David Booth, Cardinal Path - Marketing Analytics in the Age of Machine Learning

November 6, 2017

Marketing Analytics in the Age of Machine Learning

Slide 1:

Thanks everybody for taking part in this session. Obviously there have been a lot of great topics covered today and we are thrilled to be a part of this.

Slide 2:

We’re going to dive in and today, what I’d like to spend the next half an hour or so talking about is the age of machine learning and how we’re going to be able to use that in the marketing analytics function in the months and years to come. And I’ll kind of preface this by talking about different organizations being in different stages of their maturity levels with respect to this sort of thing. While some of the stuff we’ll talk about today is on the three to five-year plan for a lot of organizations, a lot of this stuff is happening today. Some of the cases and some of the examples we’ll look at are real examples of using AI and machine learning out there.

Slide 3:

I want to dive in by talking about this idea that we have entered a new stage in human kind. If we think back to the industrial revolution, this was a time when lots of change was happening. There was a lot of chaos in the world. People were essentially moving from being a farming, agricultural community into using machines. That meant a lot of change for a lot of people. For those who embraced it and move to the cities and learned how to use these machines and what not, it was obviously a time full of opportunities and there were also a lot of people left behind. I think right now we’re in the midst of what the history books or history Google glass looking goggles, or whatever we’re reading in the future, we will look back and say we were living in the middle of this digital revolution.

Unfortunately, when the next species comes and sees what we’ve done and are digging up our ruins, they won’t be finding pyramids and what not, they’ll be finding plastic and hard drives and things. But the reality is we’ve moved into this place where we’re in another age of chaos. Lots and lots of chaos, which means lots of change, lots of challenges, and of course lots of opportunities out there. An interesting way to think about this is right now the pace of change is so rapid, that when this particular article talks about it, when Dell talks about it, they say 85 percent of the jobs that will exist in just a few years haven’t even been invented yet.

To put that into perspective, in the year 2030—I have young children, my youngest boy is staring down the cold, hard barrel of kindergarten right now—this will be the year that he enters the workforce. When he enters the workforce, we have no idea what he’s going to be. Of course, as good parents, we always ask him and say, “Little buddy, what do you want to be when you grow up?” and he says, “I want to be a fireman or an astronaut.” And of course, we have to crush his hopes and dreams with this slide and say, “Buddy, you don’t know what you’re going to be when you grow up.” Again, it’s just a matter of where we’re living in an age of a digital revolution where we all are playing a part.

Slide 4:

I think in terms of the customer journey there have been some big changes as well. If we think back to olden times and we have this idea of needing a product—let’s say we lived in those olden times and I had this horrible looking anvil in my garage and all my friends would come over and make fun of me and tell me I need a new anvil. As a marketer, if you were in the anvil business, all you had to do was find a way to get a nice shiny advertisement in front of me. At that point, I had been introduced to the product and I’m now ready to go out and buy. I go over to the anvil store, I buy the “iAnvil 8,” or maybe I wait for the “iAnvil 9” or whatever’s happening in that phase of the product revolution or evolution.

Eventually I come home with my new anvil and I show it to my friends and then I become that initial touchpoint where they become aware of this product or service and go on and go forth. That’s a pretty straight-forward path to purchase. It’s very linear. It happens in places and it happens in areas where we can really control that conversation as a marketer. But then the internet came along.

Slide 5:

If we think about our very simplistic way of path to purchase these days with the online channels we have, it might look something like this. We’re first made aware of a product through something like a display ad and then maybe we interact or have a conversation around it in social channels. Eventually we’ll take a proactive search, we’ll go and use a search engine to take a look at something, maybe start putting it in our shopping cart. Then maybe retargeting ads will come get us on different sorts of platforms, whether it’s Google, Criteo, or others. Eventually, we put something back into our shopping cart, then we go back with a branded search in order to finish that transaction.

If we think about it, this is actually a really simple way of buying something. It’s not at all uncommon and there are lots of paths to purchase that include many, many more steps, but even in this simplistic version, what we’ve done here is we’ve traversed a number of different channels. We’ve looked at tons and tons of different platforms potentially and we’ve done it over different screens and devices, and that makes the marketing challenge very, very difficult. it makes it hard for us to measure and really understand the impact of what’s happening here. It makes it really difficult to manage all this data that’s being generated by all these disparate platforms, across different channels, across different devices.

