← Marketecture: Get Smart. Fast.
Marketecture: Get Smart. Fast.
Episode 130: Adam Epstein on building the Gigi AI chatbot, plus Linda leaves X and WPP gets a new CEO
00:50:49
Loading summary
Transcript112 lines
- [00:01]
Sponsor Voice
A word from our sponsors. Today's marketing leaders need more than dashboards. They need answers. That's why Zeta built Zeta Answers. It replaces guesswork with strategic control so you know where to spend and how to grow. It gives you customer foresight using predictive AI to target intent before it happens. It delivers full funnel clarity, tying media to real outcomes online and off. It puts on demand intelligence in your team's hand.
- [00:29]
Ari Paparo
No.
- [00:30]
Sponsor Voice
No dev tickets, no delays. And with live campaign insights, you can optimize while campaigns are still flying. Zeta Answers doesn't just measure outcomes, it drives them. See it in action@zetacmo.AI answers. Again, that's Z E T A C M O AI Answers.
- [01:07]
Ari Paparo
Welcome to the Market podcast. This is Ari Paparo, I'm joined by Eric Franchi and our special guest today is Adam Epstein, who is the co founder and CEO of Jiji. You may have remember Jiji? They have a really cool demo of a AI driven chatbot that helps with Amazon ads. And we wanted to get Adam here to talk about how they built it, what it is, why it's interesting, it's sort of a case study and you'll hear when we talk to him a lot of like pretty technical learnings about how he built it and why it works and doesn't work and the various product decisions and I love that kind of stuff. Eric, you know Adam pretty well, right?
- [01:45]
Eric Franchi
Yes. By way of disclosure, Gigi is a portfolio company of Apparium's. So I've known Adam for a really long time. I've had a front seat of Adam starting a business that was effectively a bet on one thing and then deciding that he needed to change. And at the same time, it was during the real ascension of AI and I have not seen a person get. And he uses this in the conversation. AI pilled the way I saw Adam get AI pilled and he just burned the boats and went all in and. And he's become very quickly a real expert on AI. So I think you're going to learn a bunch of stuff.
- [02:24]
Ari Paparo
Yeah. We'll also find out where the name came from. We also have a lot of news this week, including Linda Yaccarina leaving x new CEO of wbp. A whole bunch of AI stuff. As always, Google investing in Adex and more.
- [02:40]
Eric Franchi
It's going to be a great one.
- [02:41]
Ari Paparo
All right, let's dive right in. We're here with Adam Epstein. He is the co founder and CEO of Gigi that we profiled here on this podcast a couple of weeks ago. Sort of a talking head bot Assistant virtual friend who can help you with e commerce. How's that for a pitch?
- [03:01]
Adam Epstein
It's not quite there. I told Eric this when you spoke about us last. There is no avatar, unfortunately. It's just for marketing on our website. But there actually isn't an avatar in the app itself. It's funny if you look at our website very intentionally put an avatar there because we wanted to personify the software. One of the things that I feel very passionate about is that in this age of agentic AI, we hire software, we don't buy software, and by hiring it, we wanted to personify like what GG would be. It's like hiring a team member. So we put an avatar on our website, but unfortunately there is no avatar in the product. So you might be. Want to kick me off the.
- [03:39]
Ari Paparo
I might want to kick you off. Okay. Because I think I have feelings for Gigi. So is the avatar.
- [03:45]
Adam Epstein
So you want to know something creepy here, Ari?
- [03:47]
Ari Paparo
Yeah, of course.
- [03:48]
Adam Epstein
So we were really struggling with what that avatar would be. Yeah, we were using a bunch of AI image creation products and we kept on getting images that were very, very like clearly AI and we. And it wasn't real. And so Eric knows this. Gigi is actually named after my daughter. Her name is Gia. She's a four year old girl with curly hair. Her middle name is Cherry. Logo's Cherry. So we said in the office, we're just like, why not like, let's actually just make the avatar Gia. And So we fed OpenAI a bunch of pictures of Gia and we gave it this prompt. I actually have the prompt ready because I think it's kind of fun. Use this reference image to age this girl 29 years. She's fit, wearing chic business casual with a dark blue knit sweater, subtle gold hoop earrings and dainty gold necklace, and has a kind, helpful smile. She's looking into the camera for a photo, is centered in the frame from the chest up with some nice space above her head and there's a warm background behind her. The light is natural and the image is realistic portrait. So that's what Gigi is.
- [04:48]
Ari Paparo
Yeah. And hopefully 20 years from now, Ji will be able to use virtual therapists to work through these issues. And she'll have this podcast to hand as a training material to her virtual therapist to figure out how this all happened.
- [05:03]
Adam Epstein
Yes. Anyways, should we talk about the product?
- [05:07]
Ari Paparo
Yeah. So let me give a little bit of background here because, you know, we don't want to just hear a pitch for your product, but we were sort of Interested, because everyone's doing AI. Everyone, everyone's AI. AI, AI. And you know, the corpus of data people have varied. Media planning data, media results data, et cetera. We thought what you had built was interesting, partially because it's in a sector we don't talk about a lot, which is, you know, retail media on the buy side, Amazon optimization, and partially because, you know, the demo looked really great. So why don't you talk us through like what was the rationale? And then we want to go into some depth about how you built it.
