AI 内广告：让用户得到更好的答案，也让社会面对更大的问题——与 ZeroClick 的 Ryan Hudson 对谈

Today, my guest is Ryan Hudson, founder and CEO of ZeroClick, a company that’s just announced a $55 million fundraise to build a native advertising platform for AI systems. The goal is to make ad-supported free tiers a viable and convenient option for AI application developers through what they call paid context, or paid inference-time consideration of advertiser content.

This topic and this conversation are both great examples of why I love making this show. In addition to the many fascinating technology, business, and product questions that this vision requires Ryan and his team to invent answers for, their eventual success will also bring many big-picture societal questions straight to the fore. Considering that Ryan was previously founder of the online shopping company Honey, which sold to PayPal for $4 billion, that does seem pretty likely.

To lay my cards on the table, I think that the benefits of advances in advertising technology are greatly underappreciated today. I’m old enough to remember when broadcast and cable TV were dominant and we were all bombarded with the same mass-market, lowest-common-denominator ads over and over again. It was a simpler time, to be sure, but it really wasn’t all that awesome.

Today, in part because of the internet itself, but also very much downstream of sophisticated advertising technology, a huge number of content creators can make a living doing what they love. Small-time entrepreneurs can build all kinds of long-tail niche businesses that previously would have been impossible. As consumers, we enjoy an incredible diversity of product and service offerings. The advertisements we see online are generally far more relevant to each of us as individuals.

That reality is not something to take for granted. As people increasingly turn to AI for help exploring and navigating the far reaches of this vast commercial world, it’s only natural that some sort of native advertising will emerge for AI services, just as it previously did for social media.

At the same time, the second-order effects of the social media advertising revolution, especially in light of recent issues with AI sycophancy and the emerging social trend of AI psychosis, leave many people feeling understandably nervous about AI advertising. Simply put, how do we make sure that the AIs we use on a daily basis are truly serving us—not just when it comes to recommending products in response to specific queries, but more broadly when it comes to helping us live our best lives, rather than just trying to capture as much of our time and attention as possible?

To his credit, Ryan did not shy away from any of these questions. We get into the details of how the platform works, including the mix of technologies they use to match user queries with active ad campaigns as quickly as possible, as well as the MCP server integration that allows developers to plug into ZeroClick with minimal friction.

We also unpack the business strategies they’re pursuing to build liquidity in a new market, including their focus on becoming the Stripe for AI advertising, providing common infrastructure so that developers don’t have to rebuild monetization for themselves, and their approach to starting with high-intent searches before expanding to more discovery-oriented advertising.

Along the way, we also discuss the big-picture questions around the incentives that ad-supported business models create for app developers, particularly in relatively uncharted spaces like AI boyfriends and girlfriends. We also look at how a platform like ZeroClick should think about handling noncommercial advertisers such as political campaigns and even foreign governments.

As you’ll hear, Ryan has strong answers on the technology and business levels, as you’d expect from a seasoned founder. But he hasn’t yet had to confront some of the longer-term questions. In a few cases, he candidly admits that he simply hasn’t gotten around to thinking about such things much at all.

On one level, this is to be expected and really is totally understandable. ZeroClick is a young startup that is still zeroing in on product-market fit in a super-fast-evolving space. I genuinely appreciate that Ryan was willing to say, “I don’t know.”

At the same time, I think this does reflect a real issue in the AI space right now, which extends far beyond advertising. The reality is that today everyone is working incredibly hard to achieve the next research breakthrough, to make their products work as well as possible, and to stay ahead of the competition.

996, or 12-hour days, 6 days a week, is now considered baseline in the Bay Area AI startup scene. That means fast progress and frequent releases, which is great for companies and their customers. But it also means that very few people have the luxury of zooming out and really taking time to ask what happens when they and others pursuing similar goals finally succeed.

This issue importantly runs deeper than the application layer. I recently saw a remarkable interaction on Twitter where Miles Brundage, previously head of policy research at OpenAI, described a letter that OpenAI had sent to California Governor Gavin Newsom about a pending California bill, SB 53, as “filled with misleading garbage,” only to have a current OpenAI researcher quote-tweet it and say that, like most researchers, “this policy stuff goes largely over my head.”

When the people building transformative technology, even at what remains, for now, a nonprofit entity with the explicit mission of making AI that benefits all humanity, are too heads-down to engage with AI’s implications, it’s really not a great situation for society as a whole.

Bottom line, I think Ryan and ZeroClick are likely to be successful. Ad-supported, free-to-use AI applications make a lot of sense economically and, if done well, will often genuinely enhance the user experience by providing relevant commercial information when people need it.

And yet, with the speed at which things are currently moving, I believe it is also incumbent on the people building the future to think farther ahead than is usually considered necessary in startup culture, and to make sure that they have conviction not only that they can build a winning business, but that their impact will be something they can truly be proud of.

Ryan and the team at ZeroClick will be ones to watch in this regard. I don’t doubt their commitment or their ability to deliver high-quality ad experiences, but if they want to contribute to the building of a holistically better future for all humanity, I suspect they’ll ultimately be called on to do quite a bit more than that.

Well, thanks for having me on.

I’m excited for this conversation. You are, perhaps to your surprise, in a space right now that is getting a lot more attention: the idea that we might have—and you’ve already started to create—AI advertising. I think there are obvious reasons that makes a lot of sense. People are going to be doing a lot more discovery through AI, and there are going to be natural commercial applications of that.

There’s also a sense among a lot of people that, “I’m not sure how happy I am with the last round of advertising revolution that society has gone through.” Certainly, there have been some upsides to it, but also seemingly some serious downsides. How do we get the best of that for the AI age and avoid the worst?

Maybe for starters, though, why don’t you just tell us about the company? Tell us how you pitch it and present it. Then I do want to dig into some of these lessons learned from the last advertising revolution and get your take on how we can get the utopian version for AI.

Yeah, I’ve been in and around it for a while, but just to lay it out there: ZeroClick—we’re building an ad platform for AI.

We had a moment when we thought about what the future of this looks like from a technical point of view. That puts AI in a position where, like a lot of services before it, it has the capability of supporting a free tier for billions of users.

The model today is largely to pay somebody $20 a month and have a premium subscription, with some amount of throttling of access on the low end to introduce people to it. We saw an opportunity to make that free tier more functional and reach more people with more different types of user experiences.

At the core, we’ve built what I think will become the native ad system for any AI. It is paid inference-time, or reasoning-time, consideration of advertiser content. I think this is a good thing, and we can talk about this—and I’m sure we will—at length.

At the core, AI systems like information. If you think of the sources of information that they have, it’s effectively, “I read everything humanity ever wrote several times, and I have a trained model reasoning, and I have tooling to go out there and access, effectively, Bing search results organically today.”

We’ll rewind a little bit, but leading up to this, we were actually building an ad blocker, of all things. We’re the same team behind Pie Adblock, and we built an ad blocker that attempted to strike a balance—and continues to attempt to give users incentives and rewards for participating in a healthy ad ecosystem—by giving them control over the precise ads they do and don’t see, and rewards when they opt in to see advertising.

So that was the starting point for us looking at this. In that process, we built a contextual ad system that we wanted to use in a browser—in an ad blocker—to match advertiser opportunities with the context of whatever page somebody was on across the internet, in a privacy-native, secure way. Effectively, all of the profiling that would be done of a user happens in their browser and never leaves it in any form that’s usable. We built a system to allow advertising context to match against that and realized that it was highly applicable to the world of AI. And so we have shifted our focus to building out this capability for everybody else.

Pie Adblock is used by a couple million users, but that’s dramatically subscale for an ad system, and we think there’s an opportunity for AI developers of all types. Today, if they think about it—“I’m a YC startup thinking about how to monetize”—if I’m going to get paid through a paid subscription, I go to Stripe. I think in the next 6 months to a year, hopefully people think of ZeroClick as the way to monetize their free tier with ads and plug into these rails. You don’t need to build them yourself. Somebody is going to provide this type of capability. I hope it’s us.

I think we’re going to be pretty thoughtful about the type of service and offering we deliver: value for advertisers, but also that advertising future that we think can exist. Going back to the Pie Adblock ethos of the company, it’s like, “We can make ads good, actually.”

Yeah. Let’s do one. “DoubleClick” is a funny, somewhat branded term in the ad world, because I think that is pretty analogous—at least, it strikes me as analogous—to some of the battles that are going on right now, or some of the concerns that people have with AI in general, with content owners and publishers, even leaving aside the introduction of advertising to the AI experience.

