Name: Yext Research Live: What Actually Decides the AI Answer
Uploaded: 2026-09-08T18:16:11.132Z
Duration: 59 min
Description: Yext Research Live: What Actually Decides the AI Answer

Transcript for "Yext Research Live: What Actually Decides the AI Answer": Hey, everyone. Good afternoon and good morning. Welcome to q three Yext Research Live. I'm Javier Fabrega, VP of data strategy. Today comes down to two questions. What does AI say about your business and doesn't mention you at all? Search usually just match your listing to what somebody typed. AI decides for itself. It picks a handful of things to to say about you, and mostly, they aren't fields you've been filling in. It also picks who's worth mentioning in the first place. For example, ask for AI engines who's best in town, and they agree only 4% of the time. Two people are here today to break it down for you. First, Anthony Rinaldi. He's our VP of insights and analytics. He'll cover why four engines can agree and what the winners do differently. And then Arjun Sengwan, our director of data strategy. He'll show you what AI talks about when nobody tells it what to say. Questions in the chat at any time. Let's get into it. Anthony, kick us off. Cool. Thanks, Javi. Yeah. I'm really happy to be here today, and let me let me get right into it. My name, like Javi said, is Anthony Hernaldi. I lead the insights team here at Yext. So in the last couple Yext research webinars, I I was kinda focused a lot on our scout data, and I was focused a lot on traditional search because I've been at Yext for, over seven years now, and it's just it was my sweet spot. But we have so much more data, and today, I wanted to I wanted to do some research into something that I found fascinating, which is what I've been calling a a disagreement amongst friends. So in this research, I wanna talk about how leading AI engines disagree when you ask it very subjective questions. But before we get into my research, I wanna get a little academic, and I wanna talk about LLMs in general. So a couple years ago, if you were hot on AI in general, you were learning about LLMs, and I think everybody had a magic moment when they realized that LLMs or large language models were they were not giving you answers. They were predicting answers, and they were predicting answers in the same way that your your cell phone predicts the next word when you're texting. And I like to think of the example of something like killing two birds with one stone. And as you start to fill this in, AI or these LLMs' confidence in the next word gets gets closer to a 100%, but it never gets to a 100%. So when you start to type something in, it'll say killing, and then it might think you're saying killing time or it might think you're gonna type killing me softly. Or if you're into TV, it might think you're gonna type in killing Eve, but every word that you fill in gives it a little bit more context to what you're going to actually ask. So by the time it gets to killing two birds with 98% sure that the next word is gonna be stone, but 1% of the time, it might say killing two birds with one shot. 1% of the time, it might say killing two birds with one hand. But this is important to realize that these models are not they're not all knowing. They're not these, these omniscient beings. They're just prediction models. So what I wanted to do is before we get into the research, I wanna talk about this in theory and dive down between different question types. So this is a real question that I asked ChatGPT four times, the same exact question with the same exact context from the same exact user, and I'm asking it in four different windows. So the question is a very, answerable question. What is the capital of New York? Right? This is a fact based question, and real answers, all four runs give me the word Albany. Right? The capital of New York is a fact. That fact is Albany. Congratulations, ChatGPT four for four said Albany. But ChatGPT got to the answer Albany in very different ways. Right? The first one's talking about the history of Albany. The second one's talking about how, people might think it's New York City. The third one talks about its its geographical position, and the fourth one starts to just give you a straight answer. So in these fact based questions, you can tell that it it used its grounding. Maybe it used some citations to tell you what the answer is in a confident way, but it freestyles around that. And, actually, I think Arjun in the second half of this webinar is gonna get a bit more into what AI does when they start to freestyle. But I think this is interesting because it's a fact answerable question. Second question is a little bit more subjective. So I said, how is the weather in New York City this week? Again, I asked this in the same model, ChatGPT, and ask it four times, but I'm the same person. I have the same context. I just asked it four times. And now this is not a fact based question. This is fact adjacent. Weather is a fact, but how is the weather is more of a subjective opinion. So four times. The first one says it's comfortable. The second one says it's warm. The third one says it's nice. The fourth one says it's typical. Right? And I think they're all basically saying the same thing. They're all basically describing what it means to be, a day with this temperature and this humidity and and and these attributes. But what I found interesting