InsideAnalysis · 2026-07-01 · 51 min
Key moments - from our scoring
Substance score
47 / 100
Five dimensions, 20 points each
Alberto Sangiovanni Vincentelli, co-founder of Synopsys and Cadence and faculty at Berkeley, explores the practical limitations and strategic applications of large language models in design and commerce. He emphasizes that LLMs are fundamentally stochastic models prone to hallucinations and incorrect predictions about the future - he jokes that ChatGPT even fabricated a Nobel Prize for him. Sangiovanni advocates for a hybrid approach: use machine learning to explore design spaces and suggest novel solutions, but always validate findings against established physics and domain knowledge rather than trusting LLMs as systems of record. He also challenges the field's reliance on brute-force neural networks, questioning why more elegant mathematical functions beyond ReLU haven't been explored, and notes that even LLM creators (like Dario Amodei) admit we understand only 3% of why these models work. Mato Zambon then addresses invisible AI agent traffic, explaining how CDNs like Cloudflare and tools like Microsoft Clarity can detect non-human bot behavior by analyzing server logs, user engagement patterns, and whitelisted agent registries. For enterprises relying on analytics and conversion tracking, the conversation reveals critical blind spots: bots are indistinguishable without multi-layer filtering (server-side logs, client-side engagement scoring, machine learning analysis), and LLMs are reshaping web traffic patterns entirely.
LLMs are stochastic (probabilistic) models trained on historical data, so they cannot predict the future accurately and will fabricate plausible-sounding content when extrapolating beyond their training. This is intrinsic to how they work, not a fixable bug.
Use ML to explore design spaces and discover novel solutions you might not have considered, but always verify findings against known physics and scientific principles rather than trusting the model output as ground truth.
LLMs are exploratory tools for discovery and generating ideas; systems of record like ERP and supply chain management systems contain verified, trusted data and should be the source of truth for business operations.
Use multi-layer analysis: examine CDN server logs (which capture all requests), apply client-side engagement scoring (timing, mouse movement, scrolling patterns), and run machine learning models to identify bot behavior, then block malicious traffic at the CDN level.
GPUs were originally optimized for graphics rendering, which uses the same core operation as neural networks - sum of products - making them naturally efficient for large-scale model computation.
Our reviewer’s read on each dimension, with quotes from the episode.
The first segment with the EDA pioneer contains genuinely interesting material - Vapnik's generalization theory as a counter-argument to LLM scale, and the physics+ML design workflow - but the CDN/bot segment is competent rather than dense, and the SEO segment is thin. Heavy padding from the host's self-promotion and platitudes drag the overall density down significantly.
there is a Russian mathematician called Vapnik, and uh, he um, proved under many conditions, but that the machine learning approaches or Models, uh, that uh, generalize the best...Actually the simplest models are the ones that have more power of uh, being right on a question that has not been asked before. And what we are doing is exactly the opposite of that.
Dario Moday a few, um, months ago, ended up saying we know only about 3% why this large language model work
The Vapnik framing - that brute-force LLM scaling is scientifically inverted relative to what generalization theory prescribes - is a genuinely non-obvious argument rarely surfaced in practitioner podcasts. The CDN bot-whitelisting mechanism is practically novel. Everything else, especially the SEO and hallucination material, is recycled consensus.
Is there another mathematical function that I can use besides the relu? Because if that is the case, then maybe that we can build models that are less uh, error prone, more elegant and consume less power.
by design Cloudflare um ask to own all uh products uh all vendors uh that create um, bot or agent to subscribe to a specific form
Alberto San Giovanni Vincentelli is an authentic heavyweight - co-founder of both Synopsys and Cadence, still on the Cadence board after 40 years, active Berkeley researcher - and his domain depth shows in the transcript. The other two guests are mid-level agency practitioners with relevant but unremarkable operational experience, pulling the average down considerably.
I'm still on the board after almost 40 years in which I started the company
we are looking at how to use...machine learning...for design since this is my forte. And uh, I have particular view about artificial intelligence machine learning
There are genuine specifics - the Vapnik reference, the Cloudflare whitelist mechanism, the Spotify ML bot-blocking example, the Amodei 3% quote, and the SEO task count growing from 15 to 42 - but these are interspersed with large stretches of vague generalization and no verifiable data, metrics, or timelines.
Dario Moday a few, um, months ago, ended up saying we know only about 3% why this large language model work
by design Cloudflare um ask to own all uh products uh all vendors uh that create um, bot or agent to subscribe to a specific form that in. And uh, with that subscriber uh they have a uh whitelist
The host repeatedly pivots away from guests to narrate his own views and plug his own conference speech, asks leading questions that answer themselves, and never meaningfully challenges any claim. The interviews function less as interrogations and more as validation sessions for the host's pre-formed opinions.
And you make a, you're making a very, very good point which I'll talk about today at my speech at uh, the uh, We Make Future conference
My theory is about 50, 50. I think half of that traffic is either bots or click farms or whatever.
Computed from the transcript - who did the talking, and the words that came up most.
Recorded live at We Make Future in Bologna, this special episode brings together conversations with industry leaders, innovators, and technology experts on the latest developments in AI, data, digital transformation, and the future of enterprise technology.
Transcribed and scored by The B2B Podcast Index.
