Moz

The SEOs Guide to Agentic Commerce | Miracle Imameti Archibong | Whiteboard Friday 4k

Duration: 8:31 1464 transcript words EN summary

Video summary

The internet is currently undergoing another major transformation comparable to the mobile era, now driven by the rise of AI agents which represent a shift toward what some call "AI-geddon." These autonomous workers operate using large language models (LLMs) and tool connectivity to complete tasks with minimal human intervention. Unlike traditional chatbots that merely process text through tokenization within a limited context window, agentic AI acts more like a manager; it sets goals for itself, devises strategies, and executes actions step-by-step. However, because these systems rely on predictive models rather than true understanding, they are prone to hallucinations when data is missing or unclear. Therefore, optimizing websites for this new landscape requires ensuring that agents can efficiently process, verify, and retrieve accurate information without getting stuck in loops of uncertainty. To effectively serve data to these intelligent systems, website owners must prioritize efficiency by eliminating unnecessary JavaScript and adopting Markdown over traditional HTML. While browsers are accustomed to navigating the complexities of raw HTML where they might get lost searching for specific elements like knives or tomatoes, AI agents function much better with Markdown because it provides a clear, structured recipe that ships only the necessary data tokens. This approach maximizes the limited context window available to LLMs and reduces processing costs. Furthermore, retailers should revitalize their use of feeds such as Google Merchant Center and implement robust structured data schemas like JSON-LD. These formats allow agents to quickly locate specific parameters and understand site architecture without needing to parse unstructured text, effectively guiding the AI directly to the information it needs to complete a transaction or research task. Beyond technical formatting, there is a critical need for emerging protocols that bridge communication gaps between different systems and languages. The Model Context Protocol (MCP) serves as a universal adapter that allows agents from various sources to connect seamlessly with a website regardless of its native language or underlying tools, eliminating the need for constant translation. Additionally, updating robot.txt files remains essential to explicitly guide crawlers on which parts of the site are accessible and how data should be passed along. As these technologies evolve into new forms of log file analysis, they will provide direct visibility into what agents are searching for, highlighting any failures or parameters being called that might otherwise go unnoticed by traditional SEO metrics. Ultimately, success in this agentic commerce era depends on establishing trust within the broader digital ecosystem. Since AI models verify information against reputable sources and knowledge graphs before presenting it to users, a website must be recognized as an authoritative entity. This means ensuring your site is cited frequently, included in major databases, and referenced by other trusted entities so that agents feel confident relying on its data. By combining efficient data serving through Markdown and feeds with universal connectivity via MCP and strong signals of reputation, businesses can position themselves at the forefront of this emerging technology landscape. The goal is to make websites not just readable for humans, but fully accessible and verifiable for autonomous AI workers who will increasingly drive commerce in the coming years.