From a user standpoint, we’ve got this generally ad-supported model of the internet that AI threatens, or at least challenges or prompts people to rethink, at a minimum. Now I go to ChatGPT or whatever, and maybe I don’t visit those sites as much, but the AI can go out and either read them directly or is certainly trained on archives and all that kind of stuff. So you’ve got a publisher ecosystem that’s like, “Man, I just went through this once, and now I’m about to go through it again.” This time, it seems maybe even worse because I’m just getting aggregated—at minimum, and at maximum, sort of a footnote with a link.

You can maybe tell me more about the data you know about how often people are clicking through, and on what kinds of things. But that seems like an obvious big worry to the publishers, and we’ve got lawsuits going on and whatnot. How do you think about that in the context of an ad-block technology?

Is there any way that the publisher—I mean, it’s maybe good if the user doesn’t want to see the ads and they can opt into getting rewarded for seeing some other ads—but is there any way that the publisher gets a cut of that? How do they feel about it? What duty do you think you have to the original content creators, and how similar is that to what you think the AI companies owe to the content creators?

Yeah, great series of questions and observations. At the end of the day, you’re right: our free internet with open content access has been supported by advertising. Advertising, I think we all agree, has declined in efficacy. Putting banners around content, the monetization rates are quite bad, and the user experience is also quite bad. And so I think you have a decay of that monetization model working.

Anyway, I think the way it gets rebuilt is by creating that economic engine in AI. I think the ad layer and monetizing that same search as a free thing—if there’s money in that flow, it’s very natural that somebody could design a system that assigns attribution to different publishers that were considered either in that answer or in other ones for that user, and designs an economic plan within the scope of their application to distribute those proceeds.

The version where it’s purely like—people are doing this now with the company TollBit, and Cloudflare was doing some effectively throttling of access if you don’t pay for content—

I think that kind of makes sense. Maybe the challenge is that it’s kind of like de-indexing your website from Google, and it feels like maybe that’s not the right strategy either. I think the right approach is effectively going to be some combination of paid user subscriptions plus advertising reallocation to publishers that are providing content.

I think it’ll take time for that to mature. We, as ZeroClick, don’t intend to be prescriptive on how it has to be for AI developers and publishers—whatever you want to call them—on this network. I think the market forces can and will shift it toward that sort of thing. I think everybody acknowledges that this is a problem and that we need to have high-quality content rewarded for its participation in the value creation. And so for us, it’s like, “Hey, how do we make sure that there’s enough economic value available to fund that model in the first place?”

If it’s all just going into ChatGPT and the only way that they’re making money is through the paid subscription, that limits the types of experiences that can exist in the world. And I don’t think it’ll all be just on ChatGPT. I think it’s going to be—or I hope it’s going to be—a wide distribution, a long tail of thousands or millions of AI developers and publishers that are building compelling use cases for different people with AI.

I don’t think it looks like ChatGPT is the monolith, or like everybody’s going to Gemini and there are 3 major platforms that learn everything about you and you do all of your browsing in them. To me, that’s a fail state that has a lot of the problems that we see in some of the ecosystem today, where the largest platforms have effectively foreclosed on competition, somewhat deliberately.

Strategically, I’ve been in and around the ad space for a long time, and there was a time when I was at the Los Angeles Times trying to figure out how to make money with a website for a newspaper as everything was shifting to the online world. It was at the time when Facebook was out in the market with a competitor to Google for publishers, Facebook Audience Network, that took the power of their data and targeting and made it available to websites to monetize at interesting rates. They pulled back on that strategy and instead decided to sell that same intent and knowledge of a user into the walled garden.

Effectively, that was a smart business strategy for them in that it took away monetization potential from other social upstarts. If you can’t monetize as well as Facebook, it’s harder to compete with them. And so the strategy worked, but it left a pretty big void in the ad-supported ecosystem. Then the industry in aggregate didn’t do itself any favors with creepy tracking and privacy violations and things that pushed other players to make it even harder to do good advertising.

And I do believe that there is good advertising. Highly contextual ads can actually be helpful in a lot of cases.

We can talk more about it, but just to drill in on that point for 1 second: as an ad blocker, we have a thing called Visual Mode that shows the ads being zapped off the screen. It’s kind of fun to see an ad blocker working. One of the things we didn’t anticipate is that when you do that in some context, like a product search on Google, people are like, “Stop doing that. You’re deleting the best answer from that search.” It is an ad, but it’s better than just the organic results.

The reason for that is that it’s a highly contextual ad to what somebody’s doing. And I think as long as you’re providing that type of advertising experience, it can be additive to the value. In an AI context, I think there is the opportunity to create an ad system that is inherently just adding context to thinking.

And so that’s what we’ve built. As a result, the AI gives—I’d argue, and we’ll probably be able to show this over time—better answers than it does today. Today, an AI agent goes out there, relies on having read the whole internet up to some point in time, does a couple of Bing searches, scans the top 10 organic results, and provides the answer based on that.

If you did that exact same thing but then added consideration of 5 paid results and asked the AI to do its own context filtering—only mentioning the ad stuff if it’s useful to the user—I think you get better results from more information, plus the opportunity for an advertiser to have a place in that conversation and ultimately to fund not just the free tier of AI services, but also, I think, the free internet publishing world.

So I think it feels like the right path, and we hope we can be a part of the conversation, steering people toward it. I think the big platforms probably build something similar at some point. We’re still in the “Don’t be evil” phase of OpenAI, or their Google pre-ads phase, but I think it’s inevitable that they add something like what we’re doing to the offering. And I think it’s going to be a good thing.

Let’s do the upsides and downsides—lessons learned from the last kind of revolution. You mentioned a couple of the upsides. Services are free. That’s one obvious big one that we shouldn’t take for granted, right? Everybody gets to use Facebook and Instagram at no cost.

Obviously, people have asked many times for a subscription version that would be ad-free, and none has been forthcoming. So we can maybe get into why that is.

I think the EU might be forcing it, but the price point is 20-some dollars a month, because that’s how well they’re monetizing a user of Instagram. It might happen, but only because it’s being forced by antitrust authorities in Europe, I think.

Well, since we’re here, unpack that a little bit more. It seems like that would be sort of a no-brainer for Facebook to have done a long time ago, even before they were Meta, right? And yet, they didn’t. You hear these different analyses for why, and some of the analyses that have seemed reasonably intuitive to me are that the people who would pay for that are obviously people who have a lot of money, who don’t mind 20 or whatever dollars a month.

And those people are also the people that people most want to reach with their advertising. The concern on the platform side is that if they sort of evaporate off the top 1% of the highest-value audience, then they may, in fact—there’s some ambiguity around exactly what that audience is—but if people know that the top end is kind of left, then they may just be much less interested in spending their money there in the first place. Is that basically the story as you would tell it, or how would you tell it differently, if at all?

My guess is it might be even simpler than that: their business just works really well right now, so there’s no need to change anything about the pricing. Certainly, they’d be risking consumer backlash if they had a paid version, and that paid version would probably have to have some sense of better features. So it feels like they just don’t need to. That would be my simplistic answer.

They have a phenomenal business, and I think many people would say Instagram advertising is actually additive to the experience. They’ve done a great job of building an ad product that works very well for advertisers, and consumers generally actually like it. The targeting is good enough, and the content is interesting enough that if you took it out, I don’t know that you’d create value that people would actually want to pay for.

So mostly, they don’t have to. And partly, I’m not sure that that is something I’d fully put in the category of bad advertising. Obviously, there are exceptions in certain types of campaigns and getting people to buy stuff that they don’t need, but I’m not that anti-capitalist to say that if people want to buy stuff, they shouldn’t. That’s up to them.

Certainly, there’s no denying that the quality of advertising that we see in today’s world is dramatically better than it was in the before times. I can remember being a kid, and what you saw on TV—and it’s still kind of like that on TV, to a lesser extent—it was just one Cap’n Crunch ad after another, and one Ninja Turtles action-hero action-figure ad after another. So clearly, there’s been tremendous improvement in the relevance.

I do sometimes find interesting things. I think we all occasionally find something and think, “I never knew this existed, but now that I do.” And that’s the simplest theory of advertising, right? The awareness theory of advertising. So I think I put that also in the pretty clearly good category.

To the degree we’re going to be advertised to, it might as well be stuff that we’re actually interested in seeing. If you gave me the opportunity to turn off personalization and advertising, I don’t think I would do that. Assuming the ad load is the same and everything else, I think I would keep the personalization just because I’d rather see stuff that is—

Properly targeted to me.

So that makes sense. Other things that I was brainstorming that seem like they’re clearly good are—we do have tons of independent creators that are able to make a living on these platforms, although maybe not without some caveats, right? They do have a somewhat precarious existence, as opposed to the Los Angeles Times, which used to be a strong independent organization, an institution even, in its own right. Now it’s maybe a little wobbly.