Speaker A: The information economy has arrived. The world is teeming with innovation as new business models reinvent every industry. Inside Analysis is your source of information and insight about how to make the most of this exciting new era. Learn more at InsideAnalysis.com InsideAnalysis.com and now here's your host, Eric Cavanaugh.
Speaker B: All right ladies and gentlemen, hello and welcome back to the only coast to coast show all about the information economy. It's time for Inside Analysis here at We Make Future in Bologna. And I'm with a very, very special guest today. We have Alberto San Giovanni Vincentelli. He is quite the accomplished designer. He focuses on teaching and on design. First for microchips and then later for planes and air conditioners and all kinds of fun things. And he's uh, quite the data scientist. You might call him the original data scientist, the OG data scientist. He's been doing this stuff a long, long time and he's with us today. Alberto, thanks so much for your time today. What a crazy world we live in these days. Um, thank you for all your hard work. First of all, it's amazing to think how far things have come. And of course you help found Synopsys and Cadence, two uh, very successful companies. Now what are you working on now and uh, what do you think about these large language models shaking up the world of commerce and of human life on earth.
Speaker C: Yeah, so right now I'm working on a variety of topics. I really like to look into several uh, important problems. I enjoy researching, uh, new things. But uh, in particular now we are looking at how to use um, call it AI but really more uh, appropriately machine learning. Mhm. For design since this is my forte. And uh, I have particular view about artificial intelligence machine learning. Because as you, as everybody probably knows, uh, these models are statistic in nature.
Speaker B: Right.
Speaker C: And uh, so while the form of the language is very, very good, the content is not always right. And it cannot be. It cannot be. Especially if you ask to something that is going to predict the future. Then you know it has to do with what it knows, which is whatever has been done in the past, as we also do. And so the extrapolation is uh, iffy. So uh, there are people who believe that um, uh, this large language model, or in general AI cannot be wrong. But actually it is wrong.
Speaker D: That's right.
Speaker C: It cannot be otherwise, you know, poor baby, it cannot do what is impossible to do. Uh, but it's certainly a very powerful tool. Um, so we do know about hallucinations, right? So things uh, that are made up, but of course they Are made up because again the nature of these models is stochastic. And so if you extrapolate um, you have to take the consequence of this which is not always right. So it hallucinates. And so I love to see the progress of the large language model. Asking about myself. Mhm. First time this was uh, at the beginning of ChatGPT and Gemini and the like. I asked about myself and turns out that I won the Nobel award. Which is very nice, is the first case of, of a Nobel award given by a um, an AI tool. You can see also how these models work because it uh, it works on probability. So since I didn't win a particular award, the Frontiers of Knowledge award, and it's called Nobel Predictor because many people who won that award ended up winning uh, the Nobel award. And so I think that uh, uh, chatgpt say poor guy, I mean everybody wins it way so he must have won it. So here it goes. And uh, and also sometimes uh. Now the latest version of the, of the languages um, ended up uh, saying that I won an award that does not exist. I mean it just made up not only the fact that I won it, but this award at least amazing. It's very funny actually. And uh, so anyway, but this just to tell people watch out and to tell uh, people verify on something, you know, uh, what kind of mistake it can make. Now of course uh, the situation is improved dramatically with all kind of uh, guards. And uh. The models that have been used are uh, validated by the sequence of uh, information that they get from the Internet or from publications. So you can trace back where this information comes from and which gives some more credibility to what happens. Uh, but um. Uh. It's important that people realize that not all that is said is true. In addition they have to realize that by the way in which the answer are constructed. Uh, in the end if you keep on asking and uh, digging deeper and so on, what happens is that it will form an opinion, quote unquote of what you want to hear and then it will tell you what you want to hear. So.
Speaker B: Right.
Speaker C: Which is also very dangerous for young people. You may end up having serious situations in terms of depression even sometimes they claim that these models um, have an effect on uh, suicidal thoughts and all of that. But again it is intrinsic in the way in which they work. It's not that is a bad bad guys or bad tools or whatever. It's the way it works. Cannot be otherwise.
Speaker B: Right.
Speaker C: Because you build answers out of what you know has been asked or what you uh. Uh, as a asker are all about and so it will fit you. So anyway, but in design, the interesting thing is that what we're trying to do also with uh, this company, uh, which is Cadence, I'm still on the board after almost 40 years in which I started the company. But uh, uh, what is important to uh, to see is that I can use machine uh, learning and these models to explore the design space. So say what if I do this particular design and then you know, you don't take it for granted. So then you verify if this design is correct or not. But the interesting thing is that you may have not thought of that design.
Speaker B: Right.
Speaker C: And so the tool is actually going to be very useful to suggest you look into something. Mhm. But then don't trust it and just. Okay, wait a minute, let me verify this is good or not. And what uh, you do when you are in that situation is that you use uh, the um, standard. Standard models or what we know about the physics of the phenomenon or the physics of what you're trying to design and use that to validate. So then you eliminate, at least you try to eliminate uh, the possibility of making mistakes. So it's a combination, right, of what we know from a uh, scientific point of view and what we can do with this powerful tools, uh, which I think is the best way to, to go about it.
Speaker B: Yeah. And you make a, you're making a very, very good point which I'll talk about today at my speech at uh, the uh, We Make Future conference, which is that these models are excellent for discovery, for stochastic purposes, but they are not to be used as a system of record. And so you need your systems of record like your old systems, erp, Enterprise resource planning, supply, CH Management, for example. Those are your systems of record. That's the reality. And then you use these large language models to help explain things, to explore ideas. They're very good in that regard. Uh, but you should not use them as a system of record. Right. Use them only for exploratory purposes and to generate text that you will then vet. Is that or is that right?
