Katrin (00:00)
Welcome to Knowledge Distillation, where we explore the rise of the AI analyst. I'm your host, Katrin Ribant, CEO and founder of Ask-Y. This is episode 16, and we continue in our exploration of how agent e-commerce is changing the online marketer's world. Today, I have a guest who spent their formative years at the heart of e-commerce, and by that, I mean Amazon.
before founding a company that helps brands curate the best e-commerce experience based on cutting-edge experience optimization. We'll go into that in a lot of detail. So I'm not going to say more right now. And I'm actually not going to say more because I would just not make it justice. So Gal, could you please introduce yourself?
Gal Rapoport (00:39)
So first off, thank you for having me, Catherine. yeah, so I always joke my background in AI is way before it was cool.
I was in one of the very first deep learning research groups that some famous names came out of them. Then I started working for Amazon. My first project in my career was actually being part of developing the AI chip of Amazon, what has become AWS Inferentia and AWS Trinium. ⁓
Katrin (01:09)
That's very cool. That's some street crits
right there.
Gal Rapoport (01:13)
Nobody expected it to be that significant. My friends are still there. It's going bananas. then from there,
Mostly because of family reasons I I decided not to relocate with the rest of the team and I I moved to work on the Alexia shopping team I was the first machine learning person in the team that was focusing on personalization ⁓ My specific team was a group of applied scientists and data engineers and machine learning engineers everybody did
Katrin (01:46)
And
so just one thing, Alexa Machine Learning Ecommerce, when was that when you were on that team?
Gal Rapoport (01:55)
So I started on 2018-2019 when it was...
Katrin (02:00)
prehistory.
Gal Rapoport (02:01)
Yes, prehistory, to be honest, we did work on transformers at that time. It was already the very beginning of it. Ilya and the guys started it in 2017 with Attention is All You Need. And we already examined it, but we never released it outside. Similar to the stories you guys know about Google, now it's not a secret. Those technologies were baked internally, but were never released. But our team was working on the brain.
Basically Alexa, unlike ⁓ how many people think, is not the device, it's the brain behind the device that is helping you to understand who are the users, how to react to them, and to treat them with the best customer experience. And that's what we were doing at the time. And unlike many ways people thought, we never thought about the experience. ⁓
And I think about Alexa as the, I would say the predecessors or like the previous generation of ⁓ what we see right now in a way, much ⁓ uglier, a much not sophisticated, a much error prone type of ⁓ mechanism compared to what we do today.
Our goal was to create a relevant experience for the audiences and to create the best customer experience that we can get. We never tried to optimize for the people to buy more. It was always about creating a personalized touch with the thinking that this is going to pay off long term.
For me, when I left the company, after a couple of years leading this team, I started to, ⁓ I went back to the academy, I wanted to complete my PhD in multimodal AI, and I started to ⁓ look into new forms of data modalities, or new types of data, and one that was really interesting for me was the world of human computer interaction.
which is a niche in the academy to be honest. while doing this, I also looked into privacy and security and ethics with BI because I realized it is going to be a very big issue. The fundamentals of a lot of things that we know on the internet is going to disappear and we're gonna have more and more.
data out there, but we won't be able to use it. We will be restricted. The context on the users is going to be something that is going to ⁓ be ⁓ not so available. ⁓ And on the other side,
⁓ that solutions that exist would require more and more and more context of users. So what we're gonna see is a trend where the demand peaks, but the availability or the supply drops. So there will be some form of an intelligence gap that we realized we wanna ⁓ sort of approach. And while doing this, I was consulting at that time to one of the biggest personalization companies in the world and I realized,
that everything that I knew about personalization doesn't exist outside. mean, in Amazon, everybody was logged in. We had a lot of history on those users. Outside, we don't have any of that. Most of the visitors are anonymous. And now, with agents, it's the same exact thing. These are actors on behalf of humans that are completely anonymous. And we want to understand their intent. And when I say intent, I don't mean if they're going to buy or not. It's a whole broad set of things that we want to know about those audiences.
