Marketing Meets Science
Chris from Google Cloud explains Google are giving the very technology stack they used to build their business, to us, open source – so we too can continue to try and meet consumer expectations of better, faster and cheaper.
So, hi, my name's Chris Gale. I lead the ad tech and agency business at Google Cloud. Basically, I work with agencies and marketing technology providers, helping them have marketing meet science. So, let's kind of dig into a little bit of what that means.
To level set, this is the spectrum of services that encompasses Google Cloud. You have our namesake, the Google Cloud Platform, also G Suite, our productivity solutions, also Maps, Apps, and Android enterprise. Essentially, it's all the products that we used to build Google, so that you can run your own businesses but without having to spend $10 billion a year and millions of man hours building these things.
So, you've heard many, many times a day, and throughout the last six months, I'm sure, that you've got to be constantly innovating as a marketer in order to get the minimal attention of all of the people that we're trying to market to. We're demanding more as consumers. We want better, faster, and sometimes cheaper as well. And customers are demanding more customised interactions with them based on everything that you know about them.
At Google, here's a couple examples of ways that we're using big data and machine learning to give customised solutions to our users. So, in this one example, we know about where you work and where you live to give you faster directions to your next event, or we have information from your email about the next flight you might be taking. So, if it's been delayed as in this example, you might not have to rush to the airport and get a speeding ticket on the way.
At Google, we have tens of products across a number of consumer touchpoints. We're managing this at scale to billions of consumers all within milliseconds, and that's exactly why we built our cloud infrastructure, and our expertise in machine learning.
Back to the marketer's role is changing, we built all of that infrastructure to power our own interactions with our customers. I'm going to walk you through how Google has had to move from traditional campaign execution, you know, planning, buying, and then reverse analysis of what's been happening to a more hyper relevant real-time advertising to put the right advertising in front of the right person at the right time; everything that we're all trying to do in our daily business.
Originally the evolution of getting this to today was powered by big data analytics, and we're going to be taking this into the future with machine learning, things like you've heard about today. But before we get that far, let's level set with understanding of what machine learning is. Very simply, it's giving computers the ability to make decisions on what you're looking to get out of them without having to explicitly program them.
The four places that I think that machine learning is going to radically change all of our lives in the next couple of years is identifying customers, the target audiences that we want to find, allowing us to target those customers in customised ways at scale, looking at the budget we have and figuring out how to allocate it across all the different spectrum, and allocating our time as marketers between all of these different choices.
So, I told you we built all of this infrastructure to power Google. I'm going to walk you through one example of how we, at Google, have used this to customise our approach to lead generation for Google Cloud. So Google Cloud's largest products are the G Suite or Apps products and Google Cloud. In August of last year, we redesigned our entire G Suite look and feel to G Suite from Apps. And it came with a cartoony playful look and feel. This happened after we'd already set all of our targets as you might be familiar with. So we had to see, "Does this new look and feel impact the way that we're able to drive leads from free trials to conversions, a common practice in software as a service?"
Historically, this entire process took us about 45 days. We would plan our channels between search, display, and social, figure out how we wanted to allocate our budget. Execute the campaign. It takes 30 days to go from a free trial to a paid subscription. Collect the data of the backend, and then analyse how we wanted to work forward. Does this sound familiar to anybody in this room that executes marketing campaigns? I should hope so, because it's what I'm used to in this entire industry.
So, unfortunately, we didn't have that luxury of time 45 days to find out if this new look and feel and the campaigns that we have been provided were going to generate the results that our VPs were asking of us. So, by taking all of these different datasets, all of the marketing from display, social, paid search, double click, Google Analytics, third-party data that we knew about our consumers, information about their companies, and putting it all into a large machine learning model, we were able to actually output results within two days and understand how our marketing campaigns were performing. And that allowed us to allocate budgets on the fly and change everything that we were doing about our marketing campaigns.
Ultimately, it all sounds a little bit like magic. So, how did we do it? Well, in the original methodology, data was collected, analysed, and visualised, and then pushed back into our systems after 45 days. But in this system, we dumped it all into a machine learning model. We're able to model out what was working best and what wasn't, and then push that back into real-time media with paid search and display media, so real-time bidding, and evaluate how all of those free trials that we were getting signed up for on the frontend, how they scored in their likelihood to convert into paid subscriptions. Those paid subscriptions, we were able to tune over the entire lifecycle of the process and push back into our model, improving the model over the course of the entire process.
