 So, I know it's early and we're kind of easing into this very technical topic, so I wanted to tell you a quick story about my relocation to Seattle. So I moved to Seattle last October, and instead of being a normal person who would find an apartment or a house, I managed to find a houseboat, and I have gotten so close to some of the animals in the marina, and it was going to get weird. But these ducks have become my friends and my family, and I have so much fun with them and see them every single day. So something interesting happened right around Easter of this year, and I noticed this female behind me, I named her Linda. Linda was in a planter on the front of my boat, and it was funny because I just thought she was sort of hanging out there, and one of my neighbors was saying, no, she laid eggs. She chose you. So obviously, I proceed to just completely lose my mind, right, like this consumes me. I am like worried about her. I'm doing research on incubation and how long this is going to take, and I'm super excited. I just lost it, you guys. It was so exciting. So I, towards the end of my talk today, I will reveal what happened, and you might get to see some really ugly crying photos of me, which will be hilarious, so stay tuned for that. But basically, Linda has become my muse for some of this machine learning stuff. So I thought it would be cool to create a program that could figure out if I fed it a picture of Linda, or if I fed it a picture of my pet snake pumpkin. So I also have a ball python. Her name is pumpkin. She's adorable. I promise, I hope you'll warm up to snakes by the end of this talk. And I know this looks a little intimidating, right? It's like, these are probability scores. What does this mean? How the hell did I create something like this? By the end of this talk, you're going to have everything you need to do some of these machine learning things. And most of the time, as I'm putting some of these machine learning models together, this is live footage. Like, I have no idea what the hell I'm doing, literally. I'm not trained in computer programming. I'm not a data scientist. I don't know what the heck I'm doing. But I'm curious enough to Google these things and try to figure them out. And there are some really cool resources that will allow all of you to do this as well. So Codelabs is one of my best friends. So this is Codelabs. I do a search for TensorFlow. And you can find all of these walkthrough tutorials for machine learning that every single person in this room can do today. This stuff is accessible to everyone. A couple weeks ago, I found this walkthrough of TensorFlow for Poets. This is a 20-minute walkthrough of a machine learning model to train on flowers. So this downloads to your local computer. And you basically create this machine learning model that can differentiate, is this a dandelion? Is this a daisy? Like, what is this? So of course, I have to sort of take this stuff apart. I'm really curious. I put this thing together. And I found this folder structure called Flower Photos. And this is where I found those folders of daisies, dandelions, et cetera. I thought, well, what the heck? Why can't I just add a folder called Linda? Why can't I add a folder called Pumpkin and add photos to it? The problem is that I don't have hundreds of pictures available of Linda or Pumpkin. So I go to something called ImageNet. The history of ImageNet is actually really, really amazing. These are hand-labeled images. And they've now gotten to be over 14 million images for training models just like this. So what I was able to do is I searched for duck. And I found mallards. So I downloaded around 300 pictures of mallards. And then I found ball pythons and snakes. And I downloaded around 1,000 of those. So I put those into these folder structures. And voila. Within a 99% probability, I was able to figure out a model that could differentiate Pumpkin from Linda. Pretty cool, right? What doesn't work so well is if I feed it this handsome picture of Rand, it thinks Rand is Pumpkin. So I'm like, well, that's interesting. Wonder why that is. But wouldn't it be cool if this could also recognize people? How hard would that be to recognize a person? So Rand is also inspirational to all of my machine learning efforts. So funny. So of course, I'm a creep. I go on Google. Search for Rand Fishkin. I find all these pictures of Rand. I add this ImageSpark Chrome plugin. And I'm selecting, I think I selected 150 photos of Rand. And I bulk downloaded them into a folder called Rand in that model. I basically retrain it, reconfigure everything, plug that picture back in. And my model thinks within 65% probability that is Rand. That's amazing. This is the power of machine learning. It's not that much data, 150 photos. And this stuff can start figuring it out. It's really, really powerful. I'm so excited to talk to you today about it. But imagine what this can do for image optimization, right? Who here likes doing alt text for images? No one, right? It's the word. You like it? All right, there's one person. I'm sorry. I'm so, so sorry. One guy likes it. But it kind of sucks, right? So does tagging or titling. It's not a fun thing to do. Wouldn't it be nice if these models could take over some of this? Macy's has over 4,700 new arrivals. Do you know how easy it would be to build a model similar to that TensorFlow for Poets and pick up what these items are? Is it black? Is it a dress? Is it a top? You can easily train models to figure this stuff out and do it for you. And this isn't anything revolutionary. This is everywhere, right? Like we run into machine learning every single day. This is in our spam filters. It's suggesting things to us on Netflix. And what I'm most excited about is it's saving lives. The medical applications of machine learning gives me goosebumps. You can now take image optimization