HomeWebinarsMastering Attribution Modeling

Mastering Attribution Modeling

Overview

Attribution modeling - determining which of your marketing channels should get credit for each individual sale - is incredibly important for ecommerce businesses, especially when it comes to determining the value of your marketing strategy.

But it can also be incredibly difficult to get a handle on. With many different attribution models - and customers who, in a digital world, might interact with two or three or more of your marketing channels before they make a purchase - how do you determine which channel gets credit? 

In this webinar, host Mark Miano, VP of Sales at Glew.io, goes over:

  • Different attribution models
  • Why attribution is so important
  • Glew's first-order attribution model
  • How you can use attribution to drive more sales

Full transcript

Mark: And we are live today talking about mastering attribution modeling. Just going to give a quick check to my team to make sure that everything technical on the webinar software is working correctly. Okay, good. Excellent. Well, everyone, it's very nice to meet you again. If I haven't met you before, this is, I think, our third webinar series. My name is Mark Miano. I'm our VP of Sales here at Glew.io. I like to think of it as the most widely used, cutting-edge ecommerce and multichannel business intelligence product and platform. Wow. What a mouthful. That's right.

Mark: Very quickly, we're going to be covering attribution modeling today. We will leave some time for questions at the very end of the call. If you want to go ahead and jot down your questions in the chat window, please feel free to, and then my team is moderating and making sure that we get to as many of them as we can.

Mark: I have a feeling that at the end of this call, there will be a substantial amount of questions due to the complexity of how difficult attribution modeling actually is. Oh, and one more thing, if you're not currently a Glew user, a lot of what I'll be showing you today is inside one of our outputs, which is Glew.io. If you're not a current Glew user, feel free to launch a free trial or get in touch with me directly. My email is mark.miano@glew.io. Excellent. Or LinkedIn me, or whatever. Whatever you feel good about. So first of all, we're talking about attribution modeling. So - what is the definition of attribution modeling? Well, the textbook definition of attribution modeling is the art and science of assigning sales credits to different touch points in the customer journey.

Mark: Why is that important? Let's slow down: the art or science of assigning sales credit to different touch points in the customer journey. That is what attribution modeling is. The reason why it's important, especially for ecommerce and multichannel, is that we're typically trying to achieve goals. Typically for an ecommerce vendor or merchant, those goals are as follows: it's attracting the highest paying customer, not just any customer, but the highest paying customer and upselling those customers once acquired. The challenge is that we have a finite amount of resources to achieve those goals. Right? Specifically for attribution modeling, we're going to be talking a lot about advertising spend. I don't really care if you are Sephora, a major, major makeup company, or if you are Joe Schmo selling t-shirts out of the back of your truck. Everyone has a finite amount of resources when it comes to attracting the highest paying customer and upselling them.

Mark: So it's really important that we're understanding how we allocate those resources as efficiently and as strategically as possible. I think the last thing that's super important to talk about before we really get started is accomplishing both of those goals and allocating these resources in a way where we're beefing up our profitability. So that's a whole lot, right? It's a very tough thing to do, which is probably why we have a record amount of people joining the webinar. So let's go ahead and get started.

Mark: The problem that a lot of us have experienced is as follows. Let's say I made $1 million last year, keep the numbers round. Oh, well, okay. I want to see how much money Facebook generated me, and I go into Facebook and I see that Facebook generated $500,000 for me. At least that's what Facebook claims. And then I go into ad AdWords, and AdWords says, well, "Hey, I generated you $750,000."

Mark: Well, I ain't no Albert Einstein, but I made $1 million, and Facebook says I made half a million from them and Google AdWords says I made $750,000 from them. That's $250,000 extra dollars and thus we have a major, major problem, right? The problem is that these different systems are using different attribution models, i.e., they are grading their own homework in their own respective way. Right? And the reason why they're doing this is because of the classical principal-agent problem that we'll find in an economics class, right? For those of us who don't know what that is, think about when you go to a car mechanic and the car mechanic says, "Oh, hey, you need $10,000 of work on your car and you're like, "Hm, do I really need $10,000 of work on my car? Or do you just want to, you know, go to Hawaii on your next vacation?"

Mark: Same idea here - we have Facebook and Ad Words grading their own homework. What they're doing is they're picking attribution models that make their advertising platforms look as good as possible in front of you. So we have two problems. The first problem is - can we trust the data? I mean, I think that's a peripheral problem. The real problem is that we can't compare the numbers. To compare the amount of money that Facebook claims that they're generating for you with the amount of money that AdWords claims they're generating for you is like comparing a GMAT score with an MCAT score. Right? Both are tests and both are numbers, but both are grading different tests. So to compare those numbers would make absolutely no sense. And this is where Glew can provide an immense amount of light to help you allocate that ad spend as efficiently as possible.