Slide 6:

Again, this is a little too simple, the reality is that most of our paths to purchase on non-linear. They’re always on, they’re 24-7. And the scariest thing for the marketer is that a lot of these channels are channels where it’s the consumer conversations that are controlling the messaging and not the brands. We have lots and lots of things that we have to be doing in this pace. Really what’s happened is this is happening across channels and it’s also happening across lots and lots of screens.

Slide 7:

It used to be that we would come to conferences like this and look at the number of devices everyone had and we could play games where everyone would stand up and then sit down if you have a laptop, if you have a smartphone, if you have a tablet—eventually after four or five, there would be one person left standing and they’d be the hero. These days, we are fully living, not in the year of mobile or the age of mobile, but the multi-device age. What we’re seeing here is that you could be sitting on your couch at home watching a connected device through your television and have a tablet on one knee or you’re checking your smartphone, and that’s not uncommon.

In fact, lots and lots of people are switching between these devices. We’ve become very dependent on them. If you think about it, have you ever watched an adult, grown person walk out of their home and realize that they’re left behind their smartphone? The sheer look of panic is terrifying. They look like a young child looking up in the sky as their balloon floats away. These poor people can’t even get home because they don’t have access to GPS.  We have become so dependent on our mobile devices that they are really just a part of us and part of the digital experience for us, which means, we as marketers need to understand that and we also need to understand what may be coming around the corner.

Slide 8:

As we look into the future, we start to see that there’s been changes in the inputs we have to deal with. It came from a mouse and a keyboard, and then we did a mouse and a keyboard on a desktop and a laptop, and eventually we had touchscreens. That became a whole new input that changed the way we behave as consumers. Now we have sight and sound. We have to start thinking about the virtual systems for the in-home devices that are always on and always listening. These are more challenges that we as marketers are going to have to address and more data streams are being created every second of every day.

It’s not just there. There are new experiences and different ways to interact with media such as virtual reality and augmented reality that’s out there. Marketers are starting to play with this too, generating more streams of data. Suddenly we have a lot more to deal with, and that’s not the end of it. This internet of things has changed. Back in that conference where you sat down after just a handful of devices—I did this recently at a conference and we found that the winner had over 30 devices connected, from smart watches and refrigerators to washers and dryers to cameras and home automation. We can have lots and lots more data as marketers than we’ve ever had before. This means that there’s a lot of data for us to manage.

Slide 9:

This also means that the traditional way of marketing has broken. While many of us are still trying to make the best of it or putting our toes in the water of some of these new thing, a lot of us are still trying to do this manually. There will come a point, if it hasn’t already for most organizations, where there is just too much data for humans to manage.

Slide 10:

That’s really where we can start to look at the rise of the machine learning and the rise of artificial intelligence that we can lean on as marketers to do some neat things. I want to take a tangent here to talk a little bit about where we are with respect to AI and ML and all the other acronyms in the space and right now in the marketing and analytics industries.

Slide 11:

I want to start by looking at this idea of programming versus learning. If we think back to over 20 years ago when, in a game of chess, IBM had a machine that was able to beat the best human being we put against him. What ended up happening there is, if you think about chess from the perspective of the game itself, it is a rather confined world. There is a fixed number of boxes, rows and columns, there is a fixed number of player, and a fixed number of moves they can do. What we can do with programming is we can program computer to follow a set of rules. That’s really where we’ve been in the computing age for a really long time. We’ve created these computing rules and the computers aren’t really smart, they would just follow those rules and do what the program told it to do.

By contrast, when we look at the idea of machine learning or artificial intelligence, really what we’re talking about is this concept of the computer no longer following rules, but learning much like we as humans do. As we enter the world as newborn babies, we have some sensory inputs and we start to look around and start to understand things for ourselves, and eventually we learn. That’s really where a computer can take some great strides with machine learning.

Slide 12:

A great example of that is if we move from the game of chess to the game of Go. The game Go is quite a bit more complex than chess. In fact, it’s said that there are more possible outcomes to this game than there are atoms in the universe. That means there are not very many rules-based programs that we can right that are good at playing this game. It actually has to learn.

