How to Harness AI Content Marketing: Q&A with Mike Tamir

Artificial intelligence (AI) has already revolutionized social media and content marketing, and if you’re not utilizing it yet, you’re not keeping up with your competitors. Simplilearn spoke with Dr. Mike Tamir about how to put this emerging technology to use, to better serve your clients or make your own your company more competitive. 

Mike Tamir has been instrumental in developing Simplilearn’s AI courses, including advanced programs on AI Engineering, Machine Learning and Deep Learning with TensorFlow. Onalytica ranked him number one globally as an influencer for machine learning, and in addition to serving as head of data science at Uber ATG, Mike is a lecturer for the University of California, Berkeley’s data science Master’s degree program.

If you’d rather listen to this interview, it’s also available as a podcast shown below -

Q: What is artificial intelligence, and why should we care about it?

A: “Artificial intelligence” has become a buzzword of late. But more substantially, it’s been a study for computer scientists and, more recently, people in the machine learning field —trying to teach machines how to interact and complete tasks, whether they’re simple or more complicated tasks. In general, these are tasks that humans could do but maybe could not do as quickly or in as large of a volume as we would like. This could be figuring out what kind of advertisement matches a particular individual. It could be figuring out product-user matching. It could be figuring out what kinds of interactions will keep a customer coming back, or what would change someone from a low-engagement customer to a high-value user.

In the 20th century, artificial intelligence was rule-based and lived more in the field of robotics. In modern times, really just the last ten years or so, it’s turned into a machine learning process; we create algorithms that enable programs to figure out how to make decisions for themselves. This has been in large part thanks to much more complicated styles of machine learning algorithms called deep learning, which uses artificial neural networks in order to figure out the right answer.

Q: We hear a lot of news about bots in social media. How can marketers put such applications to use?

A: There are several ways that AI has impacted social media. Some of them are really positive. Nobody wants to get an irrelevant advertisement, so when you’re sharing something on social media or you get an ad when you’re on Twitter or Facebook, there are a lot of AI algorithms— machine learning algorithms—going on in the background in order to figure out how to match those. There’s a whole industry and ad tech in general for determining for different users what their demographics and interests are. Often this is coded in AIB categories. So this is what’s used in ad networks in order to figure out how valuable a user is—if you’re advertising a certain product that would be well-matched with a certain kind of interest. And there’s a lot of AI, or machine learning, involved in figuring out who has those interests and who has that valuable demographic fit that you’re looking for. 

A lot of problems are also caused; it’s sort of double-edged. We’re all familiar with the issue of fake news on social media. This, in large part, is something that has been enabled by AI algorithms as well.

Q: How is AI used for tracking social media traffic, not only for affinity marketing but also to monitor competitors and harness other data?

A: There’s a lot of work not just for tracking competitors but for tracking the propagation of different types of messages in general through social media networks. This centers around thinking about the social media network as literally a network, where the different individuals are sort of the nodes in this network, and the people who follow one another or connect with one another socially are the edges. Then what you can do is actually watch how different messages with different characters or sentiments or styles or other features will propagate through a network, and you can look at that to figure out and answer a whole bunch of questions.

One question would be, what sorts of messages or posts or tweets are going to propagate most widely, and what makes them go viral in that way and why? What parts of the network is it going to get posted in, and what parts of the network is it not going to get posted in? Also, for discovering malicious actors, there’s a lot of really great work being done figuring out when there are bots and social media farms that are maliciously trying to accelerate messages and create a false viral spreading of certain news items or messages across social networks. Facebook and Twitter are hard at work, trying to detect those at all times.

Q: How is AI being used for smarter segmentation and micro-segmentation?

A: This is very related to what I was saying about determining the AIB categories—the demographics and interests of users. This can happen in several different ways. One way it could happen is through interacting either through the browser or through a mobile app. You can look at the different apps that the user uses and the frequency of engagement with those apps. If you understand what those apps or websites are about with cookies, you can start to trace that back in order to get a complete picture about what the user is interested in—so, what segments that user falls into. And then this can be used to make sure that you’re getting a better match between the user and the advertisement, whatever it is.

This is a very similar approach to how we match users with products when we do things like recommendation engines. You’re looking at the whole host of interactions that the user has had with different products, other context features—any sort of demographic information that you might get for free about the user—and then you put that all together in order to figure out what’s the best and the most timely interaction for them to have at that point.

Q: Is AI going to take away marketing and advertising jobs?

