Lookalike Audience Creation with Faraday's Predictive AI | Andy Rossmeissl

David McBee: Hello and welcome to Simpli.fi TV, the web series and podcast for agencies, marketers, media buyers, and business owners. I'm David McBee. Our guest today is Andrew Rossmeissl CEO of Faraday. Andy is a seasoned software developer and multidisciplinary designer with nearly 20 years of experience in tech startups based in Vermont. He co-founded Brighter Planet, the first company offering an API for carbon impact calculation. Andy is also the co-founder of Faraday and he'll be discussing that in length in today's interview. His current focus is on AI emphasizing responsibility measures and bias detection. Andy, welcome to Simpli.fi TV. Andy Rossmeissl: Thanks so much for having me on today, David. David McBee: Happy to have you. All right. Well, let's dive into what Faraday does and the concept of lookalike audiences. Andy Rossmeissl: Totally, yes. So this is one of the things that Faraday does. We're a customer behavior prediction platform, and I think first off, when thinking about lookalike audiences, you want to understand that they're based on a seed. That's the term we call it. Some group of people about whom we're trying to find other lookalikes. So it's not helpful to necessarily think of a lookalike in general. There isn't a lookalike. Many of your people, I think listening to your show, will recognize that core group. The seed is often going to be a brand's customer base or maybe high lifetime valley customers, something like that. The next thing to really think about is eligibility. So who in the known universe of human beings do we want to try to find lookalikes within? And I think oftentimes people forget about this step and it really does matter because of the way that lookalike algorithms work. So maybe that would be everybody in the country your business operates in, or maybe it's a smaller subset regionally or demographically being clear about. That's really important. And then the next thing you do is want to try to figure out, "What you mean by lookalike?" At Faraday we think of similarity dimensions. And one way to consider this question is the difference between a doppelganger and a kindred spirit. So we've all encountered people in our lives where you're like, "That looks exactly like David physically." But then maybe you're lucky enough to have a significant other who doesn't look at all like you, but is a kindred spirit, somebody where you're very deeply similar in other ways. So similarity is not a specific thing you have to really consider, "What do we mean by similarity?" Often in the brand context, their agencies will be thinking about things like spending habits, certain demographics, maybe family makeup, if there's any sort of lifestyle attributes that matter. Life events sometimes come in key. You've got to identify what those similarity dimensions are. And of course you have to have data on all of that. So not only do you have to know these things about your customers or the seed group you're using, but you also have to know the same details on everybody within which you're trying to find lookalikes, the haystack, so to speak. And then you have to understand that there's no real such binary thing as similar or not similar. It's really a question of degrees of similarity. And we're trying to look for is, "Which people in the eligible population are more alike than not?" At the end of the day, there's not going to be two identical people. There's not going to be two people who are polar opposites that happens in the movies and things like that, but not in real life. So I would say the most difficult decision that the practitioner needs to make, because these algorithms are all very straightforward, is choosing a threshold. And to give you an example, it's easiest to think in terms of deciles. So you can organize everybody in your eligible population by their degree of similarity to your seed group. People at the top of the list are very, very similar. People at the bottom of the list are less and less similar. So you can imagine that the folks in your top decile or top percentile, however you want to organize people are going to be the most similar. And those are awesome, they're going to be great fit, but there's not a lot of them. You then maybe want to go down to the next percentile, the next decile, and as you go, you're still going to have similar people, but progressively dissimilar and choosing that cutoff point is really, that's the art of the business and that's why we love working with agencies who are really good at that kind of decision-making. David McBee: What are some of the data sources and methodologies that you use to create your lookalike audiences? Andy Rossmeissl: Yeah. So it's a good question. You have to have those data points on, not again just your customers, but everybody. And at Faraday we have that built in. You may be doing lookalike with other systems, but with Faraday we have about 1500 data points on nearly every US adult, about 270 million people. This data is responsibly sourced from companies like Epsilon who get permission and things like that in order to get data about people and compile it responsibly. And then the methodology that's used for those of you who are interested in the data science end of things is going to be a decision tree type methodology. That's what we use at Faraday and that's really what I would recommend anybody doing this kind of work to consider using. And if you don't know what decision trees are, maybe you've heard of other terms like decision forest, random decision forest, gradient-boosted trees. These are all decision tree type algorithms. Just go on YouTube and search for decision tree. It's actually an incredibly simple algorithm to visualize, see, understand, and it's like the backbone of a lot of predictive AI and you'll immediately become more of an expert just taking a look at a quick video. David McBee: Perfect. Awesome. Tell me about how the lookalike audiences benefit the businesses in terms of customer acquisition and some of their marketing strategies? Andy Rossmeissl: Yeah. I think people in your audience will often think about our two major levers of being, targeting, and personalization. And lookalikes are often thought of as really a targeting solution, but they can actually help with both. The easiest thing to do of course, is just to use targeting to stop wasting money on bad fit prospects. We don't need to pay for impressions, we don't need to pay for clicks when these people are not going to be good customers for you. And similarity, lookalike is a great way of doing that. But I think personalization is where a lot more of the opportunity is now. Everybody's using lookalike to some extent. If you're marketing on Meta or other platforms, these kinds of algorithms are built in, so it's kind of table stakes and finding alpha in the ad world