Mike King: Sandbox SEO, building programmatic templates and tools that supercharge SEO efforts. As the product lead at Uncommon Goods, he spearheaded growth initiatives, with a focus on SEO, site search, and speed. Zack is the creator of an innovative SEO tool or innovative SEO tools like SERP analytics and site speed analytics shaping the future of SEO strategy. So one of the things I talked about in my talk was the fact that Google cannot, as a single signal, identify whether or not something is created with generative AI. And a lot of his talk today is gonna be kind of about that. So presenting LLM Experience Game, How We Climbed the Animal SERPs with a Text-based Gen AI Game, please welcome Zack Notes.

Zack Notes: Hello. Hello. Hello. Can you guys hear me okay? Awesome. First off, Mike and the iPullRank team, thank you guys so much. This has been a great conference so far, especially for the first one. It’s been like super smooth, run really well. Yeah. I’m so happy to be here with you guys today. A little bit about me. I’m Zack Notes, founder at Sandbox SEO. We do programmatic SEO, web devs, and web design. Product at Uncommon Goods, if you guys haven’t heard of Uncommon Goods, we are a retailer in the GIF space. I lead a product management team there. My team mostly focuses on SEO. We also work on prepurchase, site search. Summary is I’ve been in SEO since 2011. I’ve been in product for a little bit shorter, and I’m also sort of a Web designer. Pretty much explains me. These are my three boys, Nico, Luca, Knox. And the story of this website, Animal Matchup, that I’m going to speak about starts with them.

So it was the spring of 2023, and my kids were obsessed with these books. And for any parents out there who have young children, you’re probably familiar. Who would win pits two animals against each other in a hypothetical fight. And so they set it up with fun facts, and then they move into a fight story, and then they declare a winner. And we were probably going through five or six of these books every night at bedtime. Yes, kids. I can see. So around the same time, OpenAI had just released, GPT4. And it was fun. I was having a good time with it. I was excited. But I was what was more interesting to me was the release of their API at the same time. 

You see, at Sandbox, we’ve been doing programmatic SEO for a while, and we we wanted to build a test site that combined AI content and programmatic. And in order to do that across thousands of landing pages, which is programmatic SEO, we needed access to the API. And when I finally got it, I was very excited. So I had a few ideas that we were we were juggling with, but who would win books gave me the biggest one because it turns out that the desire for animal matchups is much bigger than just these books. The Internet loves this stuff. Some of these numbers are crazy. There are 1.2 million species of animals in the world, and that’s named animals. There’s plenty more that we haven’t discovered or haven’t named. But if you do the math, that’s 720 billion unique matchups. And nobody’s searching for a cardinal versus a gorilla. Right? So 99.99% of that is no search volume. But if you just take what’s left over, that’s still 72 million matchups. And that number is probably high. It’s probably somewhere under a million. But still, as a programmatic SEO, these are the kind of numbers that get you really excited.

So we decided to go for it. We basically wanted to take the Who Would Win books and translate those into into a website. Right? So the initial launch, I I partnered with a a web developer. Six weeks it took us to get the first MVP out. Right? We started with 100 animals. We launched with 50 matchup pages, and this is what they looked like at the time. We introduced the two animals. We gave the tail the tape where we compared stats. We sprinkled in some fun facts similar to what Who Would Win does, and then we polled for the winner. And while we’re here, show of hands, in a match between a gorilla and a grizzly bear, who takes the gorilla? Okay. And the grizzly? Yeah. So the Internet likes the grizzly bear also. It’s bigger. After the poll, we unveiled the matchup, and this is 100% AI copy generated by ChatGPT. And this right here, where you’re looking at, like, this is the crux of this site. That’s it. It’s these fight stories. And at the time when we launched, this copy was static, saved in the database. And I’m going to talk about a bit more about where we’ve come since then and how we’ve evolved. So that was two years ago. In a time since, we’ve probably had 200 launches. We’ve iterated many, many times and built a lot onto it. But this is our growth chart. It’s not amazing. It’s not like blow you away. It’s sort of slow and steady, constant growth. The majority of that 84% is organic SEO. The rest is mostly direct. And in that time, we’ve been through eight core updates. And I’m sure that you guys have noticed during that time, there’s been a lot of AI sites that have blown up and then gotten smacked down. And so for some reason, we have remained. And I don’t know why that is. I don’t think anybody knows why that is. Maybe Google does, maybe they don’t. But I have some theories as to the things that we’re doing right that are counterbalancing the weaknesses of AI content. And so I’ve taken those theories and I’ve synthesized those into the agenda for today. 

