Click Fraud Blog
What Are Google’s Invalid Click Filters?
November 5th, 2007
Google has stated that it has multiple layers of invalid click detection. These layers are know as filters. These filters are, of course, kept confidential by Google to deter fraudsters from developing work-arounds. Using the Tuzhilin Report as my source, I discuss how I think the layers work.
Pre-Filtering
A number of clicks are pre-filtered before they hit the main filtering process. This is because they are know to be invalid at source and should not be charged for. Examples of these pre-filters are clicks coming from Google IP addresses. There is a function of all Adwords accounts to click on an Ad to test it works okay for example to ensure the landing page is valid. These types of clicks are pre-filters, are not chargeable and as a result are not passed to the filtering process.
Online Filters
Google run all clicks through what I call a rules engine. Each click is evaluated against a particular rule, if it fails, it is marked as invalid, and the click is not passed through to the billing process. There are multiple rules and the click must pass all of these rules before it becomes chargeable. What the filters are can only be guessed at, but it has been documented that a double click on an ad is one of the filters.
There is a finite number of rules which can be run on-line before the rules engine become a bottle neck, and the investment in technology to support this becomes too great. Therefore a number of offline filters are also employed.
Off-Line Filters
Once a click has passed through the on-line filters, it is marked as valid and a cost is charged to the advertisers account. This is not the last level of filtering. Offline filters are applied to the “click database” to query for and detect more complex types of invalid click.
Again the actual workings of the offline filters are confidential, but I suspect these are statistical analysis of the body of data on a user’s accounts. An example could be the analysis of trends such as low level noise clicks from a clickbotA or spotting anomalies on an account i.e. normal click trend is 9 - 5 local time, but a large number of new clicks are happening between 22:00 - 05:00. This is hypothetical to show the type of analysis I think occurs.
Off-Line filters come in two formats proactive and reactive. Proactive are automated resulting in credits to the advertisers account without their intervention, these tend to be done via code. Reactive off-line filters are started at the request of an advertiser after they have detected an issue (see Getting a Refund From Your PPC provider for details on how to start a request). These type of filters will be manual and a member of the click quality team will perform a review of activity and apply credits to the advertisers account if they detect any invalid clicks not already captured by automated filters.
Iterative Approach to New Filters
Development of new filters is an iterative process, as new methods of click fraud are detected i.e. new clickbots with new modus operandi,new filters will be added and inserted into the filter layers at the appropriate point.
Conclusion
Google’s implementation of the filters is not transparent so we never know if the filter are working correctly, or what they are detecting until we see credits applied to our adwords accounts. I have been the subject of click fraud attacks which appear to be very simple and yet I was granted credits to my account when I requested an invalid click investigation. This makes me believe that the sensitivity of many of the filters is set incorrectly and invalid clicks are missed. This is why I advise companies to invest in click fraud monitoring software and implement their own filtering.
Other posts you may like
Tags: Click Fraud
Posted in
Features |
January 6th, 2008 at 5:44 am
Interesting article. I have wondered about the click screening/filtering that Google does on AdWords advertising. This is quite informative, even though the actual details are hidden within the Google walls.
BTW, I visited here from your comment on Remarkablogger.