Mark: Let's go ahead and start this conversation by saying there is no perfect attribution model. If you ever have a conversation with anyone and they say or claim that they have the perfect attribution model, please run in the other direction. Attribution modeling is so unique and so important to each respective business because each respective business is A, unique and B, running a unique business process and C, therefore will have a different way of grading how effectively that unique business process is.

Mark: In terms of attribution model examples, let's first think of the Facebook attribution model, since everyone is very much in tune with it. If I go to Performance and I go to Advertising, and let's actually start looking into the tool, what you'll see is a really beefy, great page demonstrating the attribution model of each respective advertising platform. Right now, that's just Facebook. And soon AdWords will be released as well. Facebook uses a one-day click, 28 - excuse me, one-day view, 28-day click attribution model. Meaning if a customer clicks on the Facebook advertisement, and then, let's say they come through AdWords six hours later, Facebook is still going to get credit for that sale, even though they bought through AdWords.

Mark: One-day click, or - excuse me. One-day view, 28-day click. Okay, I'm confused. One-day view, 28-day click. So let me reset that real quick as we get started here. If the advertisement scrolls across John Doe's screen in the past 24 hours, and he buys within that 24 hour window, Facebook is going to take credit for that sale. Or if said John Doe clicks on the Facebook advertisement and buys within 28 days, Facebook is going to take credit for that sale. My point in that example is that that's a very generous attribution model. You can see Facebook is taking credit for as many sales as they possibly can. Again, one-day view, 28-day click. If John Doe has viewed the advertisement in the past 24 hours, and/or John Doe has clicked on the advertisement within 28 days, Facebook is going to take credit for that sale. Fairly complicated even to explain it.

Mark: Another attribution model, which we're very much in tune with, is Google Analytics' last click attribution model, which you can actually find in our Orders List. What you'll see is the order ID and then the channel and the campaign that was clicked right before that order.

Mark: You know it's a very popular attribution model. But I'll be very honest with you, I'm not exactly sure what you do with it. The problem is that orders don't buy things, people do. So if we're going to use orders as the focal point, or what I like to call the common denominator in attribution, we're going to be skewing our decision-making. And let me give you an example. Let's say that I am a makeup company and I sell eyeliner, I don't know. And let's say Jane Doe comes in through Facebook first and buys through Facebook. Well, that first order will get credit for Jane Doe's sale, in this model. But let's say Jane Doe then knows that I exist as an eyeliner company and comes in direct for the next three orders. Well, in a last-click attribution model, direct is going to get credit for 75% of those sales.

Mark: I'm not sure that's a very effective strategy, right? Facebook, in my opinion, is the one that got Jane Doe excited and got Jane Doe converted and got Jane Doe hooked on eyeliner. And now, all of a sudden, we're going to be giving direct credit for 75% of the sales for Jane Doe. I'm not exactly sure if that's the right move.

Mark: So what we're going to talk about today is going to be what I call the most practical attribution model. The most practical attribution model is what you'll find throughout Glew, 99% of Glew, which is a first-purchase attribution model. What first purchase attribution model is, is the channel and/or the campaign that got John Doe to buy is going to get credit for all of John Doe's purchases. So, for example, follow me here, I'm in Performance and Overview and I'm hovering over March 2018. Let's say I'm John Doe and I come in through CPC, Google, March, 2018 and buy for the first time. CPC, Google will get credit for that sale because that was my first purchase. But let's say we fast forward through November and you retarget me through Facebook in November and I buy through Facebook a second time in November. Too bad, so sad for Facebook, AdWords is still going to get credit for that sale.

Mark: You still have a problem, though. We're looking at revenue by channel and net profit by channel. Just like orders don't buy things, channels don't buy things, either. So I don't want the viewers on this call to be going to their revenue by channel or net profit by channel with the first-purchase attribution model and then chasing after a channel because "I made a lot of money off of that channel this month," or whatever time interval I'm talking about, for example. Think about I got a wholesaler to come in through CPC, Google and to buy an outrageous amount of product from me, eyeliner from me all at once and he came in through CPC, Google, and I see that CPC, Google is getting credit for 50% of my sales last month. Well, that's an anomaly - just because one person dropped tens of thousands of dollars with me doesn't mean all of a sudden I put all of my advertising dollars in that one channel. It was one person. It was an anomaly, right? That's what I mean by channels don't buy things.