A: I wouldn’t really even look at it as a competition; it’s more of a collaboration. Let’s say you have an email marketing campaign in which you have a whole host of different options. The algorithm might figure out what kind of tweaks to make in this email campaign—like, “Should I put in the red image or the blue image? Should we suggest rewards of X or rewards of Y? Would this person be more interested in product A, product B, or product C? What’s going to get them to open, engage, and click?” The algorithm might figure out the best match for each individual, and because there are so many different individuals out there, that’s something that humans have never been able to do at that scale. There’s some experimentation with algorithms creating their own artwork, but I would say that’s more for fun than for replacing humans.

All the aspects of design—all the different components and making the menu options for the algorithm to mix and match or to find the best connection with a particular user—those have to come from the marketers, figuring out how to get those different building blocks that then get constructed in different proportions by the algorithm itself.

Q: What are some ways algorithms can be used to help determine the lifetime value of a customer, and even for demand forecasting and sales projection?

A: This is a pretty huge area for machine learning to be of assistance in—detecting when a user is disengaging and when a user might be at risk for churn is a classic machine learning AI problem to solve. Like in all these other examples, it looks at all of the different features and data and the history of their interaction and how that’s changed over time in order to come up with an answer regarding whether they are at risk. It’s the same thing to do in order to determine when somebody is likely to be or potentially could be converted from a non-customer to a customer or from a low-activity customer to a high-value customer. These are really classic examples where machine learning can help and can monitor all of these different signals from all these different users simultaneously.

There’s also a different family of questions around automating interaction—that is, customer relationships. This problem is much newer, and really we haven’t been able to do it in the past few years. It’s often difficult or expensive to have humans there to field every single question or every single comment that you get from your customer base. So you can have algorithmic ways of looking at the history of all the ways that humans in the past have been able to successfully get that top-of-funnel interaction with users and filter it into the right place without making them feel like they are doing an automated “Dial 1 for X” experience. It’s really much more of a natural-language-processing and natural-language-understanding task. And this is something that is more properly associated with AI—deep learning algorithms that allow the machine to start to understand the language and how to respond to it appropriately.

Q: In addition to recommendation engines, how can AI help with targeted cross-selling and bundling? 

A: I would say recommendation engines are sort of the next iteration of that—when you have a suite of different products, and you want to connect the user with the ones that matter when it’s a very broad inventory. There are other situations where maybe you have a portfolio of only a few products, and the question is not so much which products to show users, because maybe you could show them all the products on the same page, but it’s more about what’s the right timing to make sure that you get the right result. Showing every product might not be the best thing to do when you have a brand-new customer, but after the customer has engaged for a certain amount of time and has sent signals about being happy with product 1, maybe then it’s time to start getting that person more involved with product 2 and product 3. I’m sure we’re all very aware that the more cross-selling you can do with with your company, the higher the likelihood is of retaining that customer.

Q: How can AI be used for dynamic pricing based on real-time demand?

A: The first thing to know about dynamic pricing is: Don’t do dynamic pricing; do dynamic discounting. A lot of psychological research has shown that dynamic discounting is dramatically more palatable. People don’t like to get a customized charge, for probably pretty intuitive reasons. But they are excited when they get customized discounts.

This is an area where you have to balance some of the risks of machine learning. I’ve been involved with organizations where we were doing dynamic discounting, and you have to be incredibly circumspect about how you let your machines do that discounting. You need to be very careful that you’re not allowing the machine to learn or adopt an inadvertent bias in how it’s providing those discounts or handling those customized interactions.

This, of course, goes for any sort of machine learning algorithm, but it’s especially important when the result of your machine learning algorithm has a monetary impact on the customer.

Q: How are deep learning and other forms of AI affecting content, not only in the curation of it but actually the generation of it?

A: There are algorithms out there that can read a stack of examples—millions of examples, typically—of a certain style of writing or on a certain topic, and then can learn how to generate sentences that are grammatical, that are written in the same stylistic way and that sound a lot like that kind of writing. This is most useful for things like abstractive summarization. Let’s say you have a very long piece of text, and you want to generate a summary of it that’s a few sentences. Historically that would mean looking for keywords and doing a summary. Now we can do a little bit better with these abstractive techniques in which you encode the content of the entire paragraph or article mathematically; coupled with this ability to generate new content in the right style; you can then write a couple of sentences that focus on what’s most important about that larger article.

Q: How can AI help content marketers with plagiarism detection?

