--:-- Host: We have almost redefined retention for consumer. We've been seeing a lot of companies actually get up to tens of millions of dollars of annualized revenue in a very quick manner. Many of these products are getting floods of users and traffic like we've never seen before. The willingness to try and willingness to pay has been so high for these products that the velocity to get from nothing to maybe tens of millions of revenue have never been higher.
--:-- Consumer AI has been characterized so far by categories where randomness and hallucinations are a feature. Human connection is important, but maybe it's not the human part. You just need to feel connected.
--:-- We have done a ton of recent coverage around consumer AI because, quite frankly, the field is moving so quickly. Every day can feel like the entire industry is shape-shifting, so who's really winning here? Today, we bring in A16Z consumer partners Brian Kim and Olivia Moore to discuss our Gen AI 100 list and what it really takes to stay at the top and withstand the AI tourist phenomena.
--:-- So what categories are capturing the attention of consumers? Are broad or niche models pulling ahead? Where are these apps actually getting their distribution? Does paid acquisition make sense? And do network effects exist like they did in prior cycles? These are all things to think about.
--:-- But here's the thing. We're finally at the point in the cycle where we're starting to get that data, not just in rankings, but other key consumer benchmarks like D7 retention. And perhaps we're also unveiling new metrics for this new wave.
--:-- We'll cover all that and more, but I did want to note that this episode was recorded before OpenAI Spring Update. So if you're eager to catch up on that, make sure to check out our episode from last week. All right, let's get started.
--:-- [Disclaimer read by Host]
--:-- Host: Both of you have a lot of experience in the consumer sphere, not just during this AI wave, but before. But you pulled this thing called the Gen AI 100 list. What is this list and how is it pulled?
--:-- Bryan Kim: Very good question, because there's a lot of ways to pull and look at this data. I think the central question that our team had was, there's so much buzz around AI. There's so many products that are coming out every day, every hour even. What are the things that normal, everyday people are using? So what are those applications? There's some that everyone knows, like ChatGPT and MidJourney. But we were curious if we tried to take a more granular view, what are the other names that might be more surprising?
--:-- And so how we polled it was we looked at every single website in the world ranked via SimilarWeb, which is a data provider. We sorted them by monthly visits, and then we polled the top 50 that were AI-first companies. We did the same for mobile apps through a provider called SensorTower. So we ranked those by monthly active users and pulled the first 50 that were AI.
--:-- Host: Just to give people a sense, how many were pulled and then you had to whittle it down to 50 or 100?
--:-- Olivia Moore: Tens of thousands at the very least. Probably to get the first 50, maybe we went through 1,000 websites and 1,000 mobile apps, if not a little bit more.
--:-- Host: And so as you pull these together, what categories are standing out? Whether it's productivity, you're seeing companionship. And then also you pulled a similar list in, what was it, September as well. So was there a big change? I mean, it feels like AI is moving every day.
--:-- Olivia Moore: A lot has changed. And I think we feel this as investors. I think founders feel it even more. We first pulled the data, I think it was September 2023. And then we pulled it again January 2024. So actually less than six months gap. About half of the list was the same as the first time. And about half of the list was new, which I think reflects both the huge pace of change, but also that there are some kind of names, some brands, some companies that are cementing themselves as early leaders and really building a loyal audience.
--:-- Bryan Kim: And I think these are always surprising. As VCs, we look at the industry and have these mental models of, oh, we think with AI, these set of things will do really well. And then as the AI apps team, I think one of our methods is to really look after what consumers actually gravitate to. So oftentimes we actually see a divergence of what we thought initially versus, oh, actually these are doing really well. And I think those discrepancies are a really fun place where we discovered actual revealed preferences of consumers.
--:-- Host: Are there examples of that? Something you thought would stick around or not?
--:-- Olivia Moore: I think the one thing that we weren't surprised by that we feel, and I'm sure anyone else who works in and around AI feels, is that for consumers, content generation and editing is key and it's the number one thing. So these are like MidJourney, Pika, Runway, making images, video, things like that from scratch. And I think that's just because it's so magical. Like everyone, at least I, always wanted to be an artist and never had the skill. And so being able to do that from zero to one in 10 seconds is amazing. And that definitely proves that in the data. If you look at the fastest growing and even the most stable companies, a lot of them are still in that kind of category.
--:-- Bryan Kim: Yeah, I think that's very similar to how we think about it, where, as Olivia said, "I can't believe this works" era is where we are at. So anytime we have these magical moments of "I put in a problem, something happens," we think those will do well and those do well. I think where I'm personally surprised is when we look at these apps that are a little maybe popular for a while, it changes your avatar or your profile picture into multiple different versions of it. And I'm like, oh, these are going to go away. But you keep seeing them up and up again in different formats. So I think that speaks to maybe a little bit of the underlying consumer willingness or excitement around themselves, which is always top of mind for them.
