Full Interview: Eric Seufert on AI and Advertising

0:28Eric Subert, how are you doing? ads in ChatGPT? Well, the saga begins where ChatGPT begins. We all knew we were going to end up here. We all knew ads were coming. I mean, I wrote a piece May 2025. I said, obviously, ChatGPT will monetize with ads. Why did I feel so confident in that? Because they just hired Fiji Simo. She's kind of known for having led the mobile news feed ads product to Facebook. That's obviously been one of the most successful ads businesses in the history of mankind. And so it seemed natural that they would bring, given the domain expertise they had acquired, that they would bring ads to bear for ChatGPT.

1:05And they have. Now, the launch looks like the launch of Netflix ads, right? And if it remains that way, it probably doesn't bode well for the business. But I think they're going to evolve it over time. And what they've said now, or what's come out, the information has had a lot of good reporting on this. They're charging on a CPM basis, which is not standard for kind of direct response ads. It's more like a brand average. advertising formulation, but so it's$60 CPM, which is exactly what Netflix came out of the gate charging. They're asking for commitments of less than a million dollars. So this is clearly like a testing phase.

1:38And they're going to offer very little in the way of measurement or targeting. So this feels like a very sort of primitive MVP. But you know, the question is like, how quickly can they evolve it into something that looks more like the Facebook ads platform? So what do you think got us to this place where you get Fiji Simo? She's done this before. She has massive connections and skill set, but it still takes eight months to roll something out that feels not as mature as something else. Is it a talent issue? Are there other regulatory concerns? Or do they do they need to worry about certain edge cases that people might not be aware of?

2:19Or is it specifically strategic to actually, is there some secret reason why they're taking this approach? So here's my hypothesis, right? And so it's not just about, you know, Fiji Simo, she's the CEO of apps, but they acquired Statsig, which is, you know, staffed by a lot of ex-meta people, that was an experimentation platform. That's the exact kind of talent that you'd want to ingest into a company if you're going to roll out an ask platform, which needs to be aggressively experimented on. Here's what I think happened. I think, you know, ChatGP, so OpenAI kind of evolved over time. It was a research lab.

2:53Then, you know, obviously they've had the drama around trying to pivot into a for-profit corporation. My guess is the investors wanted them to bring in somebody that was a little bit more sensitive to, you know, the commercial sensitivities, right? And so, you know, Fiji Simo would be a very good fit for that profile. They brought her in, and my sense was there probably was a little bit of animosity internally towards the idea of this being an ads-driven business, right? Primarily an ads-driven business, which I think it ultimately will be. Only like two years ago, Sam said, I look at ads as like a failure state.

3:29Yeah, and then he started pivoting the rhetoric. started saying, well, I like some of the Instagram ads. When they're targeted well, they're actually delightful. And in the right context, they can be good. And he was coming around to the three of our worldview, which is that advertising is great. But he did have to message that externally. And then obviously had a whole bunch of work to do, messaging it internally. And maybe that's why there was a little bit less energy internally to really push towards that. That makes a lot of sense. Well, he may have been the one that was messaged too. He may not have been the one controlling the message there.

4:00He may have been the one that was messaged too. We need ads because ads are inevitable. It's not that ads are great. Ads are inevitable. If you want to reach what I call humanity scale, ads are inevitable. If you want to monetize that at the scale it can be at its total potential, you need ads. That's how you do it. And so he may have been the one that was convinced. And bringing someone external in who had been there, done that, may have been part of an investor strategy. But putting aside my hypothesis, look, I think this has the potential to be an incredibly successful initiative. The question is, how quickly can they pivot into something that looks like the meta model?

4:32So if you look at it right now, it's – and I think, look, everything you just said mirrors what happened at Netflix. For years, Netflix said we would never do ads. Ads are offensive. They're an affront. They're an affront to my sensibilities as a consumer. That's what the leadership said. And then one day they weren't, right? And the exact same thing happened with OpenAI. And all in. Six months? All in. Don't forget all in. All in was anti-ad, and we begged them. We begged them. We begged them. And now? And I think they made the decision independently, of course, but they're running ads now. Are they always coming around?

