Name: Meta Muse 101: The Agentic Internet Is Here. Now What?
Uploaded: 2026-10-01T18:15:08.137Z
Duration: 29 min 41 s
Description: Meta Muse 101: The Agentic Internet Is Here. Now What?

Transcript for "Meta Muse 101: The Agentic Internet Is Here. Now What?": Hi, everyone. Welcome, and thank you for joining us today for human dialogue, Meta Muse one zero one. My name is Aaheli Guhathakurta. I'm a product marketer here at HUMAN and also your conversation lead today. Before we start the session, I wanna point out a few housekeeping items. To the right of your screen, we have the q and a tab. So any questions you have during the session, feel free to place them there, and we'll try to answer them during our live q and a portion. Also, check out the docs tab we have, which has a free trial of our agentic visibility solution, a blog we recently published on Muse, and a blog I published on agentic trust levels. In this... In addition, we will have polls throughout this entire event that will pop up in the tab section for your participation and sign ups for virtual events in the future. Now let's get into why we're here, Muse. So for years, most digital experiences have been built around a fairly simple assumption. Someone visits your website, moves through a journey, and eventually takes action. AI agents are beginning to challenge that assumption. Meta's launch of Muse puts sophisticated agentic browsing in the hands of a much broader consumer audience. Instead of simply helping someone find information, these agents can navigate websites and take actions on their behalf. That raises an important question for businesses. What happens when a meaningful percentage of your customers begin interacting with you through AI agents? Today with Geoff Stupay, SVP of product, and Julian Norton, staff product manager, we're gonna discuss what human is already seeing, what the growth of consumer agents could mean for analytics, and how businesses can prepare for a web that's increasingly shared by humans and machines. Jeff, thank you for being here. Welcome. Now that we've seen AI agents become more autonomous and accessible, why is human uniquely positioned to help organizations navigate this new digital reality with confidence? Yep. Thank you, and and and thank you to everyone for joining us. So for anyone who's a little bit less familiar with Human, we help companies kind of understand who's interacting with their businesses and protect that experience across kind of the entire customer journey. We spent more than twelve years understanding automated traffic, how it behaves, what it's trying to do, and then also where it creates risk. So today, we're a global team of around 400 people protecting more than 500 customers. We're verifying over twenty twenty trillion digital interactions a week across about 3,000,000,000 unique devices a month. All that to say, it gives us a pretty broad view of what's happening across the Internet and kind of how activities are changing it. Yeah. And AI agents are adding another dimension to what we see. An agent could be helping a real customer find a a product or complete a purchase. It can also be testing credentials, creating fake accounts, or doing something else that puts a business at risk. So really knowing that something is automated is really only part of the picture. So who who is it acting for? What is it trying to do? And do you want it... Do you wanna allow that? So even when the activity is legitimate, you still need to decide, you know, what an agent should be able to access and do within your environment. These are the types of questions that we're helping customers work through. We're we're we're taking the experience that we built and understanding kind of the automated traffic and then applying that to agents so that businesses can support the activities they want and take action where they see the risk. So with that, Julian, let's get into Muse. Like, what makes this launch, you know, such an important moment for for the agentic Internet? Thanks, Geoff. Thanks, Aaheli. So I am so excited to present today. I get the best slides. So, really, the most important thing is that there's a new class of traffic that really didn't meaningfully exist a year ago. So Agenta commerce and transactions are moving kind of what was speculation into now, like, provable practice. AI agents are, like you said, beginning to interact with businesses as this new class, And there's a significant opportunity and risk between, like, trust and governments and making sure that that that keeps pace keeps pace for businesses. So, just to give you, like, the most important slide that that I can kinda show today is when we think about what agenda traffic looked like a year ago, it was meaningful. It was there, but it wasn't really widely distributed. It was it was pretty concentrated, and then compare that to today, last month in September. We've seen more traffic in September. It's, like, tripling that traffic. So we've observed this across all of the digital ecosystems, muses emerging as a significant source of this traffic of measured activity. I'll get into, like, why and what. But the the kind of the key point for this is that this is, like, now observable traffic. It's not a future prognostication. It's not a future trend. Like, this is traffic that is currently happening. And this is kind of proven is that as we've been looking at the activity for, you know, the... Now years is that the Internet by default is no longer human. Most of the traffic is now synthetic, whether that's coming from bots or or, AI agents. So the assumption that, like, traffic is good if it's human, bad if it's spot, like, kinda really doesn't