I realized, know, ⁓ while today the way we captured information was clicks. Can we understand what the user clicked on, which is like snapshots on reality? We need to look at it completely different. We need to look at it like us when we speak. When you look at me, you hear me. You listen to me, meaning you're translating what you hear into some context.
you see me and what you see as a multiple meanings. It's like about the lighting, but it's mostly also about my body language. And as we know, our body language says so much about us and this type of modality did not exist so far on the digital side. And I wanted to see if we can bring it in because my idea started as, know, when you go today to whether it's a shop or it's the bank.
or it's an agency or wherever you go. And you know, you're coming in for the very first time. The person in front of you, they try to serve you. He has no understanding who you are, nothing about your history. But if he's good at his work, he will look at your body language and he will use his intuition to assume something about you. And my thinking was, can we replicate it in the digital world? And as we...
realized nobody was doing something like that. So the first two years of of Kahoona when we started it was let's build it. So we started creating a lab when a lab was not a cool thing to do. So at the very beginning, and once we discovered it works, we started to release a product and started to work with some of the largest brands today in the world. One that I can mention is LVMH group where we won the LVMH
award for best business impact.
Katrin (07:35)
I think that might literally be the largest brand in the world.
Gal Rapoport (07:39)
Yes, it's the largest conglomerate in the world, in the luxury world especially. It's owned by Bernard Arnaud, which is pretty famous. So many people probably in the audience and they own Louis Vuitton, Dior, Tiffany, Bulgari, many other, Loro Piana and many other brands that you guys ⁓ know. And we basically started working with them on ⁓ personalizing the experience. ⁓
Omnichannel, meaning cross the entire digital umbrella where it's their own site experience analyzing them. The marketing side, the CRM side, the online to offline side, everything together. And for generating the ISLY, we won this award. And just lately, ⁓ we had... ⁓
similar prize with the Caring Group, which is the second largest in the world. So I hope that we're aiming to the right direction.
Katrin (08:39)
That's amazing. We're very happy to have you. So when
you talk about the body language, the digital body language, ⁓ which is basically, if I understand well, the idea that the way it's not just the fact that someone clicks or scrolls or whatever, it's the way somebody clicks, scrolls, hovers or interacts with the pages, that that has a meaning that can be decoded. Can you explain how that actually works?
We have a relatively technical audience, so go ahead, like all the details of how this works.
Gal Rapoport (09:15)
Yeah, so think that, think of your journey when you go, let's say, to a deal, okay? And you wanna buy, let's say, a shirt or a bag. You might look at some stuff, might hover on some stuff, you want to check some stuff. If you would be looking only on clicks, it's like looking on the snapshots of the reality.
And there are so many things that are happening in between ⁓ when you think about it from, okay, think of the equivalent to the real world. Like let's say I was in a shop, okay, and would probably go into a specific ⁓ shelf, I would look at something.
I might stay longer and then I would move fast to another direction. All those small nuances mean so much and we try to replicate them in the digital world. So either it's through the track pad or through the mouse. We try to make it privacy by design. This was our goal from the very beginning when we started the company. We brought privacy experts as the very first people to join the team. People that were part of the GDPR and CCPA
committees and ⁓ the idea was ⁓ there were a lot of solutions out there that were ⁓ used to collect a lot of external sources of data, to unify them together and then to segment based on that, to understand the users based on them and use them.
doesn't exist anymore in most cases. It's actually only for users that log in, which we talk about 2 % of the users, 3 % of the users, and usually in the very late stage of the funnel.
Katrin (11:04)
So most of the, me and I think most of the people in the audience ⁓ cannot really understand something if we don't know how we would implement it. So how does it, how do you capture these scrolls and these nuances? How does that work to capture this? And then how do you derive ⁓ meaning out of it?
How do you know what the slow mean? How does fast mean? How does all of that algorithmic aspect work?