about this, again, this is not mocked up. These were real searches. What I found interesting about this is the citations, for these four questions come from two different places. The run one and run two use Bronx News twelve's website, and runs three and runs four go to the National Weather Service. So, generally, it's freestyling on what to talk about. It's answering its the question that you asked best it could, but it it totally went to different places to get that information, which is fascinating. These are citations. We talk about them quarterly too. And this this, third question is the basis of our entire research today. So I'm no longer asking a black and white question, like, what's the capital? I'm not even asking a a question that's freestyleable generally, like, how is the weather? I'm asking a a straight up subjective question. Who should I call to fix my HVAC? It's a question that I'm, unfortunately too familiar with as my HVAC in my condo has built, or has broken every summer since I've been here. But this is a question that is not black and white answerable, and I ran the same question four times on the same model. And each time, these are real results, each time, I got a totally different number one recommendation. This is what is super fascinating to me. I think my first slide about killing two birds, one one stone plants that seed of, like, you know, this isn't an answer. This is a prediction. But when you're asking it questions like this, this is, you know, asking Google the same thing three times and getting a different answer. And this is what I really wanted to focus on. So where do we get this data? It's the same type of data we've talked about in in our Yext research webinars in the past. I'm using data from YextScout, but I'm specifically using what we've historically called our subjective unbranded query type. And a subjective unbranded query is one where I don't care about the brand, so I'm not mentioning anyone. And I am not asking an objective question. I'm asking a subjective question. In this case, the question is very explicitly, what are the top 10 blank near me? Right? And the near me is going to change for every location that we power in the platform. So what this ends up looking like for this research study is we've got 3,569 different things we're looking for, and the things we're looking for are our businesses that we power as primary categories. Because a primary category was a very easy way for us to do this at scale. You are either a fast food restaurant, a primary care provider, you could be a financial adviser. These are the nouns that we're looking for, and we ask the same question with the same context, not on chat g p t four different times, but on four different models. So what are the top 10 med spas near me? We ask that to Gemini, to chat g p t, to perplexity, to Claude, and we ask this everywhere we expect to find businesses. Right? So if we know there's a med spa in this ZIP code, we're gonna ask it in that ZIP code. We're gonna ask it in that ZIP code. We're gonna ask it in that ZIP code. At Yext, we have millions and millions of these types of searches. But for this research, I held it to about a half a million because I wanted to limit it to the businesses where we ran it exactly on these four models with the exact same context before and after the search. So we have about 500,000 of the same questions asked the same way with different results coming back. And if you're on this webinar today, you may have seen, floating around LinkedIn or or some of our marketing, this 4% number. And I think that good research is great research is this combination of, what you expect and surprising. Because if it's not what you suppress expect and surprising, then it's probably wrong. And if it is what you expect and not surprising, it's not an interesting research topic. But the 4% number is what I found to be the amount of times where all four of those questions get answered by those four models with the same exact number one recommendation, which to me is nuts. Right? So you're asking four models the same question. They agree unanimously four percent of the time. The other surprising number here is 48%. So almost half the time, it's not that they're not in sync perfectly. It's that they're actually giving four totally different answers, just like my example, a couple slides ago. So they're not just not in sync. They are four for four totally different. And in the small text at the bottom here, I think this is fascinating too. So not only are they not, not only are they not in sync with one another, they're not really listening to me because when I asked 500,000 times a question that includes the phrase, what are the top 10? On average, I'm asking for 10, and on average, they're giving me back seven results, which is not only surprising, but a bit insulting. So they said, no. We got it. We know what you said, but here's seven answers. So what does this look like in practice? So, again, the next four slides, these are not mock ups. These are these are real, searches that we ran for our Scout product. So this one is, what are the top 10 auto sunroof shops in this town in California? And here's an example where we're in four different models, Claude, Gemini, Perplexity, and ChatGPT. In these four models, we get four different answers in the number one spot. And fix auto Winnetka, it is mentioned on Gemini