So I’m going to go over tips for creating AI content that survives. I’m going to attempt to inspire you to build better websites through LLM Experience Game. And I’m going to talk about some tools that you can use to scale AI content. And first, some caveats. Our traffic has not just been up into the right all day, every day for the last two years. We’ve taken some hits. Some algo updates have pushed us down, but we’ve always managed to bounce back. The other one, we’re not on this site, we’re not talking about four zero one k allocations or CPR techniques. It’s it’s not your money. It’s not your life. It’s animal mashups. And so I think it’s safe to say that Google sets the bar a little lower here. That said, let’s talk about AI content that survives.

My first tip is sort of AI content 101. Most of you have probably figured this out by now, but you can’t just copy and paste multi paragraph long form strings of text straight from ChatGPT into your website, and that’s because that copy is garbage. At least for the time being, human written or at least human edited copy is going to beat ChatGPT or other LLM copy, and that’s specific to long form. The longer it goes, the worse it gets. And I know it’s a smooth fucking talker and it’ll convince you that it’s great content. And maybe your site will rank with long form for a couple months, but then you’re going to get freaking smashed. And so instead, I would recommend prompting for small, digestible, bite sized copy blocks. And you want to want to intersperse them throughout diverse content types. 

And this next tip goes for AI content or otherwise. And there’s actually patents behind this. But images, GIFs, tables comparing facts, interactive elements, polls, videos, you can’t just have AI content. It has to be a mix, diverse content. Now, this is where it starts to get fun. People like to press buttons. They like like to feel empowered to skip past content that they don’t want to read. They like to share their opinion. Interactive elements are the features that make a website a website and not just a book. And, like, this is where I feel very strongly as a product person that these are the kind of interactions that are going to make for a great website. These are also the kind of interactions that are going to give you a better experience than you’re going get on AI overviews or you’re going to get it on ChatGPT. Right? Like, that’s our main competitor with this, is people that are going to ChatGPT and asking them, hey, I want to see a matchup between a gorilla and a grizzly. But if we can give buttons and experiences that elevate that, then we can offer something better.

Now, this is where it gets really fun. LLMs, and specifically the LLM APIs, like ChatGPT’s API, they allow you to really increase personalization, and you can take the inputs that the user gives you, use that to create live GenAI text, and that’s text in the session. Right? And you can even use the LLM APIs to create choices for the user, a generative a generative UI that gives options created in the session based on their feedback. And while I’m here, I want to talk about the live Gen AI text for a minute. This is something that Josh Degrasse Bousman, called this out when I was going over the slides with him. By the way, if you don’t know Josh, don’t follow him, strongly suggest on LinkedIn. Very, very smart guy. Josh pointed out, like, hey, having live Gen AI text on a website, SEOs don’t do that. That’s bad. Right? Because if you have text that’s created after the page loads, it’s not visible to Googlebot. Right? Googlebot can’t see this text. But that’s okay because if you do it right and you build something that’s useful, useful content, right, helpful for the user, then you’re giving them something that improves their experience. And if they can’t crawl it in the beginning, that’s okay. And I’m not saying that your entire web page should just be like a couple buttons in a form. Right? You need content for Google to crawl, but having live text can help improve that experience and win rankings.

So let’s go back to the growth chart for a minute. The green period here that’s highlighted, that’s when we really shifted to start using LLM features. We took those static fight stories, and we started using API calls instead to generate the fights on demand in the session. We also introduced the create your own feature. So this is a unique thing where maybe a user came to the site to see what would happen between a gorilla and a grizzly bear, but then after that match was over, they could see a gorilla versus a velociraptor. Right? Or velociraptor versus a polar bear. It was interesting. One thing that I didn’t really realize at the time that I saw looking back was we were shifting the site into a video game. And I didn’t really see it at the time, but it’s become more clear now. We see it in branded search, Google Search Console, people will be searching for animal matchup game or animal matchup video game. And it’s like, oops, I made a video game. Okay. 