doesn't really hold true anymore. So now when, like, we're looking at the different breakouts of the traffic mix is that there's the actors in which is actually driving that activity, and then there's what they're actually doing. So agentic activity can come from multiple categories, whether it's, large language models using bots for model training and and data retrieval purposes, cloud agents, agentic browsers, you know, browser extensions that are coming from, like, consumer devices and consumer hardware. And, like, businesses need visibility to differentiate the good from the bad, and it's not just like a simple block automated traffic allow humans because you can have good and bad traffic from each of those actors. And when we're kind of getting into a little bit of the technically, like, what are exactly these things, when we look at the left hand side, most of these are l m model training bots, and you can compare that to a real time retrieval bot. So when you're asking a question to, like, for example, ChatGPT, and you're like, hey. Summarize this website for me. That will fall under that real time retrieval where that will trigger OpenAI to have a bot go to website and try to pull that information in real time. Now that's different from this right hand side of, agents, is that when we're looking at the AI agents, these are loading a browser. These would typically show up in your analytics as Chrome or Chromium. Most of this traffic is stealth by default, meaning that most analytics solutions out there, like, they're not really fingerprinting or detecting, exactly if it's an AI agent and, more importantly, which AI agent. And I'll get into this later, but which AI agent it is really helps determine whether this might be trusted or wants activity versus not. And that, like, without that visibility, you can't really get into identifying if that traffic is malicious or not because identification is a is a prerequisite for making informed policy decisions. And Muse is in that cloud agent category. And so when we're talking just about Muse, Muse was launched, you know, earlier in September. I think it was, like, about three weeks ago, four weeks ago. Within less... Within... In less than two weeks, it reached number one on the App Store, had over, you know, 2,000,000 downloads, and it's bringing really agentic capabilities to mainstream consumers that previously just wouldn't have that exposure and experience to use it. And so how Meta... How how Muse is being positioned is it's being positioned as a personal assistant. It's across, all the users from Facebook, Instagram, WhatsApp, Messenger. You know, Meta reaches, like, literally billions of people every month, and they're positioning this as a personal assistant, travel agent, style scout, fitness coach, whatever it is, in a way that's different from traditional chat interfaces where they're actually positioning it as, like, a thing that can do things for you, navigate the Internet in, in ways that, like, OpenAI just, like, previously done support and other major model providers. And so how... Like, why particularly Muse is changing this equation, is that has an extremely generous free tier. They published, you get a 100,000,000 tokens per week. When you compare that to OpenAI and Meta free tier, like, for example, OpenAI, they don't publish exactly how many tokens you give. But when you were gonna... If you were to compare that to something like the API plan, it just, like, doesn't even come anywhere close. Like, Meta is, giving, like, a very, very generous free tier. That's number one, just on tokens. The second thing is is that you get an AI agent as part of that free tier. That's the first time that I've seen that, part of any major model provider. Like, an AI agent that actually navigates the web, can fill out form, navigate the website. Previously, you had to be on a free... On a paid tier for, for example, OpenAI to do something like that. And the third thing is is that the distribution and, like, how heavily Muse is being advertised is just, like, unlike, anything I've seen. It's it's... Feels like literally everywhere. I can't, like, open Reddit or anything like that without seeing ads for this. And and I do wanna briefly mention, DOT from OpenAI. That was, I think, released, like, literally two days ago, maybe a day ago. Like, that is kind of what they're trying to have as competitor to use. And that requires a minimum of a $100, at least today on OpenAI's plan. So you kind of compare, three versus a $100. Like, it's, that... That's kind of, like, why this is being distributed and why it's being, used so much more than kind of anything else I have right now. Alright. So in our early observations, in this early period, we've seen that the traffic has been as high as 72% of daily agentic requests. Meaning, it's it's basically, like, three fourths of every agentic request that we hear... That we're seeing is coming from use, coming from Meta. That's over five times the request volume that we've seen from Chesapeake's agent, and Muse is generated in that in that short period. Remember, like, this hasn't really been out there that long. It's already generated, like, over a 100% more requests compared to all other agents combined. And it's reached, like, literally millions and millions of requests in, like, under a two day period. And now when we're looking at, like, okay. Well, what is Muse actually doing? Like, what... Who cares on the amount of requests? Like, what is it actually doing? And so 84% of the requests from use are targeting product and search pages. It's it's really aligning itself at least when we're looking at some of the activity behind it. It's