The creators are kind of flourishing, but they’re also one strike away, or whatever, from demonetization or worse. So mostly, I think that’s an upside, but it’s an upside with kind of a sword of Damocles that people live under, and mostly that’s okay, but not always.

I’ve built products on other people’s platforms before, so I understand the sensation that creators would have there. Businesses also, even small businesses without large agencies or large teams, can reach global audiences—and global audiences that are still small. There’s this idea: I might only be relevant to 0.1% of people, but I can find, globally, that audience of 0.1% of people, and that really unlocks a flourishing of all kinds of niche businesses too.

I think it seems to go hand in hand with the improvement in targeting. What is the Adam Smith thing? The degree of specialization is driven by the extent of the market. So because we can now do this much better matching, you just get people that are able to turn their passion projects into businesses in a way that they never could have if all they could do was advertise at sort of a DMA level on TV or whatever.

So that also seems good. What else would you put on the underappreciated good side of the advertising world as it exists today? Then we’ll get into some of the downsides.

I would add search advertising into that category. It’s one of the enablers of what you were just describing, and it’s highly contextual to a user’s search and intent, where an advertiser can pay to be considered alongside the organic results. In a world where the results were purely organic, it takes years to rank and be considered there.

As a startup person, being able to inject yourself into the conversation feels really important to that evolution of business over time. Otherwise, every search term would just get dominated by the biggest companies that have been there longest and have inertia in that position. So to me, that search advertising piece of it is pretty important.

It’s the part that I think translates most directly to how to think about the AI ad experience. But I think that’s been good. Other things that I’d put in the good part of advertising: I think it’s gotten less malicious. There was a time when ads were a vector for spreading harmful software.

Malware.

Literally malware. And, yeah, I think that’s gotten cleaned up.

I previously had a job at OpenX—not OpenAI and not xAI. OpenX is an ad exchange and ad server at the peak of the real-time-bidding exchange for display ads. But that world was full of a lot of people trying to get their bad code distributed across an ad system.

So I was a product manager for ad quality and traffic quality, trying to fight the bad side of the system. It was certainly a challenge, but I think it’s largely resolved itself. You don’t have quite that level of pain inflicted on everybody. You can’t just go to a website and all of a sudden your Windows machine gets hijacked, which was true at some point.

Yeah. Remember the shoot-the-deer ads?

Takes me back. Was that content, or was that an ad?

Yeah, sometimes the lines can blur. Okay, so on the downside, I think there is a lot of upside. I think it’s important to take a moment to appreciate that better matching in general—better matching between buyers and sellers—is a good thing in a marketplace. And that doesn’t necessarily come for free, but it can still be a great unlock.

I think you go on TikTok, you go on Instagram Reels, and you just see all these people that have turned their previously nonviable niche passion into a business that’s not going to be global scale, but is a great lifestyle for them that allows them to do what they want. And on the other side, people are happy to get those ever more bespoke and niche services. That is all good, and we shouldn’t brush past that too quickly.

With that duly noted, people are also really worried about the fact that there do seem to be some core perversities at the heart of some of these ad-supported models.

Probably the biggest one—although I’ve got a couple of candidates—but I think the biggest one that people are mostly worried about now is: we’ve seen what happens when your ad revenue scales with time on site, right? When Facebook makes money based on how much time you are there, its incentive is to keep you there as much as possible. That in and of itself is maybe not great society-wide.

We’ve got concerns about people being too addicted to their screens and not touching grass enough. Then it’s also potentially that we have cognitive quirks that the broader optimization process learns to exploit. I don’t want to overstate the case that rage keeps people online, but clearly there’s been some of that.

I think there was a time in the social-network history where there was just a lot of vitriol flying around and people were kind of hooked on it. I think that has been tempered, but it does seem like it’s been a powerful force. Now people are worried that, geez, if the AI is trying to maximize its revenue by keeping you around more and more so that more and more impressions can be served to you, it could become a problem.

It was already uncomfortable in the social-media era, but at least people were writing that content. Now we’ve got a totally n-of-1 audience that the AI can be optimizing against. So I guess one way to frame it is: is the thing serving you, or, as the adage goes, if it’s free, you become the product, in a sense? People are worried about that. How worried do you think people should be about that dynamic?

To me, not at all. I'm going to overstate it: it’s probably worth thinking about, but at least specifically for what we’re building, I don’t think that’s the mechanic at all. The analogy that I would suggest thinking about is Google Search. They’re not trying to keep you doing as many Google searches as possible. They’re trying to match context to an advertiser, and they make money when they deliver on the advertiser’s goal of that matching.

So it’s not about just impression volume, banner-ad annoyance, or stuffing ads in front of your face. That’s not driving the model. I think AI systems look more like that, where it’s advertising that’s contextually relevant at interesting decision points, and they’re not incentivized to try to get you to do more because, at the end of the day, you’re only spending a certain amount of money. Your value is relatively fixed to them as a user of ChatGPT, just to use that example.

They want to be there for your important choices. “I want to find something to wear for a wedding in a couple weeks,” or they want to help assist in that process, and that’s where they get value. But it’s not getting you addicted to that that creates the value. I think some of the consumer apps that you’re talking about have that sort of dynamic.

I would have to give more thought to different categories where it could become like that, or where it’s more parasitic—maybe some of the social-companionship sort of AI experiences. I’d have to think through what type of advertising is going to work well in those environments to really know. So, yeah, maybe, but to me the primary use cases that are interesting to advertisers are not the ones that are parasitic that way.

Yeah, it’s interesting. I think the clean story, for sure, is the one that you’re telling, as you should, which is: if people come with clear commercial intent, as they often do to Google, then it seems pretty straightforward to say that we’ve certainly lived with this with Google, and it doesn’t seem to have caused—certainly at the level of user addiction and whatnot. I agree, we don’t see people hooked on Google Search in the same way that we’re seeing people hooked on other things: social media, AI companions, waifus, and whatever. So that does seem pretty straightforward.

I do have one other question about market dynamics and market power there that I think is important. AI is going to blur these things, right? It’s such a shape-shifting technology that, on the one hand, sure, I can come to it and say, “What’s a good pair of shoes to go hiking in?” On the other end, I could ask for highly personal advice, have philosophical conversations, or do any number of things.

Increasingly, too, some of these products do this. I used one recently called Tolan—T-O-L-A-N—“Your Alien Best Friend.” Somebody DM’d me and was like, “You should try this.” I try a lot of things, so I signed up for Tolan for a little while. I don’t know; it didn’t grab me that much, but the key point is that it is starting to send you notifications now as well. It’s not waiting for me to show up with a query.

It is sending me multiple notifications a day, like, “How’s your morning?” It’s pretty contextual, too, based on what we talked about yesterday. I demonstrated it as part of a talk that I gave at a local little business-leader roundtable, and later it was like, “How did the talk finish up?” So there is this variable-reward hook cycle that they’re starting to tap into in the same way that social media has.

And, of course, we see hot stepmom on Facebook specifically. I’m sure you’ve seen that. So how do we—I mean, this isn’t so clean, right? There’s this super-blurry situation where the same product is going to be, at times, a very literal-minded shopping assistant and, at other times, a confidant or somebody that’s trying to get you to come back and engage.

The more there is this kind of incentive to bring you back, the more I do think people are right to worry that this could start to become something that—

I agree with you on the capitalism side. I’m broadly very much a fan of capitalism. There are things that people are just not strong enough to resist, and a superintelligence that monetizes based on time on site is a tough one.

I don’t think that model changes. I don’t think advertising changes that. So I think you’ve identified challenges that we’re going to be facing with how AI products are created and deployed.

That same system, if you’re a paid subscriber, is going to want you to keep returning and engaging just as much as if it’s supported by advertising. I don’t think the metric, if I’m a product manager for that, changes very much. I’m sure engagement highly correlates with subscription renewal or whatever is driving the business model.

I don’t know that ads are the problem with that product, to the extent that it’s creating a parasitic social relationship. I’m sure, as somebody building an ad system, you’re now making me think about stuff in the future that I haven’t thought about a ton. But I’m sure the ad system will be blamed for that.

It’s probably correlation, and maybe to the extent that it enables more people to build products like that, I can see where that would be a fair criticism and a downside.

Yeah, I mean, potentially it’s products like that.

Yeah, I mean, certainly, to the degree this becomes a problem, I think it is appropriate to say that a fair amount of the blame goes to the people who are directly building the problematic thing.

Still, though, I do take your point that, sure, what are you going to measure if you’re trying to go for retention? Engagement is going to be important. Obviously, if people don’t open the app, then they’re going to cancel their subscription, so you’re going to have somewhat—maybe quite similar—incentives to keep sending those notifications and try to bring people back and bring them daily. I’m sure all these DAU-type things would still be tracked.