A: That’s another great use case. You know there are companies like Grammarly or Turnitin, which I used to use as a professor. Historically the way they work is just looking at old content and doing a matching. That’s something that really has to be a word for word—the same plagiarism, a copy-and-paste. With modern algorithms, you can catch a little bit more in the net, so to speak, by being able to detect not just if it’s a copy-and-paste but also if somebody has made minor but not significant changes that look like its plagiarism.

Interested to learn how to become a best-in-class content marketer? Well, check out this Content Marketing Certification Training Course now!

Q: How can automatic translations help marketers?

A: Google Translate, in particular, has become dramatically better over the last couple of years. This uses a similar type of neural network architecture as the abstractive summarization, only this time instead of reading through an article or reading through a bit of text in English and encoding that mathematically, and then unraveling that mathematical representation into English text again, you train the model to do, say, English in and German out. So now it’s encoding the meaning of what’s written mathematically, and then when it unravels it, instead of unraveling the content back into the same language (English), it would do it in a different language—for instance, German.

Q: What are some of the AI applications for things like converting text to video? 

A: This is a part of a family of techniques called generative modeling. There’s a whole host of different techniques. One of them that’s sort of classic is generative adversarial networks. To give you an example, let’s say you want to train an algorithm to generate cat pictures for you. What you can do is have one algorithm whose job it is just to tell you the difference between cat pictures and not-cat pictures. So you feed it a picture, and then it says, “Thumbs up, cat” or “Thumbs down, not a cat,” and that’s its model. And that seems rather prosaic, but it used to be a very hard problem, and now it’s at least a lot easier with image recognition algorithms and deep learning.

Now comes the adversarial part. So you have one algorithm that learned just to detect cats or not cats, and then the job of the second algorithm is to fool the first algorithm. Its job is to create—first at random, and then it learns based on what works and what doesn’t work—images that will let it sneak one by the first one. So it learns over time. First, it’s going to feed in noisy data, and then data that’s more and more recognizable as the sorts of things that you might see visually as a cat. And you flip that to the cat detector, and in that way, it will actually learn how to generate images of cats or dogs or whatever it is you want.

Once you have that general infrastructure, you can do a whole bunch of things. One of those things is to start connecting it up with a language, and what sorts of language can be associated with what kinds of images in order to beat that kind of hurdle of creating the kinds of images that match the language appropriately.

Q: How can marketers protect potentially sensitive consumer feedback data from an ever-evolving AI program? Is it safe to leave social media handles in the hands of AI?”

A: I think that the security risks are probably just as bad, which is not to say, not low. Regarding handles, Facebook recently was hacked again. Having that data sit in your internal systems is probably just as risky. Of course, it’s always important to make sure that you are protecting your data infrastructure, whether you’re using the cloud or you’re using on-premises to make sure that sensitive information—in particular, personally identifiable information (PII), or handles—is not vulnerable to attack.

Q: What are some of the ways that AI is actually a good thing for marketing consumers?

A: First, audience targeting. This has to do with the large-scale matching of users and content or users and advertisements directly, or users and products in recommendations or cross-selling. There’s the brand protection, making sure that what you’re sharing and what’s shared on your site meet certain standards, using AI as a filter to make sure that those standards are met.

Then there’s more one-to-one interaction. This has to do with automating the content or chatbots or making sure that you’re targeting users at the right time, in the right place and in the right way.

Q: Will AI make A/B testing obsolete?

A: The answer to that is very clear: no. AB testing will not be obsolete. If anything, it will become more important, because as you’re developing algorithms, while you can have offline data to simulate an experiment, it’s very hard to have a true replacement for controlled, randomized samples, and AB testing is our best way of getting at whether or not you’ve really done something that has an impact.

Q:  With all the new products coming out, how can marketers protect themselves from going down the wrong path?

A: It would be a mistake to think that people haven’t learned to use the buzzword, to say that they’re using AI, in order to communicate that they are using the shiniest thing without actually using modern technologies—or maybe using modern technologies but using them in a way that doesn’t really make a difference. In the end, both of those are ways of getting a nonveridical representation of someone’s using artificial intelligence techniques for a goal that really don’t achieve that goal.

My first inclination in situations like that is to look at what the results are. I guess it goes back to the AB test, in order to figure out if something actually provides value. The best way to do that is to test it.

Q: When it comes to AI, is it better to build or buy? What are some of the most interesting marketing-related applications that you’ve seen?