Gal Rapoport (11:39)
So first off, we do is, and I always like to think about it from an analogy to the real world, we put the script on the website. We're collecting the information, and then we're trying to train models based on indicators that exist whether on the website itself or external to the website, but based on first-party indicators. Now, think about it, how we as people develop intuition, okay?
Let's say, let's think of a seller in the shop. The very first day, I assume the seller would not have any type of intuition. Okay. Or the, it's not going to be good. Um, but over time, as he observes more and more and more scenarios, more people develop some pattern recognition on this set of actions that just happened.
might be predictive or indicate that in the future that person is X or is Y. We try to create the same mechanism. Basically we realize that ⁓ a lot of the models that we are using today, ⁓ whether it's small language models or other discriminative models, ⁓ they're really, really good at pattern recognition. They all...
bottleneck was the information that you're creating. If you rely only on the standard click stream like everybody else, it's like garbage in, garbage out. So what we start to do is to start breaking it into the most granular way possible that we can as the data. And then we start analyzing the features from physiological aspect, physics of motion, ⁓ sometimes, ⁓
even psychological features. It's the whole broad set of aspects that we looked at, obviously within the context of that specific website and its specific time. And based on that, we're able to close the loop.
So the model is developing the intuition and generates the understanding on those users within moments. So the outcome is you put a script on the website, it rains after just three or four weeks, you get a set of models that basically take a random user to just come to the specific website or any digital assets. And within just a few moments can say, Hey, this person is likely to buy or not. This is the right pricing point for them.
Is he going to buy for himself or he buys a gift? ⁓ Maybe he's interested in specific type of products. If it's a B2B case, this person is from some specific profession. Is it the omnichannel buyer? Is he going to start online and go offline? Is this person, it's a bank, a ⁓ certain set of products ⁓ better to recommend him? ⁓
for travel and hospitality. It might be about the class of the product or should we show ⁓ more affordable options for them? Is it a business travel or leisure? It can be all sorts of things. And this is what intent means for us. Intent means to create the highest amount of relevancy for generating the highest amount of context on those users. We realize that if we're gonna build the context layer for the web that was never built.
the AI agents that we're adding with ⁓ are going to have the right amount of information, not data, information that they need to do the work perfectly.
Katrin (15:30)
Great, so thank you. if I understand well, as a client of Kahoona, I would basically implement a script on my website. That script would ⁓ capture all of those gestures that the user is doing on my website, the speed at which they scroll down the page, their hover over things, the time that they spend on certain contents, where their mouse is, et cetera.
And that technology essentially was designed to read human behavior, right? But AI shopping agents don't hover, scroll, browse, or any of that. They execute sort of with surgical precision. So how does that change the paradigm for your client and your technology potentially?
Gal Rapoport (16:16)
So maybe even before I'm answering this, a quick ⁓ interesting notion. I just had a dinner with one of the CEOs of the largest brands in the world a couple of days ago. And he was telling me something very interesting. We're in an era where we see less and less and less engagement from users on
website. It's dropping over time from the day the internet and commerce started till today. It just went lower and lower when people turned to mobile ⁓ because of the lack of availability because of privacy issues and now we're ⁓ turning into agents like you mentioned. ⁓
And the need to take those small clues and to predict based on them the context that is relevant about the intent is the same thing whether it's a human or it's an agent. And so...
When we started asking this, was participating ⁓ in LVMH CIO lab about a year ago. And I was asking them the same question. Are you guys looking at it? And without going into their specific details, without going into their stuff, ⁓ my thinking was we need to approach it differently. So from one side, we first want to know when someone comes into the website, or if it's an actor comes into the website, is it a malicious bot that we want to flag?
we wanna remove outside. Is it a human or is it an agent on behalf of a human? And then this was the first mechanism that we built and we built it obviously to remove the or to improve the signal to noise ratio with helping our clients to understand how people behave on the website. But the next thing we did with it was can we actually start predicting the intent of what
the human asked from the agent to do when they arrived to the website. And we were so surprised at the very beginning to see it actually works really, really well without technology. While if we're taking the traditional approach, like everybody else who is collecting data, it's not necessarily working well. And then I realized, you know, when people are ⁓ coming to a website, ⁓
Their mind is distracted. Their original goal might shift ⁓ to a different thing because their mind was distracted. And also, they're not going straight to point. They are very noisy. Agents, on the other hand, are very efficient. They're going straight to the point to exactly what they were instructed to do. And also, from
⁓ You know, what's their target or their goal for this specific session? It's very clear. They are not shifting their mind. And so if you think about it, how we measure the error rate in every AI model, it's always about variance and bias. So bias might be different, but the variance, because there is less of a noisy environment in the behavior ⁓ compared to a human.