- [05:45]
Adam Epstein
Sure, for sure. So perhaps it's best if I tell you how we got here in the first place. So we began the company by building a product that was purposely built for buying and measuring streaming TV ads on Amazon ad tech. It was on top of two foundational Amazon ad properties, the Amazon DSP and its clean room Amazon marketing cloud. And we created a UI layer that sat on top of the two. We were doing okay. Like in our past year since we launched, we managed over $10 million in streaming TV ad spend. But like, we didn't hit product market fit. And many of our customers and prospects effectively told us, hey, like, you guys need to support full funnel Amazon dsp. We're not going to onboard a new vendor just for streaming TV alone. It's a small portion of our budget. And so we thought to ourselves, like, how can we build a demonstrably better product on top of the Amazon DSP beyond just the DSP itself? And around this time we were immersed in using AI products ourselves. Our engine team was using Cursor, we were beginning to onboard some go to market AI native tools like 11x for like outbound sales and Decagon for support. And we started to realize the like, the leaders in the AI application layer companies were really finding a job in which there were a large quantity of humans in the workforce in which those humans performed a lot of rote manual and repetitive tasks. And they used agentic AI to automate many of those tasks so that the businesses that employ those people can gain a sense of operating leverage and perform with a higher degree of efficiency and perhaps better performance across the board. Well, what better job than a media manager at an enterprise brand or agency? I don't need to tell anyone that's been in the Amazon DSP the amount of rote manual tasks that you need to do to execute a campaign or build a report.
- [07:30]
Ari Paparo
Amazon catching strays over here.
- [07:32]
Adam Epstein
You know what? Amazon would be the first one to admit this. They built an incredible business. You talk to anyone on the Amazon team they'd be the first to admit.
- [07:40]
Ari Paparo
Is that different from other team DSPs? Is it more workflow?
- [07:43]
Adam Epstein
I, I'm not here to comment on other DSPs, but from what I understand the Amazon DSP is like quite cumbersome. And so we effectively like spoke to a bunch of media managers who we were working with on our, on our old product and we said like how do you spend your days? Like literally like what do you do? How do you chunk the hours of your day? And we broke that down into tasks and so we've effectively broken all of those tasks into automations that we've now trained Jiji to run and automate so that we have this product that any agency, it's specifically built for agencies. And I'm happy to tell you why in a moment can onboard Jiji and Jiji can start to perform many of the tasks that a media manager does. So that media manager isn't spending their time running reports, updating line items, copying and pasting the same thing over and over and again, but rather understanding insights in pursuit of great client service service and then extracting more insights in a positive feedback loop thereafter.
- [08:43]
Ari Paparo
So one of the questions I have about companies like yours or solutions like yours is that if they only work on one platform, then isn't it just covering up for a deficiency in the platform? Like creating a bunch of line items with the same naming convention is something a platform should in theory be able to have themselves. So how do you make that trade off between something that's like a no duh obvious feature that you're just enabling with AI versus something that you think that you have a sustainable ability to solve?
- [09:13]
Adam Epstein
Yeah. So we have to start somewhere. Sure. We can't, we can't support every DSP or every form of media buying like that. We're a seed stage company with 11 people with a small amount of revenue and small customer base. I could give you the like PR marketing pitch of why we started with the Amazon DSP. It's one of the fastest growing DSPs, yada yada yada. But the real reason we've been building Amazon ad tech for seven plus years across our 13 person team. We have 25 plus years of Amazon ad experience and our legacy product already was integrated with all the APIs. So we started on the Amazon DSP. But many of the primitives that we're building are not unique to the Amazon dsp. They can be more broadly applied across all of media buying and all of measurements. We're just having to Start here, because that's our entry point into building this market and company.
- [10:04]
Ari Paparo
So the other interesting bet that you're making, I think, is that the way users are going to interact with AI is through a chat concept. And I think there's a lot of folks who are questioning this right now in the broader AI world where we're now getting used to chatting with things and asking them various questions, maybe that's the best way to do it, maybe it's not. We're not sure. No one's sure. Right. So tell us about your thinking in that regard. Is it easier to have a button that says like, optimize my report or for someone to type in, I need to optimize my report?
- [10:38]
Adam Epstein
Yeah, so it's a very good question. And part of what I call my red pill journey into all this AI world began when we just started invoicing our customers on Stripe. Stripe is this world renowned B2B SaaS software product. And I just wanted to know the first customers last summer that we were invoicing. Oh my God, did they get my invoice? What did the invoicing look like? And is it paid? And I couldn't do it. And I found that LLMs have actually made me stupider as a user of enterprise SaaS products. And I have an expectation now that an enterprise SaaS product will operate a lot more like ChatGPT. So that's my like, like, in general, like, I believe that to be true. And I heard the Ramp CEO say something like, who are we? How selfish have we been for the past 20 years to dictate how people should use software and learn rui? Like why should they learn our UI when ultimately many of the most efficient ways of doing this it now seems to be chat. But the problem is if you just throw someone into a blank chat interface, it's incredibly disarming. What do you say? I don't know what to say. So you can't just do that. And so we've built a host of like preset workflows that we know a media manager would do within their day in which they can initiate their workflows. Or we could have prompt based triggers that identify workflows for them based on the objectives that they're trying to have for their client. So in an ideal world, that end user is coming into our home screen and rather than clicking through a ui, sees a bunch of prompt based actions that they can take by accept, declining, accept declining, or they can interact with Jiji through a bunch of workflows that we know to be true is the mechanism that they'll do their job as an effective media manager on a day to day basis. So it is chat. We believe that chat should be the command center and not a bloated enterprise ux. But the, the, the key paradigm shift here is the UX decisions that all of these vertical AI agents are making to their domain and their industry to ensure that you're creating a positive customer experience such that they're not disarmed in just like a chat interface in which they're paralyzed because they don't know what to chat. And what's that prompt?
