Hi, my name is Josh Beck with the Ray Jay Internet Team. We're super happy to have Yelp join us today. We have David, the CFO. He's going to read a quick statement, and then we'll get started. If any questions, just raise your hands.
Thanks, Josh, for having us at the conference. We'll be making some forward-looking statements during the conversation today that are subject to risks and uncertainties. Please refer to our SEC filings for more information on the risk factors that may affect our results.
Less than 20 seconds. Well, I wanted to kind of just start off with, obviously, the big theme of tech for the last few years and probably the next few, but really kind of your AI strategy. Certainly, you have a lot of new tools with Assistant for Request a Quote, for the restaurant app. So you have a lot of kind of on-platform innovation that you're doing. I think there's probably also an opportunity to take it off-platform. So maybe just kind of high level, how do you think about the AI strategy, where it is today, and kind of where you'd like to take it in the years ahead?
There's no doubt, and I think it's obvious to everyone, that we are entering a new era in terms of the way people interact with information. Particularly around search, which in the past was all about links and clicks, we're moving into this conversation where people expect answers and to be able to take action. We've really embraced that. It's not just that we're transforming the consumer experience on Yelp, but we're broadly applying AI to everything we're doing. I think you can say that Yelp itself is transforming with AI. We're not just transforming the product. We've embraced it. We think that it presents a huge opportunity for us. We think that for a couple of reasons. First of all, in this era, human-generated content is essential: authentic, informed, trustworthy.
And we're known, obviously, for being very careful about curating our content and ensuring its trustworthiness. So that's the first part. That's obviously the core asset of Yelp, are the reviews and the ratings. Second, of course, is technical capability. And we believe that we are quite good at applying AI to practical problems. And I think this word "practical" is where we are at, and I think many other companies have come to, which is there's obviously been a lot of lofty ideas around what AI can do. But at the end of the day, can you do something practical that's beneficial, useful, and valuable to both consumers and, in our case, advertisers?
So we've really focused on that in both building out our Yelp Assistant product, which we can talk more about, as well as bringing to market two voice products, which we think are certainly at the front of what's capable now from a voice and interaction perspective. And we're doing that both on Yelp and off Yelp. So broadly speaking, we think that we can really take advantage of this moment.
So one of the things that really stands out to me about Yelp is certainly just the scale of that trusted content. I think it's almost 30 million monthly active, over 7 million businesses, over 500,000 paying locations. This is a very unique asset. Certainly, some of this is available at Google, but outside of that, there's really nothing quite like it. When I do searches on maybe ChatGPT or Perplexity and kind of get into the category of restaurants, the reviews they're sourcing are pretty arcane or not even really necessarily a review. It could be something like they're sourcing Wikipedia. No one is looking for a restaurant review on Wikipedia. But it certainly seems like there's potential there. But I imagine you've had discussions with some of these platforms. So maybe just kind of help us think about the partnership.
Obviously, I know data licensing has been an important initiative for you, but how do you see that evolving?
Absolutely. And just from a unique perspective, I think Comscore depends on the month or time of year, but we're in the 70-80 million monthly uniques according to Comscore. We publish our actual numbers annually, so we'll have that when we report our fourth-quarter earnings by platform. But there is no doubt that if you want to be in local search, or if you want to be in search broadly, I should say, a significant proportion of people's searches, and the estimates vary a lot, but we've heard estimates of 25%-40% of searches on Google have local intent. So if you want to be in the search space and you're not Google, we believe that we have the most comprehensive directory of local businesses combined with our hundreds of millions of reviews and ratings, which are, as I already talked about, actively curated for quality.
And you have to have a platform that's able to deliver those and meet the technical requirements. So we believe we're really well positioned to partner. Today, we are on Apple Maps, for instance. We're in Alexa. We're in a significant proportion of the infotainment systems for OEMs around auto. And we have quite a broad array of partnerships. On the earnings call, Jeremy talked a little bit about it. I think you could hear the enthusiasm in his voice for the discussions that are ongoing. We don't have anything to announce now. But we think that if you want to do search well, you have to do local search well. And we think that we are the preeminent partner for delivering that experience to consumers. So we feel like we're really well positioned there.
Of course, in terms of quality, there's another dimension to it, which is brand equity and brand perception. If you just get a rating and say it's 4.4, but you don't know the source, could have been referred to, but you don't know where it came from or your Wikipedia example, people don't ascribe a lot of authority to that. But when you see the Yelp logo or the Yelp name next to a rating, it conveys authority. We've done the A/B testing to see whether or not people do click through at higher frequency. We've done the qualitative studies. It's clear that that Yelp brand on another site directly next to a rating is very important.
