S4

Episode 3

AI and the Power of Programmatic

About This Episode

September 13, 2023

We dive into the exciting world of AI and programmatic advertising to learn about the advantages of AI-powered platforms.

Zeynep Akkalyoncu | Lead Data Scientist, StackAdapt

Arthur Faisman | Lead Data Engineer, StackAdapt

00:00

Transcript

Episode Introduction (00:00:00)

A very important concept in AI is called transfer learning, a model can gain knowledge in one field and then apply it in a completely different field might have been some kind of numerical optimization targeted towards a KPI like conversions, it might have been a contextual tool, like our own page context AI that will help you display your ad in the relevant context. But it’s really important to work with a highly regarded DSP that has that in house knowledge, because it’s very difficult to kind of transfer that expertise to maybe external AI.

How Agencies Thrive Introduction (00:00:32)

But then you think about the social landscape. The research data is hugely significant when we combine all of these different touch points, so that long term loyalty and then diving into the clicks to leads to sales, gotten to a point where it can drive better results in audience targeting, and really is what’s going to set you apart. You’re tuning in, you’re tuning in, you’re tuning in to the How Agencies Thrive podcast.

Sneha (00:00:53)

There’s so much discussion around AI these days, and that’s only because of its increasing significance. And today, the focus is going to be on AI in digital marketing, and specifically AI in programmatic. Programmatic advertising, as you may be aware, is a system where advertising space is bought and sold through automated auctions. The use of machine learning AI algorithms, enables them to learn and adjust their actions based on new patterns they encounter. This adaptability makes them extremely well suited for the dynamic nature of programmatic advertising. AI algorithms also possess the capability to quickly process large amounts of data. Now, with all of these factors combined, AI has a significant impact on numerous facets of programmatic advertising. Let’s learn more about this from the AI experts Zeynep and Arthur. But before that, thank you for tuning in a big hello, and welcome to the How Agencies Thrive podcast. I’m your host Sneha Suhas from StackAdapt. And I will now pass it to the guests to introduce themselves, their professional life so far and their areas of expertise. So let’s start with you Zeynep.

Zeynep (00:02:08)

Hi, thanks so much Sneha. So my name is Zeynep, and I’m a lead data scientist at StackAdapt. I’ve been working here for almost three and a half years now. And for the past year or so, I’ve been primarily leading architectural efforts, as the industry is moving away from behavioral targeting to some extent, as there are concerns around user privacy. We’ve been investing a lot of effort into natural language processing and information retrieval. And that’s what I’ve been up to.

Sneha (00:02:32)

Awesome. Welcome to the show. And thank you for joining us. I’ll pass it to Arthur.

Arthur (00:02:36)

Hey, thanks for having us. Yes, I’m Arthur. So I’ve been with StackAdapt for about a year and a half now, I’ve been mostly focused on the lookalikes audience expansion project. While I’ve been here. Prior to that, I worked at an investment company on their data science team for about a couple of years. And prior to that I worked at InLogic, which is basically kind of like Nielsen ratings for Canada. So we did TV and radio audience analytics. And I was there for about four or five years, and I was their point person for algorithm developments.

Sneha (00:03:09)

Awesome. Thank you for joining us Arthur. And a question for you. AI is a very popular topic right now. And it’s been a large part of the conversation when it comes to programmatic advertising for quite some time now. So to break it down. How would you say, AI differs when it comes to programmatic advertising? And in what ways? Is it similar to the conversation that’s currently ongoing?