5:05Yeah, I've been banging the drum since May of 2025. But I think the question is, is it six months? Is it 12 months? Is it 18 months to get something that looks like meta ads, right? And that means targeting, right, based on behavioral profiles. That means measurement. That means explaining to advertisers the sort of impact of what you've delivered on their behalf. And that means a whole lot of tech that needs to be built out, right? And that means creative optimization, all these things. is just conversion-optimized advertising is very different from what they're offering. And is it six months?

5:35Is it 12 months? Is it 18 months to get there? That's a big question, right? But that'll determine the success. And if you say, look, this mirrors what Netflix did, well, that's not really a playbook that you want to draw inspiration from, right? I mean, they had to pivot. They had to do a hard pivot. I mean, the ad steer has been alive for three years. They did 1.5 billion in 2025, right? So, you know, it's not a significant chunk of the revenue yet. I think it could be a much larger chunk of the revenue, but Netflix has other concerns, right? They've got quality concerns. They've got consumer sentiment concerns where they can't just fully embrace ads.

6:05I don't think ChatGPT is limited in that way. Yeah, yeah. Take me back to that initial launch of Netflix. You said similar CPM,$60 per 1 ,000 views. But what was the response? Were there any advertisers that were like, yes, this is great. I got my money's worth. It worked out. Was it just vague and unattributable? My perception is great performance marketers are always excited to try new channels, especially a channel that's driving high-intent traffic. So if you tell a performance marketer, hey, I've got a billion users, and I'm going to let you advertise to them, they're going to be excited no matter what the general setup is because they just want to get in there with a test budget.

6:45And you'll talk to marketers, and at a certain scale, they're like, yeah, let's try 100K here. Let's try 100K there. It's just like you're just kind of throwing around money and then you're going to double down on what really works. Right. For the initial rollout, this is how you launch an ads platform. Don't get me wrong. This is how you launch an ads platform. That's why I'm talking about six months versus 12 months. This is how you do it. This is how you bootstrap the data. This is how you get people on board. This is how you get feedback. You have to launch it like this. There's no other way.

7:12You can't launch a conversion optimized ads platform because by definition, you don't have any conversions data yet. Right. Now, that's that's another reason why I think they they launched Instant Checkout. I think instant checkout is a stalking horse. I think that's a method of gathering conversion data that they can use to then target against with ads. I think that was the whole purpose of that. Right. I don't think that was seen as a long term revenue opportunity. I think that was seen as a way to bootstrap the data. But this is how you launch in an ads platform. I mean, it's got to evolve into a conversion optimized platform, but you can't be that at the start.

7:40You can't intuit the settings. You can't intuit the tuning that you need to implement to make this work. But the question is, how long does it take them to get there? And so that we'll just see. But my sense is they've got the DNA. The information report, they have 700 ex-Meta employees. Essentially, everyone at Meta is working on ads, whether they're working as an account rep or they're working as an ML engineer. They're all sort of growing in that direction. Yeah. What's been your reaction to instant checkout? There's been a lot of debate on the fee structure. Some people saying no retailer can afford this.

8:15we were pushing back on that in a number of different cases. There's plenty of brands where you say, like, would you spend, like, 4 % of AOV to acquire a customer? I can also see the other side of it. A concern that I have is, like, you know, you're running an ad on a meta platform to find a customer. The customer then goes to Chatshubut if that's where they're searching for products, and then the brand is effectively paying twice to, like, actually find the customer, drive them, get them interested, And then at the bottom of the funnel, they're paying like another 4 % fee. That feels like not great.

8:48But what's been your reaction? Yeah, so I was in the former camp. I said that that's not a workable fee. Let me come back to that, though, because you asked a question about what happened when Netflix offered ads. They had too many advertisers sign up. They had too many. They were supply constrained, not demand constrained. They had to return money. I think that's exactly why OpenAI is saying, no, look, we're setting an upper limit on this. A million is the max, right? Because they won't have enough people to advertise to. This is just a way to test it out and to start scaling. And the problem Netflix made was they never act.

9:15So the problem that Netflix had from day one was they had the partnership with Microsoft with Xander. So they were using external tech to run the ads, right? Xander is an ad platform. And so that's the problem. They couldn't implement, you know, custom placements. They couldn't do a lot of measurement stuff. They couldn't do a lot of targeting stuff because they weren't operating the actual pipes, the actual plumbing that was delivering the ads. There's a lot of, like, data issues, just access to data issues. That's not what OpenAI is doing here. OpenAI is starting out by building their own tech.