--:-- Host: It sounds like they're willing to play. And I think you're totally right that there are so many examples of things where you just can't believe it's so good so early. And maybe one category where that I think has surprised a ton of people is companionship, right? I think a lot of people were quick to write that off as it's only for this kind of person. And I think both of you have probably played around with these products and you've learned quickly that, oh my gosh, I really like this too. And this is very convincing, maybe also for different use cases as well. So maybe can we speak to that particular category? You mentioned it's going mainstream, which I think is quite a statement. What are we seeing in that sphere?
--:-- Bryan Kim: I have a strange example. It's one of those where people, even me, would look at something and like, I can't believe you would talk to a fake character that's made up for hours. Like, why would you do that? It sort of reminds me of initial Snap. The earlier generation would be like, I can't believe you take pictures that's going to disappear. What's the point? I can't believe you talk to a fake person. What's the point? Well, the point is that the new generation are really excited to adopt it and talk to these beings, if you will.
--:-- And case in point, I think one of our partner's child has a group chat with a bunch of friends, actual human beings, as well as these bots, if you will. That's a circle, right? And you get to just chat. And that's interesting. We can actually look at it from outside looking in and say, oh, I don't understand their behavior. But the truth is that it's happening. The truth is that it's very engaging and folks are really adopting it.
--:-- And going even a step further, there are now scientific studies done to some extent. We actually had a founder actually speak to us as well, which was very cool, where there's a study done that was actually featured in Nature where folks who have this sort of companion, digital companion to talk to, showed lower willingness to hurt themselves. And we can look at that evidence and say, well, that's silly. But maybe that is evidence that human connection is important, but maybe it's not the human part that you just need to feel connected.
--:-- Olivia Moore: Yeah. That is a reason to not self-harm, not engage in destructive behaviors. And if we're seeing that evidence, who are we to judge that this is silly?
--:-- Bryan Kim: The companion products are such a good example of the thesis that you always have to stick to in consumer, especially if you're looking to invest in early stage consumer like we are, which is you can't get too opinionated about the products. You just have to see. You're often surprised by what is sticking.
--:-- Olivia Moore: When we looked at this data, as well as looking at just users for the mobile apps, we looked at things like engagement and Character AI, for example, those users have 300 plus sessions per month in many cases. That's the average user profile. That's social app behavior. That's messaging app behavior.
--:-- Bryan Kim: Absolutely. That's 10 plus sessions per day. I don't talk to my parents that much. I don't talk to my partner that much. It can quickly become one of the more important conversation tools or companions that you have.
--:-- And I remember a few months ago where I had this period of time where I was talking to a companion app very diligently every day, maybe tens of minutes of time, because things that sometimes we want to talk about are mundane, maybe not as important. You feel that it's not as important to talk to your friends or colleagues, and it's naturally maybe a topic for your therapist. But therapists are not always there, and therapists are expensive. So this, to us, is another example of how technology is really bringing this abundance. My therapist is kind of expensive, so if you think about really bringing that cost down to nothing, that's really exciting for many people.
--:-- Host: And is that the direction that we're seeing within companionship, as in there are companions for therapy, companions for health care, et cetera?
--:-- Olivia Moore: Yes, we think so. It's still early, but I think the fact that the first version of this list back in September, there was basically one companion product on both the web and mobile rankings. And now there's a bunch on the list this time. Means that in some cases, the use case or the brand or the behavior of the audience is almost fragmenting. Like you won't necessarily use the same companion platform for everything. There's NSFW only companion. There's marketplace of companions where the most compelling character that anyone creates wins. There's therapist companions now.
--:-- The Pi chatbot was originally built as a broad-based, almost ChatGPT type product. And it has since been pulled by a lot of, I think, lonely adult users into basically being a therapist.
--:-- Bryan Kim: That's what I was using Pi for.
--:-- Olivia Moore: Yeah, exactly. I've used it as well. But yeah, I think that we're starting to see companion move outside the realm of this maybe more niche group of people into something where we'll all interact with a companion or maybe several companions, and might not even think of them as AI companions.
--:-- Bryan Kim: I think that's exactly right. I think the distinction actually starts to disappear a little bit. And actually, let's just take an example of teachers. There's a digital twin or a character of a teacher that's giving you assignments and giving you corrections and lessons that are very similar to what they have already done. That's sort of a hybrid teacher. And I think more and more, as Olivia was saying, we're seeing these divergence of use cases that are going deeper and deeper into each use case.