Speaker C: Ah, absolutely. I mean this is my recommendation and that's just sure that many people agree with me. But that's the way I use it.
Speaker B: Yeah, well that's very, very smart. So what are you working on these days?
Speaker C: Yeah, so this is one. So how you use um, machine uh, learning in design, uh, and with this contrast of physics, what we know and what we try to explore and uh, you know, I always say don't throw away the baby with the dirty water. So I mean, if you know something, why do you give up on that knowledge?
Speaker B: Right.
Speaker C: And you try to rediscover something that we know. Right. Um, scientifically, no. So, uh, don't try to recover the Maxwell equation because I don't think that uh, any large language model would be able to recover it from scratch. But they are very useful to analyze and design things. Right. So we use that and keep that. Keep the models that we know are based on fundamental laws of physics.
Speaker B: Mhm.
Speaker C: Yeah, that's what I'm trying to do right now. And also I'm um, very curious about the existence of alternative model to the basic model, uh, that is used today, which is neural network. And in particular in large language models use feedforward neural networks, which are very crude approximation models. And we compensate the relatively, uh, low power of a single component with gazillion components. Right. So in some sense I don't like that as a concept to use brute force to solve a problem. So I, I'm always in love being Italian. I'm in love with elegance. And so I like to think that there is a more elegant way to go about uh, exploring solution space and all that. So, uh, we don't even know why these large language models work. So in fact there was, uh, Dario Moday a few, um, months ago, ended up saying we know only about 3% why this large language model work.
Speaker A: Wow.
Speaker C: And at least 97% we don't really know. Which is, in my opinion, is absolutely correct. There is no mathematical system that will uh, explain the behavior of this large language model. And actually if you look back, uh, in history, because again, uh, artificial intelligence is an old concept and I think goes back before the Second World War. And you know, the Turing Test for uh, machine intelligence, uh, is, uh, dates back uh, to the Second World War. So I mean, people have interrogated themselves about the potential power of computing machines and uh, and how you differentiate that with respect to human intelligence. Mhm. Uh, but they, Again, uh, the key point is that I want to understand why things work the way they work.
Speaker B: Right.
Speaker C: And if, uh, you have this very large model, you ask, uh, why are they accurate? Because they are. I mean, in several instances I, uh, would say for certain fields and I would say 80% of the times they are right. So, uh, where does this power come from? And if you look at, again looking back in time, there is a Russian mathematician called Vapnik, and uh, he um, proved under many conditions, but that the machine learning approaches or Models, uh, that uh, generalize the best, generalize the best means that are giving you a right answer to a question that the model has never seen. This is a generalization capability of a model. Actually the simplest models are the ones that have more power of uh, being right on a question that has not been asked before. And what we are doing is exactly the opposite of that. So it's very surprising. So we are going with numbers instead of structure. So my question is, suppose that I build up another approximation model instead of being the node, the neuron. Right. Let's call it neuron. But is a distant model of how our neurons work. But call it norm. Okay. Is there another mathematical function that I can use besides the relu? Reload is a typical function that people use as an approximation of what a neuron does. Because if that is the case, then maybe that we can build models that are less uh, error prone, more elegant and consume less power. They are more compact and all that. Uh, but we don't know yet. I mean it's uh, people have uh, gone down the path of minimum resistance. So this simple model and every time something is simple, people like it. So these simple models are powerful. So why try something else? But this is contrary to science. Right. So anyways, so that's uh, what I'm working on. I'm trying to work on. So these topics. One is to apply whatever we got with the large language models to explore the design space, M but combine it with physics. And the other one is to try to understand what is intrinsically the, the power of, of this model. Where does it come from? Mhm. What is the reason why this work the way they work?
Speaker B: That's amazing. So I think you're, what you're saying is that small language models will probably be more effective simpler models because we have these incredibly complex large language models that even their creators don't know how they work. Well, if you don't know how something works, it's hard to fix it, right?
Speaker C: Exactly. Well, you fix it by looking at the behavior, right? So you observe what this thing does without knowing what is inside. And when you uh, create a sort of guardrail around um, this to try to fix it that way. So you fix it from the outside. Well.
Speaker B: And uh, maybe one last point to dive into a bit. What really blew my mind when I was researching how they work is, is the fact that it is just tokens. And a token is not a word. I mean the word the could be a single token or cat or dog. It's small enough, but A longer word is a bunch of tokens. So not. It's not even presenting you a semantically correct sentence. It doesn't know semantics.
Speaker C: Exactly.
Speaker B: The probability of this letter being next to that letter, which is like.
Speaker C: Well, it's a syntactical approach. Right. To things, if you like. But, you know, this is a long debate that dates back to Noam Chomsky and other people in linguistics, because, uh, the Noam Chomsky camp says that there must be a way to explain how a language works, the origin of language and all that. So I can understand language by knowing the first principles.
Speaker B: Yes.
Speaker C: That has not yielded very, uh, good results in terms of use. It's very interesting results in terms of the general theory of languages, but.
Speaker B: Right.