- [12:55]
Ari Paparo
Would you mind pivoting the company to be GG for Stripe? Because I cannot, I cannot get a report out of Stripe. I know it's there, it's, it' there. But like, how do you do it?
- [13:05]
Adam Epstein
My God, I thought Stripe was great.
- [13:08]
Ari Paparo
But it is great. It's just also impossible to get the information.
- [13:11]
Adam Epstein
It's crazy. So, yes, so we believe that chat should be the command center. But the hard work is identifying those workflows and those prompts so that the user doesn't need to think, what do I put in to do my job better? They just come into the app and they see all of the things that they need to see. They're reactively accept, declining, accept declining, or they're going through a bunch of preset workflows based on the goals that they're trying to drive. What are opportunities across my customer base for efficiencies? Where are their growth opportunities? Where should I be reallocating budget? These are all questions the media manager should be asking themselves on a however frequent cadence every other day, day, whatever it might be that we've trained Jiji to be able to understand, to dictate to the media manager to perform.
- [14:02]
Ari Paparo
All right, so let's say I have an ad tech company with a whole bunch of data about something. You know, it's a vertical solution. It does one thing. I don't know what it does. Whatever it does. Take us through how you built this as a vertical AI agent and what advice you would have for builders out there who are looking at a similar problem.
- [14:22]
Adam Epstein
So it's been, I don't know why people don't talk about this more like how people actually like train a vertical AI agent. And so like I have naturally have been reading everything about like fine tuning, reinforcement learning and rag and all of these ways that you can do this. And so I was in San Francisco in March and I was meeting like just some very early investor meetings and I was going to our CTO and I was like, hey, are we doing fine tuning? He's like, no. Are we doing reinforcement learning? He's like, no. I'm like, so what are we doing? He's like, he's like, it's rag. And so maybe it would be helpful if I like, sure. Really dumbed it down and explained all three.
- [14:59]
Ari Paparo
So Eric, do you want to explain rag to us? I like, I like ragging on Eric.
- [15:04]
Adam Epstein
So, so fine tuning is the work that the frontier LLM companies are doing. They're like changing the weights and behaviors of the actual LLM itself. For a vertical AI agent to fine tune, it can be very risky and limit your flexibility. So if you're fine tuning one single LLM, you can't be LLM agnostic.
- [15:25]
Ari Paparo
Right. You could accidentally turn your AI into Hitler. Like some people might have.
- [15:31]
Adam Epstein
Yeah. Or some other bad thing happened, but you don't want to. Like fine tuning is very, very risky. And then reinforcement learning typically like if we live in the world of Amazon ads, like Amazon has like AI powered audience called brand and Performance plus that use like millions of signals to optimize a campaign with a reward mechanism of this is good, this is good reinforcement, this is good, go find more of this. And so that's reinforcement learning. The rest is rag. I've spoken to so many vertical AI agent founders and investors and I asked this question, like, is anyone using reinforcement learning fine tuning? There's like 99, 95% is rag. So what is rag? So like the LLMs are all trained on the corpus of data of the Internet itself, but all of these vertical AI agents leverage their domain expertise. So like we like to say we have 25 plus years of Amazon ad experience across our small team and so we leverage that by creating content like actually taking the thoughts and insights in our head and writing text and training our AI agent against that text. So that's rag. So it's additional text that doesn't live on the Internet that allows you to train the AI agent. And so when you see ChatGPT improving based on memory, the memory is the rag that you're creating on ChatGPT. So that's effectively what we've done. So like how do we actually train Jiji? What's the mechanism for doing so? So our technical team has built this infrastructure with which we, the non technical folks are able to create these, what we call like bite sized RAG documents in which we write these like 500 character documents that explain like a small subtle nuance of Amazon DSP Media buying and measurement. We see what Gigi would say prior to uploading the rag. We upload the rag to quote unquote, Gigi's brain, and then we see what Gigi would say afterwards. And we QA it to make sure that Jiji has said what we want it to say. And if not, then we change the rat.
- [17:32]
Ari Paparo
So. So like a snippet might be like, here's a metric in the Amazon report, here's what it means and why it's important.
- [17:39]
Adam Epstein
Yeah, and. And what are the downstream things that lead to it and how do you optimize against it? Exactly that.
- [17:44]
Ari Paparo
And this is just coming out of your brains. So you're basically, you're, you're taking human brains, turning them into text, and then feeding them to an AI to be a better artificial brain that's named after your daughter. There's nothing weird here. Nothing weird.
- [17:57]
Adam Epstein
Correct. But this is, but this is, this is how vertical AI agents work. And like, yeah, it has been a painstaking amount of work. Like, we've written over 1000 of these mini dogs. Like, I've spent every single night the past six months at our office writing rag docs. That seems like, so mundane and simple, but it's like a heavy degree of operational work to train Gigi. Like, this is what it takes to train an AI agent. And what you effectively described Ari. Like, it's not dissimilar to actually training a team member. Like, how would you train a team member? Like, if you hired a new media manager for your agency, how would you train that team member? You would write process docs, standard operating procedures, training videos, and you would give it to them and you would see how they are. You would monitor their performance. And if they're good, then you'd let them be. If not, then you'd train them a little bit more.
- [18:46]
Ari Paparo
So your rag process, though, is differentiated from what is sometimes called a system prompt.
- [18:52]
Adam Epstein
Rag and system prompts are two totally different things.