and basically figure out, is this person probable to have a condition or not? So this is super, super powerful stuff. OK, who here likes to create reports? Does the alt text guy like to create reports? No? Not so much. OK, great. So I think I can speak for most of us. Creating reports, especially SEO reports, it sucks. It's not something fun that we do on the weekends, right? So check this out. Now before I walk you through this beautiful dashboard, I have a question. What would all of you do if you wanted to take this dashboard, publish it, then send it to everyone in your entire organization along with a written summary? And I should also say you don't have time to write it yourself. That's exactly what NVIDIA thought. Using the boundary breaking power of the extensions API, along with the magic of AI, our partner, Automated Insights, is automatically populating this dashboard with that written summary. No work required. Now let's make a change. Thank you. Actually, I wanted to call one thing out here in this last bullet point, 18.4% rise relative to August. These are insights in our data that we might have otherwise missed. Now let's just show you because dashboard extensions are full-fledged members of dashboards. They have full two-way communication with the dashboard. So as I click on this mark, the new data that's selected is now sent, and we have a new summary that's generated. That's what dashboard extensions will do for you. So what do you think? What the shit's so exciting? Wouldn't that be so amazing? To have that just automate for you and in a meeting, I think this is the future, and this is going to help us as SEOs evolve as an industry. All right, so we're going to break down what is machine learning. Hopefully by the end of this, you're going to be able to talk about it intelligently and have the foundations and also feel comfortable to execute some models yourself. We're going to discuss applications for SEO and then the tools and resources. So what is it? It's basically a subset of AI that combines statistics and programming to give computers the ability to learn or train without explicitly being programmed. And 99% of the time that you hear the word AI, it's just machine learning. So machine learning AI has been around for a really long time. It's gone through ebbs and flows. And one of my favorite facts about it is that it was actually inspired by a woman. Her name was Lady Ada Lovelace. And in 1843, she was at a party in England where Alan Turing was as well. And they were talking about the analytical engine. And this was a computer way back in the day that was mechanically operated to do math. And basically what she said was, yeah, this is cool, but it's not showing us anything we don't already know. It's not giving us insights. And this stuck with Alan Turing for decades. And he later continued to refer back to it as Lady Lovelace's objection. Pretty awesome. Two types of machine learning. There's supervised and unsupervised. Supervised we've already talked about. This is the labeled training data you feed to the model. Unsupervised is non-labeled data that you feed to a model. And it can cluster groups of photos or data, et cetera. Three common types of models is clustering, regression, and classification. You don't need to be a statistician to figure some of this stuff out. It's actually super basic. Some real live examples is just if you were to do unsupervised clustering, you could put in a bunch of photos of animals. And it would cluster the different types. If you wanted to do linear regression, this is a great example of home size versus home price. So you can predict from there. Classification is this a duck or is this a snake? Is this person sick with something? Are they not? This is what a basic machine learning process looks like. Training data would be in the most important part. And something people don't talk about enough is you need to save a portion of that data for testing. So you train a big portion of it into these models that train themselves. And then you take that chunk that you saved and you see how well it did. You see what the prediction is like. So what's powerful about this is I don't have to say, hey, Linda has this goofy beak. It looks like this. Pumpkin is long, kind of skinny. The models train themselves. It's really, really powerful. But I think a lot of you are probably wondering, how do they get smarter? How the hell does this stuff work? It's not black magic. It's this simple. So these models train themselves on something called the loss function. So imagine if I'm training a set of data, these two points are incorrect. My model can evaluate how incorrect it is and feed me that loss function. I want to get that loss lower. I want to make the best model possible. Something you don't want to do is to overfit. I don't want to count for each and every single testing data because it doesn't account well for new data. This is a perfect example of overfitting. Like who wouldn't want a bed like that? You have to have a bed that accounts for the different types of sleeping patterns, right? This is exactly what happens with overfitting and it turns into a problem for machine learning. And you don't have to know the ins and outs, but just to be familiar with what that is, you're good to go. So if machine learning was a car, data would be the fuel. It really is. A model is only as good as the data that you feed it. That's why we see Google asking us questions like this. They want us to label this data for them. Whether it be testing or training data, we're feeding this to people. Same with Google Images, right? Is this an image of a mountain? Yes or no? It's amazing where this stuff is headed. So how can we apply this to SEOs where it gets super exciting? All right, so voice search opportunities are incredible, right? A lot of people have been