Did you know that less than 60% of web traffic is human? No, the internet hasn’t been invaded by aliens — but the ‘person’ on the other end of the line could well be a bot. And bots pretending to be people aren’t just concerning, they’re bad for business. Digital advertisers pay-per-click, trusting that clicks mean eyeballs and potential sales. Fake traffic from bots throws a wrench in that equation, essentially stealing from advertisers.
“About 20% of marketing spend is lost to digital ad fraud,” explains Sven Hezel, Founder and Managing Director of fraud detection company 24metrics. “To avoid huge monetary losses, businesses need to weed out the fraudsters, and that takes a lot of time and expertise. With ClickShield, we’ve got a solution that does it in real-time, enabling businesses to stop fraud before it happens.”
Fraudsters are inventive, but they also leave clues. 24metrics detects patterns that indicate fraud with its proprietary rule-based algorithms, helping customers identify affiliate fraud. But threats evolve, and to detect the newest schemes early, 24metrics needs to beat fraudsters to the punch.
“Many of our clients run products that come with an app or registration that’s initially free,” explains Hezel. “It usually takes weeks to figure out whether these users are real, and in that time they can rack up fraudulent purchases, which can lead to big losses. We realized that machine learning can help us predict the quality of users. That’s why we decided to enhance our solutions with machine learning capabilities.”
Excellent support seals the deal
24metrics runs all its services on Google Cloud. Using a microservices approach for maximum agility, 24metrics deploys its solutions on Google Kubernetes Engine, the core of the company’s infrastructure. For database administration, 24metrics uses Cloud SQL and Memorystore. Google Cloud also offers extensive machine learning capabilities, but to leverage those, the 24metrics team needed to build internal expertise first.
“With eight engineers in total, our team is still small,” explains John Bryan Sazon, System Administrator at 24metrics. “When we’ve encountered challenges that require fields of expertise we don’t yet have, for example, in machine learning, we have looked for a trusted technology partner to guide us through them. That’s what led us to DoiT International.”
The 24metrics team got to know DoiT through contacts at Google Cloud and quickly began talking about and testing solutions together. “From the start, the DoiT team was extremely familiar with our whole Google Cloud setup,” says Hezel. “Pricing was very attractive as well because DoiT comes with no extra cost if you’re using Google Cloud. Combined with the wonderful team and great support, that’s what really made us go for DoiT.”
In an initial call, DoiT gave a broad introduction to its platform and work processes. It then set up a Slack channel, ensuring quick access to support. “The onboarding process was great,” says Hezel. “Our shared Slack channel was a nice way to stay in touch with the DoiT team. It’s a very non-bureaucratic way of working together, which has been going really well from day one.”
Defining the problem, training the model
With no bureaucracy in the way, the fraud specialists could focus on the task at hand: building machine learning capacities at 24metrics. In initial talks, the team defined the scope of the challenge and outlined the path forward. “DoiT helped us find the right tools and laid out the relevant features of the Google Cloud solution AutoML,” explains Stavros Charitakis, an engineer at 24metrics. “After that initial session with the DoiT team, we were able to train our first model ourselves.”
Building machine learning capabilities from scratch is a learning process, and when initial results weren’t up to the standards the 24metrics engineers set for themselves, they consulted with DoiT. “DoiT really helped us analyze our results and detect potential issues with our machine learning training approach, and offered alternatives,” says Charitakis. “By following those recommendations, we soon arrived at a really well-trained model, which DoiT helped us to deploy in a cost-efficient way.”
Building ML capabilities at a rapid pace
From training to deployment, the efficiency of the whole process surprised the entire team. “We thought this project would take more than five months and involve more engineers,” says Charitakis. “But with close support from DoiT, it ended up being quite easy and much faster than expected. It only took us two months to build the machine learning algorithm and deploy the new feature. DoiT guided us through the exact steps we needed to take, ensuring we were on the right track.”
Ultimately, it’s the customers who decide whether a business is on the right track. By that metric, 24metrics has knocked it out of the park. With the new machine learning capabilities, ClickShield delivers early insights into the quality of traffic, and customers are already flocking to the new features.
“60% of our existing user base, especially in the mobile space, is already using the new ML-based services to improve the quality of their traffic,” says Hezel. “We’re always in touch with our customers, and they tell us it’s fantastic. And that really makes us happy as a team. With support from DoiT, we’ve built something that customers love, that really helps them in their everyday work.”
Stronger campaigns with less fraud
It’s not always easy to discern real from fraudulent traffic, but 24metrics’ intuitive task reports help customers to quickly see the improvements in their campaigns. “With our solution, clients can set executive dashboards to observe campaign metrics and the quality of their users,” says Charitakis. “Soon after launch, we received very positive feedback, and our customers are excited about their gains. They have seen 4.2 times higher accuracy than the previous traditional methods, and seven times faster results at the same time.”
For this to work flawlessly, ML models must be deployed in a holistic cloud environment. Because AutoML is part of the larger suite of Google Cloud AI products, performance has been top-notch. “Because we use AutoML for online predictions, latency is very important for us,” says Charitakis. “By deploying our machine learning models in our Google Cloud infrastructure, we’ve been able to achieve amazing results.”
Exciting customers with new features
As word gets around about 24metrics’ new solution, there’s growing interest from new customers. Some have chosen ClickShield over competing solutions because of this unique set of features, says Hezel: “Our ML feature is the only one of its kind out there currently, and we’ve been getting new customers just because of it. It’s added value to our already strong solution and directly improved our revenue and client retention.”
Looking back, the ease with which 24metrics bolstered its platform with ML expertise excites the team: “We thought this whole project would be a lot more difficult, and without the right support, it certainly would have been very time-consuming,” says Charitakis. “DoiT helped us navigate this new environment to the point where it’s something we don’t actually need to spend a lot of time on.”
Instead, 24metrics spends its energy on the fight against fraud. Next, the team wants to use the newly built ML capabilities to add more filters to its anti-fraud solution and adapt it to more industries and verticals, especially in the e-commerce space. With the ability to predict purchase values, 24metrics could help online stores improve conversion rates. From fraud detection to behavior predictions: 24 metrics is on its way to becoming a one-stop shop for companies who want to maximize the effectiveness of their campaigns.
“24metrics has always been about pushing boundaries and innovation, and that means we need to get our feet wet in a lot of new technologies,” says Hezel. “Partnering with the right people, such as DoiT, really enables us to get new features out faster and better than our competition. Fraud isn’t going anywhere — but we’ve got the solutions to stop it before it happens.”