Speaker C: It's not been really applied. And then there is the other camp, which is the stochastic approach to natural language understanding and manipulating. And that's what we use today. Uh, because we have the power to do it, because we have the chips. Right. Uh, and the chips, we have the GPUs. And so. And so it's so amazing. The GPUs are very good at the basic operation of a neural network model. So there is this link, right, between the model and the chip that execute the model.
Speaker B: That's great.
Speaker C: This is the strength of today's approach, is the fact that you have the capabilities of using this very large, complex things that. Because you have the chips that sustain that computation. But the chips themselves, now, of course, they have been optimized and changed somewhat, but they. Originally, the GPU were called graphics processing unit. And why graphics? Because they were used for rendering images.
Speaker B: Right.
Speaker C: And because the basic operation in both cases is sum of products.
Speaker B: Yeah.
Speaker C: And so they are very effective doing that. And so how lucky can you be?
Speaker B: That's funny. I. I often joke that we should thank all the gamers because they created a market for the GPUs, which then
Speaker C: opened the door to AI is amazing, I must say.
Speaker B: Well, ladies and gentlemen, look this gentleman up online. He's on LinkedIn, and, uh, he's at Berkeley as well. Alberto San Giovanni Vincentelli. We'll be right back. And you're listening to Inside Analysis.
Speaker A: Welcome back to Inside Analysis. Here's your host, Eric Kavanaugh.
Speaker D: All right, folks, welcome back to we make future Apollonia. And next up, I've got the smartest guy I've met today and in weeks, maybe even months. This guy. Guy gave a presentation today, Mato Zambon. He gave a presentation all about AI agents and how they're they're invisible and you have to watch out for them and you can't just rely on the old ways of seeing. I, you made such an excellent point about how a lot of the AI happens before JavaScript even launches. Right. So you have to be careful where you are pulling from and what you're pulling to even know what the heck is happening out there.
Speaker B: Right?
Speaker E: Yeah, absolutely. So thank you to having me. And yeah for me is the next challenge, uh how to track that type of uh traffic. So the invisible, the blindness tracking uh for the bot, for the AI agents because uh, there is no more uh JavaScript, uh uh execution. And so this is the big problem for the all uh analytics platform like uh, uh Google Analytics, Microsoft Cloud or Pixel Advertising. Pixel. So the big problem is that type of traffic is completely invisible for this and so we can uh, try to understand how to grab that information and how to identify if it's a good bot or not a good traffic.
Speaker D: So this is the challenge that's really crazy. And you know I, when I think about web analytics I've never really trusted a lot of the analytics that I've seen. And you look into the giants like Google and YouTube I was mentioning in my talk today, a lot of people think that YouTube is organic.
Speaker C: Yeah.
Speaker D: And there is an organic side of it for sure. But a lot of it is just pay to play. Like you pay money and they promote and those are the views you get. Like are they real people? Are they bots? Who are these people? My theory is about 50, 50. I think half of that traffic is either bots or click farms or whatever. But you are helping to solve that problem by really understanding how to set up the ingest, how to look. So maybe talk about that. What are you actually connecting to? Because you talked about using Cloudflare and some of these other tools, CDNs, content delivery networks. Right, CDNs. And your able to more better ascertain what's actually out there. Is it a bot? Is it a human, is it a good bot? Is it a bad bot? Talk about how you actually set up the ingest and then how do you persistently analyze the data?
Speaker E: There are different layers. The first one is of course the server side log. So the first touch point for all the requests in your website, uh and your E commerce is uh, in uh the, by the cdn. So the thing is uh, if you have Cloudflare as a CDN or akamai or whatever you have the log of all requests in your uh cdn. So the trip is uh, uh you can connect that logs with other tools like for example Microsoft Clarity. And uh, you have all the data, all of this information about who is uh, uh the, the bot or the agent that requests that information. That could be a specific HTML page, but it could be a specific single file. For example it's um JSON or image or XML file, uh that you have all your uh article in your xml. So there are some types of. And the thing is uh, you can uh, uh connect uh that platform with the cdn. It's just three steps by talking API. And the thing is uh, you have all knowledge and the question could be how Cloudflare can handle or can recognize the type of agent and the type of bot. It's because by design Cloudflare um ask to own all uh products uh all vendors uh that create um, bot or agent to subscribe to a specific form that in. And uh, with that subscriber uh they have a uh whitelist with the whole name. Uh this is a agent by OpenAI because they, and so they have a waitlist. If they receive the request and you are not in the whitelist you are a malicious bot, you are spam. So you, you can uh, uh get that, that request. You, you don't have the response. So this is the, the, the, the trips how cdn.
Speaker D: Uh and that's, that's really important for security but also for performance, for governance. Now are the, are you security suggesting that these big guys like OpenAI and Anthropic, they are cooperating?
Speaker A: Yeah.
Speaker D: Right. They're agreeing to work with this standard.
Speaker E: They need to insert in that uh, whitelist otherwise they cannot use uh, they can handle the request on the website.
Speaker D: Right, Right, right. Well so like you talked about the pixel too. Like that's been one of the longest, longest standing ways to measure traffic is if this pixel is called upon. Right. Look, you get an email, I hit enter, I open my email. There's all kind of stuff happening in the background these days for security. But like I've been doing tracked email marketing for 26 years I think and way back then it was great because everything was real. When you saw if it said click, there was a click. Now sometimes they'll say oh there were 7,000 clicks. I'm like are you sure? I don't think so. Yeah, I don't think so. So that's kind of what you're talking about, right? It's being able to discern between bot behavior and human behavior. And we want to know when the humans are clicking on something because then I know to call them and sell them my service.