Mark: So because of this, what we're going to do - and for all those who have worked with me, you know how much I love lifetime value - we're actually going to not make decisions based on revenue by channel or net profit by channel. This is great reporting at the end of a sale schedule, but where we're going to make decisions is going to be in Lifetime Value underneath LTV Profitability by Channel. And what I'm gonna do is I'm going to take my channels and rank them in order of lifetime value. And now everyone should be having that a-ha moment as to why the first purchase attribution model is the most practical, because it is the only attribution model which allows you to assign a lifetime value to your advertising. The lifetime of the channel begins at the first purchase that John Doe has made with you.

Mark: So let's take a look at this. I have 872 people who buy from email for the first time in the past year and these guys paid $253 each with me. I then have 782 people who buy from affiliate for the first time and these guys pay $132 each with me. Okay. I'm no Albert Einstein, but there's efficiencies to be gained by focusing a little bit more on email from an acquisition standpoint juxtaposed to affiliate, right? So in a zero sum game, if I start moving my resources, which are finite, out of affiliate and into email, what is going to happen is the email - new customer cell right here is going to rise faster than my affiliate - new customer cell, and each new customer that I'm acquiring is natively starting over with a lifetime value of zero. So that new customer acquisition is going to put healthy downward pressure on the LTV of the strategy behind this channel.

Mark: And what I always say from an economic theory standpoint, not really reality, but if this column right here is the same dollar amount, the lifetime value per channel, what that would mean and tell me is that regardless of what channel I'm acquiring my customer from, I'm getting the same amount of money from that customer, meaning I've completely allocated my ad spend over time. I'll give you an example. So I work with a company that sells home automation devices, like, you know, electronic doorbells and cameras and stuff for your home, which are all the rage nowadays. We were looking at his data and we saw that his Bing LTV was near double his Google AdWords LTV. And that is profound. Guys. He had about three times the amount of customers coming from AdWords juxtaposed to Bing, but he was able to take away two deductions from that insight.

Mark: Deduction number one, there is massive amount of efficiency via moving a little bit of resources away from AdWords and placing that into Bing. There were efficiencies to be gained there from a marketing allocation standpoint. The second deduction was, "Hmm, maybe my products are suited better for an older demographic." Older people use Bing juxtaposed to AdWords, right? And to understand that the customers coming from Bing are spending more money with me than the people coming from AdWords is also profound about my own product offering.

Mark: Once we identify that lifetime value per channel, now we can incorporate other things like customer acquisition costs. My new customer acquisition cost is going to be my advertising spend, which is this column divided by my new customers, which is this column, right? How much money does it take me to win a new customer or buy a new customer from each channel? My lifetime value minus my customer acquisition costs gives me my profit per new customer, and now you can see just how powerful and important is to use a practical attribution model and to make sure that the common denominator that we're tying everything back to is the customer. Not the order, not the channel, but the customer himself or herself.

Mark: Let's take this down the rabbit hole. Okay. We've talked a lot about acquisition and that's great, but as everyone on the call knows, acquisition is just the beginning. There's two goals: we want to acquire the highest paying customer and we also want to upsell them. And we've talked a lot about acquisition. So let's talk a little bit more about the whole entire picture. And that's going to bubble up to the surface when it comes to looking at this data on a campaign level. Right now, especially if you're using Google Analytics, you don't need Glew or anyone really to tell you if a campaign made you money or not, right? You can just go into Google Analytics or Klaviyo or whatever and and take a quick look. But how about this, how about for the campaigns that are making you money? How about answering the question: How do I use those campaigns more effectively than I was using them before?

Mark: Let me give you a couple of examples. Let's say that a campaign has revenue and that campaign has high lifetime value. Remember, we're using a first purchase attribution model, okay? And in that scenario, the revenue will be clean. Um, that will always be a last click right here. And I know the campaign made me a lot of money. But if that campaign it has a high lifetime value, what that would tell me is that that's a great acquisition campaign, because it made me a lot of money and the LTV is high. The LTV column is not just getting credit for the sales of that campaign. If John Doe first converted from this campaign, the LTV credit is going to get credit for all of John Doe's purchases forever.

