Wondering about the hidden potential of product segmentation in ecommerce? Join as we peel back the layers on this fascinating topic. We'll discuss the tension between efficiency and scale, and explore how product segmentation can help you navigate these opposing forces. Discover the relevance of the 1970s BCG Matrix in today's ecommerce world, learn about the role of algorithms in solving the fascinating "multi-armed bandit" problem, and get insights into Google's Shopping Graph and Audience Graph.
We'll dissect the intricacies of the modified BCG matrix ("heroes and zombies"), weighing its pros and cons. Get ready to see how a multi-dimensional approach, which considers a balance of performance data and off-channel data like margin, can offer potential benefits. We'll debate the limitations of two-dimensional models and the potential of AI in understanding multi-dimensional spaces. As we reach the end of our journey, we'll reflect on the challenging economic climate and its impact on marketing effectiveness. Let's challenge traditional best practices and consider the vast possibilities of product segmentation together. This episode promises to change the way you look at product segmentation!
Wondering about the hidden potential of product segmentation in ecommerce? Join as we peel back the layers on this fascinating topic. We'll discuss the tension between efficiency and scale, and explore how product segmentation can help you navigate these opposing forces. Discover the relevance of the 1970s BCG Matrix in today's ecommerce world, learn about the role of algorithms in solving the fascinating "multi-armed bandit" problem, and get insights into Google's Shopping Graph and Audience Graph.
We'll dissect the intricacies of the modified BCG matrix ("heroes and zombies"), weighing its pros and cons. Get ready to see how a multi-dimensional approach, which considers a balance of performance data and off-channel data like margin, can offer potential benefits. We'll debate the limitations of two-dimensional models and the potential of AI in understanding multi-dimensional spaces. As we reach the end of our journey, we'll reflect on the challenging economic climate and its impact on marketing effectiveness. Let's challenge traditional best practices and consider the vast possibilities of product segmentation together. This episode promises to change the way you look at product segmentation!
Welcome to Growing Ecommerce. I'm your host, mike Ryan of Smarter Ecommerce, also known as Mech. Today we're discussing product segmentation. This is a topic that has interested me for years, way back to my days in purchasing, and it's a very large topic. I'll write more about it later this year, but today I mostly want to limit the scope to segmentation strategies seen in paid campaigns. We'll talk about why is segmentation so important, why is it more popular right now than seemingly ever? How might an algorithm think about products and how is that different than we humans do? And do the guru strategies really work? All that and more so. Stay tuned. All right, let's get into it.
Speaker 1:So I don't know about you, but I've been hearing and reading a lot about product segmentation lately, one of the most popular things that we're going to talk about. This one in some detail, but it's this idea of in your paid media segment to your products along the lines of like leaders and leaders, heroes and zombies, above index, below index I see different terms for this, ghost products, all kinds of terms and it's kind of at the heart of it is a age old question in advertising, which is a sort of tension or these opposing goals between efficiency and scale. And there's a longer history behind these concepts too, like heroes and zombies. You know I would chase that personally all the way back to 1970, when Boston Consulting Group introduced what's now known commonly as the BCG Matrix I think it's called the product portfolio or a growth market share matrix. Basically they had X and a Y axis or horizontal, vertical axis, and one axis is mapped to market share and the other one is mapped to the growth rate, and the idea is that you can make kind of a scatter plot out of that and map your products on there and then you can turn that into a quadrant analysis or a four box analysis and see in which kind of corner each product is located. So, for example, high growth and high market share or low growth and high market share at these different combinations and this comes up with categories like dogs and cash cows. So that's been around for ages and really popular. Actually, a few episodes I was talking with Stefan Pires, who's an expert in Amazon ads and he loves still using the BCG Matrix and Amazon ads for product segmentation to this day Now. More recently I think we've seen that, probably popularized by Product Hero Dutch company and they kind of rebranded the BCG Matrix with concepts like heroes and zombies, and then I believe everything that we're seeing lately is probably a spin off of that. But some of these concepts go back longer, like I don't know, like 2016 or something. We had a script called budget eaters and it was designed to identify these things, and I know a lot of feed providers have focused on sleepers or zombie products for a long time, so those could be products that are not serving at all. They're not getting any impressions in your ad campaigns, or maybe they're getting impressions but no clicks, or there's different kinds of scenarios there or how you might define that, but it's a way of trying to wake up products and that's a topic that goes back years.