Arthur (00:03:32)

Yeah, I can take that. So AI is a pretty broad term. So it’s kind of hard to say exactly which specific technique is AI or is not. Broadly speaking, you call something AI, if the problem that you’re trying to solve is so hard that software developers are not able to individually write the little bits of code that actually solve that problem. So programmatic is one such problem. So when we implement programmatic bidding for a particular campaign, it’s a vast amount of decisions that need to be made behind the scenes. So there are billions of auctions that are processed per day, like many billions, and each of those auctions needs to be considered individually, and the decision needs to be made, what is the proper data amount, the campaign needs to be considered behind the scenes to make sure that it’s pacing properly, to make sure that, you know, the proper targeting options are applied, that the KPIs are being maximized as much as possible. So it’s obviously not possible for any individual to actually, you know, go through those decisions. And it’s also not possible for any individual group of individuals to write code that optimally processes all of that data and makes all those decisions at once. So at this point, what’s typically done is a technique is applied called machine learning, which is, in some cases, you consider synonymous with AI. So again, AI is kind of broad, it’s hard to say exactly, what everybody calls AI, but I’ll call AI as something that’s a problem that solves a technique called machine learning. So what is this technique? So the technique is, instead of writing those individual rules of code, which in this case is not feasible, what you do is you feed a huge amount of data through the system, and you allow this system to learn automatically, what works and what doesn’t for each individual use case. So even though I personally and no individual has, you know, gone through a bid, each individual bid of requests, like, what happens is the system figures out what works and what doesn’t on this large volume of bid requests, and makes those decisions there programmatically. And that’s kind of the only way you can really solve this huge, huge data problem and huge complexity problem you get with programmatic advertising, and real-time bidding. So the main similarity with a lot of the conversations about AI there is that since, no individual has full understanding of what’s going on with this system. And which kind of needs to be the case, for these really complex systems. You have a similar kind of question to ChatGPT or you know, ChatGPT, famously, you give it a query, like write me an essay on Darwin, and you don’t necessarily not necessarily sure whether the essays actually on Darwin or on somebody else, or what’s actually going on. And whether it’s completely accurate, because, you know, no professional has actually gone through and edited this essay, right. So similarly, in programmatic advertising, that same problem that we just described. So I think in this case, the two problems are kind of similar from the point of view of people who are kind of looking at from the outside and trying to understand what’s going on here. But programmatic advertising is more mature industry, then ChatGPT, which is what half a year old. Now, little more. So in programmatic advertising, we have given people a lot of tools to solve this kind of problem, specifically, for the tools include, like attribution, and tracking and audience insights, as well as you can get a sense of what’s going on. So that problem is kind of the same, but the industry is quite a bit more mature than ChatGPT, which is like a baby, basically, right now. And before just trying to figure out what to do there.

Zeynep (00:07:17)

I think that was a great response, I think maybe like, I would take that to kind of a higher level, conceptual level, and say, to a non-technical person, I think of AI as kind of a hybrid of individualized solutions for each user. So maybe, like, let’s go back to the beginning of the advertising industry, like, right, so it might have been possible at that point to, you know, individualize and individualize creatives, targeting strategy, talk with each client one on one, and then come up with a specific solution. But that’s obviously not the case anymore, especially in programmatic. We have lots of clients, sometimes wearing goals, sometimes multiple goals, not to mention millions of users, that we can bid on with different preferences, you know, different ones. So that’s where the AI that Arthur was talking about comes in. And then we can make those numerous tiny decisions with machine learning. So that’s why this makes it an ideal solution for programmatic advertising. I would say maybe the differences are that we need to be a bit more careful about the laws and regulations, compared to AI at large, it’s important to comply with regional and federal laws. And these are not necessarily embedded in the existing models. So that’s where the domain expertise comes in. Right? This is perhaps true of many, many other fields, but especially so of programmatic advertising. We don’t have strong pre-trained models that capture those industry trends or laws. So it’s important to have the people that kind of work together with the AI models, to deploy them to the people.

Sneha (00:08:54)

When it comes to programmatic technology. What are some of the benefits to working with a DSP that leverages AI and machine learning? For example, audiences insights, this is a question from marketers perspective.