Mark: Let me give you a second example. What if I have a campaign that has high revenue and low lifetime value? It made me a lot of money, but the LTV is low. That's a good upsell campaign because it made me a lot of money, but the LTV credit is going to the campaign that got this group of people to buy for the first time, right? The campaign I just told you about, I'll give you an example. We work with a swimwear company. Their welcome email has high revenue, low LTV. You don't get a welcome email from them unless you buy something from them first. So it's a good demonstration of what I mean by a good acquisition strategy, or good acquisition campaign, versus a good upsell campaign.

Mark: The world is not a perfect place though. Okay. So I'll give you an example. I was with one of my clients who's a jewelry client and we were looking through this and we found a campaign that had high revenue and high LTV and, oh man, very exciting, we were all talking about, "what a good acquisition campaign."

Mark: But to throw a wrench in the analysis, we saw one new customer and, like, 200 repeat customers. And obviously that wasn't a good acquisition campaign, because we only had one new customer and 200 repeat customers. But the deduction was that that one new customer spent so much money with the store that that person singlehandedly skewed the LTV to be artificially high. So the action point was immediately, lightbulb moment, "wholesale relationship." So we went to Customers, Customer List, we went to create a segment, campaign, whatever it was, and we hunted the person down, all excited about a possible of a wholesale relationship - only to discover an individual from China who had bought one of every single one of her products from her store and was ripping her off in the Eastern Hemisphere. That's what I call the opposite of the wholesale relationship.

Mark: My point here, guys, is not to get so confusing. My point is to demonstrate that although there's no perfect attribution model, I call this the most practical attribution model due to the intense amounts of insight that can be derived from a one stop shop, juxtaposed to having to go to so many different sources and spending our time getting that analysis paralysis, which is obviously our enemy.

Mark: All right. That was a lot. I expect then, hopefully, we have a lot of questions coming our way. So I'll stop there and maybe we slow down and I'll start taking some questions. So let's give it a moment.

Abby: We had someone ask - what's the difference between LTV ROAS and ROAS?

Mark: Sure. Great question. So LTV, ROAS. I'm going to pull up up a generic sticky note here. So, bear with me. ROAS equals your revenue divided by ad spend, and LTV-based ROAS is going to equal your LTV multiplied by your new customers acquired in a given timeframe divided by ad spend. It's a really good question. ROAS is a go-to metric for many agencies and many merchants around the world, and rightfully so. It's probably one of the most straightforward ways of thinking about your business. It's very shortsighted though - i's only taking into account the revenue in a certain time window and the ad spend spent in a certain time window, meaning it is for very, very short periods of time - maximum, I would say three months. I want you to think of campaign level stuff.

Mark: For this, I want you to think of quarterly or seasonal strategy. In my opinion, you should not be using ROAS for, again, longer than three months. Think of it as short wind sprints to understand if a specific strategy that you're rolling out is giving you a return on that investment. LTV-based ROAS is much more important for longer-term strategy. Okay. I want you to think more on the channel level, one year intervals, things of that nature. You can see that it's the lifetime value of the store, which is a dynamic number, right? Your LTV changes over time. Hopefully it's growing and we're multiplying that by the new customers that you've acquired in a given time period - i.e., we are taking into account all the money you can ever expect from the new customers that you've acquired and dividing it by the ad spend that you've allocated in a given timeframe.

Mark: We're actually going to be publishing a blog on this very soon so we'll make sure to get that out there to have this documented for you to review later. But hopefully that answers your question. ROAS is very good for short sprints, taking into account the revenue in a given time period divided by the ad spend in a given time period. Whereas LTV-based ROAS is actually taking into account all the money you can ever expect from the new customers acquired in a given time period, for longer term measurements.

Abby: We have a question from someone. When we say new customers, are they all customers acquired organically and through other media, or should it be specific to the type of ad spend we spend across Google and Facebook?

Mark: Can you repeat that?

Abby: Yeah. When we say new customers, are they all customers acquired organically and through other media or should it be specific to the type of ad spend we spend across, through Google and Facebook. So I think they're asking, when you say new customers, are you speaking about specifically new customers to a channel or are you tracking all customers acquired through any channel, whether paid or organic?

Mark: It's a good question. You have to keep the new customers acquired and the ad spend consistent. So you know, you are placing money into an effort from an advertising standpoint. So you want to make sure that the customers that you're acquiring with that effort are tied to that channel. What do I mean by this? Let's think of organic right here. Organic, typically from my understanding, is an SEO play, right? I'm trying to jockey my SEO and investing in SEO. So the ad spend on this guy won't be cleanly pulled out of any one API, right? You're probably paying an agency or someone's salary to do what they can on the ad spend front. So for that question, the ad spend for organic would most likely be what you are attributing to your organic effort, and then you can go ahead and see how many new customers you acquired through the organic channel, and you know, if I hover my mouse over here, we actually give you all the formulas so that you can follow it through Excel or whatnot. But I hope that answers your question head-on.