Speaker 1:There are also many other ways of segmenting products. For a long time, having margin based campaigns has been a popular tactic. It still is. There's different things you can do here to categorize your products, but before we get into some more detail there, I want to think about when we're operating in an ad platform like Google or like Meta, with their advantage plus shopping, any of these highly automated campaign types. That automation is driven typically by machine learning or AI, by some kind of an algorithm or ensemble of algorithms and you know they have a problem to solve, like some of the things why there are zombie products or sleepers, or why you have bestsellers and why you have budget eaters. These things are partly an inevitable characteristic of the channel. Like, if we look at Google Search, for example, it depends on search volume to trigger ads. So there's just going to be products that benefit from more demand and more search volume and there's going to be products that have less demand and less search volume. And in the algorithms job this is a common problem.
Speaker 1:It's called a multi-arm to bandit problem, which is yeah, let's just unpack that for a minute. You know a nickname for a slot machine in Vegas is a one arm to bandit Because you know it's got the lever that you pull on the machine. That's the one arm and it's a bandit because it steals your money. Now the whole idea is like with the multi-armed bandit problem is that imagine you've got a series of slot machines in front of you and you're trying to kind of maximize your return between these different machines, because they've all got a different probability and you want to exploit the machines that are working, but you also need to explore and look for new machines. That could be the next winner for you Generally, that thing that we try to solve, whether through our strategies, like our segmentation strategies, or through a machine, we're trying to solve that multi-armed bandit problem, solve that tension between exploration and exploitation, because the exploitation that's what's working.
Speaker 1:Those are like your best sellers and stuff. But you'll run into a ceiling there. You need to constantly also be looking around for the next thing. So you need to be exploring, but that comes with opportunity costs, because you're spending budget or a fixed resource on unknowns, on risks, and you could be spending it on the exploitation mode, like using it, using your money on a proven winner, and so the challenge is basically just making that as efficient and effective as possible. You want to use as many clues as you can to kind of shortcut that process, that discovery process, the exploration process.
Speaker 1:So, for example, google has like a product graph called the shopping graph, where they know a lot about different products and they also have an audience graph. They know a lot, they understand the audience as well and they kind of marry these two things and they look at the audiences on hand. They also have to, by the way, be in exploration and exploitation mode on audiences, not just products. So it's a pretty complex challenge and they need to be testing products. You know you've got a new product in added to your feed, a new product that you're advertising, or it's a product that Google just doesn't understand as well or hasn't specifically given a shot. And with this topic about sleepers and so on, it's tricky because also what advertisers want from Google is kind of contradictory, like people are very concerned about products that are inefficient, that are generating like a lot of clicks and no conversions. They're also concerned that there are products that the algorithm never gave a shot or or prematurely decided that this product doesn't have potential and they want to force the algorithm to reconsider that product. So that whole opportunity space, it's a risky area, you know it's.
Speaker 1:If you think in terms of like a 80-20 principle which I don't always agree with, but that's definitely in that 20% area and you're not dealing with low hanging fruit anymore, it's arguably more incremental to your business to get these sales because, like, a best seller doesn't really need a lot of attention in paid media. Sure, it needs some, but by and large it just needs your budget. It needs money to go out there and capitalize on the demand and you know it might also come at the expense of selling those products for free through organic channels or other kind of channels. So these other products we often call like the long tail, because they don't have a lot of data compared to like the head products that have our data rich and have a high volume. It's just a risky proposition getting out there after those. If you're too aggressive on them, then yeah, you are burning a budget.
Speaker 1:If you're too defensive on them let's say you're eliminating products that you think are wasting budget or that you think don't have potential then you're cutting short your future revenue, you're cutting short of headroom for growth and often the basis that you're operating on, you might think, is data driven, but maybe it's not. And what I mean with that comment is that it's so often that you don't really have a significant amount of data. I mean, you'll see strategies out there like okay, figure out the average number of clicks needed per conversion and give a product that many clicks plus X percent and Z. You gave it a fair chance and I think that's one of the better approaches out there. It's a way of thinking about it, but we often we just don't know because there's just the data really is sparse in many cases.
