B2B Lookalike Audience Modeling & Artificial Intelligence
Highlights from this Episode
B2B Lookalike Modeling and Artificial Intelligence, that’s what’s on tap for this episode of the Green & Greene show, the LeadCrunch B2B Podcast.
Hosts: J. David Green and Jonathan Greene
Topic: AI/Predictive Analytics
Subtopic: B2B lookalike audiences
Duration: 19 minutes
- The Book Jonathan Wrote about Facebook Marketing
- What are Lookalike Audiences?
- The ROI from Look alike Modeling
- Lookalike Modeling & Machine Learning – How Models Get Smarter
[0:00:03.7] ANNOUNCER: Live from deep in the heart of Galveston, Texas all the way to the gleaming shores of Jacksonville, Florida, it’s the Green & Greene Show. Here are your hosts, Dave Green and Jonathan Greene, ready to unlock the mysteries of scaling demand gen. The Green & Greene show is brought to you by LeadCrunch, which has reimagined how to find B2B customers at scale.[INTERVIEW]
[0:00:33.8]JG:Mr. Green, I presume?
[0:00:38.5]JG:Welcome everybody to the Green & Greene Show. We’re back. It’s been a little while. We had the holidays. We had the madness of Q1 and all of the fun things that go along with that kicking off. Here we are, and we’re going to talk about lookalike modeling today.
[0:00:57.1]DG:Plus, we were goofing off.
[0:01:00.4]JG:A little bit of goofing off, I will admit. I’ll admit to a little bit of that.
The Book Jonathan Wrote about Marketing with Facebook
[0:01:06.5]DG: Well, listen, we were going to talk about lookalike audiences today. I actually first learned about lookalike audiences when I read a book that you did from Amazon, and you were asking for feedback on it. You wrote about Facebook’s lookalike audiences right when Facebook was releasing them. I thought, “Wow, this is going to be pretty cool.”
If you look at Facebook revenues since they launched their lookalike audiences, they’ve basically just gone up and up and up. Obviously, there’s something there that works. I wonder if you could break down, from your experience, what is a lookalike audience, and why do they matter?
[0:01:52.0]JG: Okay. Yeah. Regarding that book, specifically, I think what I was able to see was the information that Facebook was in possession of and what the implications might be for that. The cool thing about that platform is people volunteered a giant wealth of things about themselves based upon what they interact with, whom they talk to, what pages they engage with, what content they comment on, etc. All those things end up forming what are called graph points for Facebook.
I think their initial intent was to use it to make a richer social network with some cool search functionality. One implication might be if we were looking for a Chinese restaurant to eat at and we typed “Chinese restaurant” into the Facebook search bar, prior to graph, you would have gotten very Googlesque results. You’d get whatever Chinese restaurants there are probably based on how close they are to you or whatever.
Post graph, you get rich data. Not just which Chinese restaurants are near, but which ones your friends enjoy, which ones they’ve eaten at, which ones they’ve commented about, which they’ve taken pictures of and have reviews on. It’s much richer data, but it occurred to me if they volunteer all this data about themselves, it’s going to trickle over and Facebook is eventually going to have a wealth of information about individuals, which will enable them to sort of break people down into algorithmic similarities.
What are Lookalike Audiences?
To your question about what a lookalike model is, it’s a list of people who have done something, broken down into algorithmic commonalities and then extrapolated to a large data set. That sounds complicated, but in essence, if I have a hundred friends on Facebook who like music, and I want to know which hundred people in my extended network also like music, Facebook can then go and look at what pages all those people like. Perhaps, they like Bob Marley’s page, perhaps they like Texas Flood’s page, whatever. Then they can take that and extrapolate that into an algorithmically similar audience by comparing data points with everybody else on the platform.
There are a couple important things to know about that. One, it’s hugely, hugely beneficial for advertising purposes, and we can get into that a little bit more in a minute. Two, it’s dependent entirely upon the quantity and quality and relevance of the data that you have to work with. I think that’s what’s important about that from the onset.
[0:04:48.5]DG:Yeah. Our audience are B2B marketers and not B2C marketers. Obviously, Facebook is primarily, although not exclusively, a B2C platform. The one thing that I always try to encourage B2B marketers to think about is that, from a maturity standpoint, B2C marketers are often way ahead of where B2B marketers are. While it’s very different, there are some things to learn there.
Obviously, in our view, lookalikes are one of those things that you ought to be thinking about and figuring out, “How can I do B2B lookalike audiences?” The historic way to target is pretty inexact. You end up paying for a lot of people you don’t want, and you miss people you ought to target, just because it’s very imprecise. I think this other method is way better. Do you want to talk a little bit about how you would apply this to a B2B framework?