really aligning itself around commercial activities, ecommerce, in particular. And when we're comparing the Muse traffic to other AI agents, Muse is 25% more likely to hit a checkout page compared to other agents. So it's it's very clearly being targeted on these kind of ecommerce workflows. And this shows why kind of attribution and seeing this agentic traffic matters so much to discovery and commerce and analytics and fraud teams, to really help differentiate, is this an agent that might be trying to add a thousand items to the cart, or is it a human who's... Might be conducting that kind of fraud? It really helps in in segmentation and making an informed decision around those, kinds of activity. And so... Okay. What what does this mean? So Muse, creates a traffic pattern that is fundamentally different at a mainstream scale. If left unclassified and unattributed, the agentic activity is gonna distort conversions, attribution, and performance reporting. Because remember, when we're going back to, like, is this traffic, stealth or not, like, news is showing up as Chromium. It's not necessarily intentionally announcing itself to websites that it's news. It's, it... It's something that has to kinda be, like, fingerprinted and discovered, which we specialize. So the businesses need that, verification of their intent and their identity and their behavior ultimately to kind of differentiate this traffic. Alright. So for ecommerce, retail, travel, hospitality, fintech, insurance, like, these are all relevant industries even though I was focusing on ecommerce. Like, you could think about, for example, a banking workflow and getting to a KYC know your customer, check, is that if you can know that these are actually AI agents, sooner in that workflow, across all these different injuries, they can really make a difference in terms of kind of affecting your bottom line in those downstream, metrics. In in cybersecurity, this is, like, known as, like, shift lock. The earlier you can detect these things, the earlier you can apply a policy that kinda fits the needs of your business. Because fraud and security teams need to protect accounts, payments, promotions, checkout, things that, is really important to kind of differentiate what the agent is doing versus what the agent is not doing. And then on marketing and analytics teams, is that it... It's really important to have reliable attribution and performance data at scale because it affects so much of your KPIs that that usually the the business really relies on meaningful metrics around. Alright. So considerations for your organization. Like, what are the considerations for your organization? So, like, how do you distinguish human activity from bot activity, from caller activity, from AI agent activity? Those are those are questions that, you know, I I see being asked a ton. And which agents are reaching your digital properties, your your website, your checkout flow, the the things that are, kind of business critical? And are these agents influencing your customer journey, discovery, search, product, payments? And how do you attribute those agents based off of what they're doing? Are your checkout and payment experiences, like, ready for for these agents? Like, for example, I cannot get CHPT to enter in my credit card number when I give it, for good reasons. Like, it... There's there's legitimate reasons why I'm using an open philanthropic. Like, don't wanna do that. But they all have integrations with payment providers to kind of enable those ecommerce activities. But if your if your website doesn't support those or it doesn't support some of the activities, like, it's it's, it's it's important for you just to know that. So, like, how can you apply different policies by the agent, what it's trying to do, the... Your your page paths, and the actions that it's trying to take, and who owns those activity, those agentic activity measurements across your organization. So this gets into, okay, building an agentic strategy. I I assume most people here have Internet facing websites, and so you kind of have to have a a response, a strategy around how you wanna control for this kind of activity. So it's seeing who or what is acting across that customer journey, turning, visibility into control. Like, you you can't really have a policy or control in place unless you have that visibility first to inform that policy decision. So distinguishing between human activity and bot activity in AI agencies are all, like, specializations and and you gotta treat these differently. And making the decision not just is it a bot equals blocking because there's lots of great bots that have important business impacts, turning into a trust or not decision. So some humans are doing good things. You wanna allow them. Some humans are doing fraudulent activities. You wanna block them. Same applies to bots and AI agents. And making sure that you're optimizing for what you wanna trust and whether you might wanna be blocking it or redirecting it or forcing them to log in. There's not a... Necessarily a one size fits all for what you need that agent to be doing because everyone kind of has their their different industries that you may want to allow agents to sign up in in one website, but completely block it. For example, like, signing up for bank accounts. Definitely, the... I don't think banks want that. So that you can make sure that you're capturing the upside of what these agents can be providing without absorbing all the unnecessary risking and potential fraud. K. So, now getting into, like, why this matters on the AgenTA trust side. So AgenTA trust, at least for our focus, is it turns AI agents into visible, controllable, and