is hard. Personalization I think is the edge a lot of our clients are looking for right now. And the easiest way to think about that honestly is just multiple seeds. So let's just say you're a brand, you're an agency working for a brand that has two major product line categories. Instead of one seed audience, which is going to be all of the customers, maybe you've got two, you've got product A customers, product B customers. You build a lookalike audience for each and then you know what to put in the ad rather than having to guess or just use your most popular product or be very generic. You can imagine extrapolating that out beyond 2 to 3, 4, 5. And we've at Faraday have developed I think the best way of doing this kind of very personalized lookalike approach. And that's called personas. And anybody can do this, you don't need Faraday, but it's easier with our built-in data, you essentially use clustering techniques, which is another machine learning concept. We use K-Means++, which is a modern variant of a very old algorithm. And all it does is automatically in an unsupervised way organize that broad generic customer base of your brands into call it five subgroups. And each subgroup is going to be very thematically coherent. It's going to be interesting. There's going to be commonalities within each cluster that you can use to personalize to them different groups of people. Of course, these will always depend on your customer base in general. Each of those can be a seed. You can figure out what kinds of products and what kinds of messaging creative photography each persona likes to see. Then you can use that with your marketing. I will mention though, you have to be careful when using, lookalikes too much for targeting because you don't want to suspend the natural evolution that any brand could go through if all you do forever is market to people, look like customers you've had for a long time, you're never going to find new groups, new opportunities. And that's really, I think, where a lot of the opportunity is. So make sure to always leave some random group in there for evolution. And yeah, that's how you get value out of lookalikes. David McBee: That's really good advice, especially that last part. Tell me if this kind of fits that narrative. 10 years ago, Jeep was really marketing to men. It was really all about the men. Today, a lot of women are driving these pink Jeeps, these yellow Jeeps, these teal blue Jeeps. They have a whole new market out there. So I can see how if they didn't try to reach them, they would've missed out. Andy Rossmeissl: Yeah. Jeep has the advantage of doing broadcast television advertising, they're going to get some of everybody. But let's imagine they hadn't. Thank God that they did have some of that mutation built into their advertising, so they were able to find that new segment. David McBee: Do you have any great case studies or success stories you can share? Andy Rossmeissl: Yeah. Absolutely. We have a lot of brands, thousands of brands using Faraday for this kind of thing. I would say there's an example out of the home decor space. I can't say the client's name exactly, but they use that persona seating method that I had talked about just a second ago, and they were able to do a direct comparison using an A/B test and found 2.4x higher conversion rate, 65% decrease in cost per acquisition and a 2.4x increase in ROAS. And that's all just with that basic persona seating and lookalikes. David McBee: I love hearing that. All right. Well, full disclosure to our audience, Faraday and Simpli.fi our partners. Can you share just a little of that secret sauce as to why we are working together and how that benefits our listeners? Andy Rossmeissl: It's about combining the science and the art. And Faraday is awesome at the science and Simpli.fi and agencies are, that's where the art is. We're always going to be like any great machine best used by an expert driver, and that's where Simpli.fi's partnership comes in there. David McBee: Perfect. All right. One last question before we do the questions, I ask everybody is how do you ensure the accuracy of these lookalike audiences? If I said, "This is my list of 100,000 customers." And you said, "Here's the lookalike audience." How do I know that's really accurate? Andy Rossmeissl: Yeah. The easiest way is just to try, I could tell you a million things from a super nerdy angle but try it. Put it in an A/B test mechanism and whatever platform you're on and watch the lift emerge. That's the best, most powerful way of knowing that it's working because if we were not finding lookalikes, we were doing randomness or something, then that obviously there wouldn't be any lift. But that said, every time Faraday does build a machine learning model on the backend and we're building many thousands of these every day, it's automatically goes through what's called a cross-validation process where only a portion in this case of your seed audience is used to train a model. That model is then applied to the whole eligible population, including that holdout, and we get to see whether or not our prediction about those held out people actually matched up with reality that for every model that actually happens, I think between three and five times with different random splits to ensure that the predictions we making or making are accurate. David McBee: Those were two really good answers to my question, but I especially like your confidence, "Just try it. You'll see it works." Andy Rossmeissl: Yes, try it. David McBee: "It works. Just do it." Andy Rossmeissl: Yeah. David McBee: All right. Before I let you go, I do like to ask all my guests if they have a book recommendation or a podcast recommendation, something that has positively impacted your success. Andy Rossmeissl: Yeah. I would definitely take a look at a book called A Pattern Language by Christopher Alexander. It's about architecture, but it's really about life and the idea that there are correct answers to really perplexing problems, and if you just listen very carefully to the world, you will find them. David McBee: That sounds awesome. Well, Andy, thank you so much for being my guest today. Tell our viewers how they can reach you. Andy Rossmeissl: Yeah. Go to our website faraday.ai or email me directly imandy@faraday.ai I'd love to hear from you. David McBee: Awesome. Thank you so much for being here today. Andy Rossmeissl: It's been my pleasure, David. Thanks. David McBee: And thank you guys for watching Simpli.fi TV. Please help us out with a like, comment, a share or a review, and be sure to follow or subscribe to be informed about new episodes. Simpli.fi TV is sponsored by Simpli.fi, helping you to maximize relevance and multiply results with our industry leading media buying and workflow solutions. For more information, visit Simpli.fi. Thanks for joining us today. I'm David McBee, be awesome, and we'll see you next time.