Innovators Unscripted With Cognitiv

Want more? Check out the Innovators Unscripted Bonus Round with Jana Jakovljevic.

Cognitiv’s Jana Jakovljevic on bringing real-time contextual advertising to the sell side

How does deep learning enable a more holistic approach to contextual advertising and optimization?

Jana Jakovljevic: The great thing about deep learning is that it can ingest vast amounts of data and find patterns. It can evaluate not just the context on the page using large language models, but also evaluate what’s the time of day? What device is this person on? What do I know about this person?

“We’re not just delivering more relevant, contextual targeting. We’re understanding that we also want to serve it to this user, serve it to them at this time or in this context, and we want to deliver this message.”

Paul Zovighian: So it’s not just siloing in and zooming into one KPI or one objective, it’s taking all the different elements and taking that combined approach.

JJ: Absolutely, I think in programmatic advertising, everything has been so siloed for so long. You might have your audience strategy and then you might have a contextual advertising strategy. But traditionally that’s just been an off-the-shelf strategy, or it’s done with keywords. Now we can use large language models to really understand that page with all the nuance of the human mind.

The term “real time” is often used to describe programmatic buying, but it’s not always accurate. Why is speed so important for algorithmic decisioning?

JJ: When we talk about real time, we think of it as real-time advertising or real-time bidding—you’re returning a bid within however many milliseconds. But what happens in terms of optimization and algorithms is people will say this is real time, but they’re actually doing this modeling offline.

They’re using historical data to make decisions for today or in the now. That historical data might just be a few hours old, but still, you’re taking something that happened in the past and trying to apply it to today.

For us, we’re not just doing real-time bidding or real-time advertising, we’re actually making real-time predictions.

“The great thing about deep learning is that it can evaluate that user, that context, and that bid request in real time. We’re operating these models in less than 10 milliseconds, so having ‘real time’ for performance is hugely important.”

PZ: It’s decoupling the act of the transaction, which, yes, happens in real time. You go on a page, you’re shown an ad—that happened in real time. But the motivation and the choice to go and serve that in that moment, hasn’t always been in real time.

JJ: And for the most part, it’s still not happening.

What’s the biggest unlock you’ve seen from integrating with Index Marketplaces and activating AI models on the sell side?

JJ: The great thing about our integration is that it’s truly real-time.

“We have the opportunity to deliver our full deep learning capability to any buyer in their DSP of choice through Index.”

We’re receiving that bid request, and we’re evaluating for all of our clients running on Index, what’s the probability of conversion for this bid request, for this user, in this context? We’re doing that in under 10 milliseconds, even five milliseconds.

It gives buyers the opportunity to leverage that sophistication, regardless of which buying platform they’re using.

With the increasing use of AI in advertising, how does Cognitiv address concerns related to bias and ensure responsible AI practices?

JJ: It’s an important question. I think in two main ways. One, human oversight. And secondly, transparency.

In all of our models we’re using the client’s first-party data, but we’re very careful to ensure that data is a large representation of the population. We’re always looking at the features that we’re using and how that model is valuing feature importance to ensure it’s not biased in certain demographics.

Secondly, with our contextual product, which uses a GPT-based large language model, there could be a lot of bias in large language models, especially depending on how you prompt it. So with our contextual advertising tool, we have a prompt generator which allows you to put in parameters like, I’m looking to target this product, this is my audience, this is the content that they read about.

But it’s fully transparent. So once you generate that prompt, you can see all of the URLs and you can customize it. You can say, okay, I don’t really like the angle of this URL, so I’m going to give that a thumbs down. Or, I really like this URL and this content, so I’m going to give it a thumbs up.

In real time, you’re customizing your own model, so there’s always human oversight.

PZ: So you have human oversight to set those boundaries and those guardrails to make sure it’s being used in a responsible, not nefarious way.

I like your point on understanding the output. It’s one thing to go, okay, I’m going to use an LLM to, in a more human way, get that result. But I want to know what it’s actually spitting out. You’re still giving that control to your clients to help steer and make sure it gets it right.

JJ: Absolutely. And then secondly, we’re optimizing towards the desired outcome of the client. This is someone purchasing a product. So we want to maximize that and scale that as much as possible.

Can you share a success story? How does deep learning drive performance for brands?

JJ: We now have hundreds of clients. We’re exceeding their cost-per-acquisition (CPA) goals, both with our custom models and even just with our contextual offering.

“Just evaluating context, and providing highly relevant context, can deliver performance for our clients without needing first-party data.”

Specifically, GroupM and Mediacom are using ContextGPT to reduce CPA and reduce the cost of unique reach. They’ve seen some excellent results.

PZ: Fantastic. You’re getting the full breadth of scale and not at the sacrifice of quality. And you still hit those delivery goals, which is the dream. It’s what you want to be doing.

Looking ahead, what innovations do you foresee at the intersection of AI and programmatic advertising, and how should marketers prepare for them?

JJ: I’m obviously really excited about context. But, with creative variation testing, now with AI, marketers don’t just have to be limited to three different creative versions. They could literally have hundreds or thousands.

And now with the power of AI, it’s not just about what the best-performing creative is overall, which is what we’re so used to.

“Now you can determine what the best creative is to serve ‌this person in their context, and what’s the best message to serve them right now.”

We haven’t really had the ability to do that at a large scale and make these decisions in the rapid time required to deliver that optimal message.

Obviously, Paul, you’re going to respond to something different than what I’m going to respond to in terms of creative, product, and message. But also, what time of day are we going to respond to an ad? Right now we’re at Cannes, we’re super busy. If we see an ad on our phone, we’re probably not going to look at that ad and start browsing. But in the evening, we might. Or in the morning we might, for example.

PZ: That makes perfect sense. You have all the different permutations of how you can optimize. And then to your point, join that also on the creative side, which was just challenging to do with the older technology that was in place.

JJ: And for marketers, ensure they’re working with partners that are using AI, and that have proven results in AI. Ensure that they’re not stuck in that old way of thinking of, oh, here’s my context segment, and here’s my audience segment. It doesn’t have to be that siloed anymore.

How has working with Index helped you support your partners and customers?

JJ: I can sum it up in one word: innovate.

“Index has allowed us to innovate. You’ve allowed us to be able to deliver our deep learning technology to the rest of the market.”

What I loved was when we spoke to Index about what we’d like to do, you listened and you made it happen. And you’re always there—I can Slack you, I can call you, I can email you, and I’m always going to get a response. There’s a super high level of service.

“Index really understands we’re all stronger together.”

Learn more about how you can activate intelligence on the sell side and drive stronger outcomes with sell-side decisioning.

Back to blog