So I’m going to talk a little bit more about LLM features, but first I want to cover something that’s super important for AI content. EEAT, you guys are all very familiar with it. First tip, like, we we know that AI lies. It has, you know, Fyre Fest credibility. And you have to fact check. It can be quite tedious to fact check a site with thousands and thousands of pages. So what I would recommend, focus on the high visibility pages, anything linked from the homepage, top 5%, maybe 10% percent of your best templates. But you need to take the time to fact check. I use Gemini Deep Research for this. It’s gotten better with Deep Research. It’s less manual. It’s still manual, but it’s quicker. We also cite sources as a as a part of fact checking. We’ll take the best, most relevant articles, make sure they match the content, and then we have a field in the database where we store them. I would also recommend hiring an expert author in a consulting capacity. And you don’t have to have a huge budget for this, maybe $50-$100 an hour, few hours every week at most. You’re looking for someone PhD level for most sites. You want to manage their expectations for a light engagement. And this is the actual, LinkedIn job profile that I had for our wildlife biologist that we hired. Ideally, have published work with papers in Google Scholar. Once you hire them, create a bio page linked to their Google Scholar page from the bio, link to their bio from your about page, and then coauthor a few high quality, non AI articles with them. And obviously, use author markup in that. Then you want to link to their articles prominently from the home page. And this is more for search quality evaluators. You might be wondering, like, wait a minute, dude. I thought you said this site was 100% AI. So technically, it’s 99.85% AI. We have three pages of human written content. One is our About page, the other two are co-authored articles. And then we have 2,000 of Gen AI content, mostly animal pages and matchup pages.

Okay. Back to LLM UI integration, back to the features, how we build this. Just a reminder, we’re using the API to build features like these. So the first of building block of this is prompt chains. And I’ll explain what a prompt chain is. It’s pretty simple. In a prompt chain, the output of one prompt serves as the input to another prompt. And in this case, it goes to the user. So here’s a simple UX case. First prompt, give me an apex predator in the Amazon rainforest. And let’s say the API comes with a jaguar. Right? So then second prompt, describe a jaguar emerging from a hiding spot. And then in the second prompt output, it will explain a jaguar coming out of the forest. Right? If we were to refresh this page, right now it’s a black caiman coming out of the water, or it’s a harpy eagle swooping in from the forest. So that kind of gives you an idea of the magic of LLM APIs, is you can sort of start to have something that creates on the fly, more generative. Just to be clear, this is the section that executes in the code via the API, and this is what the user sees. Let’s change this a bit. Now instead of asking the API to give us the animal, we could ask the user, enter an animal that lives in the Amazon, and then we go through the flow. This is the section that executes in the code. We can even do a generation where we use the first prompt to generate choices. Very simple example here. Generate three animals, comes up with the animals, then the user selects an animal. Their selection, let’s say cane toad, goes in the second prompt, and then the second prompt, we generate the live text, the story. Now, I give generating three animals here as an example of choices just because it’s simple and easy to understand.

In reality, this sort of problem would be much easier to solve with a database with a table of animals and randomized function. Right? It’s not the best use case in reality. We use choices as part of the gameplay. We have cards where a user can pull a card for win scenario, right, or items that the animal would use, or mutations. And so we’re using the generations here to create choices for something that’s harder to create a dataset for. That’s where choices really make sense. And with these kind of features, we can offer something called LLM Experience Gain. And now this is where we’re taking it back to SEO. With LM experience gain, you’re creating a new, unique web experience that’s not offered by existing content on the SERP, including the AI Overviews. And the the main if I if you guys take anything else away from this talk about this weird animal website, the main three things, personalization, LiveGen AI text, and user choices. These are the tools that you can use or your product managers that you work with can use to build features that improve the website and offer experience gain in the SERPs.