A great example of this is when Google’s Deep Mind was put up against a world-champion Go player. What ended up happening during this was—I’d encourage anyone to go and look this up on YouTube and watch, they have commentators for the game—and essentially what was happening was it got to this point and 37 moves into the game, Google made a move that made everyone a little nervous. It was something that was really unheard of. What they did was just put a piece in the middle of the board, and it was at that point where Lee looked at the board, got up, and left. The commentators put their hands to their ears and said, “Is that right? Google, you really want to make that move?”

It was confirmed that was the move that Google Deepmind wanted to make. Then the game kept going and 200 moves later, Google’s Deepmind won. They traced back the game and found that the pivotal point at which Google was able to win that game happened at move 37. That move that humans couldn’t even think that that was something you would do. So, to think that Deep Mind and this artificial intelligence is able to think 200 moves ahead in one of the most complicated and complex games in the world, we’ve come a long way. I think what’s happening now is we’re starting to see this make its way out into the world.

Slide 13:

When I talk about artificial intelligence and machine learning in the marketing space, what I think is a good analogy is I think it’s in awkward teenage years. I think we can tell that by a few different reasons. Right now, when we talk about AI and ML, it seems that it’s overly self-confident in its ability to know all things and solve all problems. We’ve invested in this thing for many years, just like our teenagers, but despite a few bright spots along the way that served to keep us from giving up completely on them, so far it hasn’t really lived up to its expectations. So, it’s time for it to get a job, we’ve got to kick it out of the house, we’ve got to get it to do some stuff for us.

That’s where we are right now in the marketing analytics space, it is doing some things for us. A good way to think about that is, like a teenager, artificial intelligence is out there and it’s learning to drive. It’s not out there on the road yet. We still have to have the person sitting there with their foot on the brake and we’re starting to see these things being tested and what not, but there’s also a difference between artificial intelligence and what I like to call intelligent assistance.

Slide 14:

I think the difference between AI and IA is a good analogy for the car because while a fully machine-driven car is not ready for the real world yet on all the roads to take over all of our driving jobs and take us where we need to go, we have seen intelligence assistance in the automobile market for many years. In fact, we’ve seen things like a rearview mirror when we’re about to turn into a different lane, that side mirror with have a light in it or there may be some other warning that tells us there’s someone in our blind spot. There are cars now that are common place to, when it senses that you’re texting and not watching where you’re going and you’re about to slam into the car in front of you, it will hit the brakes for you. Again, these are really neat thing that we can start to do by taking and crunching lots and lots of data and making intelligent decisions around those things. And intelligent assistance is available, and that AI driven capability is available for you in the marketing analytics space.

Slide 15:

For the rest of this, I wanted to talk about a few examples of how this is being used up in the marketing place and then wrap up with thoughts around things folks in this industry should be thinking about in the short-term, the medium-term, and eventually the long-term.

Slide 16:

We can start with things that are fairly straightforward. When you think about search and display, marketing in the digital space, there are lot and lots of different signals here that this exhaust stream of data is being generated every single time an impression occurs of a screen or a page. The potential to serve an ad, to bid on a search term, or bid in the programmatic space has lots and lots of dependencies.

What we ended up doing was we figured it out by doing things like cost per click and a click was a proxy for getting someone to your website. Then we’d look at things like bounce rates and those were proxies to see if someone would stay on our website. Eventually we would try to convert all that stuff into actual sales. Whereas now, we can really start to let the machines start to understand all these different signals that may contribute to a probabilistic outcome of an acquisition or a purchase or whatever our conversion action is that ties back to revenue.

At that point, we no longer have to do things like try to manually bid against different devices or location or audience attributes and behaviors. We can let the machine start to learn. That’s an interesting thing for us to do and we can start to see this in most of the major markets already.

Slide 17:

The bigger platforms are doing this and have been doing this for a long time. They can figure out whether a probabilistic outcome of something like a conversion is more or less likely and figure out those coefficients and what not based on any number of signals. And a lot of these signals only that platform has access to. So even if we wanted to do this manually or in a rules-based scenario, we as marketers wouldn’t be able to do that. But a lot of these AI and machine-driven campaigns can achieve results that are much better than humans. The reason why is these are learning, and they’re learning without the need to sleep and eat and make mistakes. They can go through their learning periods and they can really understand all of these difference outcomes and all of these different data points much better than a human can. When we start to look at these things, we can start to look at the models that are being built.