⁓ the agent has much lower variance in its behavior. And so our models were much more accurate. And we sort of like proved it mathematically. And then we started to capture and realizing what does it mean. And once we got to this capability, we said, OK, what do we do with the effects? Because one, yes, you can impact what the agent is going to see on the website.
You can personalize for the agent. You can understand where the agents are going. So it's not about AEO or GEO, which are trying to understand how to put your brand ⁓ upfront when ⁓ someone is asking someone. It's also, it's already about someone considering interacting with the website, discovering the website, maybe also wanting to buy. The second thing is, you know, when I ask all the brands, what are you planning to do with this?
If you go today to any website, you will see that most of them are presumably blocking the agents. Why? Because they don't have as of today a certain policy on how to treat them. At the same time, if you're to ask them, they say it's a big fish, small fish kind of a problem.
similar to Google Shopping and to ⁓ Meta when they tried to do checkouts on their platforms, whether it's Instagram or Facebook, we're not going to allow it to happen. We want the checkout to happen on our website. But if we could find a way to allow users to explore, discover.
and consider on the website in a safe way for us, we would want potentially to make it happen. And so we started to build certain policies based on the intent level or the intent of the user. So both the CIO and the CDO of every brand would allow this new existence ⁓ not just to exist, but to thrive, ⁓ to make sure their business are sustainable and are growing sustainably.
Katrin (21:22)
and we'll you in the next one.
That's really fascinating. I never thought about it from that perspective that the agent's intent is actually less noisy than the human's intent. Because yeah, you're right, we are noisy, right? We are distracted, we don't read, we think one thing, the other thing, whatever, we may see. Whereas the agent has a mission and fulfills the mission.
Gal Rapoport (21:46)
Best of luck to you all.
Katrin (21:55)
And so ultimately it might actually be easier to derive the mission from the behavior in the case of an agent. How did you calibrate that? Really interested to understand like algorithmically, how did you calibrate that you actually did ⁓ derive the right mission?
Gal Rapoport (22:16)
You mean like how do we calibrate it for humans versus to humans that
Katrin (22:20)
Yeah,
in the case of agents, how did you actually manage to calibrate that you were correct in your prediction of what the mission was?
Gal Rapoport (22:30)
So first off, ⁓ unlike humans where you need to have specific indicators, what you can do with agents, you can simulate an environment. And simulating an environment is the perfect setup for agents. Because agents, unlike generative AI that we've seen over the past couple of years, is working on reinforcement learning in its core.
There is some generative component to it as time goes by. We have developed those systems, but agents started with what we, some of us saw with DeepMind in the very ⁓ beginning of the century with ⁓ solving chess and then solving Go and so many, many other stuff. You should think of reinforcement learning as a game, and we try to optimize our score ⁓ in this game.
The way we do this is we create a simulation environment in which we say, okay, let's try to think of dozens of scenarios that user is going to do something. Then let's create dozens of types of, ⁓ you know, LLM or agents.