trusted interactions. Seeing agent actions and their intent and making sure you know what can be trusted and not trusted and make the agents play by your rules. Really, enforcing the policy that your business needs to make sure that you're capturing upside without absorbing the downside. And at least what this looks like from our side is that we have, particularly a dashboard for AI providers, the visitors, their intent, and seeing that traffic over time to really enforce the goal that the data is actionable, that it's not just in another, like, static traffic report from general analytics. Like, this is really trying to connect the data back to your business decisions across security, commerce, marketing, and analytics. And you can't you can't really do that unless you have visibility in a in a policy in place around how you want to be treating these AI agents differently from all other kinds of traffic. Alright. So key takeaways here. So gentr traffic is here. Consumers like Muse, and there's literally dozens and dozens of these AI agents. Muse just happens to be the most popular one. They're already browsing. They're already transacting. And visibility comes first. You can't really have a policy unless you know how agents are interacting with your existing properties, and then intent matters. Understanding who the agent is, who's behind the agent, what it's doing, whether it's legitimate or not. Like, spoofing is is a is a problem in the space that we have to, kind of have that attribution for. And that that all comes down to governing for trust. Allow the valuable activities that are trusted, that are attributed, that, are gonna ultimately make you and your business money while getting the AI agents out that aren't. They're gonna abuse you. They're gonna commit fraud that, just aren't the most that you want. Okay. With that, Aaheli, I I assume that we have questions now. Awesome. Thank you, Julian. We do have questions coming in our chat. So to start, do you mind clarifying? So when you say traffic observed by human and a few of those charts that we show, what traffic. are actually Specifically, AI agent... I'll just... Let me go straight to the slide here. AI agent requests. So what we're showing in this slide isn't? bot traffic. It's not looking at model training. It's not looking at real time retrieval from bots trying to access a website without JavaScript like doing a simple curl. This is the, Gentek activity that's loading the website, has a browser pointing and clicking. Like, this is called a computer use. So this traffic is, like, exclusively filtered on just the AI agents that load up a browser and interact with the page. we talk about spikes in non human traffic, we are talking about searches and discovery, or are you already seeing agents buying? Oh, already seeing agents buying. Like, I I have evidence of proof, and I've done investigation on this. Like, we... We're already seeing agents buying. This is, like, very clearly well established. Okay. Someone is asking if an AI agent logged in, added items to a cart, or checked out on a customer's behalf, would your current controls allow, block, or flag it? Yes. So there's there's a couple of different choices here. It doesn't just have to be, like, block or deny. So so, yes, it's attributed. We see that. I've I've I've seen that personally myself. I've tested it personally myself. So there's a couple choices. There could be just do nothing. Just flag it on your back end, and if you see it's a high risk fraudulent transaction with a payment provider, that can be additional information I can use to help justify, like, should Julian's logged in account being... Should should it be able to buy a 100 iPhones for, like, you know, $100,000? Like, okay. Maybe treat that differently. Maybe a mistake. Or it's something that you could... Yes. You actually don't want that allowed. You could have blocked it way sooner. So there's choices in what you wanna do, whether it's redirection, issuing it a challenge, you know, like, email password reset. Like, there's there's lots of different actions you could take based off of the needs of the business. What do you recommend as an approach to bot traffic in our new AI bot landscape? Are there easy ways to discern those versus bad or malicious traffic? Easy ways to distinguish bad from malicious traffic? No. And the and the reason why, like, there's not, like, a really easy way is that bad and malicious traffic, like, they're intentionally trying to show up as human traffic. So they they spoof, they masquerade as humans. Like, there's there's just things that they do that, a a... An unsophisticated website provider just, like, isn't gonna have the tools in place to to do. I'll I'll give an example of, like, you have a user that's logged in, you have a certain IP address, and then all of a sudden you see that same user account showing up, like, thousands of miles away in a different IP address based off of that, like, geo IP. Like, you can do things like that where a human can, like, can't teleport thousands of miles away, but, like, that kind of stuff is, just kind of not easy to deal with, you know, some of the existing tools out there. Awesome. We have a question on agent identification. You say over 90% of agent traffic is stealth by default. Does HUMAN expose any stable header, user, agent, or identifier today? If not, how can a retailer reliably distinguish legitimate news traffic from a human or malicious automation? Yeah. So so that... That's... It's a great question. It's like, if it was as simple as it's coming from one IP address or one user agent, like, you wouldn't need a company like Human. Like, that's terrific. The