- [18:53]
Ari Paparo
Can you just give us a one on one on that?
- [18:55]
Adam Epstein
Sure. So system prompts are the underlying prompt that lives against the AI agent. So I did a terrible job of explaining that, but there's a Y Combinator company that actually like disclose their system prompt. So the system prompt is like a 5,000 word master prompt that says this is what you do and this is how you execute.
- [19:18]
Ari Paparo
Yeah, it's like, your name is Gigi, you want to be helpful, you understand Amazon ads better than anybody else, blah, blah, blah.
- [19:23]
Adam Epstein
Yeah, correct. You're an AI media manager of an agency. This is how you respond to things. This is how you don't respond to things. This is how you should, whenever anyone asks you metrics, this is how you should deliver metrics. It's very different than RAG itself. And so the RAG system that we've created is very different than the system prompt. And the system prompt is something that we're always amending based on the evaluations and QAING that we're doing of gg.
- [19:50]
Ari Paparo
This may be a little bit of a technical question, but do you have to rerun all the RAG training every time it changes, or is it incremental where each new RAG is new?
- [19:59]
Adam Epstein
Each new RAG is new.
- [20:01]
Ari Paparo
Okay, so if something changed, like let's say there's a metric in the Amazon system that changed its name, you would have to retrain the system.
- [20:11]
Adam Epstein
We have a database, like the infrastructure that we've created is this database of RAG documents. So we would identify that if what we've trained GG is no longer true, again, same thing as a human. If what we've trained Jiji is no longer true, we would eliminate the previous RAG document, we would amend it and upload the new one, and that would be updated.
- [20:29]
Ari Paparo
Right? So it's interesting. So if you're like an unemployed ad tech person, maybe you have a future in writing little RAG documents for the AIs charged by the hour.
- [20:37]
Adam Epstein
That's possible. It's very possible. But this is the interesting thing, right? So like, we're building this product specifically for agencies, and we've been working with agencies for the past seven years. And what's the secret sauce of some of the best agencies in media? The secret sauce is that they've invested very heavily in the standard operating procedures, process docs, training manuals, and these are all really, really good for training an LLM. So if you think about the knowledge that Jiji's created is we have like the LLM foundational knowledge. We have the Jiji knowledge that we've trained Jiji on. And then for every agency that we onboard, we get them to build a knowledge base with us in which they give us all of the those documents. We QA all of those documents within Jiji. We work with them to understand, like, how they build retargeting campaigns. Whether they think of new to brand AUD would be how they break down line items across inventory sources for display ads. And then for each individual agency, they have a new unique GG agent. And as someone that's like sold ad tech for the past seven years, every agency that I talk to is like, hey, how is your off the shelf ad tech different than the off the shelf ad tech that you're selling to my competitive agency? And I used to give a BS answer, you can customize it however you want, but realistically it was the same software. Whereas now we're in a world in which each agency has the ability to mold their GG agent exactly how they would a team member. So it's entirely customizable to them. And that's why I think it's really powerful. And agentic AI will be a tailwind for agencies is because they've invested in this infrastructure such that they can train an LLM to operate exactly how they see fit with their level of expertise in a way that other industries and other domains can't.
- [22:21]
Eric Franchi
I was going to ask you, Adam. So Sam Alden famously said, God, it was probably A year ago, 95% of what agencies do will effectively be automated by AI. Many people interpreted that as, okay, this spells the death of agencies. And it seems to me that you're saying that, hey, if you understand how agencies actually work and where their strengths are, AI is actually like, you know, potentially like an awesome accelerant to making them that much more valuable for their customers. Am I getting that right?
- [22:52]
Adam Epstein
Right, Totally. So I think, I think that was specifically for creative agencies that Sam comment. But Zuck is saying, press a button, get an outcome. And you know what, that might be true in a few years for SMBs, but enterprise brands are always going to need a level of handholding and customization. You can't press a box and get an outcome for a large public CPG brand. And so I just, I just don't believe that to be true in the next three to five years. But the other really interesting thing is like, I think agentic AI is going to fundamentally change the operating model of many media agencies. Like agencies, like many services, as businesses have always needed to linearly scale headcount as they scale customer and revenue growth. And we think agentic AI will allow them to exponentially scale customer and revenue growth while maintaining a flat headcount. And the work that that like system of media manager, director, ops and analyst people underneath a client, I just don't think it'll need to exist in the future. And it will lead to better outcomes. It'll lead to people that are happier with their job because the task they're performing is more strategic and just better. And then two, I think it's going to lead to better client outcomes. And I think the people that are going to be at agencies are going to be paid more doing better quality work in pursuit of their clients.
- [24:08]
Ari Paparo
Super interesting. I feel like I learned something which is a good accomplishment for this podcast. All right, so let's take a break now. We have a ton of news this week week especially the Linda Yakarino and the WPP CEO and a bunch of other stuff. So we'll be back in a minute. This podcast is brought to you by audiohook, the leading independent audio dsp. Audiohook has direct publisher integrations into all major podcast and streaming radio platforms, providing 40% more inventory than what could be accessed in omnichannel DSPs. What's more, audiobook has full transcripts on more than 90% of the of all podcast inventory, enabling advanced contextual targeting and brand suitability. Audio Hook is so confident that in addition to CPM buys, they offer the industry's only pay for performance option where brands can scale audio and podcasting with peace of mind knowing they are only paying for outcomes. Visit audiohook.com to learn more. That's audiohook.com.