Slide 18:

As all of these different signals are being brought into the platforms and the capability that they have, you can actually start to model things out. You can look, and we’re not looking for cost per click and metrics that are proxies for things, in this case we’re looking for target cost per acquisitions and we’re looking for things and we’re looking for those actual conversions. You can start to say, “If I want more sales, how much will it cost me in a cost per acquisition format?” or, “Do I really want to get that target CPA down and then see what that’s going to cost me in terms of volume?”

There’s lots and lots of different ways we can model this, and of course, machine learning will start to find those areas or diminishing returns. There are lots of approaches we can use in the data science space to help us figure out exactly what that optimal point is. As the world keeps turning and more changes happen, people’s ability to react to different ads, and you put more and more ads out there in the marketplace, suddenly these things gain way too many variables. What we end up needing to do it rely on the machines to help us understand these things. Doing this manually as a human is something that will very quickly become impossible if we’re going to be keeping up with the times and doing these things in ways that our competitors are.

We can build out the bids in using machine intelligence. We can also start to look at giving it lots of things to learn from. We all are pretty aware that we should be using creative out there and we should be testing them and we should start to understand which are working better or worse than others.

Slide 19:

Lots of people get hung up on when it’s time to start putting in dozens or hundreds of thousands of ads. The major platforms out there do allow us to do this, and if we have the volume, if we have the budgets, we can have lots and lots and lots of different ads. We can use different features of the platforms that are out there, Google’s AdWords, for example, has a number of features that allows us to essentially manage data feeds. Each of those data feeds can control targeting, it can control the text of a different ad, so you can have lots of different variations out there.

This is something machine learning is very good at picking apart. It can figure out which is the best, most probabilistic, positive outcome ad to show to a certain selection of all these variables that are out there. As it continues to chew on all of this data, it can start to learn, start to do some neat things. To the point where Google AdWords, for example, has had what they call dynamic search ads out for a long time, where you can release the machine on your website and have it go through your pages and it will create campaigns for you.

As we start to see, these things have been in play long enough to get some numbers, and these things tend to work better than humans. That’s not to say there are bad campaign managers out there, what we are starting to say is that humans won’t be able to keep up with the machines for very long. Just like our best chess players and our best Go players have eventually lost to the machine, if we as campaign managers and marketers and still sitting in the interfaces and clicking around manually to try to optimize things, we’re just at a point where we’re not going to be able to do that quite as well as our competitors who are taking advantage of these sorts of things.

Slide 20:

It’s not just Google, there are lots of platforms that are out there. Over on the Facebook platform, there are plugins and tools that allow you to take lots and lots of different combinations of headlines and some copy and images and start to let these things find the right combination of audience factors and behaviors and attributes, using machine learning to do that at scale.

Slide 21:

Speaking of creative, there are lots of ways we can start to use machine learning to help inform creative. This is an example that some teams from Cardinal Path worked on a hackathon very recently. What ended up happening here was using the Google cloud platform and some of the vision APIs that are in there. If you guys aren’t familiar with that, essentially what you can do is upload an image and it will go through and calculate the probabilities of many different attributes being a part of that image. In this case, we could do this and load up the machines with hundreds, or potentially thousands, of image ads and then give it some context around those image ads.

It can start to build out coefficients using revision modeling. In this particular case, where we got some output, we found that coefficients that were positive towards things that we want people to do with an ad, like click on it or have a conversion follow or something like that, we were able to all these different labels and logos and build out these coefficients. I’m showing some positive ones here, there are also some negative ones, like “bodyman” is actually referring to the torso of a male. One of the ones that I’m not showing on the screen that had a very negative coefficient had to do with body hair. What we can figure out from this, and again this is a simple example, but the machine told us that you had better have a male’s torso in the ad, but it better be wearing a shirt.

Slide 22:

These are the sorts of things we can start to do. And of course, we can do it with static imagery, but we can also do it with sequences along video files. In this particular case, we can have the machine learning pick out the probability of different machines and then have it pick out probabilistic numbers that would help us figure out which are the best cluster, which are the best sequences of these scene, what are the best sorts of things we can do against things like click-through rates and conversion rates and whatnot. Then we can really start to understand how we’re creating the ads that we’re going to be running next.