⁓ agent types that can come to the website. And for each one of those, let's create multiple variations. So let's take a practical example. Let's continue with Dior. Let's say I wanna say, I wanna buy a shirt in Dior. Please recommend me three shirts. We can write it in a certain way to Chagy PD or to Perplexity or to Claude or whatnot. But we can write it in 20 different other ways. And...
if we create those different variations which
semantically change slightly the intent level or the instruction that you're giving to the agent. It translated into or encoded into a certain point in a semantic ⁓ latent space. What it does, it instructs the agent then to do a certain mission. And we want to create some noise. We want to add some noise intentionally into it would become robust. And it allows us to ⁓
to observe ⁓ the accuracy of what it was supposed to do. So there is a way for us to tie back and say, this specific agent came with this specific prompt.
we derive from this prompt that we just generated, where did it come from, location, type of an agent, what was the initial intent, and if there are specific flavors to it, and then we can replicate and see what were the results slash the outcomes. And so there's a perfect environment or closed loop environment to start simulating those things and start building. ⁓
Katrin (25:22)
That's really fascinating.
And could you give me an example of like two different missions for two different agents and what would be different, two different patterns of behavior of the agents where you could point at, well, this pattern basically ⁓ derived from this intent and this other pattern derives from this other intent.
Gal Rapoport (25:45)
⁓ So I think the easiest one to think of, let's continue with the same example I just gave before. I'm going to do your, I want to buy a shirt. Please recommend me three ⁓ options. This is exploration or discovery, if you will. And what you will see is that usually if the agent has no context on that specific...
user or person that just sent that agent to do the mission, ⁓ I would say there's some level of ambiguity. And so ⁓ you would see more of a noisy environment. While if you feed the agent with some context, if you could rewrite the text with filling it some context behind the scene, which is what the mechanism with Kona ⁓ allows, ⁓ you see much ⁓ less of a...
a noisy pattern, like we really measure it by a radius size of like the noise that it creates and you can see smaller radius for this. So this is for exploration. A second one is a different intent for consideration. Let's say I now made up my mind that I want one of those three different shirts, okay, because I discovered let's say a 10 of them. And I gotta ask, if you could do or Claude.
I have those three shirts, help me decide between those following ⁓ shirts what would be best for me. And what you will see in this type of prompt, the agent is going to start looking into the product details, will start looking into ⁓ the difference one, it would go back and forth. It's very interesting, you will get to see something very similar to a human behavior in a way.
Katrin (27:35)
just more precise and faster. So let's shift gears a little bit into the brand sort of realm. You work with the biggest luxury brands, probably some of the brands that are the most concerned about ⁓ their brand story and their relationship with the clients. And all of that is very, very important, obviously, in luxury. ⁓
Gal Rapoport (27:37)
Faster? Yeah, let's mind drift.
Katrin (28:03)
spilled on the idea of curation and emotionally resonant experiences, all of this good stuff. So ⁓ if the agents are doing the shopping, does the entire notion of a personalized digital experience become irrelevant? most important, it's more about how do your, really what I want to know is how do your clients currently think about that, if you can share?
How worried are they about all those aspects of the brand story, the relationship, the curation with the client, the loyalty, like all of those aspects?
Gal Rapoport (28:45)
So there are multiple answers to it. And I'll say first it goes by. Yeah. ⁓ I think it starts down with what's the company or what's the brand that we're talking about in luxury. Some of them think right now and they're debating, is it a luxury experience to go full chat to PD?
Katrin (28:51)
then it was a very good question.
Gal Rapoport (29:11)
or others, yes or no? Should we allow it, yes or no? And how we would allow it? I think that's the very fundamental question that they're asking themselves right now. And then if the answer is yes, the next question would be how? What we will allow? Now we would moderate our content because for the luxury world, the brand is the single most important asset they have. Then if you go to
You know, to the other, I would say more affordable, ⁓ brands or in fashion, or if you go outside of fashion, if you go to, beauty or pharmacy or home decor or appliances and electronics and whatnot, we see different types of, ⁓ approaches there. ⁓ what is common to everybody is that it's a big fish, small fish kind of a problem. Meaning while.
Those that are struggling or those that need to have GPT, Claude and others to bring more audience so they can survive financially are willing to open the doors to GPT to take control. Others on the other end say, we will negotiate until what point we're sharing the data. And if you ask me, it's the same exact thing that we saw with Google in the past. We saw the same thing with Meta in the past.
with other channels, we saw the same exact thing. So the old notion that everybody is agreeing on, ⁓ you know, let's think of those companies as businesses for a second. ⁓ Many of them had a physical location at the very beginning. And then many years ago, all of a sudden internet comes in.
