Click-fraud is a major challenge for online advertisers.
Here's how Google battles the scourge.

Click-fraud is a central issue for advertisers, for search engines, and for Google in particular. NYU Stern Professor of Information Systems Alexander Tuzhilin recently investigated how Google manages its system to detect fraud. He visited Google’s campus, interviewed Google employees, and learned about the company’s inspection and detection systems. His results were reassuring, but also showed some potential room for improvement. The following article is a synopsis of the report that Tuzhilin filed with the court in Texarkana, Arkansas, on July 21, 2006, in the Lane’s Gifts v. Google settlement.

 

The online advertising and e-commerce industries are growing rapidly. No company helped spur it along more, or benefited more from its rise, than Google. The search engine, launched in 1999, is now the basis of a $140 billion company. By aggregating content, allowing Internet surfers to search effectively, and providing a platform for advertisers, Google has emerged as a dominant player in the industry.

The virtue of online advertising is that users can easily track metrics in a way they can’t in other media. Technology allows advertisers to learn precisely how many people view an ad, how many click on it, and how many wind up purchasing the product or service advertised. At the same time, online advertisers – many of whom pay on the basis of the number of clicks on their ads – must grapple with another issue created by technology: the potential for what’s known as click-fraud. In cyberspace, after all, numerous parties have incentives to generate traffic on advertisements or websites that may not be legitimate.

In Internet advertising, the predominant model is CPC – Cost Per Click, also known as Pay Per Click (PPC), under which an advertiser pays only when a visitor clicks on the ad. (A second model called Cost per Mille (CPM), also known as CPI (Cost Per Impression), under which an advertiser pays per one thousand impressions of the ad, is also used by Google but is not subject to click-fraud and, therefore, was not a concern in Tuzhilin’s study.) The CPC/PPC model has two fundamental problems. First, good click-through rates are not necessarily indicative of good conversion rates; just because someone clicks on an ad doesn’t mean she’ll buy the product. Second, it does not offer any built-in fundamental protection mechanisms against click-fraud.

Google has two main advertising programs. AdWords, launched in 2002, allows advertisers to purchase CPC-based advertising that displays ads based on the keywords specified in users’ search queries. When a user executes a Google search, ads for relevant words are shown alongside search results. An advertiser has a certain budget associated with a keyword, which is allocated for a specified time period. Each click decreases the budget by the amount paid for the ad, until it reaches zero during that time period. If the balance reaches zero, the ad stops showing. Therefore, an advertiser or its partner can deplete the budget of a competitor by repeatedly clicking on the ad.

The AdSense program, launched in March 2003, is a way for website owners (publishers), ranging from The New York Times to small blogs, to display Google’s ads on their sites. AdSense for Search (AFS) lets Google place ads on publishers’ websites when users make keyword-based searches on their sites. AdSense for Content (AFC) automatically delivers targeted ads to the publisher’s web pages that the user is visiting. In both cases, the publishers and Google are being paid by advertisers on the PPC basis. All the partner sites in the constantly evolving network are periodically reviewed and monitored to detect possible problems.

Under this model, publishers have a direct incentive to attract traffic to their websites and encourage visitors to click on Google’s ads on the site. But overzealous and unethical users can “stretch” or directly abuse this system by generating invalid clicks, particularly in the Adsense for Content category.

 

Defining Invalid Clicks

To manage the AdSense and AdWords programs, Google collects information about querying and clicking activities. This “raw” clicking data is cleaned, preprocessed, and stored in various internal logs. When advertisers are billed, they receive customized reports describing the clicking and billing activities. But since the smallest unit of analysis is one day, advertisers cannot know if a particular click on a particular ad was marked as valid or invalid by Google, and Google refuses to provide this information to advertisers. Google defines invalid clicks as “Clicks … generated through prohibited means, and intended to artificially increase click … counts on a publisher account.” Advertisers are not charged for what Google deems to be invalid clicks.

Invalid clicks can come from a range of sources: individuals deploying automated clicking programs or software applications (called bots), low-cost workers paid to click on links, or technical glitches. Some of these invalid clicks are clearly fraudulent, while others are just invalid. Some are easy to detect, while others are very hard.

When evaluating the validity of a click, Tuzhilin notes that it is necessary to understand the intent of the user. Unfortunately, in several cases it is hard or even impossible to determine the true intent of a click using technological means. For example, a person might have clicked on an ad, looked at it, went somewhere else, but then decided to have another look shortly thereafter to make sure she got all the necessary information. Is this second click invalid? The intent cannot be operationalized and detected by technological means with any reasonable measure of certainty.

Nevertheless, Google’s Click Quality Team works to identify all invalid clicks regardless of their nature and origin. To do so, it employs a two-fold strategy of prevention and detection. First, Google discourages invalid clicking activities on its network by making the lives of unethical users more difficult and less rewarding. For example, it installs measures to make it difficult to register using false identities.

Beyond prevention, Google has built four lines of defense against invalid clicks: pre-filtering, online filtering, automated offline detection, and manual offline detection. It employs two main methodologies in seeking out invalid clicks. In the Anomaly-based (or Deviation-from-the-norm-based) approach, one may not know what invalid clicks are. But one can know what constitutes “normal” clicking activities, assuming that abnormal activities are relatively infrequent and do not distort the statistics of the normal activities. Invalid clicks are those that significantly deviate (mainly in the statistical sense) from established norms. In the Rules-based approach, one specifies a set of rules identifying invalid clicking activities (alternatively, one can also identify a set of other rules identifying valid clicking activities). An example of such a rule is: “IF a doubleclick occurred, THEN the second click is invalid.” The operational definitions Google uses cannot be released to the general public because unethical users will immediately take advantage. However, if it is not known to the public what valid and invalid clicks are, how would the advertisers know for what exactly they are being charged? This is the essence of the fundamental problem of the PPC model.