We’re not saying here that the computer is ready to hop into Photoshop and create the best ads for us yet, but we can use this to churn through lots and lots of data and find those insights and find those probabilistic outcomes using this machine learning to help us understand our creatives. That’s where our advantages as marketers will always be, in the strategy and the creativity.

Slide 23:

Another great example that I liked that we’ve seen a lot of lately is this idea of algorithmic retargeting or non-rules based. A lot of people know retargeting in the product space as where you put the pair of pants in your shopping cart, but you don’t buy, and then that ad will chase you around the internet until you either die or buy the pants to get it to stop. These are interesting, but these are all based on rules. Again, this is that programmatic sort of thing. We can create program that says, if this condition or this set of conditions is there, let’s go ahead and do this.

That’s what retargeting is. If this particular cookie had something in this particular shopping cart, but then has not yet checked out, then we’re going to show them this ad, and we’re going to keep showing them this ad across all the different places where we can put impressions out there on the web. That’s good and it can certainly help, and a lot of people have been helped by retargeting platforms, but it’s not optimal and it’s not learning, it’s just following the rules.

Whereas, algorithmic or machine-based retargeting can start to look at lots of data sets, pretty much anything you can feed it from an analytics platform or a CRM platform for third-party data. And it can look at these and start to pull together what each of these attributes, behavior, anything you’re tracking, what the implications is on a probability to buy. That means that you can not only cancel out of the audiences that are not likely to be impacted by the retargeting platforms, but you can also focus in on the ones that have a high likelihood of it.

You can also suppress the ones that were going to buy that even without that sort of a retargeting campaign. What we end up doing in this particular place is we start to look at all these signals, and frankly, we don’t care if he had a size 8 shoe size and a pet goldfish home as a pet, that was predictive of you being able to go out and buy this thing. If that happens to be what the algorithm and machine learning is learning, then we can go after and target those things and the DSPs that are there and buy the advertising we need programmatically against all these attributes are something that we can actually track.

Slide 24:

This is something that is out there in play. One really good case study that we’ve seen recently here, they basically set it up as they got the same numbers of impressions and same numbers of conversion, but they looked at how much they had to spend and how many impressions and retargeting space they needed to do over six weeks in order to get it. In this space, you can see that the machine was actually doing this about 43 times better than the typical rules-based or retargeting algorithms that are out there.

Slide 25:

This is a pretty simple way of doing it. These are all parts of marketing technology stack that we all have. They were looking at web analytics data and they could augment that of course with any kind of first, second third-party data you may have, but they put it into an enterprise data warehouse and inside of a modeling environment, they were able to use this machine learning and this intelligence to figure out what these outcomes were going to be.

Once you have all of these coefficients, you can load those back into a DMP, create your audiences out of that, and then push them out in a programmatic space. You can actually buy these audiences, or in the case we talked about, suppress those things out of the publisher networks. Of course, as we do this, things are changing, so we’re constantly creating new data and we’re able to repeat this cycle and do some pretty neat things. Most of us that are working in the enterprise space these days actually have almost all of this already in the marketing technology stack, so a lot of this is things we can do pretty quickly and easily.

Slide 26:

A couple of other examples, again, as we talk about impressions and suppressing audiences, we find that it’s not very uncommon to find a small percentage of users that are essentially eating all of the impression that are out there that we’re having to pay for as marketers. We can sue the same sort of algorithms to figure out exactly are those attribute that are predictive of someone being a bad recipient of our ads. In this case, this was done on the doubleclick Google stacks, using doubleclick’s campaign manager and cloud platform from BigQuery, and some scripts, we were able to pull some of this stuff out, and of course, push that back into the programmatic platforms through DBM, or DoubleClick bid manager to programmatically go after suppressing those particular audiences.

Slide 27:

You can see some pretty good results there. By not serving people ads who are not likely to be clicking on them or not likely be anything more than a waste of your advertising spend, we would answer the John Wanamaker question, which is: which half of that marketing budget is not being used effectively? Big, big gains can be made here by letting the machine start to optimize and start to understand these predictive signals a little better.