the the reason why, it's not that simple, is that, like, this is part of, like, our intellectual property and, like, how we determine our attribution here. Like, that's our special sauce. And I can get more into technical. We have published blogs on how we detected, like, comets browser. That's on our technical blog. People can look at that. There's, like, very specialized techniques to do this that isn't as simple as a user agent. Like, for example, Googlebot, you can control traffic from Googlebot through the robots.txt and through a user agent. That stuff is is pretty simple for, for, platform, owners. But for stuff like news, it's it's not that simple. And most AI agents, it's it's... It is not that simple. Okay. Looking ahead twelve to twenty four months, what security architecture changes do you believe organizations should start making now to safely support Magento Commerce? So I I think about this in terms of kind of two different workflows. One is allowing interaction and and traffic at scale. MCP is a is a... Becoming, like, a very popular framework for that. So if it makes sense for the business offering an MCP, I would call it, like, a workflow, for a couple of reasons. It's way faster for the agent, which means it's way faster for the Human behind it to do it. It costs way less, and it's also very easy to know that all traffic that comes through that MCP application is in the agentic traffic. Just... It's just very, controlled, but it doesn't make sense for all use cases. The second is testing AI agents from your own website and seeing what's happening. Usually, that means making things, accessible in the traditional sense, like well defined headers and, like, there's there's, like, very good guidelines on how agents, navigate the page and how they, like, read HTML. That is just kind of following accessibility best practices because they're trained on so much website data. But as long as you're following accessibility best practices, I get... I should be pretty good. Cool. Thank you. We have another good question. How. can businesses prepare their digital platforms to be discovered, understood, and acted upon by autonomous AI agents? So I'll start with discovery. So much like traditional SEO or AEO, like, now now they're calling it like AEO, like, AgenTek Engine Optimization, is making sure that the model training bots and the search bots and the real time... And possibly the real time retrieval bots are able and, welcome on visiting your website. It doesn't make sense for a lot of business for the real time retrievals. Like, for example, media customers, they make most of their money on, like, website articles and news and, like, that stuff doesn't make as much sense. So I would just start with at least focusing on SEO or AEO, just making sure that it's able to show up in any capacity, whether it's a summarized version or not, because that's really what these things rely on to kind of get started with their rejected activities. That makes sense? And we will end on this last question. What new security and privacy risks emerge when AI agents can make decisions and take actions on behalf of users? That... That's a big one. So I just wanna, like, repeat this back. So security risks security risks and... Can you manage to repeat it one more time? Yep. So what new security and privacy risks emerge when AI agents can make decisions and take actions on behalf of users? So security risks that that I think about as I read a really good article, and I'll try to link this, later, is being able to use Muse to build profiles based off of Facebook accounts and all kinds of things that you can do at scale to build profiles of people for good or bad purposes that just kinda didn't really exist. So I think about how much easier it is to do things like that, whether if it's good activities like shopping or bad activities like building, you know, profiles of, I I won't even, like, say stuff, but, like, you can do really bad stuff with this at scale at a level of sophistication that you just couldn't do a year or two years ago. Like, two years ago, you'd had to be relatively technical. You might need to do, you know, how do web scraping and all these kind of things, write Python script. But now, like, someone with no technical skill has so much power for good or bad. It's a tool, that they just didn't have a year ago or even two years ago. Exactly. Well, thank you so much for your time. Thanks for sticking around to answer some questions. I found that really helpful. You know, moving forward, if there's one major takeaway from today's conversation, it's obvious that agentic traffic should no longer be simply viewed as another category of bot traffic. And, you. know, as AI agents are increasingly acting on behalf of legitimate users and businesses need the ability to really understand those interactions before deciding how to respond. You know, we're speaking to what Julian is talking about, visibility. You need to know when an agent is interacting with your business, understand what it's trying to accomplish, and then apply the right policies based on identity, intent, and value. So on HUMAN side, HUMAN is already identifying emerging AI agents, including muse across billions of digital interactions and helping organizations make more informed decisions about trusted traffic. So, you know, just to close this out, make sure you're all checking out the resources tab, which includes a free trial of agentic visibility, numerous blogs like Julian mentioned from our subject matter experts, as well as sign ups for our next virtual events. Thank you again for joining, and we'll see you at our next HUMAN dialogue. Thank you, everyone.