5.2 What is attribution? - Video Tutorials & Practice Problems
Video duration:
12m
Play a video:
<v ->So, how do we credit marketing tactics?</v> How do we know which things we did in marketing actually led to a sale that uses a technique called attribution, where we try to attribute a certain set of marketing activities to the actual sale? And so you might imagine that customers are interacting through multiple channels. They have been exposed to different marketing messages along the way, and you need some method to decide what you're giving credit for, for that sale. And so if you look at this example on this chart, this is an example of single-touch. In fact, last-touch attribution. So even though the customer throughout the customer journey was exposed to several different types of marketing tactics, the thing they did last was did a search in Google, and then they bought something. And so a last-touch attribution method means that all of the credit for that sale goes to the thing they did last, that Google search. Now you might say to yourself, "Well, that doesn't sound terribly accurate. Why would I want to give Google all of the credit for that sale? Because I'm sure that my display ad, and my email and the phone call that they made had something to do with it as well. Why aren't I giving credit to other things that happened, not just that last-touch?" And so that is the decision you need to make as you're deciding what your attribution model is. And so there's basically three types of attribution model. We've already shown you what a single-touch attribution model looks like, specifically, we showed you last-touch. You could also have a different model that says, for example, first-touch. That first thing that we did that started them down the journey. We're gonna give them all the credit. So single-touch attribution gives one particular tactic, all of the credit. It could be the first one, could be the last one. That's the most common, but basically that's what it does. And we call it a rule-based method because the rule is whatever they did last gets the credit or whatever they did first gets the credit. You can express it as a rule. Now you might be interested in something better than that. Something that you think is gonna be a little more accurate, a multi-touch attribution method. Multi-touch is gonna give some credit to several different touches along the way. Now there's two types of multi-touch models. Rule-based, so that's one that can be expressed as a rule, just like the single-touch one was. And you can see an example here that says, "Hey, we're gonna give the first, and last-touches most of the credit. And then we're going to divide up the credit among the things in the middle." And that is an example of a rule. An algorithmic-based attribution method uses data to decide how to weigh each touch based on its position of when it happened first, last in the middle and what kind of touch it was. It might know for example, that Google searches are much more valuable in persuading people than your email was. And so that algorithmic model is another version of a multi-touch model. Now none of these models are perfect, but they're all useful in different situations. And there are plenty to choose from. So let's look at some of those. So rule-based attribution models are pretty simple. And remember all of those single-touch models are rule-based, but a lot of the multi-touch ones are too. And so they're very easy to implement. In fact, single-touch models that last-touch attribution method, the reason so many people use it is because that is the easiest method to implement in your web analytics system. It's very simple because you're immediately tying the thing that happened in the same session with what drove them to that session. And so it's extremely simple to implement, and that's why it's the default method for almost all companies. It might be the way your company works, but you don't have to settle for that. You can use an attribution method that does better than that by trying to collect multiple touchpoints, and knowing that it was the same person that saw it. And so we already talked about how a position-based model works. And maybe gives you most of the credit for the first and last, and it divides the rest among the touches in the middle. First and last is a variation on that, where it gives you all the credit to the first and the last, then no credit to what's in the middle. We've talked about first-touch, and last-touch time decay. That one says, "Hey, we're gonna give most of the credit to the things that happen later in the process with a little credit to the things that happened in the beginning." And linear is a simple one where we say, "We're going to give credit equally to every touch." Any of these are a perfectly legitimate way of doing attribution. It depends on your business, which one might make sense for you. So we talked about that open-assist closed model, that position-based model, where we're basically giving some of the credit to what's in the middle, but most of the credit to first and last, and you can see the examples here of how that mathematically works itself out. If you have between one and three touches, it ends up being divided equally. But when you have more than four touches, like this example here, you can see that most of the credit, 28% each goes to the first and last-touch, and then 9% each is divided among those middle touches. Another thing you have to design is what your window is? So does your window close when someone buys something or could they actually have two things that they're working on at the same time, depending on your business, you