Speaker F: Absolutely.
Speaker E: The other things is uh uh not just to analyze the server logs with CDN but put another layer to filter to analyze the data. Uh and the other thing is uh you can put a script that uh can analyze the engagement uh of uh the the click the scrolling and try to understand if there is uh uh too much uh the timing of the click is uh too lower because it's just 100 milliseconds.
Speaker D: Oh right right right right in the
Speaker E: add to cart on the other. And probably it's not a human and or probably there is a uh. I user agent that you describe. You you have a uh desktop but you don't have the. The mouse over. You have just a scroll and click and so you can handle some script to analyze with a uh scoring if the um. Yeah the right. The engagement of the in the website is really a human engagement or not. So the thing is ah you need to use a uh different uh layer to understand if uh it's about an agent uh with the server side uh filter with uh client size filter and then the next is analyze all the data uh with machine learning to predictive if something is not uh human and put that uh information directly in uh CDN or any other uh specific uh uh layer to block that uh that request. Spotify for example do this with the machine learning to analyze the data and they blocked some IP uh address or something like that because they understand that the traffic is not in the first. In the first watch in the first thing uh about but after when they analyze uh all data they understand that it's a typical bottle script. So they uh denied the future request. So it's a top and mouse uh play.
Speaker D: Right, right, right, right. It keeps going, going around and around. Well because people want to know who is on my site. People want to be able to market to those people. I mean so I talked in my speech today about how ChatGPT is very disruptive to traditional patterns on the web. People are now going to organic websites much less often than they used to uh which is a problem I mean especially if you're a media company. It's like well wait like the eyeballs are what allow us to charge advertisers money. So if you're losing that it's like whoa. So these large language models are like a neutron bomb that went off and fundamentally leveled the playing field and now everything is half as effective. So you have to send twice as many emails and there's the law of diminishing Returns.
Speaker E: Right.
Speaker D: So all of a sudden it's kind of teetering off. And the key is to be able to understand who's actually on my site now and then build in some actions to respond, respond to them, to give them offers or whatever. It's changing fast, right?
Speaker E: Absolutely. The market is moving to another uh, uh, channel, another uh, medium to response the answer. I mean uh, we have uh, the website, we have the app and now we have another uh, way to uh, correspond with the data. The LLM uh, can have uh, a uh, specific way. For example, uh, Google and Shopify create the UCP model, the UCP protocol. And if you go to the robot txt file for all instance in Shopify there is a specific pattern that you have a uh, specific file in, in the markdown. Uh, okay, that explains all the rules for the LLM. So you don't need to have just only HTML page website, but you need to create to build a specific MD M file for the LLM. So this is another way to grab the information. In my opinion, this could be a uh, really interesting approach because uh, now we have the SEO to optimize the website and the search engine. And uh, the next challenge is how to, to hack the MD file for LLM to, to trust. Because my, my products is the same products for my competitors. And where is the, the balance? Because the LLM doesn't understand uh, this is a uh, red button or the other neuromarketing stuff.
Speaker C: Right.
Speaker E: And adjusted rules just to read the rules. So if you change some rules or if you change this is the best. Uh, and uh, my competitor is not, not the best. Could be, could be a really interesting new, yeah, new, new, new approach and new services. So it changes.
Speaker D: Right, right, right, right. Well it's, you know, uh, I sometimes describe it. You may have seen these guys in the past, they stand on logs in water and they run like this and you have to like run fast to stay on the log as it's spinning. And that's kind of what we're all doing right now, trying to stay on top of this spinning log and fall into the water. And you got to get back up again. Like spinning paint fall into the water. It's like happening all the time and it's a big deal. And think of, I mean think about banks, think about financial transactions. You know, these folks need to keep that secure. And then, you know, look what happened with Mythos in the United States. I would argue that Dario got what he bargained for, cuz he said, oh it's so Dangerous. We can't even release it to the world. So they said fine, shut it down. Like what were, like what did you
Speaker B: think was going to happen?
Speaker D: Like you can't go tell the world it's going to ruin everything and then expect to not be shut down. Like don't do that. Personally, I think it's a little bit crazy. I mean if there are new vulnerabilities, we need to know what those vulnerabilities are and we need to build solutions for them. So I think, I think there is some genuine nature to his caution, but it's also marketing and it's, I think, I think. What do you think? Yeah.
Speaker E: We are in the world where the Ziggy some changing it means you have uh, opportunity. It's not just a problem, uh, because there is no more uh, advertising works, uh, works no better or other problems. You have to focus in the opportunity in this moment because uh, you can change something in the new business. You don't stay in the old business but create a new business with this opportunity.
Speaker D: Right, right, right. Well that, so I remember learning that in Chinese in, in Mandarin, the character for chaos is very similar to the character for opportunity. Right.