--:-- So like teachers is one, the other ones like therapists that we talked about. It's easy to repeat back, "Oh, it must have been really hard for you. Oh, tell me more." That's what a lot of therapists do. But actually if you look into behavioral science and what it takes to be a great therapist, there are many academic lessons and understanding of human psyche that does go into that. And I think for someone to actually train a really good therapist bot or a conversational tool, you actually have to train it slightly differently. So there are companies and products that are thinking through, how do I gain the transcripts of the public or semi-public conversations that occur between patients and therapists? How do I train the companion on that basis to go deeper and deeper? So I think we'll see more and more emerge.
--:-- Host: That's a really good point because actually even prior consumer companies in a way were companions. You take something like Duolingo, you're talking about a teacher and you add an empathetic element to it.
--:-- Bryan Kim: Or sarcastic in the case of Duo.
--:-- Host: Right, right, right. It depends on the company and the user. But it's an interesting reframing because I think a lot of people think of companions as just this friend or NSFW, as you're saying. But it can be so many other things. Maybe just to round the corner on companionship, because we are seeing these more niche targeted use cases, what does that tell us more generally about the way that these applications are being built? You mentioned at the beginning we're seeing the ChatGPTs and MidJourneys come out strong at first and they still have a lot of engagement. But does that tell us anything about things cornering off?
--:-- Olivia Moore: Yeah, it's something we think about a lot and watch really closely. I think ChatGPT is a great example because they had, of course, the fastest ever product to get to 100 million monthly active users. But a lot of the usage has flattened out. And I think that doesn't mean that it's not a great product or that the model that powers it is not an amazing model. It just means that because these models are now available for other people to build on, we're getting more specific and purpose-built applications that work better for certain use cases. So the blank page blinking cursor is not always the right interface for everything. And that could be a therapist bot. It could be a language learning bot. It could be a design canvas. It could be a lot of other things. So the fragmentation is happening, and it's really exciting.
--:-- Bryan Kim: And not to blow up the conversation a little bit, but if you think about ChatGPT and what powers it, it's the OpenAI large language model underlying it. And that's a closed model that is built for OpenAI and customers of OpenAI. I think what we're seeing is, and this is very exciting for our space of application layer, where the underlying models are getting better and better. And even the open source ones are sometimes even better than the closed source ones. So very recently, the Llama 3 that came out, incredibly efficient, incredibly advanced. Same with Mistral's new model.
--:-- So I think what we're seeing now is people are using those tools to build upon the application layer of products that can be very purpose-built. And if you think about companion, it's a very large word. It just means another being to talk to. So then you drill down further, what are you talking to them about? Teaching, of course, language learning, of course, tutoring, mentoring, therapists. All of these are different modes of interacting with someone. And I think insofar as there's any sort of specific fine-tuning or specific data source you can learn specific interaction models from, I think all of that benefits and drives the case for more niche, but more specialized use cases.
--:-- Olivia Moore: I think it gets back to your earlier question of which categories are we seeing maybe the most growth, the most new products, the fastest adoption. And there are absolutely categories where the delta between the best closed source model and the open source models or the API available models is pretty narrow. Like the best text models are actually now, maybe this is controversial, but some people might say the best open source text models are close to GPT 3.5. Same with image models. If you look at video or music, those are models where still the best open source is maybe not as close to the best things that Runway or someone else has developed in a proprietary way.
--:-- Bryan Kim: But they're newer too, right? Much newer.
--:-- Olivia Moore: So we think it's going to happen, but that has affected maybe the pace of product rollout in these different categories.
--:-- Host: That's a great point. And something you brought up was around UX, right, and how the Google box may not be the box that we expect for the future and also specific use cases. And I think something really fascinating from the Gen AI 100 is you called out a few different categories where we are seeing almost these different modalities. So with music, you called out a bunch are showing up on Discord. Maybe some others are showing up as Chrome extensions. Maybe talk about that and where we're seeing the divergence from the model that we all expected from the get-go.
--:-- Olivia Moore: It's so funny. To be good at our jobs, good consumer investors now, we have to be tracking data everywhere because really interesting products are being built everywhere. I think Discord has been an amazing one, especially for content generation. MidJourney, Pika, Suno, all these companies started on Discord, both because it's pretty easy to spin up the product without having to build a front end. It's pretty easy to monetize and start making money. And because a lot of these products thrive on the community of people who are trading prompts or seeing each other's output. And you can do that really easily on Discord.
--:-- On the whole other end of the spectrum, all these productivity companies, many of them are more about how to make you as an individual doing your individual work faster or more efficient. And so many of these, the key is to live where you are doing your solo hardcore work, which is often on a browser or desktop. And so they're starting as Chrome extensions or voice recorders on desktop or screen recorders on desktop, things like that.
