Why AI Startups Keep Dying (And What Actually Works) | Nic Baird, CEO@Koahlabs

ChatGPT is losing $4 to every dollar or something crazy like that, right? Makes it impossible to sustain that at scale unless they come up with a strong monetization method. And we saw all these companies that were getting a lot of engagement and growing really quickly just shut down because they look at the projections of unit economics going forward with subscriptions only converting at 3 to 5% and say, this is never going to make money. With Character AI, the average user engagement per day is over 3 hours. That's twice YouTube and TikTok. It's insane. Like, Character AI tried to bring in basically like a traditional interstitial ad, so like a full-screen ad that comes up once you switch talking to a different character. And they got really bad backlash from their Reddit community because that's just not the way that people want to see ads, right?

Yeah, that's terrible.

We're trying to keep it native and relevant. It's in the same font, the same coloring, not something that's like a huge banner ad. It's like flashing and trying to get your attention. It's really just trying to solve the user's problem. We're doing 2% CTR across the entire network. Traditional kind of display ads, they're usually in like the 0.4 or 0.5% CTR. And Open Evidence makes $150 million a year in revenue on a million users, which are all doctors. There's no payment option available to them. These people are having so much fun and their users are so engaged, like, why don't they deserve to live? Yeah. So if you're like an elementary school student, it's simple. It's, you know how you use ChatGPT and sometimes ChatGPT kind of recommends products? We are the layer behind recommending those products for folks that use the ChatGPT technology. It's very simplistic, but it's like, that will explain what we do. We're a recommendation system for products.

Are you seeing a lot of elementary schoolers using ChatGPT, by the way?

Hopefully we don't see a lot of elementary schoolers because we don't want to be serving ads to elementary schoolers just yet. Maybe when they're older.

The toys?

Target? Yeah. Eventually, there's lots of kids stuff for sure.

What about for fentanyl?

Um, but basically we're full service ad network for AI companies deploy via SDK into each publisher that we work with. We use their first party data combined with basically traditional ad targeting methods to find the most relevant kind of sponsored content for them. And then we show the, the end user personalized one-to-one relevant advertisement on their behalf. And then we also build all of the advertiser side of the marketplace as well. So the targeting tooling for them and the sales process for them, and we're a full service network. So. If you're an app, we deploy the SDK and then we do everything else on the advertising side as well, including the ad server.

So you're saying it's inevitable?

Basically, yeah.

If I'm building a live coding app this weekend, and when do you think I should think about monetization? Like, this is definitely not the first thing I think about.

Definitely not. I think the way to think about it is you probably want to plan from the beginning whether you want to run advertisements. As a core monetization effort. And so in that case, we're happy to talk to you or give resources that help understand like, hey, how much money could I make from advertising? Is this the right thing for me? You definitely want to think about that while you're vibe coding the app or while you're building something. But then I would say focus on the user experience, get 10,000 users that are really engaged first, right? And then figure out how you can deliver value to your users. That's definitely the most important thing to do right away. And then I think once you've figured out a way to get people really engage as you grow and as you get back into like growth mindset and okay, great. How can I get from 10,000 to 100,000 engaged users? Now is the right time to bring us in and build your sort of growth playbook around a monetization lever.

If you were OpenAI, when do you think is the right time?

If I were OpenAI, I would say the right time is as early as possible because you have to show this, but I think that's because I'm thinking of OpenAI as like a business that has to make money and survive, right? I think if I have hundreds of billions of dollars in venture money to burn, then it's okay to delay that because you're focused very much on kind of making sure that you cement your leadership in this space that's obviously going to be very big. And so in that case, I certainly don't fault them. I can really argue both sides, right? The earlier the better for them in terms of designing the product around commercial experiences. I do think that that will be a struggle to add on later for them and they'll fumble it and it'll be difficult and they'll probably get it wrong the first few times. That'll be okay. That's normal for businesses. At the same time, I totally understand if you have this huge war chest just saying, hey, we're going to focus really hard on making sure that we deliver the best possible product. Not to say that bringing in sponsored content inhibits their ability to do that. It's just a focus question.

Why did you start KOA?