Zeynep (00:09:10)

Yeah, I can maybe take this one first. So as we separate the previous question, definitely, like hyper-individualized personalization, but also massively parallel personalization, if that makes sense. So the thing to always keep in mind with AI is the high dimensionality, right? So as human beings, we are fantastic at intuition. But we can only hold four thoughts at the pace at the same time in our minds, which is definitely not enough to handle the scale of the data and the variety of data that comes in with programmatic advertising. So I think like I said, for the previous question, it’s important to kind of pair the human intuition with the high scalability and high dimensionality of AI models. And kind of to add to that as well, I think explainability is a very important point especially in programmatic advertising. So we should be using AI models, or we should be maybe working with DSPs, that leverage and models that can be used to explain the decisions they have made. So they can maybe tell you what factors went into making the end conclusion, what features were the most important, what features perhaps were not as important. And sometimes it’s just difficult with generative models. So I think, first of all, it’s very important to keep up to date with the latest happenings in the AI world, but also keeping the limitations of the programmatic advertising industry. And these changes are sometimes not as transparent to marketers or agencies. And I think that’s why it’s really important to work with a DSP that has that kind of in-house expertise and both.

Arthur (00:10:50)

Those are really good descriptions. So one thing I’d like to sort of contrast that with, is like, so if you use a DSP that leverages AI, or machine learning versus one that doesn’t, so it’s like, if it just does not leverage AI and machine learning, it’s basically dead in the water. Like, imagine, imagine you’re bidding on billions of these auctions. And like humans constructed all of the rules that are made to make those bidding decisions. So again, the humans are just not able to delve into the data to the extent that a machine learning or AI process is able to, so they are just not able to identify which of the users are like, perfectly actually appropriate for your campaign to bid highly on versus inappropriate. So we should just bid a little low on them. So if you are using a DSP that somehow just avoids using any AI in any machine learning, you’re kind of like bidding something that would be like, I don’t know, let’s imagine like some simple like flat rate bidding or some really simplified bidding process, you’ll be looking at a user, you won’t really know what how much to bid on it. And you’ll be getting outbid on users that are actually very valuable for your vertical by other DSPs, which have recognized using machine learning that this user is super valuable. So they’ll be picking up that user. And conversely, you’ll be bidding too highly on other users, which others DSPs have realized are actually not useful for your vertical and you’ll be picking them up and you just won’t be converting on them. So this is just a space that’s way too complicated for humans to effectively construct rules, without machine learning or AI to effectively compete in this space.

Sneha (00:12:32)

Okay, so what would you say are some of the limitations when it comes to the use of AI and programmatic advertising? And, you know, how does the ongoing innovation help bridge those gaps?

Arthur (00:12:44)

One of the main limitations is also the advantage, which is that the AI is taking control over a lot of the decision-making. So this advantage is that it can do really well. The disadvantage is, as I mentioned before, is that well, you don’t really necessarily have full insight into what it was doing. So you could be kind of like worried. So okay, so I set up my campaign it’s doing this unpredictable thing. They tell me there’s this intelligent machine in the loop. But I don’t really know there’s no person that’s signing off on a decision that it’s making for me. So what do I do? And how do I bridge this gap? So the nice thing is, though, about programmatic advertising is that because there’s so much data volume, and there’s so much data features also available on what is going on is that you can do pretty good attribution, on what actually happened in your campaign. So the data is more complex, and, but you also can track it far better. So even though your campaign is doing something a little bit tricky, so I would definitely like say, set up your campaign with, you know, concrete KPIs that you are tracking, I would say, you know, talk to your programmatic strategists about various campaign options that might help you, I would say set up your conversion pixels, track that attribution, and try to find a an option that actually works for you well, so this isn’t something that you will be able to figure out at first go so there are many different options that you can use to set up your campaign and this podcast has covered different techniques for setting up your campaign in different situations very well previously, so I won’t go into that into detail. But the main idea is you should try different things think like a scientist experiment. Some of these strategies which you know are kind of black boxes a little bit like they for you, they won’t work perfectly. You do need to try different custom segments you need to try whether it for your particular campaign contextual keywords work, try whether browsing audiences work. Try whether retargeting works best for you. And then after you iterate, pick a strategy that does work for you because you are able to track that attribution far better than you are outside of programmatic. So definitely take advantage of that.