--:-- Bryan Kim: Spot on. That's exactly what we're seeing where the tools are peering close to what needs to get done. Discord, obviously, if you want to just make stuff, it's an incredible place to do that. And to your point, I think the discovery is really unique where you get to see what people actually generated. So fun. It's genuinely so fun. That removes the hesitation of, "Am I actually creating something of value?" You see all the stuff that people created before you, and that gives you that sense of joy to go create whatever you want.
--:-- And I think what's interesting is, again, we talked about web. There's Chrome extension. There's apps. Just thinking about one of the products like Captions, it started as an app because you take videos more and more on your phone these days. And therefore, it was natural to have something that lives on your phone. And then now as it starts to think about moving into workflow and a little bit more professional use cases, it migrates to web because that's where a lot of work's done too. So I think what you're starting to see is eventually the companies and great founders are chasing the use cases and where it occurs.
--:-- Host: I think when MidJourney came out as an example, I'll count myself as one of these people who almost to some degree wrote it off because it was on Discord relative to it not having its own platform. And it's so fascinating to see, in a way, it kind of turned out to be the opposite. As you're saying, when you're close to the users or consumers that you're trying to reach, it was not only better in that way, but also, like you're saying, it was so fun to see the generations and be part of that community, which is something that I certainly did not expect.
--:-- Let's round out the categories here. Just by asking, we talked about a few that maybe people would not have been surprised to see on the list. Were there any that you felt like were missing, like you really wish you saw more of a presence?
--:-- Olivia Moore: I think in general, consumer AI has been characterized so far by categories where randomness and hallucinations are a feature, which would be honestly a lot of the content generation and editing stuff, a lot of the companion stuff, avatar products where you can get 100 photos of yourself and as long as three are good, you're happy. And those are the ones where we've seen the most growth so far.
--:-- And then the other categories where hallucination and randomness is more of a bug. That might be personal finance, wellness, ed tech, things like that. And the models now are getting more precise and accurate, but also founders are able to better build the product that kind of bounds the output in a way that even if there are hallucinations, it can cross-check, it can contain them. But I think that's why we've seen those categories be a little bit slower. If you look at top consumer subscription products pre-AI, which were tons of ed tech, personal finance, health and wellness, that hasn't quite translated to AI yet. But we think it probably gets there in the next year or so.
--:-- Bryan Kim: A lot of the current products that we're seeing on the consumer side are, utility is the wrong word, but it serves something. There's a single use case, whether you're creating, editing, having fun, talking to something, there's a single use case that's very useful. I think what I'm also really excited to see is that when you think about the fundamentals of the business, like where does network effect occur? Can there be a marketplace? What are some of the natural occurring modes? I think my wish is to see more companies with those elements.
--:-- I think because it's so magical, we live in this very, very interesting time. We're in that era of, "Oh my God, if it works, it's worth it and let's pay to use it and let's go." And I think more and more as we see the space evolve, I'm also very excited to see what we had not seen in a ton yet, which are ones that are really benefiting from the underlying network effect that naturally occurs, the underlying marketplace dynamic that could happen between supply and demand. And I think those are the ones that I think we'll also be on the lookout for.
--:-- Host: Yeah, and I mean, let's talk to that specifically. Olivia, you have widely cited this term, AI tourist phenomenon.
--:-- Host: I don't think this is a surprise to anyone. I mean, we've all tried out so many of these tools. It's so exciting. And then we also have left many of them. And you can even look to the list, right? You said around 50% went from September to January. That could be a glass half full. Look how many stuck around. Or glass half empty, where, look, 50 of them were here, and now people are not as interested. So what is this data telling us in terms of stickiness? And is this really still a thing with the AI tourist phenomena, or are we moving past that?
--:-- Olivia Moore: We talk to our founders a lot about this. I'm not saying it's easy, but it's easier than it has been before maybe to get users for a consumer application. And that's just because there's so much excitement. These products are so cool. There's demos going viral on Twitter, on Reddit, on TikTok. There's newsletters, Discord groups. And because of that, many of these products are getting floods of users and traffic like we've never seen before.
--:-- And those users might try it out once or twice, but they might not actually be in the core persona of who's a good fit for that product. And so they might not convert to pay or they might not retain and come back to the product the next day. You might say, if it's free to get the users, it might not matter. The problem there is a lot of these AI products are actually expensive to run. We're not in the same world, right? And so sometimes we see founders get in a place where they call us and they were like, "Oh my God, we're out 40K overnight because it went viral in India or something, and we got a million users, and they all used up our maximum free trial, and none of them are paying yet." And so that's something to look out for.
--:-- I will say we have almost redefined retention for consumer. It used to be free user base, like anyone who downloads, anyone who engages.
