#SEOisAEO: Machine Learning 101: The how, the why and … what ML means to us as marketers
- Several definitions of AI and ML
- Purna Virji explains the difference between AI and ML
- "Much of ML is an automation of what's a somewhat manual process flawed by the nature of humans." Can you explain that, Lexi?
- Are there any search-specific ML tech we should be aware of, and how are organizations using it today?
- Where do you think search marketers should use ML? Analytics, personalization, multisensory?..
- How should you go about setting up your business to take advantage of ML?
- Machine learning and the future of PR
- Can you trust the search engines own ML platforms to do the right thing for you?
Jason: Welcome to Lexi Mills, Purna Virji, and Stephen Kenwright, who are collectively going to tell us all about Machine Learning for marketers.
Several definitions of AI and ML
"Artificial intelligence is the precise or explicit description of functions and presumably hardware so that it can stimulate human intelligence while machine learning is aimed at helping machines learn on their own and perfect their operation, so it could more correctly be called machine intelligence."
"AI is aspirational. It's a moving target based on those capabilities that human possess but which machines do not."
In the article this definition appears in, they use the example that a calculator in the 70's would have been artificial intelligence at the time. It now seems very simplistic. Today we're moving forward with other applications that, presumably in 20 years time will all look like terribly simplistic. Then: "machine learning is a subset of artificial intelligence, just one of the many ways you can perform AI. Machine learning relies on defining behavioral rules by examining and comparing large data sets to find common patterns." I would add to that: instead of writing code you feed data to a generic algorithm and it builds its own logic based on the data.
I've brought together several ideas about what machine learning and AI are.
Purna Virji explains the difference between AI and ML
Purna: Of course there's confusion right around AI because it's not something that you can see or touch. It's something that people are using every day but they don't realize it; if you've ever done a search on Bing, or if you ever used a PowerPoint presentation, or dictated a text in your phone, you were probably relying on intelligent technology. The term AI essentially refers to this broad range of technologies that can perceive images and sounds and understand them. It can learn from these inputs and interactions and it gets better over time. AI is the larger, broader umbrella.
Machine learning is one subset of AI which tends to use these mathematical models to understand and reason over vast amounts of data. My colleague, James Whittaker says it really well: "AI tends to be stochastic." It works like the real brain works. It can bounce between different ideas, it can go in different areas using different neural networks, different connections just like our brain does. Machine Learning tends to be analytical because it's based on numbers and patterns. And, it is slightly different. ML and AI are not interchangeably used. So, that's a good thing to keep in mind.
Jason: What about Microsoft? I saw that you were talking about the four C's of conversational design: clarity, character, compassion, and connection.
Purna: We can go higher than that. Think about what this intelligent technology can do for us. At the highest level, it's helping us be smarter. Human beings have lots of unique capabilities, we can be creative… historically we have used technology to amplify and extend these capabilities. Look at the 1400's. At the time, books were handwritten by monks. Knowledge was limited to very few people. Then the first printing press came out. Humans used a form of technology to democratize learning. Books could reach everyone. Same with the switch from horses and animals to cars. We could go greater distances. We've always used technology to solve our greatest problems. And we are doing that again with the Internet and search. But there's so much data being produced every day. AI gives us an opportunity to make the best use of it. But how do we make the best use of it, and use it for good? I think that's a big, key question. There are so many great ways that it can make things more accessible. For example, Microsoft have the Seeing AI app that is for people who are seeing impaired and it helps them understand the world around them. If they're walking around and they hear a sound they can just press a button on the phone and it tells them what it is. That's so empowering. I think it's a great big equalizer.
Jason: Yeah, the whole accessibility question was something we looked at with Kim Krause Berg earlier in the series. Accessibility really comes to the fore with voice search. Add to that conversational search, especially empowering for the sight impaired. That's really brilliant.
"Much of ML is an automation of what's a somewhat manual process flawed by the nature of humans." Can you explain that, Lexi?
