additional reviews and checks. We have always had a review for an existing set of keywords, and we are basically taking the most sensitive ones and subjecting them to an additional level of both automated and manual scrutiny.
X: You are also announcing updates to the “risk model” that you use to identify suspect advertisers. How do you define risk model, and what kinds of things can you do to improve the model?
SR: “Risk model” is a term borrowed from the credit card industry. What we are essentially trying to do is, when we see a new advertiser who has set up an account with us and has given us a bunch of keywords, we would be trying to assess the likelihood that this advertiser has a counterfeiting history. You can look at a number of signals that help you imagine the entire context around how an account gets created. Which country is the account created in? Which country are they targeting? What is the payment history of the credit card associated with the account? We examine the full context and try to make a judgment of how likely is this to be an advertiser that is peddling counterfeit goods.
This is all a carryover from earlier systems that predict credit card risk. When you take credit cards for payment, you always encounter chargebacks; if it was a stolen card you are not going to get your money. As it turns out, most counterfeiters are not only selling counterfeit goods, but they also give us stolen credit cards. We try to update our models on a continuous basis to be better and better at figuring out whether someone is a counterfeit advertiser, through a combination of techniques like the risk model.
X: How do you flush out counterfeit advertisers who might try to evade your risk model by popping up under new names or addresses?
SR: It’s a good question. I’m not going to go into trade secrets about how we do this, but imagine yourself being an advertiser setting up an ad. Think of all the context we see. We see the keywords that you give us. We see the site that the ads are pointing to. We examine the structure of the site, not just what is visible but the invisible structure, and potentially who is hosting the site. So if you think about it, there is a fairly large number of signals that we can get. Our goal is to make it pretty hard to create a new identity out of the same set of credit cards or business addresses.
That is the power of the machine learning that we use. We can look at this incredibly high-dimensional space and see the things that are likely predictors of credit card fraud and then make accurate predictions about whether somebody is going to try to place fraudulent ads. That’s the real magic of machine learning—we are training continuously.
X: You also announced that you’re working to make the manual, human part of the ad review process faster. How are you accomplishing that? Are you just hiring more people?
SR: Some of it is hiring more people. It is also making sure that we use our offices in different parts of the globe, in places like India. The goal is to work around the clock. What is important is that we do not want to make our legitimate advertisers unhappy. There are a lot of people who just want to make an honest living out of Google advertising, and making sure that manual review happens quickly is an advantage for them.
X: If you’re getting stricter about what ads you let through, won’t there also be an inevitable increase in false positives? Are you beefing up your appeals process to address that?
SR: You are absolutely right that anytime you build a system like this, there is an elaborate process of tuning the false positives and the false negatives, and that has an effect on how many people we need to support it. These are things we look at very carefully. We have an appeals process, and typically it is the same team handling the initial reviews that handles the appeals. We have to be careful not to antagonize the majority of people who just want to make an honest living.
X: What are your metrics for success? It’s not just the number of ads you reject every month, right? That could change depending on the number of actual spammers out there.
SR: There are a number of things we look at. For example, the number of complaints that we receive. Legitimate advertisers are very upset when a counterfeit advertiser is advertising on their turf. That is one thing we play close attention to. Other metrics we look at are things like how quick are we to detect and disable accounts that do this.
X: Are you concerned at all that going public like this about your anti-counterfeiting efforts is a little like waving a red cape in front of a bull? Aren’t you daring the counterfeiters to get even more aggressive?
SR: That is a worry. But at least for me, I try to focus on the positive side of the message, which is that Google takes this problem seriously and we make every effort to make sure that the experience you have with ads on Google is a safe one, with advertisers you can trust. On the flip side, as I said earlier, there is an enormous amount of money riding on areas like counterfeiting, and we understand that this is an ongoing cat and mouse game. I would not pitch this as any kind of challenge to people. This is something we have always been doing.
Here’s a Google video about the anti-counterfeiting efforts.