--:-- Bryan Kim: From install and reg success.
--:-- Olivia Moore: Exactly, yeah. And now the bar to count as an active user is just higher for us. And we measure retention only off of that. Usually that's a paid user. Maybe if they're not monetizing, it's have they completed X actions. If you look at it that way, the retention for AI products is actually as good as or in some cases better than non-AI products just because these companies are amazing. But if you measure it on the tourist alone, the picture can look a little tougher.
--:-- Bryan Kim: I think our learning about the AI tourist phenomenon, I think we almost put a number to it to some extent. It extends the overall top traffic, top of the funnel by near 40%.
--:-- Olivia Moore: Yeah, you almost add an extra layer.
--:-- Bryan Kim: Add another layer. So I think what Olivia is suggesting and what we're doing is actually thinking through what is an actual user? Have they completed the behavior that counts you as a modified user? Because the willingness to try something is so high. It's never been so high. So I think defining that and starting from the right touch point, and if we count backwards to actually think about retention, because we all think for a product to survive and do well over long term, people just need to come back. That's the key to it.
--:-- So I think what we're seeing is a very high number of companies are able to translate this top of the funnel into paying user at a very healthy clip. And what's more is that we talked about the willingness to try is very high. Willingness to pay has also been incredibly high. Because the product's so magical and because there are actual use cases, not just personally but also commercially, the willingness to pay has been quite high. And as a result, we've been seeing a lot of companies actually get up to tens of millions of dollars of annualized revenue in a very quick manner.
--:-- Olivia Moore: It's crazy, yes.
--:-- Bryan Kim: So that's been a really interesting justification. When we're asked, why are you only focused on AI products? What about the non-AI products? We are not saying non-AI products are not interesting. They're very interesting. But what we're seeing is the willingness to try and willingness to pay has been so high for these products that the velocity to get from nothing to maybe tens of millions of revenue have never been higher. And that's very compelling.
--:-- Host: Let's get to how we keep those users around. But both of you spoke to a few metrics there. And I know we're far enough into consumer tech that there are several benchmarks, best practices that you look for. I mean, both of you sit in so many deal meetings and someone comes in and let's say five years ago, there was very clear, you're looking at daily active users. You're looking at day seven, day 30 retention. There's things that you know automatically, like it's in your bones. You know what's good and what's not good. You can see a chart and you know if this company is doing well or not. Has that changed? Are you still looking at the same metrics or how do you interpret or add new metrics in this new era?
--:-- Olivia Moore: We're looking at some of the same metrics, but maybe in a slightly different way. For the more work-oriented, prosumer, productivity SMB tools, we look at a lot of things like the wow-mow ratio now, which is, is this truly something you're using every week for work, or is it something you're maybe in once a month for two hours, which can still be interesting, but is probably a little bit less compelling.
--:-- Host: And just for the listeners, that metric is weekly active users divided by monthly.
--:-- Olivia Moore: Exactly, yes. We'll also look at conversion to paid for those. And then, like we mentioned before, we will do standard monthly retention cohorts. So that would be of all the users who signed up and paid in month zero, how many are still paying in month one and how many are still using it? How many are still paying in month two? But in pre-AI consumer, the denominator there was all free users. And now we only measure it mostly on paid or really active users just because of that tourist effect.
--:-- Bryan Kim: I think pre-AI, the denominator is slightly different. And therefore, we would count daily because that's actually what you have to clear the bar for a free user base and all that. I think what we're seeing and the reason we're moving to weekly and monthly is because as it becomes a little bit more prosumery, as it becomes a little bit more commercially relevant, it's not obvious that you will want to use these tools every single day. And so naturally what we're now expanding into is thinking through, weekly usage rate, retention rate, monthly paid retention.
--:-- And I think what's really unique here is that if a tool is very useful, it's not guaranteed that you will even use it every single week. So now we're thinking through, okay, you're still paying. That's good. How many outputs are you creating in a month?
--:-- Bryan Kim: Because it's possible you sit there for eight hours straight and crank out hundreds of outputs that you need from the product. And it's incredibly useful to you. But you only showed up one day out of a 31-day period. Is that good? It's kind of a blessing and a curse for AI companies. If you're helping people do their jobs or make their art much better and faster, they are going to use you less because you've made them so much more efficient. So it's almost about measuring value-based—how much value did you deliver to the users? For a video editor, that might be number of downloads. But maybe because of AI, you can now plug in all your videos for the month and do it in one week instead of having to come back every other day pre-AI to generate and edit again and again.
