Faster, Smarter Research: Where Synthetic Audiences Fit in Modern Marketing

Heather Riexinger: for another podcast with Michelle Ferabundo, our director of CX and Research. So Michelle– Thanks for having me back. Yes. Yes, you did such a good job the first time with Katie that I was jealous I had to interview you myself. So today, you’re here. We want to talk a lot about synthetic audiences, which is a hot topic. We’ve done a lot with AI lately. And I feel like this podcast, just one episode isn’t complete without talking about AI. So we’re going to do an entire episode on AI today. So can you tell us a little bit about what our synthetic audience is, and how are you using them today?

Michelle Furibondo: So ultimately, a synthetic audience is just an AI representation of your audience. So you can feed it whatever kind of data you have. If you have real customer data, survey data, or just general census data, and narrow in on the audience that you want to target, whether for yourself, for your clients, whoever you’re trying to understand. And then it’s an AI model that runs and predicts behaviors, predicts answers, feelings of what your audience might think. So you can run it concept testing through it, whether it’s creative or a product idea. And that’s typically how we use it today, is running creative through it to see if we’re comparing two different sets of copy or two different images. What does our audience think about them? Which one do they like better? Which one might drive action, whatever that action might be. Great.

Heather Riexinger: And how do you do a lot with focus groups and individual interviews for our clients? So how do you fit this into your research booklet? Where does this synthetic audience research fit? What are you doing? Does it pull research in sooner? What are you seeing?

Michelle Furibondo: Yeah, it’s more of an augmentation to our process. So nothing will ever replace actually talking to humans. They’re the ones that have the unique ideas, the unique perceptions. But synthetic audiences is a chance to test things sooner and earlier, and allow for more iteration throughout the process. So it’s super quick. It could take five minutes to ask your audience a question, depending on how robust your question guide is. But what we really try to do is do it early and often, so we can iterate on those different concepts. And then ultimately, in the ideal world, then you would test those final concepts with humans.

Heather Riexinger: Yeah, that makes a lot of sense. I know we’ve been doing this already for some of our clients. And have you had any instances where you have run concepts through synthetic audiences, and then spoke with humans, and gotten totally different feedback? Are you finding that the synthetic audiences are relatively accurate?

Michelle Furibondo: Yeah, I would say they’re pretty close. In some cases, they may not be as in depth. But for the most part, they’re aligned on the direction. So whether it’s positive or negative, they might call out the same pain points or challenges that they’re seeing with the concept that we’re presenting. It may just not be as specific. But at least the synthetic audience gives us something to react to, so we can kind of bulletproof those concepts before we get to the human part. Great.

Heather Riexinger: And I know there’s a lot of different ways to build these synthetic audiences. And we know you’re heading up AI here at our agency, and building out a lot of agents for us. So how often do you need to refresh the synthetic audiences you’re building? And how often do you feed them data? Or is that automatically done? How do these audiences learn over time for you?

Michelle Furibondo: Ideally, you’re refreshing them as soon as you have new data that’s relevant. It really depends on what you’re testing and the kind of topic areas. So if there’s nothing that’s changed in the industry that you’re looking at, there’s not a huge need to refresh it, because it’s really about the perceptions and the demographics of the audience. But any time you get new data that you think is relevant, push it right in so then they can just be more refined, and a little more real world scenarios. But I would say you can do something as easy as a really descriptive prompt if you don’t have an agent tool just to get your AI tool in the mindset of the persona you’re trying to target. Or you could go as in depth as a third party or first party tool that you’re creating that uses robust data sets to model out your audience. Great.

Heather Riexinger: So I’ll throw you a curve ball here, a question that just came to mind. We speak a lot about how AI, when it’s doing research, is not accurate, and how it spits out– we’ve tried to use it for competitive research. It doesn’t give us the right data. We find that we still need to do that in a very hands-on human way. So what is it that’s making these synthetic audiences so accurate, and how are they able to reflect on the industry better than when we are trying to build out some of this other data that we find inaccuracies in?

Michelle Furibondo: That is a tough one.

Heather Riexinger: There may not be a real answer to it, but I know– I ask because we’re seeing a lot of really valuable data, and we’re finding that these synthetic audiences report back very similarly to humans. So I’m always intrigued on why generative AI cannot figure out all the questions I ask it.

Michelle Furibondo: Yeah. Well, I’m not going to pretend that I know all the inner workings of it. But my assumption is that when you’re looking for something that’s out there, it’s doing, I don’t know, an enhanced Google search or something. So it’s looking for stats and data that exists out in the internet. When you’re using your audience, it’s really looking at your panel demographics and psychographics that you feed into it. So it’s not having to find information that exists somewhere. It’s actually using the model to simulate the behavior of the data that it’s looking at. So it’s a little more narrow data set. Of course, if you have it structured so that it is also using the internet, you may run into some of those issues too. But I think it’s a lot more narrow in scope. Since it’s behaving based on the personas that you fed it, it’s really just using those models to give you the answers you’re looking for.

Heather Riexinger: So it really comes down to– it’s as accurate as the work you’re putting in into building it. Exactly. Great. Anything else that you want to talk about? I think you answered a lot of my curiosity about how we’re using it today. And any thoughts?

Michelle Furibondo: Just a couple, I guess, guardrails or things to watch out for. Because having AI so accessible to everyone– you don’t have to be a seasoned researcher to use it. But you have to understand some of the research methodologies. So if you do go out and use a synthetic audience, just make sure your questions are as objective as possible. You’re not leading your panel a certain way based on how you ask the question. And also, we all know AI are– they aim to please. They have that positivity bias. And everything that you say, you’re like, oh, that’s great. That’s wonderful. But you really need to be prescriptive in telling it to be brutally honest. Because especially when you’re using it to pressure test something, you don’t want it to tell you it’s great. You want it to find the flaws. So just making sure you really are prescriptive and specific when you tell it what kind of feedback to give you.

Heather Riexinger: That’s a great point. Because we’re not using it just to tell us how great we are or to take what we’ve put in and simply add to it or amplify it. I know I asked it to be brutally honest about some creative recently. I was like, ouch. I thought it was pretty good. So I think that that’s a place, too, where we want to find that middle ground. We don’t want it to be really telling us that everything we put in is amazing. But sometimes the brutal honesty is good to hear. But we need to find that middle ground so we don’t just kill concepts that our human brains actually liked by having the platforms tell us everything that’s wrong with it. Because you’ll easily kill all your little darlings.

Michelle Furibondo: And that’s where using your expertise comes into play. So you might say that it hates a word you’re using, but you know you have to have that word in there for whatever reason. So take it with a grain of salt. Test it with real humans. Use your human brain, like you said, to really check that it’s something logical. But yeah, it’s easy to have fun with, especially because you can keep probing. You can keep asking questions in your focus groups. You can have them interact with each other. You could go in a one-on-one interview and it’s just, you know, it feels like you’re talking to a real person. It’s always got that little tinge of AI speak when it gives you its answers. But once you look through that, it has a lot of good valuable data.

Heather Riexinger: Great. Well, I’m really interested to see where you take this and what you’re able to build and what uses we find for