Speaker E: I, I, I use this, this uh,
Speaker D: absolutely right minds thing alike there. Yeah, it's very, very true. So there's a world of opportunity now, but there's also world of chaos and concern. And it seems to me that analytics is a key components to enable governance, to enable security. And the fact is we're going to need people. You cannot let these machines just run on their own. Uh, it's like self driving cars. My buddy, one of my favorite people in the world, this guy Mark Madsen, he jokes, he's like, did you notice that they always change the goal post, they always move the goal posts like oh, it's in five years. Oh it's in five more years. Oh, it's another five years. Like wait a minute, you said five years 15 years ago and like now it's 15 instead of five. So you know, self driving cars, it's not that hard to drive a car. Yeah, I mean I get that people want to uh, sit back and relax and all that. Maybe hold on to that steering wheel a little bit. And that's kind of what you're doing, right? You're holding onto the steering wheel.
Speaker E: We need to do this. If you want to do, to stay in the market, in the business, you have to dynamic mind to change and grab the opportunity. Right when the cows appear.
Speaker D: Yeah, that's exactly right. Yeah. Well who moved My cheese. Right. The cheese is moving all around, and we're like little mice trying to find our cheese now. But the good news is that if you understand the analytics stack and you work with a company like Tag, call this guy, he's good, you'll be able to stay on top of things and then better understand what's working and what's not working. That's the whole purpose of analytics, is to understand is this working or is it not working? Working. And if it's not working, what do I do to change, to make it work again?
Speaker B: Right?
Speaker F: Yeah, absolutely.
Speaker C: Yep.
Speaker D: All right, so one minute left. What's your advice to, uh, to the world out there about how they should approach this brave new world?
Speaker E: Be curious. Absolutely. Because, uh, the curious is the first move to try to understand the, the opportunity of the, the market and then the problems. So this is the, the first thing. And then the other thing is, uh, uh, don't be lazy. So please, uh, uh, invest to yourself to understand, uh, courses or whatever, to try to grab the knowledge and, uh, handle that opportunity.
Speaker D: That's right. No, that's good advice. And you know, I'm reminded of the importance of thinking and thinking through things. Don't just accept the answers. Listen to them, reason through them. Think about it. You know, think about, for example, with the cell phone, we forgot a lot of phone numbers because it's all in there. Right. But then you don't have your phone if you don't know someone's number.
Speaker E: Yeah.
Speaker D: So it's like it's important to persist certain memories in your brain and remember things and not just rely. Because when the Internet crashes or when open AI goes bankrupt, you know, bad things are going to happen. Well, folks, look this gentleman up online, Mato Zambon. We'll talk to you next time, folks. Take care. We'll be back.
Speaker A: Welcome back to Inside Analysis. Here's your host, Eric Kavanaugh.
Speaker D: All right, folks, back here at we make future in Bologna. Uh, Bologna. Fieri. Ah, Amortalia. Uh, and next up, we have Emmanuel Arosio from Data Rank. And, uh, you have talked all about search engine optimization, but now we're in a very different world with, uh, generative engine optimization or AI engine optimization. What have you learned? I'll give you my theory. I think this is very interesting stuff. In the old world of SEO, people tried to trick the system. So you come up with all ways to, to sort of leverage the algorithm and get it to come to your page. In the new world, I think it's going to be A little bit harder. I think that the LLMs first of all want to triangulate information. Let's start there. Is that your theory as well?
Speaker F: Thank you so much for having me. Yeah. We in our agency. You mean exactly. DataRank we just working on SEO now because the SEO strategy, uh, is correct in 2026.
Speaker C: Right.
Speaker F: There is not go without SEO. Right. Okay. Because in fact the basic of SEO is on top on strategy. When you approach the brand.
Speaker D: Uh-huh.
Speaker F: Uh, the brand of the website. The brand have the problem to have a good content. The good SEO tech tech SEO. The good semantic. The good. We just think about the link building on digital pr. This is one part just a little bit modify in 2026 then I mean 2022.
Speaker C: Right.
Speaker F: In 2022 you create a link building strategy because you wanted to ranking the brand for the generic keyword.
Speaker C: Uh-huh.
Speaker F: Today with a lms it's not a problem about their ranking for the keyword. But how many times the generic mention you.
Speaker D: Yeah, right. Okay.
Speaker F: This is the problem now.
Speaker D: Yeah.
Speaker E: Right.
Speaker F: Uh, the problem Every. Every market, every company, every corporate, they have this problem. We just research us to the generic prompt into jgpt.
Speaker D: Okay.
Speaker F: And we not appear. We are disappeared. We are not included into the list.
Speaker D: Right.
Speaker F: The situation because they don't have with SEO.
Speaker B: Well, uh.
Speaker D: Interesting.
Speaker F: Okay. We just starting like an order strategy by actual. Now the first approach we have is to analyze everything for the brand.
Speaker C: Uh-huh.
Speaker F: On the website. Okay. We change it. We make the SEO audit. We identify the problem. How many second your web pages is charged much about Google? Because in fact now if you. If you. If we take for example the AI overview the I mode in Google. Okay. It's a different way about Gemini Perplexity and chatgpt.
Speaker C: Right.
Speaker F: They use different kind of rms. But for repeat the SEO still now the high priority. The second one is change a little bit. It's like uh. We just calling GEO because from. From the states. The. The SEO groups like to use the acronym. But we talk about aio. We talk about to the optimization about the agentic. Not only for the new generative right engine. Uh, in 2026 to have our kind of job is more funny because every week change a little bit.
Speaker C: Right?
Speaker F: Yeah. Because if we come back a little bit like to 2021 we. We've been uh. One change of scenario about Google one time every three months.
Speaker D: Wow.