Speaker 1:So I think that this area of activity product segmentation has become more popular in recent years because these campaigns are more automated, the bidding is more automated. Even now, amazon has a more automated campaign type out now called Performance Plus. Tiktok has one called Smart Performance. These things all kind of run together in me. There's like it goes Smart Shopping, performance, mac, smart Performance, then Advantage Plus and then Performance Plus. I mean I remember hearing criticisms about how generic and samey some D2C brands are Like with their you know, sarif font, white Sarif excuse me, white Sarif font on a brightly colored background and all this stuff. But yeah, the platforms are just as guilty of samey branding. But as a result of these campaign types, you know, there are fewer promotional levers out there. There are fewer tactical levers, fewer knobs and dials, and people are looking for ways to influence the campaigns. They're looking for ways to be promotional, to be tactical, and I don't think that there's anything wrong with that. I think it's a good thing.
Speaker 1:I like product segmentation. As I told you, I've been thinking about product segmentation from back when I was on the buying side. You know I had to really understand which products needed to be like. We were in the appliance business and, for example, we had service trucks full of repair parts. I had to decide which parts need to be on these trucks and why. I were enlightened and I had to decide which parts are gonna which lights are gonna be in our showroom and why. How much are we gonna carry of different kinds of stock and inventory and looking at the overhead costs and the sell-through rate and all kinds of things. So that's a side of e-commerce and advertising that I really love.
Speaker 1:So, to talk through a couple of popular strategies in some more detail, I mean let's start with margin buckets Also because I've been accused of being, or rightly labeled as being, a promoter of this strategy and, in case you're not aware, the idea is basically that maybe you can't, from a technical standpoint or an organizational standpoint, can't get your profit data into an ad platform like Google Ads, but you wanna and you're having a hard time on the back end or you like to pay in what your shop system is, or so on and so forth. You're having a hard time figuring out if your ads are really profitable. You might also feel uncomfortable sharing that information with the ad platform directly, but a kind of a quick fix relatively quick fix is to add a custom label in your Facebook feed or your Google Ads feed that has the margin class whether it's like high, medium, low or as a percentage and then you can use this label to create campaigns and apply more or less advertising pressure, allocate more or less budget or expect more or less efficiency for products in a given margin class. Now the kind of the plus side of that where I am positive about it is that it's a step. It's a forward step toward understanding how profitable your ad campaigns are and not purely looking at revenue and not mistaking return on ad spend for profitability.
Speaker 1:I recorded a whole episode about that one time. The downside or downsides there are some this is not going to work as well as just tracking profit and actually bidding or optimizing toward profit. And I think there's a partly misconception that comes up here. Like, yeah, you don't want your campaigns to be too profitable. Like I think there's a misconception that the idea is to make the campaigns as profitable as physically possible and that's not a good idea either, because then you're going to start losing volume and that's not really the point. The point is to maximize your profit in absolute terms, not maximize your profitability necessarily, but I've talked about that before.
Speaker 1:And this segmentation strategy has the other things that the item click does not necessarily going to be the item that's bought, so it has a looseness in it because the actual basket or order of composition can really differ from the product that you're advertising, and there's workarounds for all this stuff, and I think the main thing is basically also how smartly you create your margin classes or your margin buckets, where you really need to look at the data and determine an appropriate amount of segmentation where it makes sense, based on the distribution that exists in your inventory and then also like the distribution of conversions in the ad channel. It has to work for both so that you're not just having one giant bucket and then five tiny buckets. That makes no sense. We need to structure a bit more carefully on that and the answer might be that you know if you don't have enough difference or variance in margin, that doesn't make sense. But I still think it's a step. It's a very popular option, so I want to mention it. I'm not a margin segmentation hater. I also don't think that it's a magic wand. I would rather well we'll get to that. We'll get to that.
