AdRoll CTO Valentino Volonghi gets real about AdRoll’s plans for big data.


Q: When did you start working at AdRoll, and what had you been doing before then?

VV: In 2006, I moved to the US from Italy to work on AdRoll’s real-time bidding algorithm. My background has always been working on distributed and high speed infrastructures—typically, people don’t think of the advertising space as a glamorous place for an engineer, but the reality is that advertising pulls together some of the most advanced big data and machine learning problems that make it a really interesting field to work in. We also have some of the few really real-time systems in the world, where you have to respond within a short amount of time—less than a hundred milliseconds—for a request to go through.

Q: You mention that there’s a lot of big data in advertising. How much data do we actually process, and what does that mean in layman’s terms?

VV: Every month we ingest more and more data. If you look at the 1,800 customers registered with the New York Stock Exchange (NYSE), they do about 1.5–2 billion transactions a day. Compare that to our 20,000 customers, and 60 billion transactions we process each day. The NYSE generates about 5–6 terabytes of data in a day; we generate 150 or so. In 3 days, we process more data than the NYSE does in a year.

Q: So it sounds like we’re investing a lot back into our technology?

VV: Yes. In the last year, we tripled the size of our engineering team. Machine learning is used a lot in advertising, but not a lot of companies get to the scale that we have. We had to build out our own platform to do machine learning because nothing was available in the open source market or off the shelf. The scale of data generation that AdRoll has reached was beyond what most companies deal with to this day. It took a long time, but the payback now is that our bid distribution—the distribution of the amount of money that we bid for every bid impression—is well spread across the entire range of possible bids, instead of being centered in a few specific values. It will bid very low for impressions that just aren’t that valuable to advertisers, and much higher for impressions that we’re reasonably sure users will click.

Our bidding algorithm evaluates over 100 million variables in 500 microseconds, 60 billion times a day. How recently a user has been to the advertiser’s site, how many times they’ve visited before, how many ads they’ve seen already—all of these variables end up contributing to the calculations that determine our bid on an ad impression.

Q: How has our bidding algorithm changed in the last year? What can we do in 2015 that we couldn’t in 2014?

VV: Because we’ve built all of the automation around our bidding algorithm, we can now focus on adding new features as fast as we can. In just 3 months, we did 6 new releases with new features. Most of what we’re working on now revolves around using the data that we’ve accumulated to improve campaign performance for advertisers.

We’re really beginning to appreciate the fact that every site is not a silo – the web is very liquid and people go everywhere, browse in different sites, different times of day, for different reasons. We’re building our technology to better understand this customer journey and make it easier for marketers.

Ultimately, this means that we’ll be able to attract more qualified visitors, visitors that bounce less, and better leads for advertisers.

Q: Let’s say a marketer wants to reach people based on information like where they work, or what their title is within a company. How does AdRoll’s secret sauce factor this in, and what can we do that allows us to target people just as well, if not better? 

VV: Well, you can have information laid down statically, but it’s not always true that the deciding force inside a company always goes by the same title: CEO, CMO, etc. What you see more often is that people in B2B companies who do make product decisions typically display similar types of behavior. The key behind AdRoll is that we can properly track this behavior. With all of the cross-learning we’ve accumulated across our different customers, we can determine the best time of day, place on the web, and device to show a specific message to a specific person interested in your product. A lot of people might go browse a specific document and make the final product decision, without being C-level. It’s hard to say who this person might be based on their title or demographic information alone, but much easier to identify based on their behavior and interaction with your site.

We’re keying off of observed behavior and not just guesswork. Instead of guessing that the VP of Marketing at X-company is the right person to be targeting with our ads, we’re showing the ads to the people who are most likely to convert and actually involved in the purchasing decision.

Q: Do you think there are specific variables in our algorithm that are relevant to to B2B, or do you think that across the 100 million intent signals it applies across all verticals? 

VV: Our ability to create specific segments and specific strategies work not just for short buy cycles like retail, but for very long, multi-week buy cycles for B2B or tech. We know how to apply the proper segmentation strategy; we know that conversions don’t just happen when there is a purchase event. Having flexible segmentation helps the algorithm find exactly where the high value customers are so you can efficiently reach them at the right price. AdRoll segments are open and flexible, so we can adapt for different verticals. That’s the nature of the system we’ve built, it’s not assuming that everyone is retail.

Q: Looking forward, what are the main performance metrics that we’re trying to improve? 

VV: ROI, post-click conversion, post-click CPA, and click-through conversion rates as much as we can. My mission: I feel responsible for improving the post click CPA, the ROI, and we will do everything that’s possible technologically to enable us to drive better and better ROI for your business.

To learn more about how you can use your data to drive business objectives, download our free guide: Attribution for Data-Driven Marketers.