Online advertising becomes instantly popular after Google launched its search-based ads. Google’s success comes from its high traffic where tens of millions of people visit the site daily to search for information. It also comes from Google’s large ad inventory, where millions of advertisers come to bid. Most importantly, search-based ads connects advertiser with the user’s active intent. This leads to high click-through rate, thus offers better results for advertiser and higher revenue for Google.
Given Google’s success, other high-traffic websites also want to enter this market. The question is: How to duplicate Google’s results? A more technical question is: how to conduct ad auction like Google and achieve high revenue from advertising?
Google ads are placed against keyword search. As a user searches on certain keywords, both search results and ads are displayed, with ads on the side. However, there might be hundreds of advertisers bid for the same keywords such as “cars” and “lawyer”. The key question is how to list these ads in reasonable order.
Google lists ads based on two main parameters: 1. Bidding price; 2. Click-through rate. The position of an ad is the combination of these two, such as Bidding price x click-through rate. For example, if an ad has $1 bid price and 0.6 click-through rate, it will be placed higher than an ad with $5 bid price and 0.1 click-through rate. A bidder pays for the bidding price right below his own (a bidder of $5 pays for $4.9 submitted by another bidder). This is called Generalized Second Price Auction[1]. Economists spent a lot of time discussing how this second-price auction is better than the first-price auction by removing incentive of keeping changing bids. In reality, the bidders still constantly change their bids to minimize cost or increase click-through rate. For example, if a bidder observes that position 2 (from bidding $4) generates the same click-through rate as position 1 (from bidding $5), the bidder would think position 1 and 2 is equivalent and will only bid to stay at position 2, and thus bids $4[2]. This is the reality in online advertisement auction. From Google’s viewpoint, it does not matter how often bidders adjust their price. All that matters is the total revenue from click-through rate and bidding price.
The dominant factor is click-through rate. If no user is interested in clicking the ad, it does not matter how high the bidding price is. How does a website increase click-through rate for the ads it shows? Making the ads relevant is very important. This comes back to the ranking of ads retrieved. Similar to the problem of search result ranking, more than 100 ads may be retrieved from a search of just 2 keywords. Ranking based on relevance is crucial to achieve high click-through rate.
Is ranking of keyword-related ads the same as the ranking of searched documents? It seems the problem of ad ranking is easier because the bidding price determines whether an ad will be placed high or low. Combined with the click-through rate, we have a neat formula for the position of each ad. Here we assume the click-through rate is known. But what if the ad is new and there is no data on its click-through rate? Potentially we can treat this new ad as having the highest possible click-through rate, and multiple it by the bidding price. After 1 or 2 days, the data come in and we can adjust its position to reflect its real click-through rate.
However, ad ranking can be harder than search result (document) when keywords searched by the user do not match any bids, even though those keywords may mean the same things as words bid by the advertisers. Thus Google encourages advertisers to bid for as many keywords as possible to increase the likelihood to be matched. Large E-commerce advertisers bid for close to a million keywords in order not to miss any potential buyer.
Even with the best effort from advertisers, there are still many keywords missed out from bids. This is because casual usage by users may differ from formal words envisioned by advertisers. It is hard to exhaust all possible ways people search or express their need.
The search company (like Google) is in the unique position of understanding the user’s need. If it needs to return high-quality research results, it needs to find a way to map those search words to existing common words, and be able to retrieve documents accordingly. Thus search company can apply the same technology for improving search results to ad selection. Query expansion is one of such techniques developed.
Another challenging situation is placing ads on pages not related search results. For example, a user may browse certain books and read reviews. Ad display in such case is called Contextual Advertising. There is only context, no search words. One solution is mapping the page content to a set of keywords, the all ads can be displayed accordingly.
Online advertising is a fast growing domain where many high-traffic website wants to create their own ad engine. People not only search for web pages, they also come to E-commerce sites to search for products. They go to social networking site to update their profiles. They go to news website to post comments. There is big potential for generating ads for each of these activities.
The technology for providing online advertising is closely tied to search relevance technology, but requires additional study on context, and on bidding behavior. At year 2010, this is a fertile ground for cutting-edge computer science research.
[1] Benjamin Edelman, Michael Ostrovsky, and Michael Schwarz: “Internet Advertising and the Generalized Second-Price Auction: Selling Billions of Dollars Worth of Keywords”. American Economic Review 97(1), 2007 pp 242-259
[2] At position 1, our bidder needs to pay the next lower bid say $4.5. Then at position 2, it only needs to pay $3.5 of 3rd highest bidder