Zeynep (00:14:59)

I think so I would also add, maybe first of all, human in-the-loop systems. So they have been around for a very long time. But they’re only like recently gaining attraction with more popular generative systems like charging ChatGPT. But the goal here is to kind of inject that human intuition that we referred to earlier and collaborate with the model to produce better results. So these, this can happen in more of a friendly way, or a more adversarial way, with the goal of improving the results of the model or making them more intuitive or closer to what we would expect with our domain knowledge. And the other thing is continuously monitoring the model that we have deployed. So you don’t just train a model and put it out in the world and expect it to perform splendidly. So when we, as a data scientist, when we were deploying a model, there was actually a lot of historical back testing on past data points. And then we can say, with 100% accuracy, that these models will actually do well in production. So what we do is, after deploying them to a small percentage doing beta testing, we actually compare their predictions to the actual performance. And then we always go back and update the models accordingly. So I guess there’s a recurring theme of Yes, working with the model is kind of very important, especially when we are kind of catering to people in this industry.

Sneha (00:16:18)

It makes a lot of sense. Thank you for that. So jumping into the next question, what would you say is the biggest difference to the use of AI and programmatic now versus say, five to 10 years ago? Can you draw a picture of the evolution?

Arthur (00:16:36)

I can give some high-level descriptions of a lot of the features. And those of the audience that have been involved programmatic will have seen the evolution over time. So I mean, a couple of main things that have changed over time is firstly, like, a lot of the targeting options that you’re able to use, and a lot of the custom segments available have expanded. So for example, at StackAdapt, we rolled out page context AI, just a couple of years ago. So if you were targeting 5-10 years ago, we would absolutely not have this kind of option. Similarly, we’ve expanded the efficiency of our browsing audiences. And the efficiency of lookalike expansion as well, 10 years ago, this was not available. So just if you literally go into the UI, you will be able to see more options year over year over a year. And after 10 years, like for you, as a user, there’s just far more targeting options available. There are also far more insights that you can get into the users that you actually have targeted than you able to get 10 years ago. So this is kind of apparent to anybody that signs in into the platform and just sees this change over time. Another change I’d like to highlight though, as well is there’s also a lot of invisible changes behind the scenes that you really don’t see as a user. So if you were on the exact same campaign, couple years ago, versus now you will get better return on investment, generally speaking, given the exact, you know, given the same targeting options, because behind the scenes, we continuously make a lot of improvements that are invisible to the user to find the individual bids and auctions that will absolutely benefit your campaign to bring your return on investment up. And there are just tons of optimizations that we make behind the hood. And this is very, like apparent to me, because like my bread and butter every day, like most of the features that I work on, don’t bubble up as like, you know, new buttons that you click in the UI they actually see. So I work on the look-like-expansion project. And if you click look like audience expansion, now it’s a button, you click it, the audience expands. So from the point of view of the user, is that kind of the same thing that was there a year ago, it looks kind of the same, but behind the scenes, there’s been a ton of improvements to try to figure out what is the ideal look alike audience for your campaign? How can we best expand it? How do we make sure that you still have the scale without losing the accuracy of bidding on the proper users, and there’s just a lot of work that goes on there. And this is I feel something that’s kind of underappreciated by someone that just logs in every day and sees the same thing. It’s not the same behind the scenes, there are a ton of changes, and a lot of people that work with are really involved in making these changes on a day-to-day basis.