Slide 28:

Another way to look at that is through progressive bidding. If we figure out who we don’t want to be marketing to, we can also figure out exactly how much we’re willing to be bidding on different audience segments. So not using quartiles or quintiles, or anything like that, but actually looking at lots of predictive attributes. The underlying idea here is that at the very top, if someone or a certain attribute is predictive or very likely to convert, we can bid pretty high for them. Conversely down at the bottom, we don’t want to bid nothing, because there’s still the possibility of a conversion, but we don’t want to bid that much either.

We can use the machines to figure out what are the optimal bids for each of these things. In this particular case, you can see some pretty big outcomes in the programmatic markets that are using this because they’re not using a single rule that’s supposed to fit all. They can look at all these different combinations and variable in tandem and do some very cool things. You may see that when you start doing this and the machine starts learning, it bids really high to a point where you get scared.

A lot of that happened in this particular case, where we were seeing bids that were almost 30 times or about 30 times higher than what it was normally doing. They almost said shut this thing down, it’s crazy. Eventually they let it run and they realized it was bidding high for the ones that were converting. It was doing that while bidding very low for the non-converters, and overall, it has huge improvements in conversion rates and cost per acquisition numbers and things like that. You can use this with multi-touch attribution and things of that nature to get away from that rules-based attribution as well.

Slide 29:

That’s kind of an interesting way to segway into attribution. Traditionally we have looked at last click and I think at this point, there are probably many speakers and sessions talking about that today where the age of last-click has to be dead at some point and yet, most marketers are still using it. A lot of it is because it’s still fairly difficult to use AI or machine learning driven approaches and the other sorts of rules-based approaches are a little bit of a transition period to move from what we’ve become so comfortable with in the last click area.

But we are starting to do it, and in the rules-based scenario, it’s get a little bit harder. Whereas, if we start to use the machines, we can start to get to some data-driven techniques that will help us understand what each of our channels are doing and how we can attribute value to them across the complex customer journey that we talked about in the very beginning.

Slide 30:

If we think back to a traditional report—we’re looking at a Google Analytics report here, but this could be Adobe Analytics or anything else—we’re looking at a lot of these different platforms and these different channels that are out there advertising. What we end up seeing is a lot of our conversion end up getting contributed to the things that are lower in the funnel. One thing you don’t see here is display. And this is because this was the last guy who actually clicked on a display ad. You can see that it causes him much pain and grief as his brother is kicking him out of the village apparently, his wife and kid, bad things. People don’t actually click on display ads and then buy things, right? These are typically very upper funnel activities.

Slide 31:

And in the branding space, we as marketers looked at these reports with last click and we looked at the way of attributing all of that value to the very last interaction point. We said upper funnel marketing must be dead, branding must be dead. There can’t be any use in it. Of course, we know that’s not true, but it was very difficult for us to understand what the impact was in those upper funnel activities. When we start to look at other ways of approaching attribution, what we ended up doing as marketers, we said, “Let’s find more rules that we can apply.” In this set of rules, we will apply all the credit to the last of the touchpoints on the way to the path to purchase.

Slide 32:

In the other models that are about there, we have things like first-touch, which is exactly the opposite. We would give all of the credit to the first touchpoint that we could track. We have linear, we’ve got time decay models, but again these are all very rules-based. We can get a lot better if we start to look at these things algorithmically.

Slide 33:

But rules-based helps us to at least understand that there are different sorts of ways to look at the upper funnel sorts of things so we could see, for example, in this bottom row in display, if you looked at the last click or last interaction model here—and this is a feature from Google Analytics—but we can start looking at these first and last click marketing channel reports and Adobe Analytics, and there’s lots of different tools and platforms we can apply these with. Let’s say as an agency, we would spend on our customers 4,400 dollars and we had gotten back 900 dollars in last-click attributed revenue. If we were the agency that was working for this client, that client would probably tell us we’re fired. Because again, we’ve spent a lot of money and we didn’t get much in return.

Very quickly here. We could compare that to first interaction or linear, and we could start to see in these models that we were actually undervaluing by something like 11,000 or 24,000 parents. Really we were talking about 100 or a quarter of a million dollars that came from that 4,400-dollar investment. Again, this was helpful, but it’s not right and it’s not wrong. They’re just different ways of looking at this, different rules to apply and different lenses that we can start to look at this attribution from.

