How Deep Learning Is Changing Programmatic Advertising

Jana Jakovljevic, SVP of Partnerships at Cognitiv, is of the industry’s most experienced voices in programmatic innovation. With nearly two decades in ad tech, Jana has built her career at companies pushing the industry forward.

‍Curation Conversations met with Jana to discuss her journey -  from an Australian record label to roles at MySpace and Spotify, where she launched Spotify’s programmatic business and pioneered the first private marketplaces for audio - and how she thinks about the role of AI in programmatic today.

‍

You’ve been with Cognitiv for eight years, but you actually got your start in the music industry and worked at well-known brands like MySpace, Magnite, and Spotify along the way. Can you describe your ad tech journey?

I started my career in Australia at a record label called Mushroom Records. At the time, they had a department called New Media that updated the company’s website and worked with a few small music websites and blogs. I was an assistant in the sales department, but I was interested in what they were doing, so I volunteered to help with those projects.

I decided to leave Melbourne for London and wanted to work for a company involved in music that also had a digital element. I ended up at MySpace. For those who may not remember it, MySpace was one of the original social media platforms. I worked on the ad network team, selling inventory across Europe. My manager later left for a startup called the Rubicon Project — now Magnite — and asked me to join him. I was one of the first international employees in Europe.

At the time, SSPs were primarily focused on yield optimization for publishers, but pretty quickly buyers wanted to start cherry-picking impressions. That’s when SSPs began building their own RTB structures, which we now know as programmatic. You could say I fell into programmatic accidentally, but it felt more like I was watching the industry evolve in real time.

I moved to New York after that, where I headed up programmatic at Spotify and spent a year at IponWeb before Cognitiv approached me about a role. The founders, who have known each other since elementary school, had built an AI company focused on deep artificial neural networks. When I first spoke with them, I admittedly had no idea what that meant. This was eight years ago, before the term AI was used in every conversation.

Joining a 10-person startup that I didn’t entirely understand was a big risk, but they seemed incredibly smart, and I knew I would learn a lot.

‍

A lot has changed since you started. For one thing, Cognitiv now has over 150 employees around the world. How would you describe what the company is doing today and where it fits into the programmatic ecosystem?

We are a deep learning advertising platform. Early on, our vision was that deep learning could predict consumer behavior more accurately than humans. At the time, most people in programmatic were using machine learning, and there wasn’t much variety. If you were a buyer on a DSP, you were using the same algorithm as everyone else. We wanted to give every buyer their own custom deep learning algorithm trained on first-party data.

That has always been our mission, but it was a little ahead of its time. We developed deep learning algorithms, but existing DSPs didn’t have the infrastructure to support them, so we had to build our own. From the beginning, our vision for the company was not centered around building a DSP. But when we realized we had to become one, we built it from the ground up with deep learning in mind.  

When SSPs started building out curation tools, we realized we could use curation not to deliver an audience, but to deliver a custom algorithm. That means we can train our model the same way, but a buyer can activate their custom algorithm through a unique deal ID in their DSP of choice.

‍

What is the problem that you’re trying to solve for?

I think it comes back to delivering scalable performance for advertisers. Most programmatic buying platforms were not built with AI in mind. They were built for retargeting in the era of the cookie. As programmatic has evolved and become more complex, it’s become harder for buyers to get the performance they want. Many ended up licensing third-party audience segments or layering on contextual and viewability segments. That can work, but it isn’t scalable.

We came in and said, “Hey deep learning can handle vast amounts of data and make accurate predictions. Let's give all of this data to the deep learning algorithm and let it decide which data is important and determine the probability of conversion all in real time.”

It sounds complicated, but ultimately the problem we’re trying to solve is delivering scalable performance for advertisers.

‍

Obviously, our focus on this site is curation. Can you tell us what curation means for you and for your product?

Curation gave us a way to offer our custom deep learning algorithms within existing buying platforms, so buyers could leverage our AI without changing how they operate. We started by using curation to power custom bidding algorithms. A couple of years in, we realized we were using large language models while much of the industry was still relying on keywords. We saw that as an opportunity to expand into a contextual product.

That’s how we developed ContextGPT. It allows buyers to create custom contextual segments using prompts. For example, someone could say they want to reach consumers currently in the market for snow boots, and the system builds that segment automatically. The real advantage of curation is that we didn’t need to build new infrastructure to deliver it, and we can update supply in real time. When we see a new URL, we generate an embedding, and if it matches the buyer’s contextual segment, it’s added to their deal ID within minutes.

That matters because a large percentage of programmatic inventory is brand new every day. Curation allows us to capture those moments as content is published — whether it’s around major events like the Super Bowl or the World Cup. It lets us deliver contextual targeting that’s both intelligent and immediate.

‍

What were some of the practical challenges when setting up or scaling up a programmatic curation strategy?

There are operational challenges. It’s not a set-it-and-forget-it strategy. For each of our buyers, we have best practices for setting up a deal that depends on which DSP they’re using. Our client success team is constantly monitoring the health of these deals. They look at the requests we’re sending, the response rate from buyers, and the win rate.

What’s most important to us, however, is that we’re doing everything in real time. We don’t want to train a model on historical data and then activate it later; we want to evaluate all of that data in real time. That’s why we had to co-locate within the SSPs — so we can see bid requests as they come in as they come in to predict the probability of conversion in real-time. Co-locating comes with hardware and infrastructure challenges, but it’s key to our model.

‍

Has your thinking about curation evolved as AI and decisioning capabilities have matured?

Yes and no. We approached curation with AI from the beginning. For us, curation isn’t a standalone product or workflow layer — it’s a mechanism to enable the latest in AI. From the beginning, we believed deep learning models were going to change the industry. When we first started, that was a hard sell. People thought deep learning was too slow for programmatic advertising and wouldn’t deliver any uplift in performance.

ChatGPT has certainly changed a lot of minds, but I still think much of the industry looks at AI primarily for workflow automation. That’s an important component, but our focus has always been on real-time prediction. We want to give the deep learning model all of the data we can so it can evaluate the user and the context together and make a decision in the moment.

When we talk about AI, we’re talking about making better, smarter predictions for the client so we can optimize buying and reduce waste. That distinction is important to me.

Curation is what makes that possible. It allows us to shape the opportunity upstream — before the bid ever reaches the DSP — so buyers can access intelligence in real time, regardless of the platform they use. That’s what makes it powerful. And that’s why I believe it’s becoming such an important part of how intelligence is expressed in programmatic.

‍

Before we let you go, what’s one shift in mindset or strategy that you think advertisers or agencies should be making now to build more effective, future-ready programmatic campaigns?

Advertisers need to move from asking “What inventory am I buying?” to asking “What signals are shaping the bid decision?"

There’s a growing opportunity to evaluate richer real-time signals before a bid is ever submitted. Advertisers should understand whether their partners are simply packaging inventory, or actively shaping decision logic using those signals.

‍