Zeynep (00:19:16)

Yeah, definitely. I definitely agree with that as well. I think we can also take a step back and look at AI or the evolution of AI in general in the last decade, it’s a very important concept in AI is called transfer learning. So it’s quite straightforward, as the name indicates, but it means a model can gain knowledge in one field and then apply it in a completely different field. So basically, there has been a lot of work in transfer learning in the last decade, even across like seemingly unlikely pairs of fields. For example, a model might learn to play chess, and then you can ask the model to kind of do my math operations and they will do significantly better after gaining that expertise and Chess. And this is now possible because first of all the availability of the data on the internet, there’s just so much that we put out that models can learn from. And secondly, the evolution of the physical machines as well that can handle that scale of the data. And that’s why we now talk of these generative models like GPT, that are trained on all sorts of data on the internet. And then they can be adopted in their niche domains, like programmatic, for example. And we always find that their kind of newfound general knowledge will make them more powerful in these niche domains. So that’s kind of exciting. And that’s what exciting to everyone, I guess, because it’s unclear what we can now do with that general knowledge and what specific areas we can apply to.

Sneha (00:20:43)

You know, from an advertiser marketers perspective, because it is such a hot topic right now, we know even from a product marketing perspective, terms like AI and machine learning algorithms can be a big unknown for some marketers or advertisers, especially when it comes to automating ad spend and making campaign decisions. So what advice would you give to an advertiser who is apprehensive about adopting this technology?

Arthur (00:21:12)

It’s all about like, what are your options? Right, so an alternative to programmatic, which contrasts with it quick, significantly, that I’m familiar with is linear, traditional TV and radio advertising. So a question that you might ask is like, well, do I allocate my ad spend on you know, TV, or in programmatic. So I just kind of want to say so, it’s fairly interesting to contrast it to because the TV doesn’t have many of the advantages that programmatic has. But it also doesn’t have many of the disadvantages, because it has disadvantages. They both have advantages. The advantage of programmatic is that you’re letting go of control over individual details. So for TV, the way that you typically organize your spend for a campaign is you do research on which programs or day parts contain a particular target audience that you are interested in, and then you plan that spend, you make your contract your purchase decision with a particular network, and then your ad spots run. And then when they run you kind of retroactively when they run you check up on the viewership. And the audiences that watched a particular program or daypart that you had your ad spot run. And then you look at your particular KPI of interest. So, oh, an ad spot ran yesterday, did people go to my website after did they purchase, you know, my product today in the store? And that attribution process is like, really hard. Like, it’s really challenging to say whether an ad spot yesterday leads to conversions, sorry not conversions and TV, you might just like, say, but yeah, a conversion, whether an ad spot leads to conversions for your store purchase today. So I’ve worked with some of this data before, it’s just it’s really, really hard because you don’t have that user-level information. So you’re looking at volumes of purchases, and you’re hoping that that volume goes up significantly after ad spot, and unless it’s the Superbowl, it’s probably not. So you don’t really know if you have a return on investment from your ad spots. So but conversely, you do have certainty in terms of what’s going on. So as a human, not a machine, you’ve made a very simple decision, I’m placing my ad, at this time on this program, you know exactly what happened, but you don’t know the impact. So, even though you gain some you lose some. So which world would you rather work with? I mean, I can’t make that decision for any marketer. And many marketers, you know, kind of hedge their bets and allocate budget into both worlds, let’s say from my perspective, like coming from the TV data, and now working with programmatic data, we would have loved to have the verticals and the user level information that we have now in programmatic, we would have loved to have that back in my TV days, like we dreamed of this kind of data, and we just did not have it. So the types of attribution that you can do in programmatic now are just so so powerful, so much more powerful than anybody could have dreamed of in the TV days.

Zeynep (00:24:24)