9:42And Netflix, by the way, pivoted to that. But they pivoted to that very late. That's why I think the ad initiative at Netflix was not successful and hasn't really delivered that much revenue. But going back to the 4 % fee, first of all, here's what I think is going to happen. I think they're going to drop that to zero or very low. It'll be tiny because I think the whole point of this is to just drive conversions. And the whole thing is like when you see performance in ChatGPT as a retailer who's just basically paying for the conversions essentially with the fee, you're going to want to actually advertise against that.

10:10You're going to want to control it because you have no control if it's just surfaced based on what ChatGPT believes is relevant. And at some point, you're going to have too many retailers to place into a single placement, so you're going to have to mediate that by bids. So ultimately, I think this gives way. In the main, this gives way to advertising. But whether the 4 % is sustainable or not actually comes down to how you view what those purchases are because it's not user acquisition. You're not acquiring a user. If you were, then you'd compare that to LTV. But since it's not user acquisition, you're not acquiring that user with a long-term relationship in mind.

10:45You're essentially trading a percentage of the checkout for the checkout, but that user doesn't interface with you. And in a lot of terms— So there's not going to be pass-through to the underlying vendor, you're saying? Well, no. I mean, look, you're the merchant of record, right? And you get their information because you have to fulfill the order. But OpenAI specifically says in the terms that you can't use your email address for remarketing. And Amazon says this too. So the thing is you're not capturing the customer. You're just giving 4 % for that one-off transaction. But you have no control over whether you reach that customer again, whether that customer is brought into your orbit.

11:22That's a very different proposition from advertising to someone and acquiring a customer. You acquire that customer, you get a whole stream of opportunities to remarket to them. If you're just trading 4 % for the transaction, that's all you get. So it's not actually customer acquisition. If you think about 4 % per transaction, now consider that if you thought about the people talking about, you know, when that number came out, they talked about, well, some people spend, you know, 50%, 60 % on Facebook ads. Why is that different? Because that's over the life cycle of that customer, not for the one of transactions.

11:51Or even one time they pay for the customer one time and then they come back, you know, 15 more times or they end up subscribing and they've actually owned the customer relationship. Yeah. Exactly. And, you know, think about it also this way. you know that four percent that might be all the margin you're getting as an e-commerce ever you know as an e-commerce you know operator right and the thing is that's not preventing you from advertising that's not saying okay you have to save your entire advertising but you're going to still have to advertise everywhere this is going to be a tiny sliver of your orders that you're not in control of you can't control it this could be lumpy this could be up one month down the next right you can't control that you're gonna have to continue to advertise now you have no idea to your point whether this is cannibalizing uh purchases that would have happened from the people that you advertise on other platforms, right?

12:34So this ends up just being a cost, a drag, probably on your advertising expenditure elsewhere. Can you explain attribution and the evolution of Netflix's ad product? Because if I'm sitting on my couch watching a movie, and I see an ad, and I go to my phone, it feels really disconnected. Even in a pre-ATT world, it feels extremely hard. How do they do attribution? How good can this get? And then with ChatGPT, what's the upper bound on attribution for OpenAI, given that we are in a post-ATT world? Upper bound is basically what you see with Facebook, because it's going to be direct response. It's going to be click-based, right?

13:14With Netflix, it's a little different. So with CTV, the way that companies do measurement and attribution in CTV is traditionally they set up what's called a clean room. So I set up this environment, it's a centralized environment. I push all the ad interaction data that I have. So basically when I know I showed an ad to someone, I push all the data into this clean room. A lot of times it's IP address and device information. And then the advertiser pushes that information into the clean room and I match it up. That's attribution for CTV traditionally. Now, there's other ways to do it. There's other ways to do measurement that's just totally probabilistic.

13:44This is like media mix modeling and stuff like that. But with attribution when you want to use that term, traditionally it's done with a clean room like that. So I'm just like linking the data sets together based on some key, a lot of times it's the IP address. But if you think about clicks, like clicks out, right? So that's like I'm viewing, I'm on second screen, I can look at time, like there's time elements, you look at temporal elements to do attribution. But if I'm clicking out, I've got the click, right? I've got the click. And so that carries a lot of information. I can still look at temporal aspects of that, but I can do a lot with the IP address too.