Speaker F: Okay. Now the scenario is more complicated because Google there is not the mandatory search engine.
Speaker D: Right.
Speaker F: Okay. Ah. They lose a lot of bit of research uh, the problem is uh, why Google uh introduced to the market the AI mode Because there is not the modify about only to research. It's like to have the systems like now.
Speaker C: Yeah.
Speaker F: Right. Many people just talk with the.
Speaker D: That's right.
Speaker F: It's like to. The problem is the fact is like to have a. A new virtual friend.
Speaker D: Uh-huh.
Speaker F: A new virtual assistant. You can request everything without limits, right?
Speaker D: No, that's exactly right. And people have different ways of interacting. What's your take on the importance of video for AI? Yeah, ah, awareness.
Speaker F: It's a good topic because uh, we use a lot of tool uh, from. From Google. We have a song free tool but uh, they don't provide to us the data we wanted to have. For example, uh, ah. The video like YouTube is in into the uh scenario about Google. Because YouTube is from Google and from some topic research. YouTube is more relevant than Reddit.
Speaker B: Yeah.
Speaker F: Okay. Because the people into the research. For example in long Reddit. Okay. Sometimes Reddit is not the priority source. Uh, YouTube is it?
Speaker D: Yeah.
Speaker F: Is it? And then we can analyze it now because we have a specific tool to analyze which kind of prompt.
Speaker B: Oh really?
Speaker F: The user coming from.
Speaker D: Really?
Speaker F: Yeah.
Speaker D: Interesting.
Speaker F: Yeah.
Speaker D: So it's important to know where they're coming from to understand what they're trying to find. That's very interesting and you make a really good point about Reddit. I think one reason why these LLMs like platforms like Reddit and YouTube is because there's so much organic content. Uh, and what they're doing is they're synthesizing large amounts of organic content instead of just in the old Google way, indexing and page ranking and then serving up the page. It's deeper but they want the. The richer texture of all that organic commentary.
Speaker B: Right?
Speaker F: Yeah. In fact in part is it. The other part we need to be. To be focused on the history because why Reddit is not important the situation. Why? Yeah, because Reddit is like to close like two years ago the first agreement with HPT to collect the data from.
Speaker D: Right.
Speaker F: Okay. This is why Red is.
Speaker D: Yeah. Right.
Speaker F: Relevant now.
Speaker D: Right.
Speaker F: And adopt from every kind of field you need to analyze. But there are not a lot of tools to analyze it.
Speaker D: Right.
Speaker F: And um. But the brand asked to us to analyze it why and when and when need to be present. Because it's a little bit complicated now for the brand to. To be in front of the audience.
Speaker C: Yeah.
Speaker F: Okay. Because in 2020 we are. We have had like a three maximum door to enter to the market now with a generative engine we have, I don't know, it's like a 50 indoor to connect with the customer or the possibility to customer or new new customer to come into the brand. Because the brand now I cannot. It's not easy to identify from which part uh, the customer is coming.
Speaker C: Yeah.
Speaker D: Right.
Speaker F: This is very hard now to understand.
Speaker D: Well, what have you noticed? Because I've been doing digital marketing now for 25 years, for a long, long time. And what I've noticed is that since the LLMs came out, all the old ways of doing things, they still work, just not as well. Like you have to work harder to get noticed than you did before the LLMs. It's like I would say roughly 50% like you have to work almost twice as hard to get the same results. Is that what you're seeing as well?
Speaker F: Yeah, yeah, Yeah. I agree 100% with you. It's not easy for the market. We are in the uh, in the hype situation.
Speaker D: Yeah.
Speaker F: With everyone need to have assistance uh, about our agency like our or from any kind of agency.
Speaker D: Right.
Speaker F: Because they want to have very faster the results. Yeah. Right. And the brand now there are not excuse to do what they don't do SEO in the past years.
Speaker E: Right.
Speaker F: Because we just approached a new client, new prospect and they don't have a good SEO. And we need to start from SEO.
Speaker C: Right.
Speaker F: Because there is not geo without SEO. These are main focus. There is not because you cannot, okay, you can manipulate a little bit the information. You can uh, you can go ahead with the public uh to relay some content to a exterra website side to yours. Because just gpd for example when some user research about the topic. The first step the RMS do is like to. Okay, we wanted to check the web, the entire web, how many websites talk about you.
Speaker D: All right.
Speaker F: This is the first point and this is. We call it digital PR because uh, as I mentioned before the link building was for generic key.
Speaker D: Yeah.
Speaker F: Right now.
Speaker C: Right.
Speaker F: So how many authority network are talking right to you?
Speaker C: Right.
Speaker F: It's important. It's not easy because um, sometimes the budget on the table need to be very, very strong. Because it's not easy if you are, if you are for a fashion brand wanted to have a content involved.
Speaker D: Right.
Speaker F: It's not cheaper.
Speaker B: Right.
Speaker F: I'm not expensive. But not only content. One content is enough for the strategy. We do have a good strategy. We make a plan for the client. We make uh, the plan. The strategy plan is very harder now because the SEO was like 15 tasks to do that in 10 months. Now from 15 to have a 42, 40, 43. Because a specific task for SEO NGO.
Speaker D: Wow. Yeah. And it's changing all the time to your point. Like the way it works today might not work next week.
Speaker F: Yeah. Is it? It's like, uh, we need to check every week.