Read the full video transcript

[music] >> The internet is facing another seismic shift since mobilegeddon. We're now in AI-geddon. I hope that's funny to everyone. It's funny to me. Today, I'm going to be talking you through how to optimize your website for agentic commerce. And this, in simple words, is optimizing your websites for shopping in the world of AI agents. I'm Miracle Inameti Archibong. So, first things first, I'm going to be explaining some key concepts because I like to think about things in a logical way, and I believe that understanding the logic and how these systems relate to each other will help us come up with sensible plans to optimize our websites. Now, the first thing I'm going to be talking through is AI agents. So, AI agents are essentially workers. They are agents that you send out autonomously, and they perform actions step-by-step using an LLM database. They are called workers because they can get stuck, and they need a lot of human input. The second concept is agentic AI. Now, agentic AI need less human input. They're more autonomous agents, and they use reasoning from the LLM models, and they use tool connectivity. And I think tool connectivity is something to remember because we will mention it throughout this presentation. They use tool connectivity to complete a task. So, they're like a manager. You set a goal for them, they go out, and they find ways and develop strategies to complete that goal. Now, hold on to that thought because this will come in when we talk about how LLMs hallucinate. Now, the next thing we need to understand is how these models and these databases and structures, how they process language. It's super exciting to put things into all of these LLMs and hear them like spit it back to me and spit out the perfect answer. We've heard about people having AI friends and AI therapists because of how magical it processes language and makes it feel like it understands us. But in actual fact, these models don't understand language in the way we humans do. They have to do something called chunking where when we give them a large piece of data, they take it, they break it down, and then they tokenize it, they vectorize it by turning it into numbers, and then they make the best prediction about what this number should be. And all of this is done in a small context window. And that means that chunking all of this data, processing it, understanding it needs to be done quickly, efficiently, and inexpensively. And that leads us to data. Data is at the heart of all of these systems and how they operate. And so, in order for us to optimize our websites, we need to understand how we serve data efficiently, help the models understand data, and help them verify data. Remember when we talked about LLMs using predictive models to understand data and understand language. And this is why they hallucinate, especially agentic AI, because they are managers that want to complete a goal, and they would do the best possible thing to complete that goal. And in trying to complete that goal, they would make up predictions if they don't have the right data. So, how do I optimize my website for agentic AI to make sure that it can process, understand, and serve my data well? The first thing to think about is no JavaScript. I don't think I need to dwell on this. We've been singing this song for so long. Sadly, we are still singing it because these systems don't execute JavaScript. The next thing is markdowns. This is an analogy I like to use when I'm describing the difference between HTML and markdown. Now, HTML is what we've traditionally used and HTML is for web browsers. Now, serving an agent HTML is like me coming to your house and saying, "Hey, can I have a spaghetti?" And you're like, "Yeah, go to the kitchen." And I go to the kitchen and I'm like, "Uh where are the knives? Where's the tomato? What am I supposed to do here?" Whereas markdown is me saying, "Here, this is where everything is. This is the recipe and this is how you get the best spaghetti." And it is the most efficient way of serving AI your data because it uses less token. It doesn't It only ships the data that it needs to and that helps it Remember context window? That helps it perform the most task in that context window. The next thing I'm going to talk about, and this is really important for retail sites, is optimizing and feeding the data you already have. Remember feeds? They're sexy again. Remember Google Merchant that you've forgotten, you never optimized? They're sexy again. And why? These are all concepts that are sexy again because they're efficient ways of serving data to agents. And they're already there. You have them in your toolbox. Dig them out, polish them, and make them usable. And then finally, structured data. I mean, this is easy. We've talked about how these systems don't understand language, but they understand structured data. So, why wouldn't you use it? Now, we've been able to serve them our data in an efficient way. What do we need to do next? Help them find that data. Now, I know that this is a hotly contested topic about what's important. And the key thing is, and that's why I started out explaining this key concept so that we can logically get to where we're going to. So, the way I think about this is thinking about mobile first. Remember when mobile first came out, all the things we needed to do? People had to build mobile versions of their site. Remember AMP? I'm sure nobody remembers is anymore because at that time, you had to optimize your websites for search engines. And eventually, they got cleverer. They begin They were able to then optimize themselves to access your content well. And then it became one big happy web, and we could do away with all of these things. This is where we are with this technology, because it's emerging technology. So, there are emerging structures, there are emerging languages, and all of these things is to be at the forefront and make your site accessible to all of this new technology. So, things you can do is agent permission the JSON, LLM text. These are essentially structures that tell the LLMs or the agents where to go on your site, how to pass your data, how to find data efficiently. So, that's all they are. You can also update your robot text files. I'm sure before I finish this talk, there'll be other emerging technologies that will come up that will be there just to help all of these agents understand your data. The next thing is MCP. MCP is essentially an adapter. And what it does is it helps with the connectivity between your site and all of the tools. Remember when we talked about AI agents needing tools to complete a goal? What MCP does is say, "Okay. Say my site is in German, and then I have different agents coming. I don't need to translate it into so many language for all of these agents." It says, "This is one universal protocol. Connect to this, and I will interpret the data for you." And that's what MCP does. What's really exciting about this is that this is going to be like the new log file analysis. Because here you can see directly what AI agents are looking for in your site, what parameters they're calling, any kind of site failures. And I know this will traditionally not sit with the SEO team, but speak to your engineers, speak to the information architects on your website, because they're probably looking into this or they're probably building this out. And then finally, trust. Now, we know that all of these systems are powered by large language models and they need to verify the data that they are giving out. And what do they do when they need to verify data? They go to the knowledge graph, they go to reputable sites, and they try to verify the data from there. And so, the final thing you need to do is to make sure that your site is a trusted site, that you are in the knowledge graph, and that you're being referenced other people are citing you, and making sure that they are sending signals that you are a reputable site so that you could be included in the database. I hope this has been helpful, and thank you for listening.