OK, that's very helpful. Maybe going back to some of your earlier points about practical AI, where is the kind of assistant today versus maybe what you have on the roadmap for the next year? Certainly, it seems like for the Request a Quote flow, it's really quite efficient. It has a nice summary at the end, and then you can kind of request a list of pros. So it certainly seems very additive. And then certainly, it seems to be layering on more so kind of in the restaurant tab of, would you like to understand if this food is spicy, or how spicy, or how bad is maybe the bathrooms, or something kind of random that's not high level. It seems like it's kind of adding another layer. So kind of where is it today? Maybe what has the kind of customer engagement been?
Kind of where do you see that headed?
Yeah. So we developed Yelp Assistant for specifically our Request a Quote product, which, for folks who aren't familiar, enables people to find a service pro and ask a few questions in order for us to match that consumer with that service pro. And we think it's a better experience, but also it's able to elicit more useful information for the matching algorithm because it's dynamic and it's context-[audio distortion]. So we've invested heavily there. One of the things that's interesting about AI is that the first time you use it, it has this huge wow factor. Amazing. I can't believe it sounds human. It's surfacing interesting information. Now, it's not always reliable, obviously. So it takes a lot of work to make it reliable. But the other thing is that's just the starting point of creating the right user experience.
What we found is it takes a lot, a lot of experimentation, iteration, and thought around that UX in order to create a useful experience. We spent basically the past year doing that, and now we're rolling it out to all categories. This idea that you can have a conversation, that it can prompt you with questions, I think is very powerful. It provides a lot of opportunity to do in-product marketing. We are rolling out, and we're actually in testing now for an assistant that works across all of Yelp. Not just on the business page, which is where you see it today for things like restaurants, but actually for all categories. That's in limited release. We're continuing to do experimentation around that. We're just excited about the way that we can engage with folks going forward.
As I said, I think we clearly all believe we're entering this new era.
How about taking the Yelp Assistant off-platform? It certainly sounds like it's going to be more pervasive on-platform as we go through the year. How do you see the off-platform opportunity?
We think there's a significant opportunity there. It can take several different forms. So there's our traditional licensing business, which is surfacing Yelp content. And that can be done through API. We also now have a capability for that API to call our assistant. And I think it's still an open question. What does the future hold? Is it an era in which an assistant calls other assistants and identifies that it's calling other assistants? Does it just surface that information? Do people prefer to go to each assistant and use it independently? I think it is still emerging in terms of what the consumer preference is going to be. Clearly, on the app side so far in the United States, or since the iPhone came out nearly 20 years ago, the experience has been people still choose a specific app for the thing they want to do.
There's no super app in the United States. I don't think we are going to end up in a place where there's a super assistant. And I just use one assistant to do everything I want to do. I don't think that's actually consumer preference here. And so my expectation is that you're going to find that people want to engage with the specific assistant within other assistant experiences. And that could be on a variety of different platforms, for instance, say, in Alexa or something. So we are building all the capabilities in order to power those experiences, but it's just still very early.
OK, that's very helpful. And then maybe some of the voice products, the early kind of reception, well, or feedback on Host and Receptionist, what has been some of the early learnings there?
Yeah, it's a really interesting emerging space. And there's a couple of attributes that are pretty essential to these voice products. And then I'll talk a little bit about sort of the limited qualitative feedback that we've gotten so far. Since it's still very early, the product's only been out for a few weeks. The first thing about these voice assistants, to avoid someone just hanging up the moment they hear it, it cannot sound robotic. It has to sound human-like. And it needs to convey that in the first seconds. So the demand for this is very high: low latency, human-sounding, interrupt-tolerant, and able to handle a broad array of the phone quality could be poor, as an example. It needs to be able to handle a lot of different voice conditions.
So we've spent a lot of time on the technical side in order to create that first moment of engagement and experience so the person doesn't just say, "Transfer me," or "Hang up." That's the first hurdle for anyone that wants to build these products. And there's a lot of discussion about voice-to-voice, voice-to-text-to-voice. And I think there's a lot emerging there. But we feel really good about that capability that we've built in the first place at that front end. Then, of course, it has to do something useful. So it's got to connect to lots of other systems, and it has to do that quickly. So this is not a trivial product to create that people, consumers, are willing to use. And businesses aren't going to buy something unless consumers are willing to use it because otherwise it's not delivering value. So that's where we've been very focused.
Initial reactions have been positive. People find the voice quality very impressive, and they see that it can handle a broad range of questions, and it has a lot of knowledge associated with it. We've done a lot of work to tune that, and I would also say it's early. There's a lot of iteration to go in making sure that that experience is effective, and one of the things that's pretty clear is that you do need the service pro or the restaurant owner to engage to make sure that the content that they have and the way they want to convey that content has been input to that tuned AI for their business. They have to participate.