14:13And so that's, you know, if you look at a CAPI, that's what CAPI is. A CAPI is just a way to get this data on the back end so it bypasses the apps you know the bypasses the mobile device it bypasses the browser so there's no way for you know apple to interfere with that with itp or atp or att and so you're able to get you know it's still probabilistic in that way if you consider you know the the decay of the validity of the ip address over time to induce probabilistic measurement but that's traditionally how it's done and then a pixel as well but i mean that's what open ai has at its disposal it's got the capi and the and the pixel because it's going to be click-based.

14:48So there's a high degree of fidelity that'll come with the measurement there, the attribution. That's way higher than what you can do with CTV. What do you think the prospects are for the number of companies that are trying to build LLM ad networks? I got pitched at least a few companies that were maybe a year or so ago saying like, hey, LLMs are going to be free. They're going to insert ads. We're going to build this kind of network to serve ads across a bunch of different apps. I was super bearish on those, specifically just because I thought that OpenAI would want to own the full stack. Obviously, Google would own the full stack.

15:23And I think we're heading towards an oligopoly of sorts in consumer AI. And where's the scale really going to come from? But any kind of takes on that front? Well, there'll be a long tail of agents that are going to want to monetize with advertising. So I think there might be space for an ad network to exist that services them. i don't know why that wouldn't be an existing ad network um i don't know why that wouldn't be or you know just sort of dsp like i don't i don't i don't know why that wouldn't be the trade desk to say um but but yeah i mean i think there's probably an opportunity to monetize some of those long tails of agents with uh with ads but if you think about like a chat of course they're going to build their own i would you know perplexity is apparently restarting its ads initiative um i think these companies it would be silly to not build their own i mean you want to have that ad stack so you can make everything bespoke right and and and and maximally perform it right If we look at meta earnings, I mean, that's exactly what delivered that big beat.

16:18And I think what people forget about the prospects of this sort of like AI enablement with advertising in digital advertising is that these effects compound over time, especially when you're talking about direct response based ad platforms. Right. So like with meta, I've been banging this drum since Q1 2025. These effects compound like people point like 3.5 percent or 5 percent or whatever increase in click through rates or increase in conversion rates from Andromeda or from Gem or from Lattice. Who cares? These are tiny. No, but first of all, that's every quarter they're noting these performance improvements, but also they compound over time.

16:50If I'm targeting 90 % a 90 day recoup on my ad investment and 120 day, 110 % ROAS, what am I doing after 120 days? I'm reinvesting 110 % in advertising on your channel. And so if you're giving me 3.5%, 5 % bumps every quarter, that's just going to end up getting reinvested. That's why we're seeing the growth reaccelerate. Growth is re-accelerating going into Q1 2026. That's amazing. That's why Facebook was up 10 % last night, because growth is re-accelerating. They're going to see growth rates they haven't seen since the post-ATT doldrums. That's incredible. That's incredible growth. And anybody who doubted the CapEx historically has to accept that reality now.

17:29And they did. They did yesterday. Yeah. Yeah. What people, like the criticism is like, Zuck is being so ballsy with CapEx. Specifically on Gen.ai, which they haven't shipped Gen.ai products. And the criticism is like, hasn't launched a hit Gen.ai product yet. Our take earlier on the show was like, look, this guy is serving more Gen.ai content than almost anyone on Earth. Even if he doesn't have like a breakout net new product, his product has massive tailwinds because of all this. Like he's in a perfect position to decide, yeah, I'm going to be one of the biggest spenders in this category and plan however many years out in advance.

18:10They're making more money on AI than anybody except for Google. Look, the idea that they're not utilizing generative AI, have you seen the ads on Facebook? Those are all generative. There was a revolt. There was an advertiser revolt. There was a scandal recently because Meta was being too aggressive with the ads generation. The idea that they're not implementing generative AI is absurd. I mean, first of all, generative AI is not just what you see in the output. I mean, they talked about pairing LLMs with ad ranking. that's really fascinating cutting-edge research in advertising using llm to sort of give you feedback on the ad does this resonate is this something that someone would click on like pretend you're this demographic there's already research that shows that that has a really beneficial impact on advertising it's not just like what you see in your feed but also the stuff that you see in your feed is generated the idea that they're not utilizing generative ai in advertising that's absurd.

