payment per clickthroughs at position i)
Sponsored search: Revenue Model (cont'd) The search engine total traffic is a function of its overall quality The overall quality of the sponsored portion is the average relevance score of all paid listings (Σ i α r(i) /k) To model market sensitivity: λ ( 1, higher leads to a greater reduction in demand) Aggregate user traffic attracted by the search engine: (Σ i α r(i) /k) λ Search engine's placement revenue is:
Allocation Mechanisms v ranking: highest payers at the top v * a ranking: relevance and bid jointly determine rank a ranking Posted-price mechanism PS: we do not consider either external editorial control
v ranking Slots allocated based on company's WTP For a certain keyword a company j makes a bid Bj (payment per click) Second-price auction: the highest k bidders win r(i) is the firm with i th highest valuation the search engine gets a price per click v ri+1 for slot i Overture
v * a ranking For a certain keyword a company j makes a bid Bj (payment per click) Rank based on the product B j *a j, where a j approximates the expected number of clickthroughs for the listing) (A kind of) Second-price auction If company r(i) wins slot i the payment is: B r(i+1), if B r(i) >B r(i+1) the leas amount B i s.t. : B i a r(i) B r(i+1) a r(i+i) p i = v r(i+1) The price paid: or p i = v r(i+1) r(i+1) r(i)
a ranking Select the highest k bids and rank the bidders by their expected number of clickthroughs All winners pay the highest rejected bid price paid: p i =v r(k+1)
Posted-price mechanism The search engine sets a reserve price for each position for a certain period Companies are ranked based on the k highest bids and pay the reserve price The search engine determines the k reserve prices by computing the expected revenue potential for each of the k positions.
Example Two sponsored slots available for a certain keyword Four advertisers (A,B,C,D) Amount they are willing to pay: v i = (0.8, 0.7, 0.4, 0.2) relevance score a i = (0.3, 0.7, 0.8, 0.2)
Example: results First Slot Second Slot winner payment winner payment v ranking A 0.7 B 0.4 v*a ranking B 0.4 C 0.3 a ranking B 0.4 A 0.4
Bibliography Implementing Sponsored Search in Web Search Engines: Computational Evaluation of Alternative Mechanisms. Juan Feng and Hemant K. Bhargava and David M. Pennock. INFORMS Journal on Computing Vol. 19, n.1, pages: 137 148, issn:1526-5528, 2007 Equilibrium bids in sponsored search auctions: theory and evidence. Tilman Borgers, Ingemar Cox, Martin Pesendorfer and Vaclav Petricek (Unpublished), 2007