So the implementation ends up being important and making it easy for them to enter information that can inform the AI for their business to create the right experience is also essential. So you can see, again, you have this I think we all have this perception, "Add AI, magic happens." But the reality is it takes an enormous amount of engineering and product work to make these things useful. And that's, I think, where we are focused. And I think broadly, the industry is coming to realize that that quality of experience is pretty essential to adoption.
Yeah, I think it's a great point. I think we all have a lot of baggage from a bad robo call. And if you hear that, if you hear that same tone, you instantly want to get a human on the line. And that's kind of been the last couple of decades. You have to kind of unlearn this and have a better.
Zero, zero. Transfer me, transfer me. IVR, I don't like this experience. And we're right on the cusp of these products that people are like, "Oh, yeah, this was useful and engaging and answered my questions and occasionally seemed a little humorous.
Yeah. To your last point there, how much data do you need about these businesses? You said there's certainly effort on their behalf required.
Yeah. So one of the beauties of Yelp, of course, is we have this directory. We have a lot of information already about the business, and we have all the reviews. So we come with this advantage compared to some others in that we're already pre-populating all this information into the experience. So when we demo the product for a specific business, it's already specific to their business. But there's a lot of things they might have preferences about. First of all, they get to choose the voice itself. And in time, I think we will offer something that enables people to, if they prefer to record their own voice and create the experience with that. That's technically very feasible. But again, getting that set up takes some work. So that's one aspect.
I want it to sound like me" is something that's coming, but I think it's important to people, but it's much more about, "I want it to answer this way," or, "I want its tone to be this way," or, "If it's after hours, I want to engage with the person in some way." You have to enable the business owner the ability to customize that experience for their preferences, and you have to do it in a way that doesn't require additional engineering work. That's the trick of it.
What does the adoption curve? We're kind of in this awkward part of the AI curve where we've had tremendous gains in intelligence, and some areas really take off. Customer service in this kind of voice area seems like a real natural power alley, but it also does seem like it requires quite a bit of work to make the product-market fit to be better or at parity with a human, so is that something you see happening in 2026? And then 2027 is maybe when the adoption inflects, or how would you think about that?
Obviously, it's a little hard to predict. I think in the S-curve, we're sort of in that flat entry point. And then there's the scaling moment that's definitely coming. Is that in eight months, 16 months, 24 months, 32 months? I think it's a little hard to say. But I'm pretty confident within three years, it's going to be ubiquitous. But it's not ubiquitous today. We're not quite there. There's still a lot of learning to do. And there's still consumer preferences. People are not yet attuned to, well, let's start with people. All of the AI products basically can have error associated with them. And when you do the surveys, what you see is large numbers of people use these products, but with low confidence. That's a little bit of interesting insight that we've gotten from our qualitative research recently.
It's relatively low that the person says, "I both use it and trust it." They say, "I use it a lot, but I sort of need to verify it." And that's actually where Yelp comes in because the authority is, and here is the quote from a review. You can present the evidence. And I think probably the other dynamic that's at play around the search products is there has seemingly been a reluctance to link out from the products in the experience. And I think that's pretty much across the board. And yet people want the link. They want to go see for themselves. They want to verify that the information is accurate, or they want to increase the amount of information they're getting around that particular question or topic that they're exploring.
And so again, we think that we're well-positioned to provide the evidence that the answer that they are getting is accurate or trustworthy.
If there's any questions, just raise your hand. I'm going to shift gears a little bit towards the financials. Within the services business, obviously, it's been a real standout, going low double digit or high single digit more recently. What are the kind of key drivers in that business in the next year or two? How do you think about the kind of durability of growth there?
Over the past years, as folks who have followed us know, we've really shifted the share of revenue from services up dramatically. We're closing in on about 70% of ad revenue coming from services, although I would say broadly, consumers still know Yelp for restaurant reviews. Part of the success there has been our ability to differentiate the experience between the Restaurants, Retail, and Other side and the Services side. Request a Quote has been a big part of that. I would say this shift to the conversational interface, the assistant interface, really is that next phase. It's not only important, as I mentioned earlier, in that we can get more relevant information and fewer questions dynamically through a conversation. I think we've all had the experience where the assistant asks you a follow-up question at the end of the interaction.
Can I do this for you? Would you like to know more about that?" I really think this is the next phase, which is driving engagement. "Hey, I saw that you hired the plumber. How'd it go? Would you like to leave a review?" Or, "Hey, I see that you did some landscaping work. Have you done your gutter cleaning this spring?" Or, "Did you realize did you go to that Italian restaurant? Did you know Yelp does services?" Or, "Do you have a service need?" Or, "Do you need a plumber?" Or, or, or. That's just a totally different dynamic era, broadly speaking, in the user experience that we are definitely focused on. So if you ask me, where do we head with services, it's at the end of the day, a service pro wants a highly qualified lead. That's what they value.