Lexi: Yeah, of course. If I said Elephant App... you know what an elephant is. You know what an app is. So putting those two concepts together, a human immediately gets a good idea of what I'm talking about. That’s two things. Expand that to four or five things and the human brain just goes “nope, I'm not doing that”. We simply can't put more than three concepts together at once. Machines can. And that's a great challenge. It's a challenge for many applications. Commerce through to medicine - there's some amazing algorithms that help match up kidney donations, for example. So when we see things our brains aren’t capable of working out, we can get the machines to do it for us. So it allows us to kinda be superhuman. Which is pretty exciting.
Jason: Brilliant. Superhuman, I like that. My friend Dave Clayton told me how he teaches A-level students, and he said, "I always have three points. Even if I've only got two, I make up the third one because three always really gets into the kids' brains." I’ve used that technique a lot in this series, and I do find putting things in threes is very effective. Not just for the people I'm talking to, but for me too. It makes it very easy for me to remember what it is I'm going to say. Super duper. Now we're gonna move into the practical part of how we as marketers can use machine learning to our own advantage. I saw this, which was tweeted by Stephen a few days ago. "Machine learning improves conversion rates from 40 to 200% in our research." Obviously, that's very global statement that doesn't really mean very much, but the idea that we can use machine learning to really improve our own performance as marketers is very interesting.
Are there any search-specific ML tech we should be aware of, and how are organizations using it today?
Stephen: Mark's tweet earlier was from Hero Conf. He was saying he’d seen between 40 and 200% increase in conversion rate based on using a specific proprietary PPC platform. Anyone from Microsoft through to Google has plenty of machine learning tech in their stack and they use huge amounts of data in different ways to provide you with better indications of what could be going on, and allowing you to make that inference for yourself. There are a lot of tools available that claim to use machine learning. Not all actually use some form of machine learning. And although most search tool providers are using some form of machine learning, I think the most interesting machine learning technologies in the industry are not in the industry, so to speak.
To really get to grips with SEO specific machine learning tech, you are gonna have to look outside of our industry where there are plenty of tech tools available that I love to use in an SEO context. A simple example: There is an email platform called Phrasee that does ABC multivariate tests on email subject lines to determine which words are most likely to result in a click-through or an open. That's particularly interesting because these are people that is your target audience because they've opted in to receive your email comms. What is cool is that you can use that tech to then create PPC ad copy, for example - You can improve click-through rates from your target audience because you've tested this on that type of person. There’s a nice example of where my agency background helps search - we can cross-pollinate using that kind of technology. Another really interesting application could be around forecasting. Probably the biggest challenge for SEO's is determining what's likely to happen next. We tend to reach for data that is very much within the silos that we work in our organizations. You're too often looking at data in that silo of the analytics platform you're using. But there's a huge amount of data that is not factored into that one silo - you're not factoring in for example, whether you were on TV at that point, or whether you had an outdoor campaign in that particular area, or what the website looked like. A lot of that data does live in other silos, such as your CRM. You can definitely lay other data on to your analytics data - such as behavioral or demographic - to provide better forecasts. A tool that we tend to use to create those forecasts would be DataRobot. DataRobot is a platform that just uses different machine learning models to try and tell me what's gonna be the most accurate forecast based on what's come before it. Going further, you can use all sorts of data, lots of which freely available. For example, in the UK we have the Office for National Statistics, that allows us to lay data such as: What does the economy look like? Do people have money in their pockets? Are they spending at the moment? What's the weather like? Then think about layering on everything from your TV schedule to what's going on with your email campaigns and print campaigns going on. You can't do that manually. Exactly as Lexi says. You're looking at maximum three data points in your head. You need a machine to be able to actually combine that data in a meaningful way.
Jason: Brilliant stuff. I like the email example. I think I might actually use that for one of my clients so, thank you very much. Also, I tend to look at things like this and try and figure them out for myself. I'm obviously wasting my time and would be better to get some machines to help me :)
Where do you think search marketers should use ML? Analytics, personalization, multisensory?..
Purna: I loved Stephen's examples - those are brilliant examples of turning analytics data into insight. At a higher level, just finding clues in the massive amounts of customer data signals, is only possible with AI enabled analytic solutions.