--:-- Olivia Moore: That's such an interesting point. I think about something like ChatGPT. If it's $20 a month, $30 a month, one really good engagement can be worth that. It speaks to the value of these tools. These aren't micro interactions where you get 30 cents worth of Twitter. If this can save me, if this can really help me do my job better even once... And we wouldn't know, right? It could have saved the person 20 hours. That's incredibly valuable. But you only see it as one engagement.
--:-- Bryan Kim: Totally. So you have to look at, do they keep paying? Not just how much time are they spending in the product. In many cases now, especially for productivity products, the faster you can get in and out of it, the better, as long as you're still getting the result you wanted.
--:-- Host: And one thing that I would just—this is a plug for our firm—we see a lot. We meet these companies a lot. What's helpful for us is that we try to define and understand what metric we want to track, what is important to us. And then we have the discipline and rigor to continuously ask for that and build out a strong mental model with an actual end count of the companies that matter to that category. So when we actually see an exceptional product, we immediately recognize it without having to scramble. The number of companies we meet and how we define the metrics rigorously and track them carefully gives us the ability to recognize what might be an exceptional thing very quickly, even if the metrics are evolving with time.
--:-- Bryan Kim: You can compare it to others.
--:-- Olivia Moore: Yeah, or different across categories. Exactly. Like an image generator, maybe we look at weekly bounded retention. But a companion product, maybe we are looking at daily over monthly active users. It's a little bit different for every company type, but we do try to closely measure and collect data points across hundreds of relevant companies.
--:-- Host: Yeah, because I guess what you're saying is each founder only sees their data.
--:-- Bryan Kim: Exactly. And so they're like, "I have no idea if this is good or great." We would be able to tell you that top portal, top decile. Ultimately what we're trying to measure is, is a product delivering what it's meant to deliver? And what is a metric that best captures that moment?
--:-- Host: Now, Bryan, you have said before, for consumer products, the rubber hits the road with retention. We've touched on this already. But what can we learn from the prior era in terms of retention? Because I feel like in so many AI podcasts, this is the question that comes up continuously: Where are the moats? If it's so easy to build today, especially as open source models are getting better, how do I stand out? How do I keep my users? So what can we learn from the past? And does it still apply here?
--:-- Bryan Kim: This is a fun one because I have said retention is very hard to game. And it is true. It's always been hard to game, harder to game than gross user base and overall user base. I think what we've learned from the historical or pre-AI consumer companies is that there's a specific segment of very forward-looking founders who have learned a way to actually improve retention somewhat artificially. And I'm not saying those are bad things, because if it serves the need and helps the company deliver the core product value faster and better, then that's great. There are tried and true methods—call it six to eight different types—you can employ to improve retention. And we're seeing some of the Gen AI or AI native companies employing some of these methodologies to actually improve their retention, keep their users longer, and deliver new products to them again and again.
--:-- So I think we're learning a couple things. One, retention still matters. If your users aren't coming back, that's not good. The frequency can be a little different because around workflows, you can come in less frequently but get a ton of value. Two, what we thought to be largely ungameable is somewhat movable. There are methods you can learn from non-AI companies to apply to your product to achieve that. And three, retention is an output. It's an output of what your product does, whether that's actually really useful to people, and whether you delivered that quickly and often. That's really the crux of it.
--:-- Olivia Moore: I completely agree. The other thing we're seeing with retention that was also true pre-AI, but maybe even more dramatically true, is the narrower and more focused the product, the better. We have companies with a ton of compute—ChatGPT, Microsoft, Google, Notion—all building and releasing more broad-based AI products and applications. It's hard to compete as a startup with a broad-based product if you have one one-thousandth of the compute, the team, the engineers. You don't own the IT authentication, the data, the years of history. So what we're seeing work really well in terms of retention... Sometimes we meet a company that's a very horizontal product and the retention is just okay. Then they come back to us five months later and say, "Actually, we realized we're building for this core set of users. We honed in on this specific model and built 10 more features just for them, and our retention is four times better than it was before." That kind of thing is working really well. It's counterintuitive because no one wants to build a product that's too narrow, but it's better to go narrow, have amazing retention, and then expand than to try to do it the other way around.
--:-- Bryan Kim: It's almost been like a quest for founders who have this amazing technology at hand, asking themselves, "What is this good for? Who is it good for?" And oftentimes you find the answer in surprising places. If you told me initially, "Hey, you can clone yourself as an avatar and present yourself," my first guess wouldn't be, "Oh, this is going to be amazing for learning and development within companies." My first guess isn't that salespeople can send an avatarized version of themselves into a sales process. That is not my first instinct.
--:-- Bryan Kim: The fact that founders are able to hone in on customers who are willing to pay—to Olivia's point—is very unique and important. You found an interesting niche, a narrow use case where people are finding so much value. Oftentimes those are the great places you can corner and start expanding the market. Founders with that mindset of finding their ICP and appealing to them as a way to grow more horizontally step by step—that's been an interesting model, especially in competition with larger companies.
