[VIDEO] 2023: The year of Deep Learning AI… and what it means for e-commerce

This is going to be the year of Deep Learning for e-commerce, a topic that almost assuredly is going to affect each and every one of us in this room. Let’s talk about what this means, how we got here, and have some tips for marketers.

So first, a question: which one of these is not like the others? And I’ll tell you what these four pictures represent. We’ve got the Nobu restaurant, the real one in Malibu. It’s expensive but it’s a scene. We’ve got a liver transplant, not nearly as fun as Nobu. The Bing search engine, and the Google search engine. Which of these is not like the others? Think about that. It’s really about the reason why that makes this so fun. The answer is the Google search engine, because the other three have a waitlist. That’s right, the Bing search engine has a waitlist to use it. What planet are we on? If I had told you four months ago, the Bing search engine would have a waitlist to access it, you would have said, “You are crazy, Jason. That’s impossible.”

Even Bing admits that most of the people using their search engine are using it through syndication powered by Bing, maybe not on Bing.com. So what’s going on here? This is a reflection of Bing’s decision to beat Google to the punch in integrating a large language model such as ChatGPT and AI into the search engine. It’s creating unprecedented demand for their product, and it’s potentially changing the search world. And I bet there’s at least one or two of you out there that drive some of your website traffic from search, so this could be very important to the crowd.

Let’s talk about this a little bit more. Why is this so effective and so important? Well, a legacy search engine – I like to think of it like a parking permit. You ever had to get a parking permit in high school or your office? A parking permit does not give you a parking spot, all right? At least those of us in the Los Angeles area, parking can be tricky. It is the right to hunt for a parking spot. That’s what a legacy search engine is. It gives you the right to sift through a bunch of links and try to find what you really want, and sometimes that works and sometimes it doesn’t.

When I needed to come up with ideas for that first slide, I’ve got some ideas, but I don’t know everything. I’ll be the first to tell you that right now. Most of you probably know way more than me, but I wanted some help from the internet. So I put in a very specific thing: list examples of things that have a waitlist. What did I get back from a legacy search engine? Well, I got one link, and it talks about being on a waitlist for Harvard and colleges, and then I got what other people searched for, which I really don’t care about. I have a very specific question, and then I got “people also asked,” which seemed terribly duplicative. This didn’t seem to be a great use of real estate by a company that’s supposed to be organizing the world’s information.

That means there’s probably some kind of opportunity to do this in a better way. And, in fact, if you look at what people also search for, I think it says “What is a waitlist in a restaurant?” I know English is hard – “wait” and “list” – it’s, I mean, it’s two words together. If that’s what other people are asking a legacy search engine, and that’s informing my results, I’m not sure I really want that kind of information.

Let’s look at the alternative for a moment. ChatGPT uses a Deep Learning AI to actually answer the questions that we pose. Is it always 100% accurate? No, it’s not. And, in fact, sometimes it’s very overconfident, and it likes to think it’s always accurate when it’s not.

But nonetheless, in this specific example, I put in these same exact words. There’s no hocus pocus here. “List examples of things that usually have a waitlist.” Look what I got. I got a list, I got examples, and I got high-quality content. And I got it right away without having to read a bunch of pages or scroll through a bunch of links. This is revolutionary.

If anyone out there generates content, generates code, generates anything that’s text-based or can be a derivative of something text-based, this is going to change the way you do business. By the way, if you want to access these tools, you can go to OpenAI.com. It’s still a free beta research-y sort of mode right now, which is great because there’s no direct cost. If you need it at scale, there’s an API, and you can sign up for a very reasonable price. I get no commission from these people. I just think it’s a really cool product.

You might ask, “Well, what about my PowerPoints? I spend a lot of my time making PowerPoints. Can ChatGPT help me with that?” Absolutely. When you ask it what to do – and again, these are real screenshots that I took; you see my picture there – no way I could fake that. So it starts off, “I’m a text-based AI model, so I can’t actually make a PowerPoint. But I can do the hard work for you. I can show you what you might want to have in that PowerPoint.” And it generated slides and bullet points and everything.

Well, here’s the funny part. ChatGPT is just getting started, and there are a lot of people already out there hacking it or tricking it. There’s a great hack called “Dan, do anything now,” which gets you around ChatGPT’s rules, such as not being biased and not being crude. You want it to be biased and crude? It’ll be biased and crude. You can Google search something like “Dan Reddit jailbreak ChatGPT.”

But another way that you can use it without formally jailbreaking it is asking it to generate code to do something it doesn’t think it can do. It just told me it can’t generate a PowerPoint.

Well, not exactly. If you’re a coder like me and use a language called Python, which can do just about anything. You can ask it, “Why don’t you generate a Python script that will make a PowerPoint for me?” Oh well, of course, I can do that. Here you are, and you can then edit that. And this PowerPoint that it made, I’ll show you here. It’s not going to win any Design Awards, I’ll tell you that. But again, this was a six-minute process compared to spending hours making a PowerPoint. You could install a nice template, you could use your brand colors, you could use logos, you could use images. This was just something I threw together for example purposes here, to show the power of a large language model extends beyond large language. The other thing is, you might ask it to generate code to do an image. I said, “Well, what if we tried that? Can you make me an image with hexagons in the style of a certain artist? The hexagons are usually a little rounded and very colorful.” Again, ChatGPT is very confident. “Sure, happy to do that.” Well, what did it create? Are those hexagons, ladies and gentlemen? Not in my book. Might be a diamond, might be a double diamond. I didn’t specify colors. These colors are horrific, especially if you’re red-green blind. So, occasionally it will fail. So, it’s not going to take your job right away. That’s the good news. There still needs to be a lot of oversight here around ChatGPT. That was a fail.