An interesting example is Pizza Hut, who are trying to analyze all of the different channels to see where customers are super happy, where they are unhappy, and figure out what needs to be fixed. What they initially found was that people who order pizza from their mobile apps had a low customer satisfaction rate. If I had been in their shoes, my instinct would have been looking at what I need to fix in my app. But they pulled lots of data and put it into an AI application… and found that people were unhappy from that channel because they didn't know how long they'd have to wait when they picked up pizza. Waiting times were too unpredictable. Nothing to do with the app itself, it was more to do with the experience at the store. Based on that discovery they were able to create e-tickets to tell people exactly how long they would have to wait… and satisfaction rates went up. That is really brilliant. They would not have been able to figure that out without an AI-enabled platform. As search marketers, we're going to become much more data-centric. But that doesn't mean we're gonna have to spend a lot more time running all the reports - the analytics tools are getting much smarter and they will do a lot of these things automatically. We may even be able to talk to it: “hey, what was my pick up rate for this campaign two years ago? What was the other factor at that time?” The tool could pull in different trends.
Another good example of data insights is the Microsoft Graph. Great AI needs great data. For years, Microsoft has been building a repository of data in order to develop a rich understanding to drive a better set of services for the end user. It includes things like world knowledge, work knowledge, people knowledge based on things like search history, brand preferences, LinkedIn data, information from your Office 365 software etc… It pulls it all in and ties it together to get a well-rounded idea of a person and thus develop messages that are personalized and highly pertinent to that person. Looking at our low tolerance levels for irrelevant advertising. We're not willing to put up with that we had as recently as five years ago. We want things to be about “me”. We want it to be helpful and not distracting. And that is evolving so fast! So as marketers, develop understanding and you can deepen your customer engagement, you can increase brand loyalty. If you understand the customer, you can understand how to speak to them, and then AI even gives you the tools to reach them. Lastly, multisensory. Think voice and image. It's not just those old days where you typed into a search box. It's gone WAY beyond that. I could be walking down the street and take a picture of the shoes somebody is wearing, and ask “Where do I find them?”. We have that at Bing. Google has something similar now with Google Lens, but within Bing, you can search détails within an image, and also pull up similar results so you can go ahead and quickly find your path to purchase. Those are just a few different examples of ways that AI can really, really help you. Two useful tools, off the top of my head - Power BI, Dynamic Search Ads. They both work on the fly based on the comprehension that search engines can have.
Jason: It's nice that we can now use Dynamic Search Ads on Bing. As an SEO professional, Dynamic Search Ads are absolutely perfect. I can optimize the site for the SEO, apply that to Dynamic Search Ads and save myself absolute loads of work. I've been using that on Google for a while and have been super-keen to open up an additional channel for that with DSA on Bing. Quick plug: it's going incredibly well on BingBack to multisensory - Bing you can go to a detail of a picture, Google Lens isn't doing so well with that… but I heard that the other player is Pinterest.
Purna: Pinterest has Shop the Look. And then Snapchat has Snap to Buy. And on Amazon, I can shop by image. They are training us as an audience to shop by image.
Jason: Oh Yeah, this technology is changing the way we act. Then, as you said, we're all very impatient when an ad comes up for something we're not interested in. A great example is when I was looking for a microwave oven. I then bought one online. But I still get ads for microwave ovens. That drives me absolutely crazy. Especially when it's ads from the company I actually bought the microwave from.
Purna: You want to start a collection of microwaves, Jason?
Jason: I love cooking, but not that much :)
Lexi: I click on those ads out of principle. If someone targets me I'll click them a few times.
Jason: Brilliant. I'm gonna start doing that. That's a great idea. Back to Stephen.
How should you go about setting up your business to take advantage of ML?
Stephen: There's a great article on Harvard Business Review which is titled, "If You're Not Good at Analytics, Then You're Not Ready For AI." That's probably a good way to view the world.