- [25:10]
Eric Franchi
All right, and we are back with the refresh after taking the week off for July 4th. What's on the docket? We have a lot so big exec moves, exec changes, including at X&WPP. We have some interesting Google scoops around their reinvesting in helping publishers sell ad inventory, a ton of AI stuff, and if we have time, one or two other things. Let's take it from the top. Ari, you want to go with X? You want to go with wpp?
- [25:39]
Ari Paparo
Let's go with X.
- [25:40]
Eric Franchi
All right, so Linda Yaccarino, longtime media and advertising executive who is named CEO of X a couple of years ago, announced that she's stepping down yesterday. This is after a week where Grok, which is part of xai, which is now the overall company that X is connected to, went crazy and some other rocky times. What's your take on this whole thing?
- [26:06]
Ari Paparo
Yeah, sure. So hasn't been the greatest run here. So first Linda was I worked with her at Comcast. I don't know if that's relevant, but she took the CEO title at a tech company, presumably because she thought it was a big career move up. In my opinion, she was never the CEO specifically, Elon said at the time she won't have product engineering reporting to her. For a tech company to not have product engineering is sort of a sham. So she was really the president of commercial operations. The two years she was there was pretty inglorious. We have some data that was provided to us from mediaradar. Mediaradar is an independent company that tracks spending. According to them, spending on ads on x is down 27% year over year to about $1.3 billion in the trailing 12 months. It's been down basically every single month for the last 18 months. In addition, you have the dismantling of Garm, which was at the hands of Elon, the continued lawsuits. I think there's still a lawsuit against Unilever and some other folks for not advertising on X. And then the reported allegations that the sales team is using political and legal bullying to get people to spend more. And then finally the company gets absorbed by xai. So Twitter is now a division really of xai, not the main focus or the main driver of the evaluation. So not good overall.
- [27:38]
Eric Franchi
To clarify what happened with Grok this week, and again, I don't think these two were connected. Like her leaving the day after Grok kind of went viral itself on X. What happened?
- [27:49]
Ari Paparo
Well, Grok has. We talked about system prompts earlier. So Grok has been vacillating effectively with different system prompts. And the vision of Elon that he said is that he wants Grok to be unwoke and also kind of punchy and telling the real truth and being fun as opposed to what he perceives as the other AI engines which are overly woke or overly staid in their system prompts. And the problem is that it seems like every time an engineer changes one of these system prompts, Grok like goes crazy and either goes really, really liberal or really, really conservative. And this week it effectively started saying, well actually Hitler had some good ideas and that's really not great. You really don't want your AI to start sounding like Drill the Twitter account or he also was doing fantasizing about raping people. I'm saying he. Grok is not a he. Right. Grok is an id. But it was about the worst case scenario for what your AI could do.
- [28:47]
Adam Epstein
To be fair, just I saw Grok News this morning that GROK hit all of the best benchmarks for any LLM Frontier model LLM. So I think like this is just really a transformation of the business. Like this business is not an advertising business. This business is a proprietary IP data business to feed the LLM. And it's a consumer business in which the extent to which it gets advertising will be entirely self serve without humans selling those ads.
- [29:15]
Ari Paparo
Well, I think maybe Grok is the ultimate brilliant asshole, right? That's always the question. As a CEO, you have some engineer who's the best engineer, but a total asshole grok. Is that in the AI space? So do you want him or do you not?
- [29:30]
Eric Franchi
What then is the future of X from a business model and commercial perspective? Is it subscriptions, data licensing, subscriptions to xai?
- [29:41]
Ari Paparo
Is that Basically I think that they are going to have a hard time replacing Linda for someone else who's an ad exec who wants to run this thing. They have a new head of product, this guy Nikita.
- [29:55]
Eric Franchi
Nikita Beer? Yeah. One of the best followers on X by far.
- [29:58]
Ari Paparo
He's funny as hell on X. He's really funny and so he's likely going to have to grab the reins here. My guess is that they try to be totally, as Adam said, self serve, programmatic, direct response, bottom of the funnel. And I don't know if that's going to work because the problem X has is that their data is not commercial in nature. It doesn't have a lot of value. Someone tweeting about politics has no commercial intent. And actually most of my understanding is that the majority of their revenue or the highest quality revenue all comes from more branding stuff like on the NBA channel on the Discover tab. And that's against AI. That has nothing to do with AI. So I think Nikita has some real problems. You can call me if you want. I have some ideas. Maybe I'll write in the newsletter this week. We'll see.
- [30:47]
Eric Franchi
That's a good one. Okay, let's move on to wpp. So we talked about this one a few weeks ago. Who the top candidates reportedly were. There was the drum ass Patty Power, a UK betting platform on making odds for the next WPP CEO and nobody got this one right.
- [31:05]
Ari Paparo
She wasn't even on the list.
- [31:07]
Eric Franchi
She wasn't even on the list. And maybe we have a little bit of a US ad tech skew. Brian Lesser was the name we banded about. But it's Cindy Rose and I very quickly got up to speed on Cindy Rose and she seems like an all star. So six year board director for wpp. Former Microsoft exec, media company background, like she's legit.
- [31:29]
Ari Paparo
She seems perfect. I mean she was at Microsoft in charge of AI solutions for enterprises. So she has sort of a customer orientation. She was on the board of wpp, works at Disney. I think she's European also. I don't know where she's from. Yeah, dual citizenship.
- [31:45]
Eric Franchi
Yeah. In U.S. yeah.
- [31:47]
Ari Paparo
Couldn't get someone better out of central.