The ROI from Look alike Modeling
[0:06:00.6]JG:Yeah, sure. Let’s segue quickly into how effective it can be in the B2C space, because I have seen advertising initiatives that have cut their cost of acquisition, I kid you not, by 30%, 40%, 50% by introducing lookalike targeting, specifically demand generation initiatives in the B2C space on Facebook where we were paying $20 a lead, $25 a lead. Then, when we began modeling people who had previously bought the products through Facebook’s lookalike modeling, that cost of acquisition went down to 10 bucks, 12 bucks, sometimes 8 bucks. It’s extremely effective.
In more cases than not, I’ve had that impact with lookalike modeling. The problem has always been, as a B2B marketer, we just don’t have that embarrassing wealth of data about people in their positions, and where they are and what they’re looking for, like we do on Facebook. One of the eggs that had to be cracked in order to make this possible was how to go about getting past the old stale targeting methodologies that B2B marketers use, which are heavily reliant upon firmographics and contact level data which is gained in a bunch of quasi-nefarious ways.
Somebody had to think through that problem and arrive at a solution in order for it to even be possible in the B2B space. My expectation is that, to the extent that that process works, we’ll begin to see similar results in the B2B space as I have seen in the B2C space. The cost of acquisition is going way down because it essentially cuts all the fat out of targeting, if that makes sense.
[0:08:01.2]DG:Yeah. I think that’s a good way to think about it. It cuts the fat out of the targeting. As I was saying earlier, there’s a lot of the audience that you’re paying for who are really not a good fit. They don’t tend to convert. They don’t tend to even respond. Then, there’s a lot of the audience that you’re not getting because you’re not targeting them properly. You don’t have that option.
[0:08:24.4]JG:Yeah. Some of the problems are industry data is grossly inaccurate. It’s developed on codes that were not expressly intended for that purpose. A lot of what people who tried to solve this problem in the past have done has been to approach lookalike modeling based on firmographics. But firmographics are imperfect.
In a way, it’s like trying to shove a circle into a square hole or vice versa. Somebody had to come along and take on that problem of B2B data. That’s one of the things I’m most excited about in working with LeadCrunch. Those are ostensibly the people who have enough courage to try to crack that nut. It’s not necessarily an easy thing to do, but we’re beginning to see the output of that be pretty effective now.
[0:09:23.0]DG:Yeah. If you’ve ever worked with traditional industry codes, they come in three flavors. There are SIC codes, NAICS codes, and then there’s a do-it-yourself model. If you go onto LinkedIn, and you use LinkedIn navigator, for example, there’ll be a pick list of industries that they’ve created and decided those are representative of what most people do. Usually, you look at that and there’s something there that doesn’t quite fit with what you’re looking for or who you are, but you have to jam yourself into one of the categories.
The problem is that those things often don’t reflect what you are actually doing in some significant ways. The first is a really common problem. They conflate what you sell, or who you sell to with what you do. It’s very common for a software company that sells software for hospitals to be labeled as part of the “healthcare industry”. Well, no, it’s a software company that just sells to a market. Who you sell to is not the same thing as what you sell.
The other thing that happens is that companies are almost always doing more than one thing, yet it looks there’s a static picture implying all they do is this one thing. How often do you see a gas station that’s got a fast-food restaurant hooked up to it. Then you get to the big companies like Google, and they’ve got restaurants inside of Google to feed the employees. They’ve got software that they make like Google Docs. They’ve got a big, giant ad business with search. They’ve got another ad business with YouTube. You know what I mean? It’s a very diverse company. To what degree it is any of those things is often important. Those are just a couple of ways where I think the problems with the foundation of data for doing lookalike models is not that great. Trying to address that, I think, is really key
[0:11:28.5]JG:How good could the outcome possibly be if it was based upon really, really flawed data? That, I think, has been the problem up until now. I’m really pleased to see that our data scientists, in particular, but the data science industry is starting to come up with ways to address those problems and dial that in using some really, really smart artificial intelligence technology.
I’m hopeful that we’re going to be able to extrapolate this in all kinds of uses. I think for account-based marketing, the implications are profound. If we can take a list of companies that you’ve already successfully worked with and extrapolate the algorithmic similarities between those companies and the rest of the market at large, return to you hundreds and hundreds of similar algorithmically similar companies, and then enrich that at the account level, we can essentially give you account-based marketing lists that will far outperform any general targeting that you would be able to do by leaps and bounds. I also think that you can go further than that. You can curate display advertising with this technology. You can do all sorts of really cool things with it.
[0:12:48.4]DG:The other part about this is that trying to figure out who your ideal customer is, the characteristics of that customer, and who you should therefore look for in the marketplace absolutely changes from company to company. I was talking to somebody who had an HR solution. The thing that really moves the needle for them is when companies are growing. They started adding headcount and however they were handling their HR stuff starts to break. They start looking around for alternate ways to manage people through the HR systems that they have.