So basically the story is that when we were— my co-founders Mike and Eric and I were all members of South Park Commons. And within South Park Commons, we were looking at all these consumer apps that were being built and people were leveraging it. Kind of what we talked about before, you've got all these people who are doing children's storytelling and all these people who are vertical search for construction workers and all these people who are like trying to build product recommendation tools. And there were all these different kind of AI, core AI experiences, these generative experiences that were popping up and people were, they would get 40 or 50,000 pretty engaged users and then they'd realize how expensive it is to maintain 40 or 50,000 engaged users when your core experience is generative, right? That's a much more expensive business to run, especially when your users are very engaged than traditional app or a static web app, right? Because you have this inference cost. And so we saw all these companies that were getting a lot of engagement and growing really quickly just shut down because they look at the projections of unit economics going forward with subscriptions only converting at 3 to 5% and say, this is never going to make money. We're only just going to lose money forever. And I think even some folks were throwing in the AdMob or the AppLovins of the world and trying to make that work. And sometimes it was okay. And sometimes you could do a dropshipping thing and you could kind of squeak by, but it just, it was clear that there was not the right monetization solutions available for these consumer developers in the market for AI, right? And so when we saw that, we started to try to solve that problem. And originally we thought, can we get better at subscriptions? Can we increase the conversion rate from 3% to 4%? That was the original way we started thinking. Thinking about it, but simultaneously we were exploring more on the marketing side. And so I have a big network in there. My wife works in marketing. And so we were asking a lot of marketers what was top of mind for them. And the answer was always the same. What do we do about AI? What's going to happen in this world? How do we prepare ourselves? How do we take advantage of the opportunity to win in this new channel? And so we really just put two and two together. We're like, wait a minute, advertisers want to get into AI and AI companies need to figure out how to get paid. This seems like an obvious solution to just bring them together. And so that, that's basically how getting started. So this is one of our publishers, Liner, and this is just a great example of who are all these apps that exist. Liner is an AI search tool for students. And so you can see on this agent side, on the thing on the side, they're really like aimed a lot at like higher level education students and things like that. And that's their main core demographic. And we power the advertising behind Liner and help them monetize their search inventory. And for example, just to be clear about how things are non-intrusive, we're trying to keep it native and relevant. Let's say. And you can see some of my demos on the side here. Let's say I have a prompt coming in and I have, "Ask me, how can I deal with that?" Something that, like, let's say, like, a high schooler would ask. And there's going to be this organic response that comes together. And then we have our ad format here for one of our advertisers, Honeydew. And so the idea here is you can see that it's very native to the platform. It's in the same font, the same coloring. Like brand associated with liner, not something that's like a huge banner ad. It's like flashing and trying to get your attention. It's really just trying to solve the user's problem, right? It's relevant to the query. It's bringing in some personalization, right? It's mentioning Prom in the copy. And so it's trying really hard to make sure that product is being recommended to the user is something that they actually find value in. Right. And so if you look at some of the organic sources on the side, right, it's like, these are informational sources, which is very useful. But very few of them are going to be commercial sources. And so this is an area where Liner's organic responses are great for learning. And this is what most people use AI for. But when it comes to actually purchasing a product or like solving the solution immediately, this is where our recommendation system will come in and bring in the right sponsored experience.

So it's kind of standalone. It's not embedded in the chat. Like one of the new 5 bullet points is one of it. And it says like a little sponsor tag on it. It's not like that.

That's right. And in fact, we actually, some publishers wanted us to bring into one of these and test that. And we looked at it, but the fact is that, um, you know, one, I don't think we want to get into manipulating the response itself and showing people that we're bringing brands into organic response. People want to read this information and hear what the model has to say. And so I think creating a separate standalone experience that is clearly demarcated as an ad, but something that is very highly relevant and hopefully useful to the user is both the best way to actually monetize that surface area and get paid as a publisher. But also the best way to make sure that you're giving the right recommendation to the user that doesn't feel like it's degrading the product itself.

How do you find good ads to show the users?

Yeah.

It has to be relevant, right?

Yeah, it has to be relevant. And so this is a search problem. And so it becomes really interesting, right? Because really what search is, right, is you say, I have this incoming request and then I have all this information. How do I surface the most relevant information? And that's basically the same way that we think about it. We have this request coming in. We have to determine and interpret the request, and then we have to figure out against our database of information, and in this case it's advertisements, right? What is the right thing to show? And is there a right thing to show is the big question, right? And basically like, you break it down into a traditional sort of like, really it's a search question across all of it, and then you kind of sprinkle in like the AI components. So there is some embedding involved, there is some semantic search involved, there is also keywords involved and static stuff involved, and you've got different columns that you're pulling against. So there are disqualifying columns, like if they're not in the right country, right? We shouldn't show an advertisement for United States based home insurance to somebody who's in the Philippines asking about. So location is a big piece of it. Demographic information is a piece of it. There's a buy publisher targeting that we can do. Some publishers blacklist advertisers. And so there's a lot of complication there. And then there's a semantic component as well, right? So basically as the query comes in, we run multiple layers of analysis on that query to understand one, is it a commercial intent query? Does the person have any interest in actually seeing a product right now, or are they just in learning? Mode? And if they are in learning mode, is there something different we can show them? The next thing we ask is like, basically, is there any reason why we definitely shouldn't show a sponsor experience here? Is there profanity? Is it content that people, brands wouldn't be excited to be associated with? So there's some brand safety and protections there. And then the third component is basically, great, let's go see if we have a relevant match. And if we have a relevant match, then we generate the copy for it and through one more layer that basically checks before the copy goes out, is the copy going to sound right or make sense given the rest of the content answer, right? And so we do all of that basically in many cases, as you saw in the liner demo, before the answer loads, because we can do most of that based on the query alone.

Why shouldn't I, or why can't I just build my own ad system?

Yep.

Here's what I'm thinking. If the problem we're discussing is related to user experience, can I just take the AdMob data and then make it look better myself?

Yeah, with AdMob in particular and some of the mobile ad networks, The simple answer is you cannot because they won't just give you like the advertisements, right? The way that you would have to do this if you want to build it yourself is you would have to build a connection to the programmatic demand side, right? And so is this super hard? No, you could probably do it in a couple of months. Great developers could do it in a couple of weeks. There's documentation public on prebid and sort of like the OpenRTB spec, and you could definitely build out publisher-side ad servers and all these things. The problem is that while you might be able to stand up something that, you know, sort of works, there'll be a lot of issues around brand safety, a lot of issues around verification, a lot of issues around kind of even what we're talking about, like hallucinating. So it's not just the format piece. It's a lot of the targeting ad server piece as well. That is a very complicated and deeply technical product, right? So there is a component of just, it is possible, but it also, and it might take only a few weeks to stand up a simple product, but to do it and to be able to justify the CPMs that we have and bring in really high quality advertisers, you need to have a very strong product, right? And that's something that takes, I mean, it's taken us a year plus to build something that is functional in the way that we have now. And that's our sole focus.

How many people on the team?

So we have 12 on the team now.

So it took 12 people a year to build it.

That's right. That's right. Yeah. The other piece is, so there's a technical reason why, and then there's of course a go-to-market reason why too, right? Like