19:02You can see it. Just go on Facebook right now. Yeah. I want to move on to other companies, but I want to have one more question about opening. Do you think that there's any, um, any sort of risk to like push back of like the creepy ads, like being too well-targeted. Uh, Facebook went through that with like the t-shirts that said your name on them basically in like your whole career path. Uh, my, my assumption is that eventually like you're only getting ads that were specifically generated for you. Like this idea of like, historically you'd have, you know, one ad that a brand would make and they'd run it at the Superbowl.

19:38They're like, we're going to send this to all of America. And then eventually with, you know, on platforms like Meta, I would assume that every ad is like a one-off generated and it knows, it knows exactly how to position a product, what color, what environment to put it in. And it'll be insane that we used to have ads that we'd make one piece of creative and then run it to like 10 million people because it was like good and it's like no this will all just niche niche niche niche down uh more and more yeah what do you think look i mean it's like the idea that everyone would see the same feed when they opened up facebook is absurd now right i mean like look if if people truly hated creepy ads facebook user numbers wouldn't you know monotonically increase every quarter right and uh you know europeans would be lining up to buy the subscription in europe and when in fact only one percent of users in Europe has taken that offer.

20:30Is that an offer to eliminate? Is that like an ad free like meta offer? Is that what you're talking about? Because I know they're talking about rolling out. So they've got the less personalized ads option in Europe as a result of the DMA compliance. So they have three options. They have a full on subscription, no ads. They've got the less personalized ads and they get the normal experience. 1 % of people opted into the subscription. Like people hate ads. They just hate them less than every other monetization model, Right. People love ads. And if you look at, you know, demonstrated behavior, people love ads.

20:59People love access to free tools. People love access to sort of like endless capacity. Right. You start capacity constraining things or you start putting up paywalls. You know, obviously, by definition, those are going to get used less. Right. Or most likely by definition. And so the thing is, like the idea that like ChatGPT is going to run into this problem that's unique to chatbots, I just don't buy that. And first of all, maybe they will because maybe they'll design this the wrong way. But I give them the benefit of the doubt. I think they can do it in the right way. And one way to do it is to just not tether the ad at all to the chatbot context.

21:33You could just say, look, this is a display ad for what we know you're in market for based on Cappy, based on the pixel. And we're going to make sure it has nothing to do with what you're talking to the chatbot about. So you never have to even question whether the chatbot has your best interest at heart. That's a very easy way to sidestep that problem. But I think they've got very skilled people there. They know exactly what that trap is, and they're going to avoid it. Do you think we're going to get ads in Siri in the next 24 months? No, just because of the optics of Apple doing that. I think Apple would probably love to do that.

22:05I think there's probably people inside Apple that are pitching that, but they won't because it's not in alignment with their optics around advertising. But nonetheless, Apple is also increasing. We were talking, like, I think people will be normalized to ads in Apple products. They obviously are in the App Store today. German was talking about ads in Apple Maps coming in the near future. And then I think from that point, if people are starting to use Siri as a search engine, there will be ads eventually. It will just be to be able to serve ads like basically natively in the UI to iPhone consumers will just be too tempting.

22:43They got to do ads in the alarm clock app. I want to wake up to a different ad every day, get out of bed, start buying, get online. It's time to check out. Yeah. Morning to night, you should be consuming. I don't know. Maybe. I think Maps podcast for sure. But the thing is, I wrote this piece when Apple introduced ATT. I said, Apple robbed the mob's bank. I said, look, this is just an opportunity for them to shift some of the budget that's going to meta into their own ads ecosystem. They did that because they expanded their Apple search ads. They just doubled the number of impressions per search.

23:18They've got two placements now in Apple search ads, and they're also blurring the lines between ads and organic results there. So the idea that Apple hates ads, I think, is mistaken. I think what Apple needs to do is they need to sort of walk this tightrope of appearing to hate ads while also benefiting from ads because they are such a, you know, an accretive sort of margin-expansive way to make money. Right. So maps for sure. I think podcast is another prime piece of real estate for ads to be inserted into. I don't know about Siri. That feels like maybe a little bit too on the nose, but who knows?