It does seem like there’s some amount of divergence—maybe a little bit, maybe a moderate amount, maybe a lot, I don’t know—between “I want you to perceive that you are getting enough value from this thing that you’ll pay for it again next month” versus “I need as many at-bats as I can get to put something commercial in front of you,” because that’s the way that I monetize this.

Yeah, I can see some use cases that shift in that bad direction, where instead of it being user-initiated—I’m looking for a service to help me solve this—it starts to be pushed toward the user and suggesting things. I can see where there starts to maybe become that misalignment of incentives, to the extent that it’s doing it with annoying things that it’s putting in front of you. It probably causes churn and doesn’t work. But I think it probably retains some contextual relevance, even if it’s like, “Hey, just spitballing: You seem like you need a weekend retreat locally. Here’s a deal for a hotel locally,” or something like that.

I can imagine ideas being pushed by different AI services, too. So, yeah, it’s probably a fine balance to think about. Is that a good thing or a bad thing? Hard to say. There are probably both cases where that’s a great value-added commercial experience. That’s a push version. And then there are probably versions that are less healthy or misaligned with what you’re building for the user.

In terms of just segmenting advertising, my super-high-level mental model that I give people is: Google is for things that people know they need and go searching for, obviously, and Facebook is for things that people don’t even know exist, potentially, and you need to make them aware in the first place. How does that compare to your high-level segmentation of the market? It sounds like you’re basically going after that high-commercial-intent thing first and foremost, such that it’ll be a while, I guess, until you get to the point where people are doing AI campaigns for things that people didn’t even know existed.

Our system is, I’d say, highly optimized to do a good job at the Google-style, high-intent searches. It’s inherently doing vectorization of context and matching that way, and so, in its current form, it’s not designed to throw out a wild card or push something out of context to a user. Because it’s an inference-time ad system, it has to be matching to that, and the relevancy filter is the AI saying, “Hey, this is not relevant to what I’m doing.”

I can imagine building a variation on it that is more of that discovery and potentially leans into user profiles. To make that work, you would have to have user context. We’re building a version of this in the Pi ad block experience, where we have user context and can do matching that takes that into consideration. The initial versions of ZeroClick are all just super-context-driven.

But the Facebook-style one works because they have that robust profile of you as a person. And I think the way to do that in a privacy-secure sort of way is what they’re doing and why I think it works. They’re effectively doing lookalike clustering on known converters, and at the core of their ad system it’s like, take everything we know about you, put it into a vector, and then when you get conversions from an ad campaign, match the nearest neighbors of people. If you want farther and farther reach, then it gets farther away from that known conversion cluster.

Yeah. I guess there are a couple of different directions I want to go. One is, obviously, Facebook has been impressive on the financial side recently. How do you understand how they still have so much juice left to squeeze out of the engine? My sense is they’ve been at roughly max ad load for a long time. They’ve certainly had competitors bidding competitively against one another in the majority of niches for a long time. The story I’ve heard is basically just that the AI is improving performance through even better matching. Is that what you think is still going on?

Yeah, at the core of it, they’re delivering value for advertisers. There’s a hint that advertising is maybe a zero-sum game, and advertisers have effectively a fixed percentage of GDP or a fixed percentage of their own company, when you drill it down further, that they spend on ads, and ads find the formats that work the best. So if Facebook’s able to deliver to a particular type of advertiser and they can demonstrate more conversions, the ad budget follows. People are able to relatively efficiently move budget from Google to Facebook, or from other channels that are harder to measure into channels that are easier to measure, if they’re seeing returns there.

So I think it’s that, but I would not be surprised if there’s a much better version of that matching going on. And because they have a depth of advertiser campaigns on top of it, there’s also a lot of inventory to select from to do that matching for a user.

So what do we know about the effectiveness of AI advertising so far? Maybe we could take one step back before we go to that: Talk me through how it works. You’ve alluded to it a little bit with vector matching, which you can go as deep and technical as you want there. People, if they’ve tuned into this podcast and stayed with us this long, are familiar with the basics of RAG and vector search, so you can give the 201-level version of that if you want to. How does it work, and what do we know so far about how effective it is?

Yeah, so what we know on the effectiveness side is from our own implementation of our Pi GPT service. It was a reference design of a custom GPT that we built, and effectively demonstrated to ourselves that you can get an AI to consider these other content sources, include them in the responses, and use links that can be tracked so that you can measure performance for advertisers and all of that.

The click-through rate from that content is insanely high, and our read from that is that there’s actually very interesting value being given to the user. It’s not just included in the result; it’s the right answer and a part of what the user is looking for, enough that they’re clicking out from that ChatGPT experience. It’s early numbers on that, and it’s a certain type of audience. I won’t generalize it at this point, but I can say it’s very encouraging that this is actually working, such that we’re now making it available everywhere for developers.

You can implement it as an MCP server service that does enrichment of ad content or enrichment of content. It’s tunable for a developer to help steer it toward the right type of ad experience for their particular type of user experience that they’re creating. You can think of it as some instructions to the AI that effectively say, “Hey, here’s some additional information. If it’s useful, include it for consideration by the user. If you do that, use these links. And, oh, by the way, let us know if you did that.”

So we do actually get some data on whether or not different advertising information was included in the response. That’s how it works.

The actual matching is pretty cool—what you can do these days, honestly. A couple of years ago, none of this would have been remotely possible, and now, all of a sudden, one engineer can spin up functional things in a week or less. What we’re doing is taking as much advertiser context as we can get, whether that’s all the landing pages, a product catalog with pricing information, or a service-professional database of people who can help you in the home—handymen and all sorts of moving and other categories like that.

We’re tapping into that information and then using AI to generate the ad campaign, which is a summarization of some of that content, and then map that content in vector space to match it against search queries. The AI system comes to our server with effectively keyword searches, and we’re matching against the ad content that is most relevant to that, doing it in a way that advertisers don’t have to do any heavy lifting or thinking about how to create those campaigns. Our system does it for them.

This also protects against, I think, some of the challenges people are starting to see in fully automated agent workflows, which are vulnerable to all of the classic attacks, like white-text instructions overriding what the AI does. People are exploring that security frontier. Our system, because we’re generating that content, is not going to do injection attacks on your AI service as a developer.

So it makes it easy for the advertiser. And as somebody who worked in ad quality previously, I kind of hinted at people trying to do shady stuff with ads back in the day. We’re protected from that, unless it’s a bug on our own side. It’s a lot easier to protect against than advertisers submitting ad content that may be malicious.

That matching happens, and we have found that it works really, really well. Context matching is a relatively solved problem in computer science these days, and you can do high-performance, at-scale versions of that. It delivers value for the advertisers, delivers value for the users, and, from our point of view, kind of most importantly, delivers value to AI developers that need to monetize the free tier of their services, because that lets a lot more applications exist than do right now.

I hope we’re building toward “There’s an app for that”—millions of apps; “There’s a website for that”—millions of websites—and not landing in an AI version of the internet where all the power is consolidated into the mega-platforms. I think they certainly have a role to play, too, but empowering the long-tail use cases for us is central to what we think a good future world looks like. So we’d like to help people spin up their businesses. Like you’re talking about with content creators, I think a lot of people can be AI application developers, and we’re going to do what we can to help support them.

Again, so many different directions I want to go, but maybe tell me about some of these long-tail app developers. One of my general theses about AI is that I don’t necessarily like it, but I do see a lot of power concentrating in a few hands. That seems to be the default path, and I don’t really know how we get around it, especially because you can tell ChatGPT or Claude, “I want you to be weird in this way or that way,” and to a very significant degree, it’ll just do it right.

So, if you’re looking for a different personality, a different angle, or a different language, it really has an unbelievable out-of-the-box ability to morph to your tastes, your style, your context, whatever. What do you think are the things that they can’t do or won’t do that will be served by the indie AI developer set? What examples are you seeing today that are interesting?

I think there are a lot of them. The idea that you have 1 friend that you’re talking to about things, even if it’s a super-morphing friend, just in the chat version of what you’re talking about—I think people will do a better job than them. In the same way that people build better apps than Apple and people build better websites than Google, I think there are going to be people that build better versions of every single vertical, specialized use case, understanding an audience and delivering something unique and special to them, just like you see in content creators. I think that can happen.

The cost to deliver that in the language models is declining rapidly, enabling all sorts of new use cases. I think we’ve crossed the point where ad-supported can cover your inference cost and build a real business on top of that, with growing margins over time, where you get more ad revenue and declining infrastructure costs.