Speaker 1:I don't want to get ahead of myself. I want to spend a bit of time talking about this modified BCG matrix where you've got like heroes and zombies, for example, or bleeders and leaders. My reservations here like first off, I like this approach. Let me just say that the good stuff I like this approach. It's simple, it's easy to understand. You basically, yeah, on one axis you've got efficiency, on the other axis axis you've got volume. Then there's different setups, but this then creates outputs like high efficiency and high volume and in principle you want to maximize those. The reality is that if there's already high volume and the efficiency is still really high, then they're probably already pretty maxed out and they just have naturally a high efficiency level and that you're not going to be able to push more. But then there could be, of course, hidden gems and cost budget years and all kinds of things in there when you categorize them this way. So it's a really attractive logic. It's simple, it's easy to understand. That's also a shortcoming. We'll get back to that. But to stay on the positive, it's easy to implement, it can be a quick win. There can be once again a bit of an 80-20 principle here, where you've gotten a good effect without doing a lot of work. There are free scripts for this that you can use and modify. There's a lot of options.
Speaker 1:Now to get back to that sort of more negative side, it's that it is two-dimensional, that it is so simple, because your inventory, your business, your campaigns these things are likely not so simple. They're likely multi-dimensional. Like, let's say that we want to consider margin. Earlier we were talking about margin and it's probably not right to have that as the standalone logic, that one-dimensional logic for campaign segmentation. I think that's not the best way. But let's take volume and some of the in-platform metrics, like conversion rate, for example, and then if you imagine that that's a square, that you've got those on two axes, so they make a square, they make that four-box model. Then as soon as you add a third metric in there, like margin, imagine that as kind of a z-axis that's going into the distance. You've created a cube, you've created something that's now three dimensional. Then you might also want to consider off-channel data that Google doesn't know about but that's important to you, like the sell-through rate. If a product is behind on sell-through rate, maybe you want to make sure that it's getting additional advertising pressure. You might look at the organic ranking. Maybe the organic ranking is not good and it needs more advertising pressure. Then you've now you're up to like five dimensions and you're connecting lots and lots of dimensions.
Speaker 1:Ultimately, we can't as humans we can't imagine a shape like that. We can't imagine a seven-dimensional object or an n-dimensional object. There's no limit to this. This is an advantage of machine learning or AI is that it can handle these spaces well. We can understand that there's a multi-dimensionality here. We might have an intuitive logic of that or we might deploy some maths or some formulas to try to solve it. But it is a limit of the way that we think about things. And if you look at, let's talk about I hate to do it, but let's talk about chat GPT for a minute, or AI, it is able to yeah, it is able to understand things in many dimensions. It's able to kind of graph things in a multi-dimensional space pretty successfully. We can do that too, implicitly in our unconscious mind, with different concepts too. But let's not get into the weeds here. In certain use cases we can do that. It's hard for us to do that about a product, right? So before I lose the plot here. Just to wrap that point up.
Speaker 1:These models are inherently two-dimensional and they're missing a lot of complexity and a lot of potential. Sometimes they're not even two-dimensional, sometimes they're more like one-dimensional. I've seen specific implementations where one axis will be return on ad spend and on the other axis is conversion rate. This is messed up. I'm sorry, but that is not a correct way of implementing that In my professional opinion. Reason why is that these two metrics are heavily correlated with each other. They're arguably dependent on each other. Rather, roas has a dependency on conversion rate.
Speaker 1:There are three main things that go into your return on ad spend. That would be the cost, the unit cost like cost per click, for example, the conversion rate and the order value. And when you have two dimensions that are so highly correlated or even dependent, you're not even doing a two-dimensional analysis. You're back to a one-dimensional analysis, like the margin splits that we talked about earlier. I don't think that's the right way, especially since the platforms are pretty good at taking care of return on ad spend. At that point you're competing with the algorithm. You're not really supporting it. You're double optimizing, if you will. I don't think that's the goal here. I think really a better way of doing that is to find metrics that are purposely uncorrelated or even opposing to each other and rather have a tension between the metrics, so that you're triangulating toward a balance of goals.
Speaker 1:Ideally, I think it's great when you're adding things from your first-party data, and I don't just mean first-party audiences. I also think we have all kinds of first-party data, and it's limiting to just talk about first-party audiences. You've got all kinds of product data that Google is not aware of. That's not in their graph, like the margin that we already mentioned when you're freeing, or if you know that a product is associated with lifetime value, or that a product is better at acquiring new customers than another product. There's anything, any number of things, going on here. If you can bring that data in, you've succeeded because you're supporting the algorithm. You're leveling it up instead of competing against it or trying to play with the thermostat, if that makes sense.