They were going out and looking at things like, “Is headcount changing?” Or, “Are there words similar to ‘growth’ on their website?” There are signals out there that can begin to tell you that these people have the propensity to be a good customer of yours. How much of an HR staff do they have relative to the overall population of people? There are a lot of things you can do to hone in on who your ideal customer is based upon who you’ve sold to in the past and who tends to be very profitable for you to work with.
[0:14:18.3]JG:Yeah. There are a lot of ways to look at it, too. Techstack is another way to skin that cat. Startup organizations tend to have certain software they work with. As they reach scale, some of those software become problematic. We can just dive into people’s text stack and tell, to some extent, how big this business is, how much revenue it’s doing, how many contacts are circulating because they had to ultimately switch from Constant Contact to Infusionsoft or HubSpot. Now they have a Salesforce instance.
All those things are telling, in some ways, of the scale of the company. More importantly, when you get all these different, we call them vectors, but dimensionalities to the data, it starts to be really, really revealing and really accurate. We’re not there yet, but my next prediction around artificial intelligence is that because of the nature and the sources of the data and the agnosticity of them in the B2B space, once somebody has fully cracked this egg, I think that, eventually, B2B lookalike modeling will actually be more effective than B2C lookalike modeling. A lot of that B2C data is self-volunteered and therefore suffers from the Hawthorne effect. People offer up data about themselves that’s not entirely true and therefore skews the modeling.
That’s not the case in the B2B space. We actually have an opportunity, if this is fleshed out and taken to the nth degree, where this could be much more accurate and much more effective than B2C lookalike modeling in the long run.
[0:16:01.2]DG:Yeah. I think there’s huge promise with this. How are we doing on time?
[0:16:07.5]JG:We’re probably about 15 minutes in, so we should we wrap it up pretty soon
[0:16:12.6]DG:All right. What are your concluding thoughts about B2B lookalike audiences, Jonathan?
[0:16:20.4]JG:I think you should try, if you don’t believe the hype, give it a whirl. It’s pretty easy to do. For instance, our company offers this technology on a CPL basis. You can do a pilot program with a certain number of leads, testing the efficacy of them, and see how it performs. There’s really very little risk involved with it, but I think, as this technology continues to mature, you’re going to look like a genius for being on the front end of this and not on the backend of the adopting curve.
Lookalike Modeling & Machine Learning – How Models Get Smarter
[0:16:51.7]DG:I want to add one thing to what you said. The really important part for people to understand about lookalikes is that you can tune the model. The way you tune the model is that you generate leads and you feed all the leads back into the model with however far you got with them with your sales organization in your nurturing. You can tune the model and say, “Okay, I see. We have some characteristics that we used that we probably shouldn’t have. Let’s throw those out, and we need to weigh these other characteristics in a different way.”
Like everything in B2B, it’s going to take a little bit of time because you can eat enough of a sample to reach validity. That is a part of it. It’s like going on a digital transformation. That’s not an event. It’s not like, “Okay, we got Salesforce and HubSpot. Now it’s over.” No, you’re going to get a data management platform at some point and this, that, and the other thing. It’s a journey. This is a journey, too, and you’re going to need to prepare everybody that you need to stick with it awhile if you really want to get the fruits out of it. In your case, the journey you took on the B2C side, how long did you invest before you got to these insane conversion outcomes?
[0:18:22.2]JG:Yeah, the learning which data points to pivot upon, etc. Also not for nothing. If you plug great targeting into the bad marketing, it still doesn’t perform. You have to know what you’re doing with messaging and ad construct and placement and all those things. They’re on different levels. In 10 years, this is going to be the only way it’s done. You might as well get your head around the technology now and start learning the platform.
[0:18:54.0]DG:The other side of that coin is you can have perfectly awesome content and you get it in front of the wrong eyeballs, and it’s also not going to perform. You really do have to hit it on both sides or you’re going to be underperforming.
[0:19:10.8]JG:Yeah. Every demand generation marketer needs to be an expert copywriter and an expert targeter if they want to have success.
[0:19:18.6]DG:Now actually a podcaster and a video person.
[0:19:24.8]JG:You don’t even have to be good-looking for that. I’ll just get a camera
[0:19:29.9]DG:Yeah. We’re going to get some lookalikes who look better than us for the future episodes folks. We’ll put ourselves into the AI machine and good-looking, handsome, younger guys are going to pop out.
Anyway, Jonathan, hey, thanks so much for your thoughts today. Really appreciate it
[0:19:46.6]JG:All right, guys. That has been another successful episode of the Green & Greene Show. Thanks, ladies and gentlemen. It’s been real. Have a good one.
[0:19:51.8]Announcer:Thank you for tuning into the Green & Greene Show by LeadCrunch. Green & Greene think differently about B2B and want to start a movement to transform demand gen. If you have ideas for topics or would like to be a guest, send an e-mail to firstname.lastname@example.org. If you’d like to find more customers, visit our website to talk to one of our demand gen guides, www.leadcrunch.com.[END]