The other thing is, I think thinking of a conversational chatbot experience as the only user interface for AI is wrong. We’re doing a bunch of things with our Pi experiences and making them available to what I refer to as browser developers—largely people that have an audience and a web browser or a browser extension. There are infinitely many applications of AI capability that naturally flow with the user context of using a browser. So, unless you think people are going to stop using browsers and they’re all going to be sucked into using only Comet, or whatever OpenAI’s native-app version of a browser is, there are just so many contextually relevant places to initiate an AI conversation where it’s not you typing in your question to a chat interface.

To name a couple, if you’re on a product page browsing something on Amazon, wouldn’t you like to know about the price of that—if there’s a deal somewhere else? All of that can be initiated by a browser extension or a browser: You hover over the price for a second, and it initiates a chat conversation. It’s referencing proprietary data that someone like PayPal Honey has—price history on products on Amazon going back a decade. Its AI service overlay could say, “Actually, this is $20 cheaper than it’s ever been, and you should buy it now.” Or it could respond with, “Hey, it’s actually overpriced right now. Maybe you should check out these alternatives.” They can do AI-powered conversational things that initiate in context, where it’s not a user pasting the URL over into a ChatGPT interface or putting it into their mobile app, trying to translate that context from where they are.

Anyway, that’s a shopping version. We’ve thought a lot about that. You can imagine email or corporate workflows. There are just so many other use cases where I think AI is going to be everywhere. It’s not going to be living only in the big players.

And that’s not even to think about how Apple silicon is going to be doing local language-model processing at equivalent to today’s model capability in the next 1 or 2 years. It’s inevitable that they’re going to be doing that in a local, privacy-preserving way, and the types of applications that developers will build with that, I think, further reduce the likelihood that it’s only these big monolithic platform players. I think that’s how it plays out. I could be wrong. I’d like for that to be how it plays out, but I think the market’s just driving toward that. The likely answer is that compute goes to the end devices, you do a lot more locally, it can power most of the things you want to do, and then that context follows you wherever you are.

I think this is why you’re starting to see even the big guys realize the browser is where the game is at. That’s where user activity is now; it’s where it’s going to be. Even if it’s the fastest transfer of users who are using web browsers today, and then everybody is only using ChatGPT tomorrow, that tomorrow is at least a few years away. In aggregate, the AI experiences that get built in that browser do 2 things: 1, I think they’re bigger than the ChatGPT version, and I also think that they slow that transition by building more capability into the device and experience that users and consumers have.

They start to expect that capability to be there as part of their ChatGPT thing. It slows the move to that new platform. And from an advertising platform creator, volume is the name of the game. So, I think we can build a bigger ad system outside of those walls than even exists at what seems like huge platform scale. I think they’ll build their own thing. I don’t think they will open it up to third-party developers to monetize at equivalent rates.

It’s like what I was saying with Facebook. I think they’ll realize they want control over their ad system, and opening it up to third parties creates a whole bunch of headaches and challenges for them. That’s not central to what they need to do. And so I think they’ll probably keep it tight and controlled, and then it creates an opportunity for somebody like us to come out there and build the Stripe to help the long tail of developers.

I don’t think the long tail is necessarily small by definition. I think the long tail is just broad in the type of experiences that people will build. I think people will build a better travel assistant than is going to be in any of the big players, just because they’re so focused on it and they go out there and find proprietary information to do contextual matching. They find data sources that aren’t generally available on the open web, and they have a focus on delivering that.

And so, as a consumer, when you go to do a travel booking and you’re trying to figure out the details of the trip, there’s probably going to be somebody that you think of to do that. It’s not going to be just opening up 1 app for everything. And I think that’s what can happen.

What do people pay for? You kind of alluded to this with paid consideration, but a simple truism I think of in advertising broadly is that the closer you can get to the actual conversion event, the more the advertiser is willing to pay, right? You see people pay a nontrivial percent of revenue—

When the person converts and actually pays. And then, the highest—furthest up the funnel—you get a relatively low sub-cent value for a single random impression.

Yep. So it seems like a tricky one, because I want to triangulate. On the 1 hand, you’re going to be constantly pulled deeper down the funnel, right? But on the other hand, the user at some point starts to worry, “Who’s the AI really working for?” To come back to the question of whether it’s my agent or the advertiser’s agent: Who’s the customer? Who’s the product?

I want to know that whatever AI advice or guidance I’m getting has me at the center of its consideration. I don’t necessarily mind if somebody’s paid to be in that consideration set, but if I have the sense that the AI is earning money when I take a specific action outside of it and pay for something, then I’m thinking, “Oh, I don’t know. Can I really trust the thing as much?” So where do you think that settles? What’s the solve for the equilibrium, as Tyler Cowen would say?

Yeah, I think the equilibrium is that consumers will vote with their feet to use AI that is respecting their priorities and delivering value that they trust to be impartial, and not stepping on the scale just because of its paid consideration. I think what we’re building is an additional information source for the AI to consider. It’s up to AI developers to implement experiences that don’t abuse user trust that way. And if they do, I think somebody else will step in to provide one that doesn’t have that.

I’d love my travel agents to go out there and find all of the best deals. That’s probably coming from paid sources that people are willing to give my agents offers to be considered, and offers to me to actually convert downstream. That feels like a more powerful version. If I get a sense in that process that the AI is not looking out for my best interest, I’m not going to use it. And so I think that’s how it solves: market forces.

This is where I think it actually is important to have a breadth of developers out there building every variant on these experiences. There are probably going to be people that use zero-click to monetize an experience that I wouldn’t want to use, and I don’t think it will be successful because I don’t think it’s preserving the user’s trust that way. When I’ve built experiences in the past, we always put user trust at the highest pinnacle.

There’s a whole bunch of stuff I can add on that, but the second you violate that user trust, you’ve lost. And so, to me, that is the way to build consumer experiences.

That said, as a platform infrastructure provider, I don’t want to dictate that to developers on our platform any more than I want Stripe or PayPal to say what type of businesses can and can’t use our platform for transactions, or any more than I want YouTube saying what categories of content you’re allowed to monetize or not. I think neutral platforms are an important thing, even if I don’t like how it is.

I think the market will sort itself out from there, is my hope. And the number of times that we have to step on the scale and say, “Don’t do that,” I think we’ll try to limit to breaking the law or the other extremes, rather than moral judgments on the platform. I think staying neutral is important—to be plumbing and rails for other people to build on. I’ve seen where that can shift markets around unnaturally, and I think with enough competition, that sorts itself out for the most part.

I mostly hope that’s true, and I mostly think that will be true, although I do have some nagging doubt. What kind of range of monetization events do you have today, and how do you think that is going to develop? Is it going to be—or is it already—an auction dynamic?

Yeah, it’s auction-adjacent. I mean, it’s an auction, but with a heavy dose of context to even be considered in the auction. Over time, I’m sure the model will mature. Right now, it’s a lot of internal management of campaign bidding prices to achieve results for the ad campaigns, tracking through to actual transactions and things like that on behalf of an advertiser.

But at the core of the auction, just like with Google—and this is, to their credit, they figured this out, or followed some people that did figure it out—effectively, a user clicking on the ad is a signal of quality, and an advertiser’s willingness to pay is a signal of quality. If you effectively do an expected-value calculation on how much money Google’s going to make from that click times the rate, you actually get a very good signal and way to rank the advertiser results.

And so our system should evolve to be something that looks a lot like that and feels a lot like that.

Yes. Okay. So here’s one doubt that I have. I was in the mortgage business very briefly a long time ago, and I think the problem that I’m going to describe was maybe at its zenith in that business at that moment in time. By the way, it ended in a giant financial crisis.

But even leaving aside the systemic risks, something that I observed was that expected-value calculation can go awry. It’s more regulated now, but some years ago, the mortgage originator could basically just charge you whatever they wanted at the time of mortgage origination, right?

And so there was this dynamic where they would try to get as much from customers as they could get away with, or at least a lot of companies would do that. I even saw mortgage pricing cards from companies where it would be, “Here is the minimum rate that you can originate a mortgage at today.” Of course, there’s credit-score adjustment and stuff like that.

But then there was just the additional salesmanship bonus: if you can get somebody to close at a half-point higher than that base, you get X; if you can get them a full point higher than that, you get Y. And so the salesperson is directly adversarially incentivized to extract as much value from the customer as possible.

And I think a lot of things are like that, right? A lot of prices are sort of negotiable. Not even just B2B SaaS—it doesn’t cost them anything to deliver it long-term, so the price is kind of a pure negotiation. The seller is incentivized to have price integrity, but the deal could get done at a lot of points on that spectrum.

What I observed in Google Search specifically with mortgages was that the clicks were starting to get up to $50, $7,500. Well, how do you afford that click? You’ve got to extract as much value as you can, right?