And while we were doing that, enable us to find our customers at scale which is good for me because I sell Google Cloud. Optimize our cost performance modelling so that we were spending money in the places that were driving results and save our time because we only have so much time as marketing and sales organisation to reach out to our customers. So, this is great, right? Google used machine learning and some really sweet tech to be able to optimise what we were doing. How does this apply to everybody else in this room what we can do? So I want to show you how you can take the infrastructure that we used to build Google and execute these programs at scale with your marketing campaigns and the people you're trying to reach out to.
So I'm going to give you a couple examples right now of people that I've been working with here out of the UK. The first one is Ocado. They are the largest online delivery grocers in the world. They're doing an order every two seconds, and they use Google Cloud to power their infrastructure. As most of us are familiar with, they have a customer call centre that gets inbound emails about a plethora of issues. "I'm really happy with your products and services. Thanks for not making me go to the store. I have two kids." Two, "I'm really upset because my eggs were broken, and can you refund me for this order?" Everything that you're used to, so feedback, out-of-home, all of these types of things.
So they thought, "Over the last 10 years, we've been an extremely data-driven marketer from not only the way we do our marketing, but also the way we interact with our customers. Can we apply this to our email systems?" So, being very data-driven. They had collected all of their emails over the past x number of years, and their customer service centre had tagged them with things like, "This is feedback. Customer is happy." Adding that dataset for training with our natural language processing algorithm that we have where you can just use it as an API and you don't have to actually train it, they were able to get the sentiment analysis out of the emails as well. So they were able to take their corpus of old emails and use it to train this model so that they didn't have to actually manually deal with every email coming in.
So right off the bat, they were able to respond to the most upset and angry customers four times faster than they were before. They were able to save £10,000 a month in terms of time and effort in managing those emails, and we're getting a 10,000% ROI in terms of dealing with all of their email systems. And that sounds really good, but from a business perspective, that was just the tip of the iceberg. By looking at all of these emails and understanding how they were interacting with all of their customers, they were able to understand at scale what products were more likely to go out and drive customer unhappiness and change the way that their fulfilment centre was dealing with their customers. That saved them millions of pounds per year. They were able to understand which products like this one, you see some eggs up on the top left, might go out using the Vision API cracked and case issue before they went out of the fulfilment centre. So they can actually stop the email from even coming by routing this one to "this is probably broken" and "repack this grocery item." And they were able to start to understand which of their products were driving positive user sentiment and negative user sentiment so they could adjust the types of products that they serve as a business.
So ultimately this initial project that they had to try and understand their users from the emails was able to impact their business at a much larger scale than they ever thought was possible, and it didn't take a whole bunch of effort because they were just using APIs and all of the data they had already to execute on this.
Another one of my favourite examples is MultiChoice. They are the largest pay-TV provider in Africa, serving the lower 49 countries of Africa. They have about eight and a half million customers, and their marketing manager was hired in and was really, really interested in understanding who her customer was and what they were doing. They had 35 different channels of consumer touchpoints, and no way of bringing those together and understanding that this person in email was the same person that called in the other day. They actually were only able to understand their consumers at a billing level, at a household level, and that didn't really give them an understanding of how their users were using their services.
So the marketing manager who doesn't have a tech background went to the CEO and said, "I would really like to build a unified marketing data warehouse that marketing owns in collaboration with all of the IT groups." And our CEO said, "I don't know if I really want you to do that. It sounds like a three-year project that's going to cost me millions of dollars and not probably going to get us anywhere." But she was able to get the buy-in from a couple of groups and said, "Look, we'll do a prototype. Give me a month, and we'll bring three different verticals of information together. And if that works, we can go on forward." So a couple weeks later, she was actually able to work with a partner of ours to develop this unified marketing data warehouse, please all the different constituents, and… Oh, it says one. Okay, she was able to please all the different groups, and… She was able to please all the different groups, and she got the ability to go ahead and build the rest of the marketing data warehouse.