- [31:49]
Eric Franchi
Casting, especially at a time where they issued preemptively a bad revenue outlook for the year. Revenue down 3 to 5%. So packaging that with what seems to be a really good CEO Candida is the right way to do it.
- [32:03]
Ari Paparo
Did you, had you ever heard of her before?
- [32:05]
Eric Franchi
No, this is, and this is my bad.
- [32:07]
Adam Epstein
One of the things that I think is going to be an interesting bit is that since she's been on the board for six years, she knows the business well and it's clearly a business that needs a lot of change immediately. It's not like one of those businesses in which you can hire a new CEO. Six months to review, six months to plan, six months to change. I think she's going to make some changes very quickly and it's not lost on me. Like I hate to say this stuff but like it's not lost on me that Microsoft just let go of 9,000 go to market employees. So unfortunately I think there's going to be a lot of AI transformation and a lot of change and that's going to lead to the transformation of that headcount of the company while they go through that. But she's going to be in a position to execute it and execute it quickly.
- [32:47]
Eric Franchi
That's a great point actually. What would a outside CEO that's not close to the business do? They would take their time, they would talk to customers, they would interview employees. Steve's probably coming in with a plan after being very close to the business. That's actually a really good point Adam. We will keep an eye. Wanna talk about Google?
- [33:05]
Adam Epstein
Arif?
- [33:06]
Ari Paparo
Yeah, I thought this article was interesting. So Adweek had a scoop that Google was investing in its sales team to help publishers on Adx monetize better. So these are the people who glad handle advertisers and agencies and tell them why Adex is the best solution. They should do private deals there, there don't use the competitors for those private deals. So I think there's like the surface level and the deeper level. The surface level question is like it's interesting that Google has been sort of flat footed for the past year or two on this that Google has the highest rates in the industry. So if you're doing a private deal having high rates is a real detriment. You know you want to go where it's cheaper, especially a programmatic guaranteed deal. And they haven't been very aggressive, they haven't been the easiest to work with despite having the most market share and they've been getting beaten by Pubmatic, Magnite index, et cetera. And there was a quote in this article that said publishers get some of the Lowest prices on the ad space sold by Google Ad Manager compared to when they sell through rival ad advertising technology firms like Pugmatic and Magnite. Two publishers said so that's right there. They're just not getting the results and all the SSPs are staffing up here so that's good. The second question, the more deeper question is as always reading the tea leaves, are they trying to make make adx bigger, more profitable so when it's spun out it looks more valuable because the adex spin out is the top of the list of the remedy requirements from the DOJ and other parties. So you don't want a shrinking market share spin out that wouldn't look as good. So maybe this is just trying to firm up the ship before it gets sent off to sea.
- [34:47]
Eric Franchi
Yeah, that's a very interesting point. How much of an impact can a bunch of sellers on the street helping publishers learn how to use ADX better versus Magnite and PubMatic really make a difference? Is the product at parity and there just needs to be more attention and commercialization. Do you think it can make a difference Art?
- [35:07]
Ari Paparo
Well I think these folks are selling to buyers, they're going to agencies and advertisers and saying you should prefer adex. And they probably have some pretty good evidence around match rates and other things that Adex has advantages is there has been a movement over the past five plus years for preferred SSP relationships sometimes called supply path optimization agreements. So not having that salesforce certainly hurts.
- [35:32]
Eric Franchi
And then they had this offer wall ad format product that you thought was interesting as well. So basically user goes to a publisher site rather than being able to access the link that they clicked on to get to the article or piece of content. User needs to to watch a video or input some data and then the publisher gets a little bit of data and probably a rev share off of the ad as well. How big do you think this is?
- [35:58]
Ari Paparo
It's not that big. It's nothing new. I mean this has been pretty common in the mobile space for a while but it's good that Google's doing it. I made the point, I tried to spin this into my favorite subject which is I get asked all the time why does it matter that Google has a monopoly in ad serving? Why should the government be involved and how does this help publishers? And I feel like this example counterintuitively makes my point point which is that this is a fairly obvious innovation and it's 2025 and Google's just coming out with it. If there was A competitive market for ad serving. You might have all kinds of companies that service publishers offering all kinds of solutions that don't need to be double tracked and cut and pasted tags and integrated and whatnot. You can just have natively innovative ad serving. Good for Google. But I think it just kind of makes my point point.
- [36:44]
Eric Franchi
All right, let's move on to the AI stuff. There is so much going on here and I'm happy we have Adam to give us additional perspective as my AI pilled founder friend. Let's start with Cloudflare. Speaking of publishers. So Cloudflare and we missed this one because we took a break last week. This is news. Last week they launched what's called the red button. On July 1st. They apparently had a party in New York City with a bunch of publishers who symbolically press the red button on July 1. What is the red button? The red button effectively lets publishers block all AI crawlers at the click of a single red button, but also has more elegant solutions where they can offer like a pay per crawler option basically to help publishers monetize where they see fit, rather than some of these one off licensing deals so publishers can stop the crawlers. This is a good thing and this is a bad thing. It's a good thing because publishers now have a little bit of leverage or a lot of leverage if they block at all. Some of the increased costs of the AI crawlers goes away, but then they're also not feeding the crawlers that have their content be surfaced in AI answers. And there's also some controversy over whether this red button will block search crawlers, which is the main source of traffic for Google's AI. What do you guys think of this cloudflare thing?