Speaker D: Wow. Yeah, that's pretty funny.
Speaker F: Is our milestone M to be on the market every week? Wow.
Speaker D: Well, I think the answer is that companies need to realize you still have. Have to do traditional outreach. You still have to do all the things you used to do. You have to add this on top of it and use. Use the LLM awareness to show you where you need to focus your attention. Is that about right?
Speaker F: Yeah. Yeah, you're right. Yeah, you're right. You're right. The focus, the topic is it that some brand don't understand? Uh, now, because they are a little bit in SEO or the school. Okay. We give you one K by month.
Speaker D: Yeah.
Speaker F: Okay. But it's not a problem with 1k. Maybe some time is enough, maybe sometime is not. Because it just depends which strategy.
Speaker D: Right.
Speaker F: You want to create.
Speaker D: Right.
Speaker F: You are on the market now. No, you are not. Okay. 1K is not right. No, you need to split, move from 1k to 6,7k. Because if your business wanted to improve internal MS, you need to be more focused now than in five months.
Speaker D: Uh, I have one more question for you. This is the big one. I think that what the different large language models are doing, they must have some kind of reference architecture to know who talks about what. And I think this is where influencers come in. People who have a trust in voice, an established name in a particular space, like analytics. So if you have a YouTube channel with hundreds of videos about analytics, that's going to be good for your awareness for the LLMs. Not that it can't change, but if you're consistent and you focus on a particular area that is going to get noticed by the engines and then you will be kind of pigeonholed in that space. So it's good to know what they think of you and then work from that instead of trying to. Trying to create something brand new. So my question to you is, do you think it's important for brands to work with the influencers in their space who understand their technology and, and use those folks to help tell the story?
Speaker F: Well, it's a good question. How many m minutes we have to. How many hours you got? 3 minutes left to have a feedback. Ah, yeah. Uh, the brand are not. They don't have the grip on the market. Uh, the, the brand asks you like agency offering like consulting. Uh, which is a good strategy for day.
Speaker B: Right.
Speaker F: Okay. Because uh, the brand don't understand now the scenario. They the only. The only thing they have is like we wanted to to be.
Speaker D: Huh. Right.
Speaker F: About the strategy is up to you. It's not my. My problem.
Speaker D: Okay.
Speaker F: Friends like it's not like disinteresting. It's like because some people working to the brand have a lot of list tasks to do if day by day.
Speaker D: Yeah. Right.
Speaker F: And you need to to be very close to the brand to understand why they need to be present in which channel.
Speaker B: Yeah.
Speaker F: Why and again which strategy. How many budget. And the other thing we missed to spoke a little bit is I like about the competitor. Sometime your budget is not enough. Not because uh, it's not too big the budget, but because the. The competitor move faster than you.
Speaker B: Yeah.
Speaker F: I think we just analyze a lot of um, prospects. Okay. Okay. We just serving some analysis because we, we. What we saw is like we check the competitor are too faster. Too faster than the compet. The brand are older. Like to be in 2021 now.
Speaker D: Right.
Speaker F: Okay. It's like the gap of the years is very high.
Speaker D: Wow.
Speaker F: It's very high. And if they wanted to approach the market, they need to have very, very powerful. But not about the budget. But they need to understand which strategy.
Speaker D: Yeah. And which channel to explore. So uh, Emmanuel Arosio, look this gentleman up Data rank. He understands that you have to have your basics together and use the large language model models to better understand which direction to go, which channel to focus on. And you just got to keep on going and going and going. Right. It never stops.
Speaker F: Uh, absolutely. We can stay here all days to discuss about LMS because it's good topic sometimes for our client the prospect we serving the training.
Speaker D: Yeah.
Speaker F: Because they understand a little bit because they read some.
Speaker D: Right.
Speaker F: You know, because for SEO all the markets are likely to be in the situation to be still.
Speaker C: Yeah.
Speaker F: Now everyone read about AI with the search. They try to research in JGPT and Gemini. Our brand is not in what we need to do. Every market and every category of brand are influential about the act of search.
Speaker B: Yeah.
Speaker C: Yep.
Speaker D: Got to keep it going well.
Speaker F: Grazia Mille, thank you so much.
Other episodes covering the same guests and topics, from across The B2B Podcast Index.
Eric Ries on Why Good Companies Go BadPodcast Archives · on Cloudflare92 / 100
AI Agents, False Productivity, and the Sales Team Reset with Gabe LarsenMake It Happen Mondays · on AI agents91 / 100
Why a $1.2B exit felt like his biggest failure, and the customer-obsession thesis behind AgencyThe GTMnow Podcast · on AI agents86 / 100
Ep 90: AI Pioneer Jürgen Schmidhuber on the State of AI TodayUnsupervised Learning with Jacob Effron · on Large language models85 / 100
Is Your Business Invisible to AI Search? (And How to Fix It) ft. Ray YoungRevenue Science · on ChatGPT85 / 100
Episode 018: Season 2, the $75 Consult and the Frankenstein StackAI Tools for Practicing Lawyers · on ChatGPT84 / 100Live from Accelerate: The Future of Data-Centric InfrastructureLive at We Make Future: Exploring the Next Frontier of AIBuilding AI-Ready Data for the Agentic EnterpriseAI Agents at Boomi World - Governance, Knowledge, and AutomationGoverning AI and Building the Intelligent Enterprise