So let’s look at the SERPs. Right? Within the context of animal matchup, let’s say the query is gorilla versus grizzly bear. Current content, content types. At the top of the search, there’s an AI Overview. Throughout the search, including the AI Overview, there’s opinion pieces, scientific facts. People love to debate about hypothetical animal matchups. It’s crazy, there’s so many opinions on it. Like, Joe Rogan had a post about it. Forum debates. Reddit and Quora, really popular here in these SERPs. There’s images, there’s videos, YouTube, TikTok. But with an LLM, now we can offer something new and different to this SERP. Right? Personalization. They can create their own matchups. Live Gen AI text. Fight stories that display in this session. Choices. They can choose the environment, the win scenario. Right? 

Let’s leave the animal kingdom for a second and look at a different SERP. Pasta carbonara recipe. Current content types for this, pretty simple. Right? It’s recipes. With an LLM, personalization. We could offer recipe customization based on time availability. How much time do you have to cook this pasta dish? 20 minutes, 30 minutes, an hour? What about dietary needs? Is this gluten free, vegetarian? Choices. Allow users to request alternate ingredients. This can all be in the website that you create as part of the experience and you can do this all with LLMs. Live AI text. Once you’ve collected that information from the user, generate a custom recipe, offer interesting modifications, suggest side dishes. Right? That would crush so many so many, you know, static, simple, frozen recipes that really can’t personalize as well to people and their individual needs and tastes.

Another query, things to do in Kyoto. Current content, AI Overviews, experiences, attractions, guides, events, a map pack, Reddit forums, of course. With an LLM, you ask the user, when are you going? Who is going? How many people? Are you bringing kids? How many days are you going for? Are you already staying somewhere? Right? Choices. Which are you interested in? Temples, performances, restaurants. And the LiveJan AI text could build you a custom itinerary. You could do this like, they can do this now in ChatGPT, but you can make the experience so much better in web format with choices and buttons. Because you can make it more structured. They don’t have to type as much. Again, people like to press buttons. The chatbot is a very exhaustive format of UX. You can build something better. The ideas here are the hard part, right? There’s some SERPs where there’s probably not a great idea, but there’s some SERPs a lot of them, I think where there are good ideas to be had. So I built a custom GPT to ideate for you, and it’s on the slides. There’s a link there. You just give it a keyword and then it’ll give you ideas for features that it thinks would help build on the SERP to give you LLM experience gain. Okay. A few more minutes.

I’m going to give you all some tools to scale AI content. We’ve built a lot of images for animal matchup, probably like two or three thousand of them. And the tradition like, if you right now, before we built the CMS, the process was to go back and forth between our dataset and ChatGPT or MidJourney we used, and our prompt list, image compression, format changing, like there was all these different steps. Instead, we built a CMS for image generation. So in the CMS, the prompt’s saved here. It connects to the database. So there’s a dropdown where it says hippo. We can just choose the animal we want. We hit Run. It hits the ChatGPT API. It generates an image. If we’d like it, we hit Save, and then it’s written to the database. And we can even do bulk, right? So we can upload 100 animals at a time and just pop out 100 images. I’ll reject usually more than half of those, but it’s a lot easier than going back and forth between all the different systems.

For our stack by the way, the site is built on Firebase, React, and Vercel. That’s my favorite stack right now. Our database is Firebase. It’s cloud based, easy for nontechnical people to edit. And the reason that I bring this up for scaling aside from over other databases is that it just has a really easy UI for nontechnical people to get in there. And I use that for fact checking. I hire VAs to help with fact checking. And it’s pretty easy to find the fields, edit the fields. It’s great for that. Also, I’ve heard this mentioned a few times this week. GPT for Sheets is great for scaling the actual copy. I use it to scale the emojis as well. One specific tip here, I keep the prompt and the output in separate columns. You can use the concatenate function to wrap it in so that you can call GPT. The reason for this, you’ll have easy access to your prompts later. You don’t have to store the prompts in a separate file. You also won’t accidentally rerun ChatGPT, which will add costs and overwrite good output. So your prompts are there, your output’s there, and it’s kind of frozen, so it won’t accidentally rerun overwrite. This is a great tool. Love this. Someone else mentioned you can use GPT. Jori Ford, you can use GPT and Sheets as an AI, a checker on, like, AI overview exposure. I think that’s it. I want to thank you guys for your time.