The spotlight's on ad fraud, in our new series dedicated to all things CAAF.
The coalition against ad fraud has been a great success since launching in September, and we want to shine the spotlight on our partners who are positioned to make a huge impact in the space.
In this edition, we spoke to Dennis Mink, VP of marketing at Liftoff, to find out more about their approach to taking on mobile ad fraud, and how they had been working with the coalition so far.
When did Liftoff first begin to notice fraud making its way into the ecosystem?
From the day we started running user acquisition campaigns, we included a layer in our bidding system to exclude any suspicious or irrelevant bid request that we get from exchanges. Over time, we evolved our systematic exclusion as we discovered new signals of fraudulent behavior.
When we started in 2012, our systems were designed to exclude all anonymous traffic, bid requests containing blank or invalid device IDs, and “low intent” publishers which showed abnormally poor click-intent metrics and downstream conversion rates. In 2014, we began observing that certain users that we served ads to had unrealistic amounts of click activity. That year, we added the logic in our system to exclude what we call “hyperactive devices”. In 2015, our system was enhanced to exclude traffic originating from data centers, since this represents traffic from machines, not humans.
In 2016, we worked with an increasing amount of our clients to isolate and remove traffic that was due to abnormal user activity. These are often characterized with extreme values of click-to-install rates, or click-to-install time.
How has Liftoff dealt with ad fraud since then?
There are 3 things we do to limit the impact of fraud on the campaigns we run for our customers, including (1) optimizing campaigns against a cost-per-action goal, (2) systematic exclusion of bad bid requests, and (3) identifying abnormal user activity.
CPA optimization: Our most important defence against mobile fraud is CPA optimization, because it ensures that the client’s spend is aligned with real, long-term user engagement. When optimizing to a CPA target, our ML (machine learning) automatically detects publishers, user segments, and individuals which are not hitting our CPA target, and will automatically stop spending on them. This approach circumvents entire categories of fraud, like impressions, clicks, and most installs generated by bots.
Systematic exclusion of bad and irrelevant bid requests: Before we bid on an impression served by an exchange, we automatically filter out what we consider to be suspicious or irrelevant bid requests.
- Low-intent publishers. This list is generated based on abnormally poor click-intent metrics and downstream conversion rate metrics across all of our advertisers.
- Traffic originating from data centers. For instance, Amazon AWS and other hosted services. This represents traffic from machines, not humans.
- Hyperactive users: Users we've observed who represent an unrealistic amount of advertising activity (e.g. who click on hundreds of ads) and are thus fake.
- Anonymous traffic: Traffic with removed or invalid device IDs.
Abnormal user activity: When user activity deviates from the norm, we isolate these users and analyze them and look for evidence of fake advertising flows. Some of the metrics we look at include:
- Click-to-install rates: low click-to-install rates in a publisher means their clicks are very “low intent”, either because the clicks are occurring invisibly in the background, or the publisher's ad UX is spammy and generates lots of accidental clicks. We automatically downweight inventory sources with very low click-to-install rates.
- Click to install time: on average, users install an app within minutes of clicking on an ad. If the delay appears abnormally longer than this for a given publisher, the click is considered “low intent” and the user may be installing just by coincidence (i.e. they were really an organic user). On the other side of the spectrum, unrealistically short click-to-install times indicate a fraud technique called click injection. In either instance, we blacklist inventory source with abnormally short and long delays between the click and install.
What motivated Liftoff to join CAAF?
Mobile fraud is a pervasive problem in the industry that impacts both the advertiser and the audience. And while many companies are working to address the issue, we felt a coordinated approach with other ecosystem players makes good strategic sense. Within the ad supply chain, Liftoff is in a unique position in that we have end-to-end visibility of data from the exchanges (bid requests) all the way to the actions of the audience (eg. booking a ride, making a purchase). Our hope is that by working with our industry partners, we can coordinate our efforts and exchange best practices to combat ad-fraud together.
What impact has CAAF had on Liftoff’s business? Has there been a change in the way you’ve developed your technology?
We are continuously making improvements in how we detect, visualize and filter out suspicious and fraudulent traffic, and, through collaboration with other members, it is reaffirming to see that CAAF members are doing so as well.
And with the CAAF Slack channel, coordinating with other CAAF members has become a much more streamlined process.
Has a member of Liftoff’s team attended a CAAF event? What was your experience?
Yes! Liftoff-ers were present in all of the CAAF events in San Francisco, Tokyo and Berlin. We enjoyed meeting everyone and had productive discussions with the industry leaders in attendance.
Though previously not often talked about, fraud hit the headlines with high frequency in 2017. What do you think will happen with fraud in 2018?
We are optimistic that the increased exposure and awareness in mobile fraud will reduce bad supply in the industry. This will, in turn, benefit both the advertiser and their audience.
Want to join the CAAF? Get in touch with our partnerships team today to find out how to get started.