And then all of a sudden, they say, ⁓ cool, we might need to have ⁓ our own website at some point. ⁓ And website at the beginning meant almost nothing, right? It was just like having a business card in a way. And then ⁓ search comes in, and they understand there's starting to be something called an intent economy. And then commerce comes in. Amazon introduced it, obviously, but then everybody else followed after.
And then we started to see social, the discovery economy. So we see multiple eras within, I would say, or eras of the internet. And each one of them sort of introduced another door, sort of speak, to this physical store of that business that we discussed before.
You should think, and many of those businesses think of the website as another physical store, a big one, one that drives a lot of the attention, one that drives a lot of the potential customer, the prospects, the actual buyers. As time goes by, it drives a lot of the revenue, but it's sort of like...
They're transitioning ⁓ into a new ecosystem. And as of now, some of them see the agents as a complete game changer, that it's going to change completely everything. Some of them are saying it's going to serve some of the people, some of the time, and some of the scenarios. And I'll give you an example. ⁓ Would you agree to buy?
⁓ shirt through a chat GPT agent yourself.
Katrin (32:43)
Me personally, no problem. Yeah, absolutely.
Gal Rapoport (32:47)
I'm not sure for myself. But if it's like a standard shirt that I buy every day, like you know, a black shirt just because I need 10 black shirts, I don't mind. Then goes with socks or whatever, okay? So it goes down to the specifics and how I would say deep you want to go into analyzing what you want to buy. It's a trust factor.
I'll give you another notion that people didn't realize and we saw it at the time in Alexa. It was the same thing. Would you be open ⁓ that an agent would describe to you 10 different items like when you go to a category page on a commerce website today? Like describe them one by one by one by one. Right? So there is a lot of stuff. ⁓
Katrin (33:36)
No, that sounds boring.
Gal Rapoport (33:42)
⁓ in the i would say
the human interface that we're to have that is likely to change and it's going to apply a lot of, I would say, different outcomes on the way we're going to design those systems and we're going to allow them to work and we're going to ⁓ see different types of problems that we haven't encountered in the past. I would say that the best example that we saw was Alexa, where I was working for Alexa Shopping because it was one of the only, I would say, former versions of agents that allows you to
to buy something through them. That was the whole notion of it. ⁓ you know, it mostly worked if you look at it, and Amazon is a public company, so you can speak about it very openly, it mostly worked for groceries, because we realized that in groceries people are okay with this. In fashion, in beauty, appliances, in home decor, and buying an insurance, it's still a trust factor.
Katrin (34:17)
Yes.
Gal Rapoport (34:44)
And then trust, it goes ⁓ by the person. So some people are more techie, they trust the tech and they are good with it. Some of them are more conservative. But it's clear that we're seeing here a trend and like every other ⁓ say adoption curve, we're right now seeing the innovators or maybe ⁓ the early adopters. don't even think, I think we're actually ⁓ before the big peak of the...
Katrin (35:12)
Yeah,
I'm sure we are. I think we are very, very, very early stages. But do you have a sense from the data that you're exposed to through technology of how fast growing that curve of adoption is?
Gal Rapoport (35:25)
So what we have seen so far, there are two very interesting, I would say, topics. ⁓
most of the, I would speak about the brands that we work with, the bigger ones and the one that we're ⁓ on boarding with our prospects. I'm talking about bigger brands ⁓ and bigger hospitality and travel companies and banks and insurance. ⁓
and telecom, I think what we have seen is that everybody is thinking about it and designing it right now and they're in an experimental type of phase. But everybody understands that if they're not going to cope with this change, they are potentially going to lose. So there is a mix of fear.
and experimentation, it still feels to me that they're very early. They realize that their infrastructure is not well set for the agentic world. It's good for us. Yes.