- [38:02]
Adam Epstein
I think a lot of people are looking at that Reddit licensing deal with OpenAI in which right now Reddit's licensing of their IP I think represents like 10% of their revenue and over 50% of their profits and are like, how can I replicate that? But so many people don't operate at the same scale as Reddit. So how can all of these publishers effectively monetize their content in a, in a world in which, you know, maybe open Internet advertising and search advertising ceases to exist for them in the not too distant future? Like everyone's reading the tea leaves of what's next for LLM search and output put. And this is a step in a direction I think people are going to try. I think that it might be effective. It might not be, but as you mentioned, Eric, it might just give People leverage into exercising licensing agreements with some of these LLMs and saying like, hey, like we're going to just turn off and it's going to be paper crawl or you're just not going to be able to access our content at all. The challenge is though, like that might work for like mid tier publishers, but like the long tail publishers are still going to be like unable to monetize their content in the way that they used to. And I think it's going to still be a challenge. But arming publishers with some ammo in this fight with the LLMs is always probably a good thing right now.
- [39:16]
Ari Paparo
Yeah, I think that the controversy over search is the interesting part here because anyone can change the robots Txt to ban crawlers, but Google is claiming that they will only will accept one robots Txt entry for both search and AI. And this has already come up in the DOJ case. I sound like a broken record, but this is kind of where it's being litigated. It's obviously an interesting issue because no one wants to be excluded from search, but you might want to be excluded from AI. And I think Google's pushing the limits here in terms of fair use and what they can do here. There was also, I don't know if it's in our notes, there was a legal finding recently that supported I think perplexities fair use claims on text. But I'm not an expert on that, so I can't say much more about it. But this is a very active issue issue.
- [40:07]
Eric Franchi
So much going on. Right. So you've got the cloudflare red button and then you've got the LLMs themselves launching their own browsers. So we talked about this when the news came out a few months ago. Perplexity launched Comet this week. It's only available if you're like a Perplexity Enterprise Premium user, like at the highest tier. Presumably it'll come out shortly for everyone, but it's their own browser with, you know, the search built into it, designed to keep you there. And then perhaps even bigger news, Reuters had an exclusive that an open AI browser is coming over the course of the coming weeks, which basically is going to use AI to fundamentally change perhaps how users browse the web. So while Google certainly holds the cards for the crawlers of today, both kind of AI and search, what happens if there's a change in the way people operate on the web with companies now owning the browser? It's a pretty wild technology.
- [41:08]
Ari Paparo
Yeah, I think that owning a browser gives you obviously the ability to interact with the consumer directly. So instead of searching for say a vacation travel and getting five links and then clicking each one to an open tab and then cutting and pasting the results into a Google Doc to say where I should go on vacation or what product I should buy, you can quite easily just see that the user visited those five web pages and instantly create a summary for them about what they should buy and then potentially make money when they click on it and buy the so that's a big transformation. It also sort of skirts around the robots issue because you can see the page as the person browses it. You can also see what ads are there. There's a lot of data. I think data is incredibly important in this case. Also, we didn't mention the DIA browser which already is in production. So the DIA browser is doing this right now and it's worth checking out. I think it's a pivot. It was a different browser and they switched to being AI first recently. It's from the browser company I Once again, I'm going off the top of my head, so I might be wrong, but this is already kind of a battleground and do I have to talk about the doj? Do I have to bring it up again that the main remedy in the search case is the spin out of the Chrome browser. And while the spin out of the Chrome browser seems like this radical change to Google, it's sort of maybe happening outside of the control of the government.
- [42:32]
Adam Epstein
It's just clear that everything's up for grabs. Like every single consumer interaction that is not AI native right now now will have an AI native winner. Like one of the topics that you guys wanted to talk about is the meta pricing of many of these OpenAI researchers and the hundreds of millions of dollars that they're paying for everyone.
- [42:51]
Eric Franchi
I like how you're saying researchers are being priced. So for those who haven't read the article, we'll link to it. Meta's been reportedly giving pay packages of anywhere from 100 to $300 million, with the 100 million being a sign on bonus to the top AI researchers, many of whom worked at OpenAI and Zuck is cold texting and emailing all of the researchers directly.
- [43:22]
Adam Epstein
It is wild. Of course you're building a super team. And let's just break it down. Everything is up for grabs. I am a mere founder of a vertical AI agent. So retail media and media buying and measurement is ostensibly up for grabs until there's an AI native company that can can emerge as that leader. But if you look at all of the trillion dollar consumer opportunities, whether it be E Commerce hardware search. I think I'm forgetting one like effectively it's up for grabs and Zuck is effectively saying, I think that I saw a stat that 33% of all of Meta's revenue is now spent on CapEx for AI, which is the biggest bet that he's ever made in his career. And that's not inclusive of the $15 billion scale AI acquisition that he just did, not inclusive of the Nat Freeman Daniel Gross Venture fund that he just did, and not inclusive of the $100 million signing bonuses that he's giving to all of those engineers. Like this is a king's game in which he believes that consumer Internet and consumer hardware is effectively up for grabs for him to win. He is not winning right now. He's not winning either of those. And he wants to. And one of the things that you learn when like working in this space and we're not at the frontier model, we are mere uploaders of rag documents as I mentioned. But one of the things that you start to learn is that like no one knows the way to do this. Everyone is figuring this out on the fly. And so if he can hire some top OpenAI researchers that have this incredible domain expertise and knowledge of how OpenAI got there, there he has the unfair ability to level himself up and level Meta up so that they can start to attack some of these trillion dollar opportunities. So maybe it's worth it.