talking about the featured snippets. Google released Google Actions recently and this is available to everyone. You don't need to know any programming to get started with these three templates. So you could create a trivia, a flashcards or a personality quiz. And basically what this is doing is you can communicate those things to anyone with a Google Home or a Google Voice device. Really, really powerful stuff. If you wanna get your hands a little bit dirty in some of the code, Google will send you to Dialogflow, which is fascinating. This prompts you to fill out something called Smart Talk or Small Talk. And this is our first glimpse as SEOs to keyword research for voice. It's amazing, it's hilarious and it's super interesting. You get to see what are the most commonly asked questions to voice systems. It's really, really powerful. So I quickly created a quiz. I actually, I just used the template, that trivia thing. And I created this SEO quiz and was rejected dozens of times. Google did not like this because they didn't like the word Moz phonetically. They thought it sounded like mom or MA apostrophe S. I then finally decided to get rid of the branded name and it was approved. So, super fun, super easy to do. If you wanna play around with it, you just, I won't say it out loud, but you ask Google, talk to SEO quiz. It's pretty fun. But any of you can do this stuff. It was just plug and play. Recommendation models are evolving, right? And these are getting better and better and they should be. So last year, I talked about how at Moz, our blog was pulling suggestions for next posts from the three different categories. This wasn't proving to be a really successful model because it wasn't specific to the content that someone might have been on, right? So I fought for this to go to just that first primary category and the results were amazing. Pages per session on our blog went up over 11% and over the previous period, I think it was 4.5. Really powerful stuff, but this is gonna get better and there's no reason you can't take some of these models and apply it to real live users on your site. What did they visit last time? What are they interested in? What have they already read that you don't wanna suggest? This stuff is gonna get more and more powerful. It's Melinda, she's so crazy. I don't have time to go over this, we're gonna skip. Writing, okay, who here likes writing meta descriptions? Does alt text guy like writing meta descriptions? What? It's gonna be all day, huh? All day, sorry. No? No, yeah, it sucks, especially for really big sites. How do you scale it? It's kind of a pain in the ass, right? So I was looking at Moz and we have the community forum and what's frustrating is Google typically will pull meta descriptions for us, kind of, thank God, right? Because this system, this platform is a little old and all of these pages actually have the same meta description. There's not a whole lot I can do about it until we switch to a new updated platform. So it got me thinking, I was on Algorithmia, this is a website that has pre-trained models and I found this, advanced content summarizer and I thought, this is interesting. What if I can use this to do some things, right? So I go back to that page I showed you earlier in the SERPs and I pull the question text in that first answer. I plug this into the model. This is just on their website, on the front end and the output I got was amazing. I thought, what would this look like if I cut it down to 150 characters? It is way freaking better than what's Google pulling, right? The first part's basically the same but the fact that the model up at the top says the short answer is no, do not change a URL that is currently, that is amazing. That is automatic. Imagine what you can do for large sites that could use stuff like this. And so it got me thinking like, okay, I know that there are these things and we could do this but I'm not technical enough to do it myself. So I found the help of two people way smarter than I am. JR Oaks and Grayson Parks are unbelievable. Highly suggest you guys give them some shout outs. They basically helped me put this thing together in the last couple days for all of you. I said, would it be possible to take, just allow SEOs to give a URL to this program and to pull the text of a page, send it through something like that and generate meta descriptions for us. And they did it. This is insane, right? This is amazing and while it's not perfect, this is gonna help our industry evolve to have a bigger perspective on strategy and more powerful things. It does incredibly well. I'm super impressed with what they were able to whip up. And unfortunately, I wasn't able to share this particular Google sheet with all of you because we would all pull the APIs and basically crash JR's AWS but I have all the steps you need. So all you need to do is you have to find a developer familiar with AWS to go through this process of setting these things up on a server. If it took them a night or two, every single person in this room can find a developer to do this as well. And then the best part is us as SEOs, we get to kind of look like badasses in Google Sheets. We just connect everything. We copy and paste this function into our script editor. You literally, you just go to Google Sheets, tools, script editor, paste that in, and then within one of the cells, you can put page description, A2, 150 characters, whatever Google decides is the meta description this week, right? Really powerful stuff. Is that cool? Are we excited about that? This is awesome, thank you. Super excited, these guys worked really hard and I think this is where things are headed and I couldn't be more thrilled. So yeah, it's, I felt like this. I was like, what's up Power BI? Taking your lunch money. I'm generating the ship for you. What's up, Will? And what was funny was