A: The question of build or buy is age-old because you have to keep asking it, as every case is different. Some of the factors that I would consider in purchasing a chatbot company versus building are the age of the technology and, of course, performance. The technology that we have keeps getting better and better. The different algorithms that are used, and the sorts of infrastructure and platform apparatus that have been built in order to support that AI, are things that might take a long time to build on your own. So the question to balance would be, how advanced is the technology, and is it leveraging everything that it could work in whatever year you are making that decision, versus how much time it would take to build out that infrastructure, which is often nontrivial and could take quite a while.

To be specific, another really important use case that I’ve been working on quite a bit recently is brand protection. Where do you advertise, and who do you allow to advertise on your site? This has a lot to do with the fake-news problem, making sure that you’re flagging for malicious or sensationalized or conspiracy-theory content, which might have a negative impact on your brand. That uses a lot of techniques similar to what we were talking about before with machine translation and abstractive summarization.

Q: What are some of the associated costs and risks of developing an AI system or even bringing in a third-party vendor?

A: Whether you build or buy, it’s going to take the capital. There also may be costs in terms of risks that we mentioned. For instance, introducing hidden biases in your model is a huge risk that could have could be detrimental. There was just an article in The New York Times about Amazon and their struggle with trying to automate résumé filtering and finding out that it was detrimental to their efforts to have an unbiased filtering system. So, being very vigilant in making sure that you’re not introducing those hidden biases, or at least making sure that the people who are managing and developing your AI are vigilant, is a risk and a potential cost if you don’t manage that risk.

Q:  Are there some key questions you can ask vendors to make sure they know what they’re talking about?

A: Some of them really require an expert in the room in order to make sure that they are giving the right answers. Certainly, ask questions about how they measure performance and make sure that their measurement of performance is something that makes sense. The example I like to use, and people do this in marketing all the time, is that you can really fudge your metrics for performance. If I can build a really great predictor of whether or not a user is going to click on a banner ad—it’s a 99.99 percent accurate predictor—that’s pretty darn good. And the algorithm works like this: It always says no, the user is not going to click on the ad, because most users don’t click on banner ads. And so this is a great example that just picking something like accuracy as your only measure is not going to be the right way of measuring performance. It’s important to interrogate a potential vendor to make sure that they’re using the right kinds of metrics to really represent if they’ve created something that improves performance over a “no” predictor.

Q: How can we determine if an AI project has been a success or failure? What are some meaningful key performance indicators?

A: That really is going to be tied to whatever the goal is in order to know what the key performance indicator for that goal is going to be. Certainly, this goes back to AB testing too, and it’s part of why I was so confident in saying that it’s not going away in the end. The KPI that really matters is the one that you can move in AB testing because that’s going to give you the most veridical representation of values.

Q: What are some other ways you can use AI for competitive analysis, such as detecting consumer complaints about others’ products? 

A: There are a lot of modern sentiment and intent detectors that can summarize that text and look for when you’re getting a high volume of sentiment about a particular concern with a competitor. There are a lot of things that can happen in competitive intelligence that should or should not happen. One example is pinging a competitor’s site and figuring out reaction times and things like that, which I hate to recommend, because it may not be the right thing to do, though that does happen.

AI Master's Program Banner - CTA

Q: Any other words of advice for people curious about using AI in marketing?

A: I think maybe the most important thing that came up in this conversation is dissuading people that this is a competition of machines versus humans, any more than if you’re in construction thinking that a power drill is a replacement for a human because now you’re not going to be twisting the screwdriver. If anything, it’s going to make your job easier and quicker and hopefully much more effective.

Learn the Latest Intelligence on AI

Whether you head a marketing team, or you’re a practitioner seeking to branch out and help future-proof your career by upskilling in the latest technologies, check out Simplilearn’s new online course: an Introduction to AI and Machine Learning. Developed under the supervision of Mike Tamir, this course provides an excellent overview of AI concepts and workflows, machine learning, and deep learning.

About the Author

Dan BiewenerDan Biewener

With 15 years of experience teaching and developing instructor-led training and video-based e-learning curricula, Dan is currently Director of Training Research at Simplilearn where he conducts and compiles research on the latest content and training best practices. Backed by his degree in Speech Communication and numerous certifications in Digital Marketing and aviation technologies, Dan brings insights from both sides of the training process.

View More
  • Disclaimer
  • PMP, PMI, PMBOK, CAPM, PgMP, PfMP, ACP, PBA, RMP, SP, and OPM3 are registered marks of the Project Management Institute, Inc.