Slide 34:

When we start moving into the data-driven, then we can leverage the power of the machines and let them go through these algorithms and applies things to understand what the true impact it. I’m looking at a data-driven screenshot here from Google Analytics, but what we’re doing here is we’re actually applying data-driven cost per acquisition, data-driven ROAS, data-driven conversions. These will actually come down to using these algorithms to attribute the number of conversions that should be attributed back to these channels.

Slide 35:

The way they’re doing that is by taking into consideration hundreds and thousands and tens of hundreds of thousands of paths to purchase and conversion paths and non-converting paths. It looks at every one of these and it starts to figure out in each of these similar paths, when one of these channels was not present, what is the implied lift? And it does this across lots of different touchpoint and whatnot to really take that ability to leverage and work with that much data that human being can’t do and figure out exactly what we should be spending our money on across all of these different channels, across all the campaigns that we as marketers are running.

Slide 36:

Those are a handful of examples. I’ll end up here by talking about what I believe to be the short, medium, and long-term of AI and machine learning and the marketing analytics space. Like I said at the beginning, for some of us, this is already here and we’re already taking advantage of it and doing some neat things, but I think those who are, are in the minority right now. Because of that, just like those who embraced the machines early in the Industrial Revolution, it provides tremendous competitive advantage. We are seeing that many of our clients who are embracing these things are starting to see huge gains, are starting to see the numbers that we’ve looked at through this last half an hour.

But it’s a very complicated landscape right now. There’s lots of different vendors out there, all with different value propositions. There’s really no set standards with respect to attribution quite yet, and we are still kind of in that phase of: do we really need to be so focused on the creative parts or the upper funnel part or branding, and things of those natures. I think what’s going to happen is, in the very medium term, we’ll start to have standardization. There will continue to be mergers and acquisitions, there will be new vendors that pop up and become part of a growing platform. there are probably just a handful out there that have, in a very short order, established some rules, some norms and some standards in this space.

At that point, it’s not just you who’s going to be using machine learning and AI to do your bidding and create creatives and work through all the different things we do as marketers, it will be everyone. If everyone is using the same AI, the same MO, across the same platforms, suddenly it’s not a competitive advantage anymore. It’s just table stakes. That’s what you’re going to need to do. As a result of that, your job is login every day to these platforms and try to do anything like this manually or changing your bids or trying to run some manual split tests and things like that. That job’s probably not going to be there for very long. Then just like the farmers move d to the city during the Industrial Revolution, we will have to learn how to control the machines rather than try to compete with the machines.

I think that’s where we go in the longer term. Is it really that strategy out there? Is it that the value of the creative and the strategy will be the competitive advantage? It’s those who can drive the machines better because it will be the same machines that are capable of the same things for you and for your competitors. Of course, pretty soon we’re looking into the spaces where AI and machine learning will start to figure out better ways of doing that AI and machine learning and they’ll start training and working to do things that can surpass humans in that learning environment.

Slide 37:

I’ll leave you with this and then we’ll wrap things up. The one thing that you’ve probably seen as a theme throughout this is that machine learning provides a very large and very diverse data sets. It’s learning like a child, like a newborn child. It needs to make mistake during its learning period and it need to figure out if, like a child that stand on the stove and realizes it’s hot, we as adults know that, we look at that child and say that’s a thing a child does and how it learns. Machine learning during its training period, consider those budgets not performance budgets, but testing budgets. Let it train, let it learn, and eventually it will be doing things faster and better than we as humans can do.

The last thing I’ll leave you with is to compete in a world of tomorrow, I think that we’ll need to do is figure out how we’ll build those machines to figure out what our move 37 is going to be. So when we look back 200 moves and realize why we won, it will be because we’ve guided and trained these machines better than our competition.

With that, I will say thank you and I’ll tell you to enjoy the rest of the sessions for the day and thank you for being a part of this event, and particularly this session.

Slide 38:

Thank you so much.

Previous Video
Krista Seiden, Google, Looking Forward to 2018 and Beyond: Where the Data is Taking Us
Krista Seiden, Google, Looking Forward to 2018 and Beyond: Where the Data is Taking Us

Krista covers three areas of focus to make data more attributable, simple, and useful.

Next Video
Charles Farina, Analytics Pros - Actionable Uses of Google Analytics
Charles Farina, Analytics Pros - Actionable Uses of Google Analytics

Explore ways to incorporate AI & ML into your marketing—or risk being left behind.