So, very, very quickly, this gave her new super powers really as a marketer. She was able to understand her users not at a billing account level but at a user level whether they were looking at TV on their mobile phone, on their tablets, on their computers, or on their TV set-top boxes. She was able to start to understand what different regions and what different types of consumers were doing. They were able to use machine learning to extend out their CRM database. Is this a woman or a man? What languages do they likely speak based on the entomology of their name? All of these types of things. And it was a good thing they did, because shortly after they did this project, there's a large streaming movie company that came to Africa and started picking off their customers in the higher bandwidth areas. You guys might be familiar with Netflix.
And this was a bit of a problem for them, but because they had this unified data warehouse, they were able to take a look at the types of customers they were losing. And they understood that they were losing these, because they're paying for couple hundred channels and watching about 15. And they interviewed a couple of them and found out, "Yes, that's why we're upset. We're spending all this money and not getting what we feel the value is." So they're able to craft custom packages for each one of those and use their customer service centre to proactively call look-alike types of customers and get them to buy into these new packages that were customised just for them.
They were also able to look at the gap in the market of what they didn't have that a competitor was coming in, and they built their own content marketing, Showmax, to compete directly with Netflix. And lastly they were looking at the overall market of Africa in noticing that they had customers in all the affluent areas. So there was an opportunity for them to build a lower cost, lower serviced option that serviced the parts of the market that they weren't able to access before. And, for them, it was great, because they didn't have an IT team and they had no data maintenance and no team that had to make sure that everything stayed up and running because they were using a cloud service provider.
The last example I'm going to walk you through today is Zulily. They're a large ecommerce retailer based out of the west coast of the U.S. This one is awesome. They're an entirely e-retailer. They have a million plus products that they sell, and they give it to you in a tiled format on your phone or iPad device. Historically, they had sales managers that owned different verticals like house moms, or young women, or young men, that type of thing that would build custom hand-designed just, you know, sets of products that you would get to buy, because they have a million products and you can't service all of those on a tablet of nine tiles. And they wanted to bring machine learning into this because they understood that there was, A, more divisions than they could have earlier, you know, manually done. And there was also more customisation that they could provide as a retailer.
So they looked at personalising this entire makeup based on everything they knew about you from apps data, clickstream, analytics across the website, third-party data, and their CRM systems. And after that, they were actually able to, within the subdivisions that they already had, serve you different pictures depending on what was more likely to get you to convert even if you were in the same group. They were able to build a whole plethora of new groups underneath those that had customised products that they were offering them. And this was able to give about a 0.3, 0.4 percentage uptake in revenue per visit for every one of their customers which – well, it sounds like not very much – is 7 figures when your business is as large as this. So the ROI on that was immediately paid back within a couple of months of the project being executed.
The more important thing is by aggregating all of their data and starting to do these types of projects. The CEO said that they're now able to answer questions that took months and weeks in minutes and seconds. And they're also able to answer questions now that they never could have possibly done in the past.
As marketers, we all have to figure out ways to delight our customers. They're more demanding as ever before, and they want brands that are able to predict and assist with those needs that they have.
Big data machine learning is able to help with every step along with process, but it's just a tool. Is machine learning the Holy Grail for marketing? I'm sorry. Not this time. Marketing still needs the science. You've got to have the ideas and the answers that you want to solve, and then big data's going to allow you to execute on them at scale in a way that you never could have in the past. I'm working with marketers and ad tech companies trying to help them move from these giant bins of Legos where you have to figure out what do I plug together to execute on these, you know, demands that they have as marketer to packaged up solutions so that we have APIs that help you out with things. We have entire marketing analytics starter packages to help you have all of the components that you might need for unified marketing data warehouses, things like that.
We also have a host of machine learning models that are prebuilt for you, everything from the Cloud Vision API that helps you understand all of the objects in an image. Sentiment analysis: is somebody happy? Are they sad? We have the video API that helps you understand everything that's happening within a video, translate, speech-to-text, everything that you would need to build a system that interacts with the world like a person without having to know all of the deep machine learning code that you'd have to build those things.
So, I hope what I've shared with you has been a little bit inspirational of what you can do as marketers. Not everything is already built for you, but that's the exciting thing from my perspective. Whatever you need to deliver to your business is now a lot easier at scale with the new cloud technology that's on the marketplace.
My name's Chris Gale. If you want to talk, give me an email or ask some questions.