Filtering Efforts

Most of Google’s efforts are focused on the second approach: detection. Google employs a series of filters. Tuzhilin found that certain clicks are removed immediately from the logs before they are even seen by the online filters. After this preliminary stage, the next three lines of defense against invalid clicks include online filtering, automated offline detection, and manual offline detection.

"Thanks to the quality of the inspection tools, the high levels of experience and professionalism of the Click Quality inspectors, and the existence of certain investigation processes, guidelines, and procedures, these inspections are generally successful."

Online Filtering. Several rules-based online filters monitor various logs for certain conditions and detect the clicks satisfying these conditions, then mark them as invalid and remove them. The invalid clicks are removed only at the end of the filtering process; therefore, each filter sees every click. However, each invalid click is associated with the first filter in the pecking order that detected it. It turns out that the vast majority of invalid clicks are detected by the first few most powerful filters, and the last few filters in the pecking order detect only a small portion of invalid clicks that have not been detected yet by the previously applied filters. The Click Quality Team constantly works on the development of new filters and the improvement of the current set of filters.

The Click Quality Team provides only indirect evidence that Google filters perform reasonably well. For example, newly introduced and revised filters detect only a few additional invalid clicks. A recently introduced filter managed to detect only 2 to 3 percent of its invalid clicks not already detected by other filters. And the offline invalid click detection methods (to be discussed below) detect relatively few invalid clicks in comparison to the filters. This observation does not provide irrefutable evidence that the filters work well – it could be that the offline methods work poorly. But the low ratio of the offline to the online detections provides some evidence that the online filters perform reasonably well.

Automated Offline Detection. The next stage is the offline detection and removal of invalid clicks that managed to pass the online filtering stage. First, Google deploys alerts, which are used for detecting more complex and more subtle patterns of invalid clicking activities. Since these clicks cannot be safely removed by filters, the filters pass them as valid, and alerts identify them in the offline analysis stage and pass these suspicious clicks to human experts for manual investigations. Alerts can also check for various conditions more complex than those used in filters. These alerts take into consideration a broader set of deciding factors and can monitor these factors over longer time periods. When alerts are issued, they are manually investigated by the Click Quality Team, based on their priority.

Manual Offline Detection. The Operations group of the Click Quality Team conducts manual reviews of potentially invalid clicking activities. Investigation requests are generated from various sources: from advertisers noticing unusual clicking activities, from alerts, from customer service representatives who might notice something questionable, from publishers, and from an automated system that examines publishers and determines whether they are spammers.

Google’s goal is to conduct proactively as many of these investigations as possible. Another goal is to investigate the suspicious publishers in the early stages of their inappropriate activities before they receive payments. The basic idea behind most of these investigations is to discover unexpected behavior of the entities being investigated. Based on experience, the investigators look for the deviations from these “normal” behaviors using the inspection tools. Once such deviations are discovered, the investigator drills down into the problem and uncovers the reasons causing these deviations and, most likely, the source and reasons for the inappropriate activity or a set of activities. Tuzhilin has personally observed several such inspections. Thanks to the quality of the inspection tools, the high levels of experience and professionalism of the Click Quality inspectors, and the existence of certain investigation processes, guidelines, and procedures, these inspections are generally successful.

 

Evolving Eco-System

The offline invalid click detection methods detect relatively few invalid clicks. Again, this could be because the offline methods perform poorly. However, the Click Quality Team puts much thought into developing reasonable offline methods. Therefore, even if they did not perform that well, the low ratio of the offline to the online detections of invalid clicks would still provide some evidence that the online filters perform reasonably well.

The Click Quality Team provided additional indicators that led Tuzhilin to conclude that the click detection system performs reasonably well. Since late 2004, the number of inquiries about invalid clicks for the Click Quality Team has increased drastically, but the number of refunds for invalid clicks provided by Google did not change significantly. Since each inquiry about invalid clicks leads to an investigation, this means that significantly fewer investigations result in refunds. The total amount of reactive refunds that Google provides to advertisers as a result of their inquiries is miniscule in comparison to the potential revenues that Google foregoes due to the removal of invalid clicks (and not charging advertisers for them).

This evidence doesn’t provide proof beyond a reasonable doubt. And there is room for improvement. For example, Google could make greater use of classifier-based filters based on well-known data-mining methods. Data-mining methods allow for the construction of statistical models based on past data that can classify new clicks as either valid or invalid and also assign some degree of certainty to this classification.

Still, Tuzhilin believes that the indirect evidence provides a sufficient degree of comfort to conclude that these filters work reasonably well. This does not mean, however, that any particular advertiser cannot be hurt badly by fraudulent attacks. One simply should not generalize such incidents to other cases and draw premature conclusions. Also the Click Quality Team realizes that battling click-fraud is an arms race. The Internet is a constantly evolving eco-system, and Google is making efforts to stay “ahead of the curve” and to get ready for more advanced forms of click-fraud by developing the next generation of online filters.