Speaker 1:Our approach to this is to use machine learning. Our model is able to evaluate hundreds of attributes, and those are standard feed attributes, but other data as well, and also it's open to first-party data that can be integrated. The point is to then model product neighborhoods. It also has a predictive value at that point because, instead of just being based on historical data and most of the products, by the way, don't have any data when you create a product neighborhood, you can assess the performance potential of an item, even one that doesn't have data or only has a little bit of data. Everything is better together. You've got a more nuanced approach, so that's something I just think it's a better way of handling it.
Speaker 1:And a last point that I want to make for now about product segmentation is that there's no kind of magic number. Like I've said this before on the, I believe in hammering home the message. I've said it before on this podcast, I've written it online many times there's no magic number to segmentation. We know that the way people used to segment, the idea like granular granular was such a buzzword back in the day. I should really pop that in Google friends and see what that chart looks like. But I was so sick of the word granular and every time I just saw like granulated sugar or like table salt and it was such it was the strangest buzzword. Granularity was King or Queen back then. And now people have moved toward more consolidated approaches and that's good, and it's another reason why product segmentation is really top of mind for people because they're trying to find the right logic and they're looking for a mental model or a quick win here, and I think that there is. There is potential for that.
Speaker 1:But back to the, to the message I want to hammer at home. There's no magic number for the amount of segmentation that you have. What what's important is that you are finding and addressing meaningful difference, and maybe you use machine learning to help you find that difference. Maybe you use simple maths, maybe you use a four box model. I think what matters is that you're trying and we just want to look for meaningful difference, in other words, meaningful two ways. There is enough of a difference that it's worth segmenting and that there's enough volume in the output in the segments that are created in the partitions. It needs to be meaningful in both of those ways or significant in both of those ways, because if there's not enough difference between the segments, then don't segment. If there's not enough volume in both of the segments, one of those is just kind of it's just going to get cut off and die and you're not even addressing, you know, a meaningful amount of your revenue or meaningful amount of your costs.
Speaker 1:And I think when you look at these, these, these simple frameworks, they're very appealing at face value, but when you look at the output of them, I think you sort of have to question the quality of the of the segments that were created. But I don't know. I said there's no magic number. I believe the magic number is greater than one. That's what I think I've, and I've seen the data to support this, by the way.
Speaker 1:But as soon as you are segmenting at all, you're doing something good for your business, you're doing something good for your campaigns. I believe in that firmly. I don't. I don't even have to believe. I know that it's important. But then, yeah, realistically, I think three to seven campaigns it depends. If you're massive, you know, have a massive amount of products or massive amount of spend or whatever the case might be there, it could be merited to have more. But three to seven campaigns is kind of a normal range and that would be sort of the interquartile range. When I look at segmentation, I've seen up to like 90, probably too much. I think I mentioned a few episodes back or one time. It still blows my mind. I still forgot to look and figure out why they're doing that. But maybe it's totally merited. But, yeah, let's to wrap this up.
Speaker 1:I would say that, yes, segmentation is important. Product segmentation is important. Your first party data is important, and I don't just mean audiences, I mean your product attributes and the things you know about your products that nobody else knows and frameworks. I don't want to diss the frameworks. They are a starting point. They're a starting point. Use it to help you. Use it as as a mental model, but not as an endpoint. It's not a destination. In my opinion, I think you need to test, learn, iterate, customize and maybe that doesn't mean that you're going to build a custom machine learning model or work with someone like us who has already built one. But the importance, the important thing, is that it's a journey and that you you keep moving forward on that.
Speaker 1:With the economic climate being what it is, this is recording. Time is Q4 and 2023. People have been talking about are are we going to be in a recession or not? For like over a year now, and now there's all kinds of headlines about student loans starting to need to be paid back and what's happening with all kinds of topics. But it is just a different environment than it was a couple of years ago and people are more concerned about their marketing effectiveness and not just maxing out volume or things like that, and I think that's appropriate. And I think product segmentation is a good tool in your toolkit. I'm a believer. I just want to invite you to take it as seriously as possible and not just take best practices at face value. So I hope that helps. Thanks for listening to Growing Ecommerce and if you enjoyed this podcast, please consider sharing it with coworkers, friends or within your professional network. We really appreciate it. This podcast is produced by a smarter e-commerce, also known as Mech. To learn more, visit smarter-ecommercecom.