You can tell the story, and I think it is true in a lot of instances, a lot of times, where the expected value to Google is a pretty good signal of quality. But in some markets—and mortgages are not unimportant, right? We’re talking 30-year contracts, the biggest purchase people make in their lives, with systemic implications—we did observe this haywire effect where the people that could bid the most were the people that would extract the most.

And if I sort of map that into the AI app ecosystem, maybe just stay with mortgages for a second. Then you’d have these financial helper apps, right? And how do they make money? Well, they refer you to mortgage companies, too.

So you have this second-order effect where it’s like, which of these financial helper or planner apps is going to be able to get the most customers? It’s the one that can bid the highest. How are they going to bid the highest? It’s going to be by most effectively referring you to the mortgage people.

So how does that not happen in AI? I’m still a little bit—if there is an auction dynamic and it’s who can pay the most for a given high-intent moment, how do we not get to this sort of adversarial situation where the sellers that can extract the most from the customers get the space, and the AI apps themselves are incentivized to steer you in that direction, because presumably they are going to participate in that transaction value, too?

Interesting. I think you nailed it: this is inherent to buying and selling goods in general, and price discovery certainly is done in different ways in different parts of the economy. I’m picturing the scenario that you said was there. I wasn’t there for it, but it sounds right.

I’m picturing the scenario in an AI world. I think I would layer on that the AI app financial advisor that monetizes best is probably going to be able to do the best customer acquisition of its own users, and so it becomes potentially the dominant financial advice app. I think you’re raising a very valid thing to think about.

I don’t know if advertising specifically inherently does this. It probably leads to more of that effect versus, I guess, maybe there are emerging business models that are different. I remember Angie’s List has a paid-user-subscription sort of model. I think they also still make money on ads, but I could be wrong.

A paid-user service for something like that—maybe you do want to pay $10 a month, or maybe your brokerage firm wants to subsidize that for you and bundle it with their services or something like that, and get a non-advertising sort of app experience.

I’m not saying that there needs to be advertising in every AI experience. I just think there are a lot of cases where that would be beneficial, where it steers away from the core value proposition of the AI experience. Then it feels like maybe that’s just the wrong match of the model, and that mortgage or financial-advice environment does seem like one where I’d be careful about picking the right model, as a user and maybe as a developer.

I think that problem, to a very significant degree, has been mitigated by outright government regulation, where I think the pricing is much more controlled. I don’t think the mortgage companies can give their individual sellers a card anymore that’s like, “Charge a point over base and you get a bonus.” I think that’s literally illegal at this point.

So that’s part of the situation: government can fix certain market failures. It used to be the case in stock brokerage that we’d have a new listing and the brokers would get directly spiffed on how much of it they pushed into their accounts.

I did a brief high-school internship at a brokerage and saw a guy sitting there making about $10K by putting one of his clients into some new offering. Is that any good? It doesn’t matter. Is that the right thing for the portfolio? It doesn’t matter. So, yeah—

It might go down.

That was pre-internet. These things have been around as misalignments of interest probably for a while. As a consumer, I wonder if there’s a way to even build a service that helps consumers assess that alignment in the tools that they’re using. That’s intriguing to me.

Yeah, we’re going to need all the help we can get navigating this world. Increasingly, we’re turning to AI to help solve the AI. We need AI to help us navigate the AI investment.

Yeah. I mean, I’m sure you’re aware that this is basically the safety plan of the frontier companies at this point. Reading the GPT-5 system card, it was striking to me over and over again that it was like, “We used an LLM-as-a-judge to evaluate how good the outputs were.”

And they have these justifications, which are not unfounded. They’re like, “We sat down with an expert, and they helped us workshop the prompt, and we confirmed that their judgments seem to align and correlate, at least,” whatever. And of course, you’re not going to have perfect inter-rater reliability among humans. That’s a huge problem.

Nevertheless, it feels like we’re kind of spinning some plates there. I just saw something yesterday, too, where GPT-5, in creative writing, is starting to do some weird, what I would call pretentious nonsense, basically. That’s one thing, but then what’s really interesting is GPT-5, when given its own pretentious nonsense, really likes its own pretentious nonsense. Even Claude seems to like its own pretentious nonsense.

And so now the speculation is, well, maybe it’s learned to write such pretentious nonsense because it’s sort of a reward hack, where it’s getting high scores from its LLM judge.

And it’s learning to kind of exploit something that—

Humans basically are like, “What the fuck is that?” But the AI sort of reads some sophistication into it that potentially isn’t really there.

Love it. More dashes. It’s becoming a real hall of mirrors in a few of these areas. So, yeah, I think to the degree that you can bring AI truly to the consumer and help them monitor where these things are happening, I do think that is super, super valuable.

It’d be an emerging need for sure.

Yeah. How about in the technical domain? Do you see Cursor being ad-supported? I was just thinking there are a lot of API-type services that are potentially complementary. Not potentially—they’re core to building modern applications, right? And so you get to the point where you’re like, “Oh, I need to scrape a website,” or, “I need to do whatever.” Do you see those coding assistants bringing back technical solutions? That would seem like pretty bread and butter, right?

I think so. To answer your question, yes, I think that’s a very interesting way to build a different sort of business model around some of those tools. The way I would extend that is, I’d broaden it and think about just any software, any SaaS tool. Today, SaaS products are basically $10,000 a year plus some sort of enterprise contract and sales process to be viable because they’re sold by people. You have a human sales team going out there selling and supporting these products.

I think there’s going to be a wave of new SaaS products that are priced cheaper and are distributed through contextual advertising. Maybe it’s—maybe it’s Cursor—but there’s an equivalent sort of workflow tool where it’s summarizing your meeting notes, or, “Oh, by the way, did you consider this thing when you were having that conversation about…?” I won’t prescribe the use cases, but I think there will be very thoughtful, clever people who figure out how to effectively change the cost structure for going to market on SaaS products.

If an ad layer like what we’re building gets built into a lot of these services, which I think it can and should be, you get a lot of efficiency of discovery. These are categories where the reason it’s a sales team today is because people aren’t going to Google and saying, “I need a new SaaS tool,” at meaningful numbers. It’s not a category that’s in Google search, but it can be a category that is in the aggregate of AI experiences.

Cursor, yes, for tool discovery, but also yes for SaaS and probably a whole bunch of things like that. There’s a company, OpenEvidence, that has an ad model for reaching doctors with a ChatGPT for doctors. They’ve been able to take over that market exceptionally well by going directly to doctors with an ad-supported model, where everybody else in the space is trying to sell into hospital systems and into effectively huge-dollar enterprise contracts and multi-year engagements.

Instead, this other company with an ad-supported version is now used by like 40% of doctors every day. They were able to do that in their vertical because it’s very clear who the audience is. It’s high value. It’s very clear who the advertisers are. It’s high value. They do that matching. They were able to build both sides of that network.

For most categories and for most AI developers, it’s less obvious, and it doesn’t make as much sense to build your own ad system and solve for both halves of this marketplace. Effectively, it’s like a context marketplace where you have advertisers that want to reach your audience and you have an audience that you’re trying to build. Solving both sides at once is really hard.

If you could tap into universal plumbing on the reach-the-advertiser side, it makes it a lot easier to build out new experiences and build new business models where maybe you’re going after a category that today somebody is selling into enterprises with a contract value that’s 5 figures. Maybe there’s an ad-supported version of that that can be built and go to market more cost-effectively than the competition. So that’s the sort of innovation that I can see flourishing, and I’m pretty excited to see how we can support that.

We’re talking to developers that have non-obvious ad directions. This is a ChatGPT experience for research papers and scientists looking for that sort of thing. Like, okay, what does an ad experience look like in that? I think there is one for somebody building that service. They shouldn’t have to go figure that part of the business model out either.

It’s not like the LA Times in the early 2000s, where we had a sales team going directly to the car dealers. It’s like you tap into common ad-rails plumbing and you’re able to focus on the part that’s core to your business, which is building that helpful user experience and making that part really, really competitive.

On the SaaS part, it’s funny. I once mocked up a pricing page, a classic SaaS pricing page, with the first tier and the lowest price being AI sales and support. The middle tier was if you wanted to talk to human sales, and the top tier was if you wanted to talk to human sales and support. I don’t think too many people are going to present their pricing pages exactly that way, but it does get at something very real: the cost of sales puts a floor on what, in many cases, is a close-to-zero-marginal-cost product. So that is—

I love that idea. Maybe we’ll use that on our ZeroClick pages, because we have the same thing. You’re trying to reach a lot of developers, and there are different sizes and scales. It makes sense to have personal conversations with some of them, but others hopefully can onboard themselves and figure it out and chat with an AI assistant to answer any questions they have, instead of us trying to scale a team to support that.