we were actually working on this and JR goes, Paul Shapiro came out with this article and I don't think I noticed it and neither did I and this was recently. But he basically shares another way for you to do this. He also shares the GitHub repo and the steps it takes. It's a little bit different than the way that JR and Grayson did it, but definitely check it out. So this is just the tip of the iceberg and this is why I'm so excited to share this stuff with all of you because I know that people in this audience are gonna think of better ways to apply this stuff. You guys are gonna think of more applicable things for SEO and machine learning if you can understand how it works in the fundamentals. It's gonna bring us light years ahead as far as SEO in this industry. So super excited to see where some of this stuff goes. All right, what are some tools and resources? How to build your first ML model is pretty easy. The hardest part is just cleaning your data, right? That is gonna take the most time for all of you, is just pulling enough data or images or whatever it might be, cleaning it and having it available to train. The rest is just a couple lines of code. And those lines of code are readily available to each and every one of you through code labs where, again, you can do searches for machine learning or for TensorFlow and Google has this awesome setup. It just walks you through every single thing. They've also implemented machine learning crash course, which is amazing. I took it when they released it and it walks you through the basics. I still don't feel like I'm super proficient in some of the math and the programming, but that's okay. At least you get an understanding of how this stuff works. TensorFlow actually has really great tutorials as well. Eager execution is sort of the new trend in TensorFlow right now. It's gonna speed things up. Super exciting. And something I won't shut up about is colab notebooks. So is anyone in here familiar with iPython notebooks or Jupyter notebooks? No? Okay, that's, I heard a clap. I heard one clap. These things are amazing. So Jupyter notebooks is basically where we were doing machine learning previously on our local computer. And it was sort of a check and balance system of these models. And Google has said, hey, we're gonna create the Google sheets of this that will be available to everyone. So it's really, really powerful. Not only can I collaborate with people like JR, but it allows me to use GPU for free. So that little red box, it automatically connects to GPU making this way faster than things I could run on my computer. And I don't have to download different libraries like Python and TensorFlow. This stuff is evolving so quickly and it is readily available to everyone. It also comes completely ready to do interactive graphs. So you can do three-dimensional graphs. You can do all sorts of things. It gives you the programming that you just copy and paste to your data to visualize some of this stuff, which is really powerful. Kaggle is one of the world's largest data science competition platforms. And it's fascinating to keep an eye on this stuff because you sort of get an idea of where things are headed, what are people working on, et cetera. TSA has put up a million and a half dollars to figure out can someone create a model for us that can recognize a behavior that we could flag? And I need to say this just, I know I'm running out of time, but just super briefly, there is an ethical issue in AI and machine learning. These models are only as good as their data, right? So if we are feeding it racist or biased data, it's gonna cause all sorts of problems. That's why we need more people in this field and we need to be aware of some of these implications. Something sort of funny is Quora is trying to de-dupe all of their questions through Kaggle and Google is trying to understand what all of their videos are about, right? They're trying to automate some of these large scale things. The hardware has evolved so quickly as well. So this has basically followed Moore's law. We went from CPU to GPU, TensorFlow at IO this year just released TPU, a TensorFlow processor. So this is gonna speed up some of this computing to the extreme, it's really exciting. And what do you need to get started? These are the very basic resources that I have found to be really, really helpful in understanding some of these things. If you sort of already have that background, these are some of the advanced resources. Regardless, this zip mystery will just blow your mind. If you're bored one night and wanna check something like that out, it is amazing. All right, so what did we learn? Machine learning is just statistics and programming. That's all it is. A model is only as good as its training data and that loss function helps us improve models but we don't wanna overfit that. Every single person in this room can go to Codelabs and create a model today. 20 minutes, probably not even. Machine learning is gonna help scale some of these SEO tasks and we are gonna evolve and be able to work on higher level thinking. All right, how's everyone doing? Good. All right, what the hell happened to Linda? What happened to these eggs? You guys aren't gonna believe this. Three of them hatched and I completely lost it. I ugly cried all day. That's what I usually look like. It was so amazing. I lost my mind. How cute are they? One sort of fun sad fact, that's not a fun fact but a sad fact, these ducklings in the marina, they die a lot because they don't, they can't get out of the water to rest because they're in a marina. So my boyfriend and I built this ramp that is basically a duckling ramp that they can walk up and down and they loved it. They, it was amazing and all three ducklings are still doing very, very well. So thank you so much for sharing this with me today. Appreciate it. Thank you.