Katrin (36:28)
Can we talk about that a little bit? Can we
talk about that a little bit? what do you see in terms of the infrastructure is not well set and what would need to be done?
Gal Rapoport (36:41)
So one is ⁓ obviously I'm a little bit biased here, but they don't have. ⁓
Katrin (36:47)
That's the
reason why we have you here, it's because you are biased. So go for it.
Gal Rapoport (36:50)
⁓ So what
is, as they understand they don't have ⁓ information or the right amount of information to fit in through those agents to do their job?
Katrin (37:02)
But it's obviously
also really not easy to get that information these days. Like you don't get that from your GA4 data.
Gal Rapoport (37:06)
That's
That's why we started
the company for that. Like this is one of the key reasons we realized where the market is heading with this. realized the context layer and the memory layer are going to be fundamental as the engine getting better if you don't have the fuel.
to run it, it's not gonna work. You can have a Ferrari, but you won't be able to drive the Ferrari. You can buy a Fiat and you can buy a Ferrari, doesn't matter, it will do the same. The second thing ⁓ about, other than having context and context in ⁓ that regard,
It's not just about ⁓ where am I. It's more about the persona, the person who I am, what is my current intent at the moment, not what I know about the person from the past, but what's intent at the moment because intent is very timely. The second aspect is security, privacy. ⁓
those two and ethics. think ⁓ we see and we speak a lot with CIOs. are now informing, you know, they're facilitating protocols. They are trying to realize where they should prioritize their focus because
I mean, potentially there is so many things that should be solved, they still need to see where the world is heading because it's still like a mystery to them as well. they're trying to say, okay, what are those first principles?
Katrin (38:38)
everyone.
Gal Rapoport (38:44)
that we are already certain about, we should already approach and we should already tackle. Then it's also about analyzing those agents, analyzing how the ecosystem from like their internal ecosystem, the way their company is working, the way they harness tool or what tools are they using, should completely, completely change. They say,
that those tools ⁓ has, ⁓ like I would say the previous generation tools were built under some first principles that become ⁓ somewhat irrelevant to the new world that we're heading into. And they're questioning themselves ⁓ if they should ⁓ go one way or another.
And it's very interesting to see it's like those places where they're certain about ⁓ they're willing to, ⁓ you know, to start a partnership and they can commit for a long-term partnership while in other places they like, we want a very short partnership ⁓ or a short period of time so we can opt out. Luckily we're in this, I would say safe.
Katrin (40:02)
Yes, that
makes sense, right? People are going to dip their toes and measure the uncertainty, but from a measurement standpoint of you, specifically, ⁓ for people who are currently set up with traditional tools, with GA4, with Adobe, with just basic traditional measurement tools, what would you say they should be looking at in their data to have a
notion of how this phenomenon exists for them or not, how it's growing and what proportion of the traffic is agentic. ⁓ Maybe you can give them tips and tricks to derive intent from their own agentic data in GA4 although that's probably not really that easy. ⁓ But yes, so what could you tell our audience about with the tools of today, with what you have, what can you actually do?
Gal Rapoport (40:59)
So I would look into two trends if they exist. One, ⁓ it's very easy to look for a declaration.
⁓ of specific agents. when a Chattipiti agent and a Tropic agent, come to the website, there is some footprint that you can track and you can recognize it in Chattipiti very easily. And I will say the very first thing that you want to see if there is a specific trend of it increasing over time and by specific geographies, okay? Because you will see it's very, very different by different countries. ⁓
It's very interesting to see the maturity level.
So far it's completely completely different by the region. Like when we work with some of the global brands, you would see in the US one trend, ⁓ in France, for example, another trend, in Japan and Korea another trend. So that's one. And the second thing that I would want to look into ⁓ and to check if there is a correlation or even an inverse correlation between the two ⁓ is the amount of time people are spending on the website.
it dropped because one thing that everybody's seen is that agents are doing a lot of the exploration side for us. So when we come to the, to the website, eventually as humans to complete the purchase, ⁓ the time we spend on the website is shorter. It's usually just to validate the work of the agent. And, we're already coming with a higher level of intent. The conversion rates of those audiences tend to be much,
⁓ because of that. So we do see that ⁓ those agents are bringing back the same humans ⁓ much more baked if you will. ⁓ They already...