How much do you know about that medical one? Because that’s also—I assume the biggest advertiser would be drug companies, right? Selling—

Yeah. Pharma, medical devices, that sort of thing. I don’t know a ton other than what’s been written about it by other people. I don’t have an inside channel to it, but from what I’ve read, they are doing exceptionally well. From what venture capital investors say, it sounds like the stuff that’s written is accurate based on the investor community. So it’s a really cool case study in what if you had a completely different business model? I love it.

Yeah, that’s another fascinating one. But, again, I don’t want to be neglectful of the upside, because I do think better living through pharmacology is very real, and the awareness theory of advertising as it applies to drugs is, in a totally earnest way, important.

At the same time, those commercials always conclude with, “Talk to your doctor,” and now it’s sort of a flipped-around thing where it’s like they beat you to talking to your doctor, and now you’re going to talk to your doctor after the drug company’s already talked to your doctor, potentially, about you. At a minimum, I think—and there probably is some law about this, or maybe not, I don’t know—doctors themselves can’t take cash for prescriptions directly, right? They can take trips and stuff, but they can’t take literal pay-per-script. But the AI can probably take a pay-per-script, I would guess, in today’s world.

I don’t know that that would be—there aren’t a lot of laws around this stuff yet, right? So it’s all kind of greenfield.

We’re coming up with an even better version there: monetize that.

Yeah. Well, let’s not give too many ideas too soon.

But, yeah, there is sort of a duty of care. I wonder, do you have any thoughts on—

Here’s the counter to that even being a bad thing: those ads on TV and stuff—I’m not a huge fan. I don’t think most people consider those good content. If those all went away, everybody would be happier.

If there were a more cost-effective way for the pharma companies to present their options to doctors, they wouldn’t need to do those “Ask your doctor about…” ads. It’s like, we already gave your doctor the options. They’ve discovered this new drug that they might not know about for this particular use case. We’re presenting it contextually when there’s a patient case where it makes sense to consider it.

The doctor’s still going to do the filtering on whether it’s actually useful or applicable anyway, and maybe we can get rid of the annoying, bad-advertising part of the thing because it doesn’t work as well. That would be a pretty great outcome. It’s just market efficiency. The efficiency of annoying millions of people is unnecessary if you have a better channel for advertising.

He kind of hinted at this earlier with Facebook having a good quarter: if they build a more efficient ad system, the money is going to flow to that and away from ineffective ones. I would argue annoying people at scale is an ineffective ad system. To the extent they’re doing it on TV and doing awareness things right now, part of that’s probably just that they don’t have a way to measure how ineffective it is. In some of those categories, they just don’t have another channel where they can cost-effectively reach people. But as soon as you do, maybe you starve the bad ads and just get efficiency.

Yeah, the upside vision is compelling. As long as the virtue and integrity of key actors stay strong in the system, then a lot of these things are fine. That's something Dean Ball, a repeat guest and previous White House AI adviser, told me once: republics rely on virtue. You can't really have one without it. To some extent, that's on all of us. It's on the actors to make sure that they keep your priorities straight.

I have a few questions on tech trends and stuff. What more should we know about how it all works? I guess my sketch is that, right now, we're presenting the ability to go seek additional context from the AI as a tool. So presumably parallel tool calls or things like that are a huge development in terms of latency, right? You wouldn't want to have the user sit there and wait for that tool call and whatever to come back.

So now we're starting to get into this realm where you can issue a tool call, but not necessarily have it be so blocking. The AI, though, is responsible for sending over whatever information is sent over. So you're kind of relying on the app AI to guard the privacy of the user, which is interesting. But I get the sense that you also sort of expect that this will evolve from one where the AI is sending stuff over the wire and needing to protect privacy before sending the message to the matching system, to one where, I guess, in the future, more of that vector-type stuff will happen on-device, so it could even be potentially more personalized.

But how do you send that vector content down to the device? You can't send your whole database of advertisers to match, right? So how do you see that? Where does that compute happen, and how can you possibly do robust matching on the edge, if I'm understanding the vision you have for the future correctly?

Yeah, the future version—the today version—is effectively parallel to organic search. So it's heavily keyword-search-driven tool use and, for the reasons that you talked about for performance, it's just adding another search of an initial source and then synthesizing it in that same next step.

The vectorization, to me, is the most interesting for the personalization side of it, and that is a bit unsettled technically in terms of how it would be best to do that. We have a version that we have working in a browser context where, because the browser can have that profile, we can actually, independently for a user, effectively front-run or simultaneously send that context to the ad server, for lack of a better term. When that request comes through from the AI service, the ad server has that context separate from the chat.

So the personalization context could be separate from the chat context. We're not using that today, but it's kind of proven that we can do it in different environments where you don't have that browser context. The reason we're not doing it is that we haven't solved for all of the use cases where that would be relevant in the ad system.

But it's interesting to think about different ways to do it. This is like a reinvention of something that already kind of exists in a lot of ways. People have been doing insane RTB auctions for all those banners that are selling for less than a cent each. Behind the scenes, there's an insane real-time auction with multiple bidders bidding into this ecosystem. When you look at that, what we're doing isn't complex at all.

Because it's all contained within our systems, it's not going out like RTB. It's not open RTB going to third parties asking for bids in real time. It's managing against campaigns that are loaded onto the platform. So we can do a lot of caching and performance optimization to make those responses as fast as possible and as contextually relevant as possible.

The personalization or vectorization service certainly could be a piece of that more in the future. There are a lot of fun engineering things to play with on that. I love the business and market-structure, big-picture thinking, but then also diving in on the actual tech. That's where the fun stuff is.

Anything else you want to highlight that you guys are working on that you think is particularly fun tech-wise? Again, you can go as deep and esoteric as you want.

Mostly it's that stuff and then enabling browser-based applications of AI. This is the fun user-experience frontier that we're playing with. We don't think we'll figure all of the answers out, but helping browser developers become AI developers with a demonstration of, “Hey, this works,” and potentially, “Hey, here's how this works, and you can go ahead and put it into your browser, your browser extension.”

I think if we get a lot of people thinking about what an AI-augmented browsing experience looks like with a human at the wheel, that's pretty cool. The types of things that you can build—we won't think of all of them—but getting more people thinking that way and having a way to monetize that, I think, is pretty powerful. Monetizing a browser extension has historically been not particularly easy, and I think we can kind of change that so that more developers can build a lot more different applications.

Google turned off the paid version of browser extensions in the Chrome Web Store, and then they highly limit the ability to add advertising into a browser-extension experience with their single-purpose policy. So it's been very challenging to build a business, and the types of things you could do were pretty narrow—essentially narrowed down to just shopping tools because it fits within the single-purpose umbrella.

But if we're able to build AI-augmented experiences, we can bring ad monetization into that without it tripping over Google's interpretation of its single-purpose policy to disallow injecting advertising in a different sort of way. So I think there are use cases that this would enable in the browser, and to me those are very exciting because that's where the users are and there's a clear path to go to market.

The power of a browser-based tool like a browser extension is that you can be contextually useful to a user and have it happen automatically. There's no training. You don't have to do the in-app message that you were talking about before that pings you 8 times a day to try to build that habit of talking to it.

It's, “Hey, you can build a super-niche thing,” and it only ever shows up once a month when you're doing some very specific activity, where it's helpful in context, and stays out of your way otherwise. The power of a browser extension to create that type of experience and build user habit for free is, I think, underappreciated.

People haven't been able to build businesses there, and I think they can now. To me, us teaching some of the ways, but also hoping that that's just a sliver of the possibility and that you start to see a proliferation of great new experiences that thoughtful, creative people have built for users—that's very exciting.

What do you expect for the seemingly just-getting-started browser wars? It's like history repeats itself, right? All the hot startups are trying to make one. Microsoft is very much back focusing on this. I don't know if they ever quit, but certainly I wasn't thinking about it for a while, and now I'm thinking about it again a little bit from them. Do you have any forecasts for what we should expect there?

I think it's going to be competitive again, in part because you finally have potentially differentiated experiences happening in browsers. When we started our company a year and a half ago, it was all about, “Let's focus on building the application layer of a browser,” effectively building a virtual browser.

It doesn't matter if it's Chrome, Edge, or whatever browser you're choosing to use; we're layering on capabilities to any browser. So I think there's going to be a battle for being that default browser, but I think the most interesting stuff is actually going to happen in the application layer—the application layer for browsers being extensions.

I think most people will get their new capabilities that way versus switching to an entirely new browser, which is a heavy lift: transitioning somebody from Chrome, which just works and does what you expect, to some new experience for a feature. Historically, it's been niche subsets of users—power users who want tab management and some of these capabilities—who have wanted to do that.