The difference between AI and ML is that in ML, you train a machine on data sets and it allows you to spot patterns and empowers you to make decisions more readily... Whereas AI tries to make those decisions for you. But then, if you're going to let a computer make decisions for you then you are going to have to get a lot of data in one place talking to each other. But that is rarely the case. Businesses who have mastered this is less than 0.1%. There is so much disparate data in every single different marketing platform, even in the biggest organizations. You have CRM data that doesn't talk to web analytics that doesn't talk to risk databases and other databases… In that situation, allowing a machine to make decisions on your behalf is never gonna work. The logical starting point is to start pretty small. And you should be looking at one data set and then looking at other data we can add to our data set. That means one of the questions is, "Has anyone on the panel leveraged deep learning from marketing, PR, SEO, customer service et cetera?"Maybe take web analytics data, then layer in CRM data - the percentage of our customers who are a particular demographic, live in a particular area, use a particular device - and use that to create personas. Taking that further, you could start to layer with third-party data as well; TGI media platform, for example. Here we are at a stage where we take that web analytics data, find your best conversion audience, then layer in your CRM data to find out who those people are. What is their family status? Where do they live? What kind of device do they use? How often do they convert? How much did they spend? Are they frequent purchasers? Did they always purchase from you or is this a one-time thing? You have an audience of frequent converters and you can start to layer in media preferences. For example, for each converter, 10 read this type of publication and they're interested in these particular hobbies, and they watch these particular TV shows. Then you push that out to PR - our campaign should look like this and it should go to these sorts of journalists. And we are potentially avoiding an argument about domain authority and how do we prove the value of the link? We can establish the value of a link because we know that “X” percentage of people who spend “Y” money would read this publication. It is liberating. Regardless of what Google or Bing wants, we can say that people who genuinely buy from us read this publication. That makes it a lot easier to avoid loads and loads of complicated SEO questions that no one can really answer.
Jason: Yeah. If we use ML within our agencies to help with our marketing efforts, it's helping us to make decisions without having to argue with each other on opinion based stuff. It's very much like GPS helps stop couples arguing in the car about which way to go.
Purna: I find this is a very common confusion point. People ask “how do I start using it?” But they are probably already using it, but don't realize. If they’re using any kind of automated bidding system, such as Target CPA bidding, they're already using this.
Stephen: That's a really good point. There are plenty of tech solutions that are incorporated into solutions we use and we just don’t see it. But what's really interesting about this whole marketplace is “do we have this data”. If we do, we can start with ML... And there's usually some form of tech out there that does that particular task for you.
Jason: Yeah, we've all got loads of data. Applying machine learning to it is gonna enable us to make decisions more easily, less costly, more accurate... and will help avoid us having arguments with our colleagues. That is great.
Machine learning and the future of PR
Lexi: I specialize in PR and SEO, so I've kinda worked both sides. I love data, but I also love people. There's a point where you have to recognize where technology is not necessarily the answer. Know when to say no to tech and recognize that the human touch is required. The role of the human in a world of artificial intelligence is a beautiful discussion. But that discussion is for another day. In the PR industry, our job is to talk with journalists. As Stephen said, knowing who reads what publications is really helpful. What's even more helpful is knowing how they use language. An example is SwiftKey. We did a PR stunt with them. We sent them all the Queen's speeches, and we made a version of SwiftKey that allowed you to text like the Queen - basically translating everything into Queen-speech. But more practically, you could do that for your customers and target them with their own words. Or even learn how a journalist speaks by looking at their writing and target a journalist with their language. There's an argument about whether that's manipulative or not. From where I stand, if someone's speaking French and I could speak French, I would speak French to them. It's a matter of politeness and I think we're gonna see a whole etiquette change come out of this, which is incredibly exciting. It is a moral dilemma. And there is an etiquette change. But also something really beautiful. It is really interesting when you look at publishing because now you can speak to a journalist and a journalist can speak to their audience by using machines to understand what each of us wants. A lot of publications already use AI or machine learning to tell them what they should be covering. What keywords? What are people interested in? What gets clicks? It goes quite far. We have journalists automating article creation. That really excites me and we use this a lot internally. We're not directly using AI within the agency to do this, but we try to understand the AI that's being used by others. We now know that a large proportion of articles in business and finance, especially in the US, are now auto-generated by machines. And that means that we have to pitch machines as well as humans. In our PR, we work out what feeds the machines and we make sure we deliver our news in that way. Look at Newswire. SEOs don’t like Newswire because they've been so badly discredited. But they're gonna have a little bit of a rebirth because AI is a way to pitch. From my perspective, getting AI to write up articles in finance and business makes sense. Those press releases are fairly standard in terms of their structure. So, telling a machine what it needs to look for and to pull different bits together and automating an article is not rocket science. A manual process that's been automated.I'd say maybe 20 to 30% of all our strategies have some element of smart answer built in and our clients are ready for that.