Katrin (42:54)
Yes, I
love that. That's amazing because like, know, time on site, time spent is one of the metrics that everybody laughs about because it's one of those vanity metrics that really nobody cares. But now we found, finally we found a good use to it.
Gal Rapoport (43:11)
Yeah,
and I will say the next thing once you see this trend try to look at it by specific categories on the on the website itself because sometimes you will see that well ⁓ Your homepage might be good enough For the agent equal there are specific categories that are not well built and maybe just to go back to your initial point ⁓ my thesis is that
Some users or some people are still gonna buy offline in the physical stores because that's how they are going to build for some of the products and some of the time. Some people are going to continue with the traditional e-commerce website because they wanna feel, they wanna experience it before.
decide to buy. They see shopping as an experience, okay, or whether it's like choosing their next ⁓ vacation, whether it's a hotel, where to apply and everything. ⁓
And there will be a new door into the store, which is going to be the agents. And this door is similar to e-commerce is going to grow over time in terms of ⁓ revenue volumes, traffic volumes, and everything that we can see. And so you need to pay close attention into it as if it's another subset of audiences or humans or actors. That's why we call it, we call it actors from now. said there was a.
or there is a different actor which is an agent on behalf of a human. And so when we build our own models now we treat them completely differently, separately, because there are specific nuances that are only ⁓ relevant for...
And the second thing I will say, agents are not built to consume context and information the same way as humans. They love markdown. Those agents love text. They do have the capability to ⁓ somewhat understand visuals through taking screenshots and trying to analyze them, but...
You know coming from my background I was I think I did research in all of the domains of AI. It's called saliency areas of entrance in in the image that Those agents are not so clear about when they are instructed with the text at these days, so
you will see that, and this is my bet, that we will see a new form of website that is going to be specifically tailored for agents that is going to be prevalent as a... ⁓
the whole ecosystem will mature. Meaning you will see some mechanism when you go into a website that would technically ask, is it a human or it's an agent on behalf of a human? And if it's an agent on behalf of a human, actor, this actor would be redirected to a different version of that website, which is more customized.
Katrin (46:32)
think
that makes a lot of sense because trying to optimize for both is really problematic.
Gal Rapoport (46:38)
Yeah, yeah, 100%. It's like separate behaviors, and we already see it. And it's going to be a very interesting ecosystem, and that's why we're...
helping now, one, the brands to think about it, and second, to develop those policies and how they can enforce them from an experience perspective, from an information perspective, and also soon to think about it from, I would say, ethics. It might be also going all the way there, but for sure about privacy and security and what data to share.
Katrin (47:11)
Yeah. Well, great. Thank you very much, Gail. This was really amazing. I certainly learned a lot. ⁓ So for people who want to talk to you, where do they find you?
Gal Rapoport (47:21)
So ⁓ I'm trying to be available basically in any channel. So feel free to approach me over LinkedIn, over email. It's gal, G-A-L-R-A-P-O at kahoona.io. ⁓
I'm more than open to have those deep conversation and to speak more about what we do. And I really encourage you to take a deeper look into what Kauna is building. We're basically reinventing. ⁓
the way the web is working. We're creating the context layer that does not exist so far, context and memory layer. And ⁓ yes, we're working today with some of the largest brands in the world. So I'm happy without going into their specific details to keep it very anonymous, to keep their safety first, ⁓ to share some best practices. And ⁓ we'll take it from here.
Katrin (48:21)
Great.
Thank you. We'll put all of that in the show notes so people can find you. Well, thank you It was really amazing. And that's it for episode 16 of Knowledge Distillation. If today's conversation made you want to experiment with AI for Analytics, visit us at ask-y.ai and try Prism. Thanks for listening. And remember, bots won't win. AI analysts will.