Abby: I think so, yeah. Let me know if that didn't explain your question properly.

Mark: Cool.

Abby: I was going to ask you, Mark, to go into a little bit of detail about something we talked a little bit about while prepping for this, about specific industries where an attribution model like this might not apply. So, maybe an industry where there's a really long sale cycle and maybe you can expect them to make just one purchase. What attribution model might they use?

Mark: Absolutely. It's a great question, Abby, I'm glad you brought that up. When we're thinking of lifetime value, you know, normally I'm thinking - I'm going to bring up a lifetime value slide here - actually, can you see this? Yes. Okay. Normally, I'm thinking of an average purchase frequency-heavy type company, right? I'm thinking of anyone whose goal is to get the said customer to buy more than once. And that's a huge, huge majority of the ecommerce world today as we know it. Makeup, food, consumables, toys, you know, even phone cases. And the long tail of things is actually an average purchase frequency-heavier type type of strategy. What Abby brought up is more of an average order value-heavier type strategy, where maybe I'm Carvana, which is that new hot car company or I'm Sonos and I'm selling high end stereo systems.

Mark: And the whole point is, I'm buying a soundbar that's $5,000 so I don't have to buy another soundbar for another 20 years. Right. And I just had a discussion with a company is making extremely high-end air purifiers. Okay, these things are like $4,000. And the whole point is that the sale and the relationship with the customer really ends at that final sale, that one and only sale. So that begs a question, is lifetime value even relevant anymore? And frankly, in my opinion, it's not. All of a sudden, if the average purchase frequency of my company is one, right? Like Carvana, where I hopefully am only buying one car. Well, LTV, then, actually ends up becoming the average order value, making the whole attribution modeling, the whole first purchase attribution model, maybe not the best one, because the whole point of the first purchase attribution model is the ability to assign a lifetime value to a channel and to see that grow over time.

Mark: Right. So in Abby's question, what we need to think about is what are the customer touchpoints in that customer's journey, right? What are those touch points and how do we meaningfully assign sales credit to those touch points? The biggest challenge is going to be the fact that John Doe is anonymous in this particular situation. Let's think of that for a moment, John Doe is anonymous. The reason he's anonymous is because he hasn't bought anything from you yet. He hasn't gone into the store yet and typed in his email address so he can get that receipt from that purchase. Right? So I have a first purchase attribution model or last click attribution model. We have the email address. So the challenge that Abby is bringing up is how do we unanonymize John Doe? And there are many methods of doing this. Okay. One method that we've outlined is using something like Hubspot, where hubspot plants cookies.

Mark: You plant these cookies, anonymous cookies across people's browsers and then one day you ask them to maybe download a white paper or something, where they have to feed you their email address. And thus, we unanonymize John Doe, and then have the ability to start understanding the customer journey prior to that first purchase. The rule of thumb that I want you guys to walk away from this with is, as long as the data is available, we can assign sales credit to it. So the challenge isn't assigning the actual sales credit. The challenge is using the technology available to unanonymize John Doe. So we can have meaningful analytics at the end of the day and not anonymized analytics, which is what I call dirty data, you know?

Mark: I want to end on that note, if there aren't any other questions and we have some time. I want to pull up an example of what I'm talking about. So, really quickly, this slide, I'm not gonna share anything proprietary, but basically think of this visualization that we're getting, we're framing out as we speak in this way where I have an activity or event in the dropdown menu, the channel in the dropdown menu and the campaign in the dropdown menu. Basically what I'm trying to understand, and what Abby's question is, is okay, X percent of my customers maybe downloaded the white paper and bought, but maybe Y percent of my customers downloaded a white paper and bought through direct. X will be greater than Y. So, again, it's just about understanding what it is that you're trying to measure, unanonymizing John Doe and then cleanly assigning sales credit in a way that's logical.

Mark: Cool. Alright, well, I hope this was a blast and not too confusing. If you have any questions, please let me know. Email us, start a free trial. Typically for attribution modeling, a one-on-one conversation is more appropriate, due to the uniqueness of your business. But hopefully this has added a lot of value and we're very excited that you stopped by. Thanks so much.