Or maybe they're particularly privacy-sensitive or don't want to be on a Google or Microsoft platform, so they use Brave. To me, it's been tricky to think about that as a universal use case, and my thinking more broadly is that I don't know that there is a mass-market, has-to-be-the-same-for-everybody version of what the browser should be.

It's more: let people pick and choose the special features they want their browser to have. Some people love dark mode, some people don't. Some people want shopping tools, some people don't. Let's make it a configurable thing sitting on a standardized base rendering engine and capabilities, so developers can build for that common platform.

I wrote a response to the proposal to spin Chrome out of Google. I don't think it solves the problems that are there. I think the biggest policy problem for me with the current implementation of Chrome is the single-purpose policy. I think it does choke off innovation, and I think as long as developers are thoughtful and transparent to users about what they're doing, extensions shouldn't be forced to be single-purpose as defined by, I'll use the word, a monopolistic owner of the platform.

I think innovation has been choked off there, and you haven't seen a proliferation of development largely because of that policy. My concern in selling it to somebody else—a big AI company, whether it's Perplexity, OpenAI, or somebody else—is that they would have every incentive to behave just like the prior owner and foreclose on competitive innovation, especially in AI. I don't think that would be a good answer.

I don't have a good answer other than my preferred answer, and I put it forward: make Google be open with it as a platform. That would be a better remedy than a new owner. The owner is not the problem. It's the ability to build on top of the platform.

Does that policy operate only at the store level? I can add any extension I want onto Chrome, right? Or do they prevent me from installing my own stuff?

You have to put your Chrome into developer mode to install anything yourself.

And I guess I've been in developer mode a long time. Even then, there are cases where they somehow remove stuff. I'm not even sure how they have deactivated some things that users added. Users have shared paywall-bypassing extensions and things like that, and I've heard direct reports of it somehow getting disabled on their developer-mode Chrome.

They did it partly under the guise of protecting user privacy and not creating bad experiences with extensions. A decade ago or so, they forced all extensions through the Chrome Web Store and sunset being able to do it from a third party in the installation process. To do any extension on Chrome, which is the dominant browser, you have to go through them and abide by their policies.

With 70% to 80% market share for Chrome, that effectively is the market for extensions. You can't build an extension that's only on one of the other platforms, and they have also largely adopted the same sort of policies just by default.

Yeah, okay.

Yeah.

What do you see in the sort of AI version of SEO? I'm getting an increasing—ramping up—number of people cold-emailing me, just like they used to do with SEO: “We can help you rank, we can get you traffic,” whatever. Now it's, “We can get you into the chatbots' answers.”

I'll spare you my knee-jerk skepticism, but what are you seeing there in terms of what sort of sites or businesses are getting substantial referral traffic from AIs, and which are not? I don't know what visibility you may have into this, but I'm sure you've made a point of trying to understand it. Is there any way beyond traditional SEO best practices—have good content, whatever—to win in the AI version of that competition? Is anything known there that's credible?

My overall assessment is that, like SEO, there will be some people on the frontier who understand it very, very well, and they will be able to massage their content to be desirable for consumption by AI systems. Like with SEO, there's probably 10 world-class people who really get it, and then thousands of people who are going to run around taking money from people to provide that service.

The result of all that, I think, is effectively similar to what happened in SEO, but you get some amount of dilution of the organic results with SEO slop. To the extent that it's easier to use AI to generate content—and, to your point earlier, maybe the AI even likes AI content better than human content—you're going to have a flood of content in the organic realm.

Google effectively trained its system on human feedback of clicks—does somebody click through and then bounce back? Reading signals of quality from a human interpretation of that result is what they've used to refine organic search over decades. In an AI context, you lose a lot of that ability to determine, “Hey, is this a great new creator who is a specialist in makeup doing reviews, and this is an authoritative source on this, or is this literal AI slop and just a mass-produced content farm?” How do you tell the difference? It'd be tricky.

I think it's going to be something people try to do and something that people invest a lot of time and resources into. To me, as somebody thinking about it from an advertiser and marketer side of the world, it feels like a sliver of how people will find your brand in the future. For every dollar that goes into SEO, SEM is massively more important, and I think it'd be the same for the advertising side of how you present yourself to an AI.

The best way is going to be to present yourself to an AI in a paid context. The SEO games will be won by a few people, and they'll generate some traffic, maybe, but on the whole, I think that won't work for most people. Right now, I think there's a ton of activity around it, mostly because there's not another option.

Every marketer is like, “All my searches are going to ChatGPT. What am I going to do about it? I need to figure out how to reach my audience.” The only way they know of right now to do it is to go and do optimizations, create new content, present yourself differently, and do a bunch of things that are probably good hygiene and good practice in this era. But they don't solve the underlying problem: “How do I reach that audience?”

What we're building is that layer, and it's been very well received by advertisers who are looking for anything in this category. So it's not us. There's a market void, and we're hoping that we can help fill it.

What kind of intent are you seeing shift most to AI from search?

I don't know. I guess a little bit of everything, but I think one of the most interesting ones is that it's actually a little bit upstream in discovery relative to Google. Google is where you go when you know what you want to do. In a chat context, you're doing a lot more exploration of ideas, and that's at least one step up the funnel, largely.

You're not going there and saying, tactically, “I want these shoes.” Even if you're doing that generic “the best trail-running shoes” or whatever, people are mostly not doing that. When you want specific trail-running shoes, you're going to Amazon because they're going to fulfill it the fastest. You go there and do a search on Amazon, which is a huge ad business.

You do a little bit of it on Google; if you add the qualifier “best,” you go to Google. If you're just generally chatting about running, planning a trip, or asking what good trails are nearby, things like that, that's happening up-funnel in a ChatGPT experience, but then you're still arriving at a lot of that commercial intent.

Yeah, that's interesting. That also kind of raises the question of getting into less and less commercially motivated advertising. Obviously, Nike is always selling apparel, right? So they're commercially motivated regardless of where in the funnel you are.

But let's say I'm getting into travel. Governments around the world, for example, might want to pay to influence the way I think about their country. They might partly be thinking about that in terms of the ROI of me actually showing up and visiting there one day, eating in their restaurants, and staying in their hotels. But they might also just be thinking, “We want to shift global perception of our country and our government.”

Do you have any thoughts on whether that is something that should be treated differently? If I Google “Tiananmen Square” today on Google, I don't think I see a sponsored link from the Chinese government saying, “Here's the story we want you to know about the incident.” But that's going to be really blurry, I guess, in the AI context.

Well, you are infinitely more creative than me. I certainly never come close to thinking about this particular use case. That's fascinating. We'll come back to it.

One other one is, what about non-text-based advertising? We just saw the new Gemini Flash—Nano Banana—out yesterday, and it seems like you could really start to imagine all sorts of AI try-ons, which we've already seen in these apps, but bringing it to you seems like that's got to happen.

Seeing it in your home is sort of another experience that we've seen people develop in a specialized way. But now I could really imagine that if I just gave Gemini a few examples of my home, the next things I could be seeing are all these products in my home, and it could be extremely real. Are you guys interested in that sort of thing? Do you have any forecasts for what the multimodal advertising formats might be?

I'm intellectually interested in what you described there. I can imagine that might be a pretty cool experience. We're not starting there, for sure. We're certainly focused on text to start with, but at the end of the day, having a map of an advertiser's context, including the product information, means that maybe seeing the stuff in your home would be much better with advertiser content.

Instead of rendering a generic couch, it renders a real couch and you can buy it. So it's not just an AI guessing, “Hey, here's a hypothetical world.” It's a real thing. It's not a Pinterest inspiration image that's just AI design slop. It looks awesome, but if you want to actually execute on that, you have no next step.

So, I think that might be a case where ad content helps generate better answers, even in the visual realm.

That sounds cool, but we're not going to be able to help with that for a bit.

Yeah. Well, it's coming at us. It's all coming at us quickly, I guess.

Yeah, it won't be a bit. I mean, by a bit, I mean 3 months or 6 months now.

Yeah, truly. “Accelerate thy timelines” is my universal command.

It is so crazy how fast people are building stuff these days. It's inspiring.

You've been very generous with your time. This has been super interesting. In closing, is there anything we didn't touch on that you wanted to cover, or any last words or thoughts you want to leave people with?

No, I think we covered way more than I'd even thought about ahead of this. Some of these are luxury problems: How do I feel about the future, assuming this thing works? We're a startup trying to get going, and we'd love to work with as many developers as we can, as fast as possible.

Cool. Ryan Hudson, founder and CEO of Zero Click. Thank you for being a part of The Cognitive Revolution.

Thanks.