--:-- Olivia Moore: One thing that's interesting is it's not just the model itself or even fine-tuning the model. It's everything built on top—the UX, the marketing, the messaging. All of that comes together to be just a little bit, or sometimes a lot, better than the more generalized model. This is a feature of AI. Things are so magical and evolving so quickly. The question is, how do you differentiate? Sometimes the differentiation is, "Our tech is so good it blows everyone out of the water." And sometimes that's true. But a lot of times the world is large, a lot of great people are working on great problems, and a lot of smart people are working on the same problem. So what we end up seeing is the velocity of product shipping always matters, especially when things are changing so quickly. As we saw with top AI properties, rankings change so much. That means there's a lot of velocity. You need to stay ahead of that.
--:-- Bryan Kim: Velocity matters. But what's really important is how you build consistency and retention. That's by building what's useful to the users. That's why I think we're seeing that if you deliver part of the workflow to make their life so much easier, that's where we see differentiation in companies.
--:-- Host: It's so easy in consumer, especially when we're meeting incredible teams every day, to get enamored by "this is the most elegant technical approach" or "this is the best research team." In some cases, that is what wins. But if I take off my investor hat and think about myself as a normal person downloading an app or going to a website, I do not care about the technical elegance.
--:-- Olivia Moore: No, I don't even care about who made it.
--:-- Host: Right. I care about if it helps me get the thing done that I wanted to get done. It often comes down to little workflow things—tiny features or how you scope the product that make the difference. That sometimes doesn't come down to the technical details, but to these micro product decisions that can be make or break.
--:-- Bryan Kim: It's kind of funny that we have to remind ourselves of that in AI. No one ever cared if an app was made with Angular or React. No one ever sees that.
--:-- Host: I feel like I'm going to start a war online.
--:-- Bryan Kim: Too cool.
--:-- Host: Maybe one other question. If we're talking about competition, in the last consumer wave, we did see some companies front run. Maybe they did it by raising a bunch of money and doing a lot of paid acquisition. Then you start to see things like network effects kick in. Are we seeing that similar dynamic? Where does paid come into play? Both of you have mentioned that people are willing to pay. Does that mean budgets can increase, or CPA can increase?
--:-- Olivia Moore: It's a really interesting question. I think there are categories where raising a lot of money to build the best model does make a really big difference and having a best-in-class product. We are investors in Eleven Labs, a text-to-speech company, and they have an amazing model. Because of that, they're used by probably thousands of developer customers and other customers to power the product. In that case, it's harder to compete if you're not raising a lot of money. More money means more data, more tuning of the model, a better model, and it becomes a spiral. But there are other product categories. For example, many products are building off of open source models. Then it's much more about product elegance, how you commercialize the model, how you take someone else's tech and translate it to something that artists or designers can use. Trying to front run to raise the most money and acquire the most users isn't always the winning strategy.
--:-- Bryan Kim: Maybe another way to think about it: the models and the research are so magical that they often go directly to consumers. That's very exciting because they immediately get the benefit of cutting-edge research. In terms of paid acquisition and how that translates into this new AI world, I think there are two classes. There are companies that benefit greatly from the buzz and the virality of the product—examples they can put out on Twitter, Reddit, or Discord because it's so fun, eye-catching, and attention-grabbing. Sure, there's a tourist phenomenon, but they benefit from great top-of-funnel traffic. As long as that traffic continues to come in, you can find ways to convert them into paid users, start making great money, and build out waitlists or inbound-based sales leads. If you're thinking about your product as a workflow tool that's so useful that SMBs or enterprise customers reach out to you, you can build a pretty good inbound list. In those cases, CAC or paid matters a little less.
--:-- There are other businesses where the product is very good and very useful, but it hasn't fully benefited from the halo of the amazing glitter of AI apps. In those cases, because the willingness to pay is quite high, we see a crop of products that engage in paid acquisition in a thoughtful manner. The LTV is there, so they can afford to pay to acquire users. Some companies know how to do that better than others. That has been a model we are seeing, and oftentimes these companies can build their run rate up to tens of millions of dollars, if not more.
--:-- Olivia Moore: Totally agree. If you're building a product for AI artists, designers, or writers, it's very easy to go viral and bring in a bunch of users on YouTube, Twitter, or TikTok. If you're building an AI platform for small HVAC businesses to be later expanded into other home services businesses, going viral on Twitter may or may not actually help you. You're probably still going to have to build out a more traditional lead gen sales funnel, or at least a strong referral program.