- [45:05]
Ari Paparo
Yeah. I listened to the six hour podcast about Zuck from the acquired podcast and that may sound ridiculous, but you can listen to it on 2x speed on a plane. So that's only three hours. And my God, that guy. They really make a case that he learned everything he knows by playing Civilization as a teenager. And it really kind of shows that he doesn't do the thing most CEOs do, which is have a single bed and put all their wood behind the arrow and try to execute the plan. He always has three different little schemes going on at one time and he puts a lot of wood behind all three arrows and then one of them works, one of them doesn't. He turns off his crypto and then he puts more money to VR and then, then now he's putting money into AI. The guy has no preconceived notions. He just plays the table in front of him. It's really impressive.
- [45:55]
Adam Epstein
And he has the money printing machine of Meta's advertising business to allow him to play these games. So great for him.
- [46:02]
Ari Paparo
So I think the Solution here is a seller cap and a draft. It's time to bring together all the tech companies and you get David Stern out there and you say this PhD coming out of MIT. What's the top, top offer?
- [46:20]
Adam Epstein
I'm not sure if you've seen but the TBPN guys, it's also NBA free agency in which like these NBA guys are signing $300 million deals and they're making like updates as though these OpenAI researchers just signed a $300 million deal for the Chicago Bulls. But they're going to meta it's. It's very fun.
- [46:40]
Eric Franchi
And don't forget he has super majority on the the board voting so whatever he says goes. Where do you want to go next? Source Point.
- [46:50]
Ari Paparo
Yeah, let's do SourcePoint.
- [46:51]
Eric Franchi
Yeah. Ben Barroka. Shout out to Ben. Shout out to Brian. Those guys have been at this for some time. And SourcePoint was acquired by Do. I've heard of Dr. Have you? Yeah. Or Didomi seems to be another, another source point which is you know like consent management and you know, privacy centric data collection. You know, basically building, building a larger company to compete.
- [47:17]
Ari Paparo
I think the story here, and I don't want to be a downer, is that privacy technologies haven't been a breakout in advertising, that the companies that focused on that ended up being acquired by bigger companies were not huge numbers. And this maybe is laid at the doorstep of Chrome and, and the sandbox failure. Or maybe it's just the market wasn't ready and the advertisers didn't care enough to spend the money. Or maybe it's that the incumbents caught up. I'm not sure exactly what it is, but it definitely felt seven or eight years ago or whenever SourcePoint was founded this was going to be a new category that consent management platforms was going to be the new SaaS and that every publisher would have to have them and you'd have all this valuable data flowing through your pipes and you can make money by charging people in various ways ways. And I think it turned out to be a small opportunity.
- [48:09]
Eric Franchi
SourcePoint raised 43 million over the years.
- [48:12]
Ari Paparo
They had a bunch of up and down I think in that.
- [48:15]
Eric Franchi
Yeah, yeah, I think and that's. It was a very helpful framework of like what may have led to privacy. You know, technology companies not having generally great outcomes. I think the number two at the end of the day, brands really not caring or not prioritizing this at the top of the list versus other things that are just more central to the business. Business feels like it's the thing. Right. Because if it was, all these companies would be crushing it from a revenue and adoption standpoint.
- [48:43]
Ari Paparo
Yeah. And I think the clean room story is pretty similar, which is that publishers would adopt clean rooms because it gave them more media. But publishers are a bad way to make money. They have small budgets, they're very price conscious. You make the money by having big brands adopt you and implement you at an enterprise level. And that just didn't really happen for clean rooms in any meaningful way. They became add ons like Liveramp buying Habu or WPB buying Infosum. And I think SourcePoint is similar. Their business model effectively was get publishers to adopt this, make a little bit of money on the publisher side and then flip the model and charge advertisers for some privacy safe thing. And the advertisers didn't really care that much. And you know what, the same thing's happening in carbon. Same thing. So Scope three and other Carbon companies go publisher first because publishers are the ones who need the money. They can't make a lot of money. They try to flip buy side, buy side doesn't care air, the market collapses and you end up having to pivot out of carbon.
- [49:37]
Eric Franchi
The other side of this is when you get it right, if it's time where you organize the supply side and then it's actually something that brands value and this could be everything from viewability to brand safety to contextual. You can build tremendous businesses.
- [49:54]
Ari Paparo
Double verify is and moat. Right. All giant businesses.
- [49:57]
Eric Franchi
Such a matter of timing. Yeah, such a matter of timing. We had Jonah on Open Market kind of recounting some of this stuff and why he's actually starting starting with the supply side with his bet on advanced contextual. It's pretty fascinating. Should we end it there?
- [50:11]
Ari Paparo
Yeah, let's end it there. This was an amazing conversation. So much good news, so much good information.
- [50:17]
Adam Epstein
Thanks for having me guys. Appreciate it.
- [50:19]
Eric Franchi
Thank you so much. Appreciate you. And we will see you next week everybody.
- [50:24]
Ari Paparo
Yep. Thanks for being here.
- [50:26]
Eric Franchi
Bye bye. Thank you for subscribing to Market.
- [50:30]
Ari Paparo
New interviews are added every week at Markitecture TV and your favorite podcast app. Thank you for listening to the Market podcast. New episodes come out every Friday and an insightful vendor interview is published each Monday. You can subscribe to our library of hundreds of executive interviews at Markitecture tv. You can also sign up for free for our weekly newsletter with my original strategic insights on the week's news at News Market tv. And if you're feeling social, we operate a vibrant slack community that you can apply to join@adtech God.com.