What we want to do is connect consumers with the right service pro. We want to continue to engage them by having connected them with the right service pro. So I think there's benefit across everybody. The other thing that is certainly true is this user experience when it's a conversation can be monetized at a much higher rate without a perception that you're increasing ad load. So when people today go through the traditional Request a Quote flow, which is a question tree, they are shown typically four businesses who are all paying advertisers. Those advertisers charge to click to be shown as one of those four when the results are presented in our Message Center. When you move to this conversational interface, it's very likely. Well, it's Request a Quote. So we have that same 4x in terms of clicks.
But broadly, that enables you to present businesses that are paying for advertising. But there's a real constraint around that, which is you still want to be able to surface these organic results. And consumers are not going to be satisfied with only seeing businesses that are paying advertisers. So we remain very focused on the consumer experience across all of this because we want to deliver value to folks. And you need to iterate around all that. So even though there isn't a perception of higher ad load, you still need to ensure a quality experience and discovery. So we're all learning how to do that. I don't think it's figured out yet by any means across any of the platforms what is the ideal experience. And I expect there's going to be a tremendous amount of experimentation.
Just as in search converged over probably a decade on what the optimal experience was, I think that we're going to see the same thing with this AI conversations.
Have you seen conversions improve with somebody going through the assistant process versus the tree? I mean, certainly, that's kind of been the message from the broader search space. When you go through one of these longer queries, you have more context, maybe less clicks, but better conversions. Are you seeing something similar?
So one of the stats that we shared on the Q4 call was the 400% increase in Request a Quote requests through Yelp Assistant. Now, that's partly because we are putting it in more entry points. But it's also a reflection of people choosing to use it. And so we feel like we do get better engagement. We do get better information. Interestingly, when we first did it, the matching algorithm performed poorly compared to the question tree. And we're very surprised by this. This happened last year. We absolutely had better information from the consumer. And yet the matching was poorer. Why was that? The reason was that the algorithm was tuned to the question tree, not to this information that was being elicited from the consumer that could make for a better matching.
So we actually had to rebuild the matching algorithm to use and more accurately match with the better information. That was a surprise. That's just an example of the unexpected interaction between these different systems that occurs when you're creating these new experiences. There have been a lot of moments like that where we learned and iterated and improved through experimentation. So we're really at that same phase. And we feel very optimistic as we are able to roll this out, as we learn. Yelp really does approach product development through experimentation-driven agile development. That's our approach. And I think it's been quite successful. Our product velocity is much higher. We were really proud of the release that we had in the third quarter. You can see it in the shareholder letter, all the things that we're doing. Product velocity is high. We've embraced this.
So a lot more to come.
How about on the margins side? This has been a big topic as apps transform to maybe have more of this assistant component. How do you kind of manage the margin structure? I've seen some of the companies that I follow that have maybe acquired an AI-native app. They've seen a lot of pressure on margins to some degree. Is this something that you see as somewhat of an investment cycle that creates opportunities? How do you think about the margins?
We haven't had that experience. Interestingly, our actual spend on models has not grown dramatically. It's up, but it's not grown dramatically. What has actually gone up a lot is we're on AWS. We spend a lot more on data than we did in the past, not on the models. Also, the way that we've architected these systems, we've done a couple of things. One, we can swap any model in. So we're not dependent on a single model. We're always looking at both open-source and paid models. And we are stratifying the compute among models. So you can use a less expensive so there's a supervisory model. When you first type in the query, "Hey, is this a valid query? Is the person asking something inappropriate?
Is it on topic?" There's a very long list of things that you need to check around that query to ensure that it's a good query, a safe query for us to process in the actual model. So that can actually be done with a pretty low-cost model. Then when you actually have to answer it, it's different on the voice side versus the text side, and then you really want, in essence, to minimize the thinking tokens. You want the thinking tokens to be on the heart of the query, and everything else, you want to be low-cost tokens, so we're not over-focused yet on expense because it just hasn't been an issue, but the whole industry is going to end up stratifying tokens across these different parts of the experience to manage cost.
There's also been a steep decline in the costs of these models because of competition, which we obviously welcome. There's a lot of models in the market, paid and open-source, and so we've also been able to take advantage of that, so I have heard this for other platforms where they're seeing significant cost. We've been so far, I think, very effective in managing that cost.
OK. I think that's a great point to end on. Thank you, David, for the time. Thank you, everyone else, for joining today.
Thanks for having us.
Thank you.