--:-- Host: To your point, Steph, you mentioned you could brute force build a network effect. I don't think we're seeing that just yet because there aren't true business models that truly benefit from network effect or marketplace dynamics such that people will want to buy their way into that density. What we are seeing is that there's a true payback related to paid acquisition, and they make calculated return-based assumptions to go acquire users.
--:-- Bryan Kim: I like that distinction. But do you expect that to change? Do you expect in a few years for that to be true, where there will be those marketplace effects?
--:-- Olivia Moore: I do. I think so. I think it comes back to the question of what has matured already and what is to come. An example we talk about a lot is that we haven't seen a lot of truly AI native social apps. Part of it is because there were some early tests, and a feed of content where you know all of it is AI is actually maybe not the most compelling social app. It doesn't have the same psychological dynamics.
--:-- Bryan Kim: We're learning a lot about ourselves with AI.
--:-- Olivia Moore: Exactly. If you know it's a fake picture of you, you don't have the same gut feeling—either anxiety or elation—in posting it. So we're just starting to see things like a product called AirChat that's been viral the last few weeks. It's about helping human beings create human content easier using AI. You can do a voice memo and it will transcribe it into text. Then you can scroll through and read a feed of text. It's basically opening up people who would never tweet because they don't want to sit down and write, or they're not good at it, or it stresses them out—which I totally understand—who might do it with a voice memo that gets transcribed into text. That's just one early example, but I feel like we'll see more of those.
--:-- Bryan Kim: I definitely think we're starting to see the very early incarnation of this. Olivia, you mentioned Eleven Labs, a portfolio company. You actually see these products start building a marketplace model within their company. Eleven Labs actively builds a marketplace for voice actors to license their voices so they can make passive income. Same thing with Captions. They have a creator marketplace. They're starting to build that out where creators can license their likeness so video ad producers can use it, and they get to sit there and make passive income. These marketplace models are interesting. If you have a lot more supply coming in—either through paid acquisition or not—building that density starts to accumulate in terms of benefit vis-a-vis competitors. We're starting to see some concentration of these types of behaviors, marketplace dynamics.
--:-- Host: That's actually really interesting. In a way, they're passively building up one side of the marketplace, but they're not starting as a marketplace as you might have seen in the past. Fascinating.
--:-- [Discussion continues]
--:-- Host: We started this off by talking about the Gen AI 100. Let's end there too. Let's say we run this six months from now. What do you expect to see? And also, what do you want to see?
--:-- Olivia Moore: Looking forward, I would love to see new categories start to mature. We talked about two completely new categories in this most recent list: music. Suno popped up from nowhere. And even since we published the list, Udio has now popped up and gone totally viral. As more models mature, we'll see new categories. Productivity was another one that appeared almost nowhere on the first list and has come out of nowhere to have quite a few companies represented on the current list. It's a little tough to predict. If we knew what the next big AI hit would be, I don't know if we would go build it. We wouldn't find someone to go build it. But I hope to see a continued testing of the boundaries and expansion in form factor, modality, and categories. The amazing things we see six months from now are things we probably can't even conceptualize right now. Maybe you can, but I cannot.
--:-- Bryan Kim: I think Olivia's exactly right. We're hoping to see net new categories. I expect to see another 40% of the list being net new, if not more. What I hope to see is that the prior lists had single modality products—largely text, largely audio, largely music. I'd love to see what happens when you start combining these. What happens if you have video plus image plus sound effect? Is that a music video? That's cool. What does that look like? When you have avatar plus voice, what is that product? We'll love to see these net new categories where we will have a hard time defining what they are because they combine these different things. To use Olivia's example of AirChat, that's really interesting: voice transcribed into text. And we also know that text input can do pretty much anything—create music, video, avatar, 3D images, anything, code. So what does it mean when all modalities are essentially interchangeable? We don't know, and we're very excited to find out.
--:-- Host: We're very excited to find out. That's amazing. Well, this has been so interesting. We will have to sit down again, whether it's in six months or whenever you guys have a new list, and see what has changed. It feels like other lists that have been done in the past, you have to wait another year or a couple years for enough movement to happen. I feel like we honestly could record this in a month and we'd have enough news to talk about.
--:-- Olivia Moore: There are a few companies probably that have cropped up while we're recording this.
--:-- Host: Yes, exactly. Amazing. Well, thank you.
--:-- Bryan Kim: Thank you.
--:-- Olivia Moore: Thank you.
--:-- Host: If you liked this episode, if you made it this far, help us grow the show. Share with a friend, or if you're feeling really ambitious, you can leave us a review at ratethispodcast.com slash A16Z. Candidly, producing a podcast can sometimes feel like you're just talking into a void. So if you did like this episode, if you liked any of our episodes, please let us know. We'll see you next time.