So, images though are a critical part of everything we do, especially as marketers and as people who try to drive sales through fundamentally a visual media – a website. This is a painting by the real Dali, the Salvador one. And it has certain stylistic elements to it, and you might recognize it as a Dali. It’s got a sparse landscape, it has subdued colors, it has kind of melting clocks, which is one of his themes. This, however, is a painting that was completely generated by AI. That means you can use it potentially – there’s some legal issues here – without royalty payments. You don’t need to hire a designer, you don’t need to hire an agency to make this for you. This is the future of creative output because it’s already being used in static images and migrating even into video content going forward. So, this is another way that Deep Learning is changing what we do as marketers and as people who create content. By the way, the prompt for this, if you’re like me and you live in Southern California, you probably recognize this. This is the iconic view of the Santa Monica Pier just south of the pier looking northward. And that is exactly what I asked it to do.

There’s the prompt. Again, OpenAI.com if you want to play with this. They give you a hundred for free, then there’s a small charge if you need images after that. You can see it generated a few things with these kind of Dali-esque elements to it.

So, now that we’ve had a little tour of the power of Deep Learning, let’s talk about how this affects marketing and a little bit how we got here. There’s kind of this pendulum that swings in Tech. You have periods where you consolidate around winners, you have periods where you break everything apart. We are definitely breaking everything apart. That’s exactly what Deep Learning is going to do as we defuse search and other previously consolidated technologies.

So, what is that going to mean for marketers? And I’m speeding this up a little because I do want to stay within the time. Let’s talk about how we win in this exciting new world. First thing is, there’s been a little bit of a development behind the scenes in how ads are bought and sold, and that’s the transition from a second price to a first price auction. To summarize the importance of this for marketers, you can now have multiple agents acting on your behalf in an auction without loss of efficiency. That means if you want to have two or three partners all trying to win an ad for you, you no longer lose efficiency because you used to pay what the second partner would bid. Now you just pay what you bid. It’s much easier. You bid six dollars, you pay six dollars. If you have 14 other bidders in there and they’re all under that, it doesn’t change your efficiency. It doesn’t change the power of your other partners. This is a wonderful thing for the Deep Learning revolution because it allows you to be more experimental. It allows you to try a lot more different things, and it allows you to find solutions that work for your entire population of users.

I’ll give you a specific example of that that’s sort of near and dear to me as someone who’s worked in website retargeting for many years. Here’s an example of all your website visitors. Some of those visitors, they show up, they make a purchase. Others are completely different. They have a very complex path to purchase. I’d advise everyone go into GA or your attribution tool, look at that multi-channel funnel report, pull it line by line. You might be really surprised that you’ve got some users that needed 8, 10, 12 site visits to buy something, especially if what you sell is somewhat considered. Those are cases where at least our research shows that Deep Learning, it digs in there, it digs into those users that have the highest AOV and the most complex path to purchase. It doesn’t mean that legacy solutions aren’t useful.

Keep using those as long as they’re delivering a ROAS as you need, but Deep Learning can add to your stack. And you can now confidently build those stacks because of the way ads are bought and sold on a first-price auction. So, action number two here is incrementality. Who’s called a vendor and said, “I want to talk about incrementality”? You know what you normally hear, “Okay, I think I’m losing, yeah, yeah, I’m going into a tunnel or something.” Right. It’s not something that most vendors want to be super front and center about. And there’s a very simple reason for that. When you try to drive incrementality, and this problem has been around, this guy was born in 1838 just to demonstrate that this is a difficult issue. I’m going to talk about why Deep Learning and incrementality actually mesh very nicely. This image was generated through DALL-E. Someone called it terrifying. I thought it was great to use for this presentation.

Deep Learning doesn’t just learn from ads that are shown. Most legacy systems show an ad, predict who’ll click on it or engage with it, and then predict who will convert. That’s fine, that’s awesome. That’s much better than the old way of just bidding three dollars CPM and showing your ads out there. That’s a great solution. However, it doesn’t know if those end conversions are incremental or not. Some of them might be, but you have to specifically measure it and test it. An incrementality that is generated is a byproduct of the generation of attributed sales. It is not directly a goal. In Deep Learning, it understands everything that machine learning understands because it looks at who gets ads and what happens, but equally, it looks at who does not get ads and what happens with those users. That is the foundation of incrementality – the understanding of what happens in a world where I show ads and a world where I don’t show ads.

So, Deep Learning tends to drive better incrementality because of that conceptual understanding that it has that a legacy or machine learning approach doesn’t have. It also ingests more data, and there are other technical reasons. And then finally – sort of the last minute here – a lot of people are asking, “Are you worried about this?” Yeah, of course. There was a list of the top six jobs that ChatGPT could do, and I think I do three of them as part of my job: data science, data analytics, and content creation. I think there was journalism and some other things. We all do things that are going to be replaced by this. We can let it replace us or we can cannibalize ourselves. And what I mean by that is, get ahead of this, start using these tools, learn about these tools, use them to superpower your own job. Then, instead of being fearful of them, you’ve become an ally with these tools. And I think you’ll find that you can go a lot farther than perhaps you thought initially.

When you think about all of this, when you think about that pendulum swinging towards innovation, when you think about the SSP changes, when you think about the need to address all your very heterogeneous users and the need of a stack of solutions, I think it all adds up to 2023 being a great year for Deep Learning, for e-commerce, and for marketing. Thank you for your time today.