Jason: Brilliant. Great stuff. Speaking the language of the person with whom we're trying to dialogue makes sense.
Lexi: And you can illustrate it. You can identify the deviation between how you speak and how they speak. And then you could take that factor and translate it forward and see how your client-base is being miscommunicated with and have their emotional response.
Jason: Brilliant. I've got the meeting with the client in three days. Do you think I could build something that will convince them?
Lexi: I work in PR. If you can make it look like it's working, it's working.
Purna: That's the best answer I've heard all day :)
Can you trust the search engines own ML platforms to do the right thing for you?
Stephen: The art of PPC is to limit Google’s abilities to make decisions on your behalf. Because Google makes the right decisions for Google every time. You're serving an ad that might meet the metric you're looking for, but it also is meeting other metrics that Google's looking for as well. Look at Isaac Asimov's book “I, Robot”. The first law is “never harm a human”. The second law is “protect yourself unless it conflicts with the first law”. So, Google’s ML is based on “Google is number one, do what's right for Google”. The short answer is, therefore “no”. I would be pretty careful about letting the search engines’ own ML platforms just run. It's absolutely advisable to use them - they are genuinely a lot smarter than we are - but it's also absolutely advisable to keep a really close eye on them.
Jason: That sounds like very, very good advice.
Lexi: From a Google perspective, they are looking for efficiency. But sometimes efficiency isn't actually healthy, so some of the searches may be unhelpful or unhealthy. And I think that has to change. Talking about cancer survival rates. Google gave me an efficient answer. But it wasn’t healthy. I ended up having a panic attack. Google gave me an efficient answer. But it should have said, “call a friend” and given me a list of my friends currently online. It wasn't healthy. My relationship with Google wasn't good at that point. There is a balance - maybe not give people what they want as fast as possible, but maybe give them what they need. A slightly different approach. We should NOT trust the search engines to give us what we need.
Purna: We have to remember is that this is just the tip of the iceberg - we're just scratching the surface. This technology is not there to replace us. It is been designed to make life easier so you can really focus on what matters. There are human gains - I'm not spending hours managing bids or writing ads for a million different things - I have the time to focus on what my goals are. AI is not there to replace us, it's there to help us.
Jason: Yeah. Changing tack, I read somewhere that ML is a little bit like cooking. With cooking, you've got ingredients and kitchen tools.... but you need the talent to actually put those together to make something delicious. In AI or ML that would be maths, data, and intuition. Intuition is central - it all depends on the person who sees how a particular application of a mathematical formula on a specific set of data will be useful. So, my understanding of that is that within ML the human aspect is never really so far away. Maybe that's overly simplistic and optimistic of me.
Purna: No, that's a great analogy.
Jason: Just to be clear, that analogy is from somebody else - it's not mine.
Lexi: I also think we have to give AI and ML a chance to grow up. It's in its early stages. I remember my mom giving me some paint and some paper and left me to play. And she said, "Paint your face." She meant paint a picture of my face. Well, when she came back I had painted my entire face. Was she wrong to say “paint your face” and leave me with the paint? No. We have to forgive ourselves for some of the learning that has to happen when it comes to building this technology. We have to have some level of forgiveness whilst we're steering a whole new level of technology. It's not gonna be perfect straight away. The kid is going to paint its face and not the paper now and again.
Jason: Great. I love that. We need to forgive ML and AI to allow it to grow. Super duper. A great place to end. Thank you, Lexi, Purna, Stephen. Next week, we're back on the safer ground for me “how does RankBrain affect query results today” with Omi, David, and Doc Sheldon.