Yeah. Also, like another perspective we can offer is so sure, we have been talking about AI recently, a lot. You see it in your news feed in the morning, you see it in TV shows, but it’s not exactly new as a field of study. So put it in the context of the first AI program was actually written in 1951. So chances are if you’ve ever worked with a DSP, it’s quite likely that you have in the garden some ad spent our ad revenue from some form of AI so it might have been some kind of numerical optimization, targeted towards a KPI like conversions and might have been a contextual tool like our own page context AI that will help you display your ad in the relevant context. But the point is that it’s not new, you’ve probably used. And you’ve probably seen a lot of improvements over the years, as Arthur said, as well. So the exciting part is, obviously, the AI has come a long way, as we’ve talked about, you know thanks to data availability, better machines, and we can absolutely do completely new things that were maybe unthinkable a few years ago. So I would say, this is not so much a question about whether to use the AI, we absolutely have to, as Arthur said, I don’t think it’s quite a choice. But it’s really important to work with, you know, a highly regarded DSP that has that, again, in-house knowledge, because it’s very difficult to kind of transfer that expertise to maybe external AI, AI workers as well. So it’s really important to work closely with like marketers, data scientists, data engineers, as well as people that have expertise in setting those campaigns up, and maybe even vertical-specific knowledge as well.

Sneha (00:26:08)

Amazing. Thanks for that. And looking into the future, where do you see AI in programmatic heading by, say, 2025? And will we get into the territory of generative AI for ad creatives? Or do you have anything? You know, far more exciting than that that you see?

Zeynep (00:26:28)

So I think anything we say about this will be pure speculation. We don’t know we have no idea what’s going to happen tomorrow. But yes, I think generative AI for ad creatives is definitely on everyone’s mind. So maybe we will be able to create on-demand creatives to complement the user’s interests, or we are already working on very powerful internal tools, that with the goal of, you know, saving companies, hundreds of worker hours, that we can hopefully invest those in more creative pursuits that are better suited to human beings, and perhaps builds, you know, highly complex predictions with models that enable multiple KPIs at the same time. Again, going back to the theme of like high dimensionality. But, you know, looking back on this podcast, in 2025, we will probably have a completely different perspective on this. And I don’t think anyone can claim to know what will happen in programmatic in the next couple of years.

Arthur (00:27:27)

Yeah, I definitely agree with Zeynep. This is one of those questions that you just can’t, you don’t have that crystal ball. So you can’t see the future, especially when it comes to things like policy changes, or like new AI developments, like ChatGPT came out of nowhere, like half a year ago. So even you know, experts in the industry did not expect this level of accuracy from a chatbot. To come out like so soon, right. So it’s very hard for anybody to predict the future. So I won’t try to do that. But I’ll give kind of like a day-to-day view from let’s say, what we are working on, which I know for sure was happening, which is again, that invisible, but very important improvement, and in the overall accuracy and efficiency of your ad spend. When you are running that same campaign, from last year to this year to next year to the year afterwards, that same campaign will find your users and will manage your targeting better than it did previously. So working on many optimizations, and I can’t speak about any of them. But they are all behind the hood. And trust me, we’re very, very hard at work picking out those users trying different techniques. This one works, the other doesn’t pick it up again and try another one. What about this machine learning model? What about the other one? What if you segment users this way, or the other way, it’s all proprietary, but we’re all very, very hard at work on trying to get your KPIs up, and your CPA down every day.

Sneha (00:28:57)

Always fun to hear informed predictions. So thank you for that. And I think whatever you share today was super useful. Both of you made the topic of AI and machine learning very approachable for marketers out there. So thank you both for joining us. And to you, the one who stuck around till the very end. Make sure you subscribe to the podcast or listen to the new episodes right when they drop, like the podcast, share it with your teammates. It could be a cool resource to post on your office work chat as recommendation. So go ahead and do that. And if you want to get in touch, write to us at academy@stackadapt.com That’s academy@stackadapt.com we have episodes releasing every alternate Wednesday. So stay tuned. Until then, this has been the How Agencies Thrive podcast. See you in the next episode.

Episode Outro (00:29:40)

Thank you so much for tuning in. This has been the How Agencies Thrive podcast. If you like what you heard, then there’s three things that you can do to support the show. Number one, subscribe. Number two, leave us a review. And number three, share our podcasts on social media or with anyone who might find value in this content. If you have questions or feedback or just want to learn how agencies and brands work with StackAdapt, you can us at StackAdapt.com. Thanks for listening, and we’ll see you next time.

S4