need to answer questions like that. But what about those algorithmic models, don't they seem the best? Well, maybe, but they're actually the most complicated. So there's a couple of reasons why you don't just dive into an algorithmic model. One of them is that they need data scientists, but an even simpler one is that they need data. So why does it need data? Well, if you're on a last-touch model, because that's the default, and nobody ever changed it, well, you are not even collecting any of the data around those other touches. So you don't even have any historical data about what other things the person might have been exposed to before whatever that last-touch prompted them to convert. And so you don't have the data to even do an algorithmic model yet. So if you're on that single-touch model, the first thing you have to do is to go to a rules-based multi-touch model first, then once you've started collecting the data, now you can start to move to an algorithmic model if you think that that's worth it for you. Now, the data scientists will do some clever things. Markov chains can test removal effect. So suppose you had hundreds of journeys that each contained in them a display ad, a search, an email touch, and a social touch, but then you also had hundreds more that were exactly the same, but they didn't have the display ad. Well, by comparing the conversion rates against those two groups, you'll know what the removal effect was of the display ad, and that'll help you understand how important the display ad was to increasing the conversion rate in those second set of journeys. Similarly, you can use game theory to test adding something. You could use that exact same data, and you could look at all of the journeys that had the search, the email and the social, and then compare them against the ones that now added the display. And so both of these are doing pretty much the same thing, but they're using two different techniques that have to do whether you're testing removal or addition. Now you need a lot of data in order to be able to do this. That's why you have to be collecting that data for a while with those multi-touch models that are rules-based first before you get into this, but these are the kinds of things data scientists can do that will not only help you understand whether the position of a touch matters. So was it more important for that search to come first, last or in the middle, but also they can test the type of touch it was. So are display ads more important than email? Are searches more important than social touches? And so this is really what these types of algorithmic approaches can do. And as you can imagine, they can be a lot more accurate, and helping you understand the return on investment for different types of marketing tactics. So these are the most sophisticated attribution models, and can help you the most in your spending for your marketing mix. And so you might try and compare single-touch to multi-touch models. The biggest difference between single-touch, and multi-touch models is not only that the multi-touch models are more accurate, but because they're collecting data for all sorts of touches that didn't immediately result in a conversion. You now have data about what types of marketing tactics resulted in an abandoned situation. So you not only know which touches resulted in conversion, you know which resulted in people not converting. And so that can help you understand that maybe something like a display ad is just something that people see all the time, and it doesn't have much impact on conversion. So yes, there are lots of people that convert, but they're also people that see the ad, and don't convert. And so they're collecting all of these failed journeys that helps you to know a lot more information than the single-touch models do, because they're only collecting data from journeys that succeeded in a conversion. And so you might compare which attribution model works the best. So single-touch is easy. It works out of the box with your analytics system, multi-touch takes some work or requires collecting that data. It requires you pulling together different touches so that you know it was from the same person. So it has more complexity, and requires a little more investment, but it lets you customize the rules, and it gets you to a more accurate result. Similarly, the algorithmic multi-touch is the most expensive of all. It's the hardest to do, but it also might be the more accurate, and give you the most valuable information for your marketing mix. Now, how are everybody else doing attribution? If you think you're the only dumb kid on the block that's doing last-touch attribution, it's not true, but you might look at surveys like this one that say, "How are marketers doing ROI? How are they figuring that out?" And you see here, this number that says, "Holy crap. Look at this. Over 40% of people are using data science methods for attribution, we're behind everybody." Well, you're probably not. Because what's really happening here is these are people answering a survey question, and what it means is that they have some data scientists somewhere that's been looking at this, or they have some analytics person that's been looking at this, and they are doing some kind of pilot. There are very few companies that are using algorithmic modeling for all of their attribution methods. Almost every company is still using the dreaded last-touch attribution in many, many of their areas of their company. And so don't be too concerned about being behind other people, really focus on whether this is the value that you need to solve the marketing analytics problem that you have.