and is mentioned on ChatGPT, but it's not it's not in the one spot so that at least they know that this is exists. But there's two other things that I find interesting on here. One, if I animate this, is that there are certain businesses on here, like Cook's Corner Auto Body and Paint that are mentioned in these models. But on these models, they are spelled differently or they are capitalized differently or they have different words in this business's name, which is not a new problem for AI. It's the same reason that well, one of the same reasons we started Yext in the first place was to make sure that these different publishers or these different listings companies have, consistency with your name, address, phone number, things like that. But it's nuts that in perplexity, they call it Cook's Corner Auto Body and Paint. But on ChatGPT, they call it Cook's Corner Auto Body and Frame or Cook's Corner Auto Body. And this to me is is just a sign of, one of the many reasons that they might not even have the same word bank across models. The other thing I think is fascinating is that I mentioned they're not giving me 10. None of these gave me 10. Claude and Gemini gave me five. Perplexity and ChatGPT gave me eight. So in total, they gave me 26 results, and there are 19 different businesses named in 26 different results. And that's something that I want you to think about as we get further into this presentation because it's not just about not agreeing who's the best. It's these models not even agreeing with who they should consider. So I'll fly through the next three, but as we go through these, they're going to start to converge a bit. So here's an example. What are the top 10 waterproofing services? Waterproofing is a primary category that we work with near Sedalia, Missouri. Easy for me to say. And in this example, affordable foundation repair shows up number one spot on Claude, shows up at the number one spot on Chat UPT, shows up in the number two spot on Gemini, shows up in the number four spot in perplexity. This, you at least have what I would call is is pretty strong agreement because Claude, Chachi, Pity are in agreement with number one, and Gemini is pretty close at number two. But, again, if I animate this, you'll see there are plenty of examples where there's mismatching names and mismatching, inconsistencies with things like KC waterproofing and ChatGPT versus KC waterproofing and foundation repair in perplexity, and in 27 results. So this time, we got Gemini to give a six instead of five. But in '27, you results that came back, we got 14 different businesses. So even though, they were a little bit more consistent in who they recommended, number one, There's still a lot of, disagreement in who to even consider. Here we get Valvoline in three out of four, in top 10 oil change searches. Valvoline's also number three in perplexity. I animate this, and we get Goodyear Auto Service and Goodyear Auto Service with the city, Jiffy Lube versus Jiffy Lube Findlay. These are some more consistencies and potentially why these businesses didn't get, didn't get ranked or or have a little bit higher sentiment across these models. We get 28 results in 19 different businesses. And finally, in a search for what are the top 10 tire shops, we get discount tire, discount tire, discount tire, discount tire. So here is a search where they all agree on number one where we still get this agreement on how to, how to spell or how to capitalize certain businesses. And the most shocking result, I think, is despite having perfect agreement across this model with who to who to recommend in the first spot, we get 22 different businesses in these 29 results. So those four slides, I hope I didn't go too slowly through them. These are not these are not mocked up. These are real examples from Scout, but with progressing, consistencies and recommendations. So if I bump that stat, the stat that I opened with was 4% of the time they all agree with number one. If I bump that stat to three of the four engines or four of the four engines, agreeing, then it's not 4% of the time. It's 18% of the time. It feels a little bit better that they're a little bit more in sync. But we looked at this not just overall. I looked at this by density. And if you were here for my ELO presentation in January or February, we talk about competitiveness and density together. So not just as a market rural or highly dense, but also do they have a lot of competition for the keywords we're looking for. Example I used back in the other webinar was a surf New York City in general is a very, densely competitive area, but a surf shop in New York is a bit more rare. So we take all of that into, we take all of that into account when we look at things like the numbers progressing exactly as you'd expect. These models are way more, in way more agreement in the rural areas than they are in the highly density areas. We can also look at this by industry. If we look by industry, you see some things like retail in all markets and in the rural markets are much more often in agreement than food. And there's some surprising things here like services, but we'll get into that. But it's interesting here to see that the range is from rural retail of 30% of the time these models agree in three of the four recommendations to 8% in a highly dense food market. And there's a lot of reasons we're gonna get into why this might be, but I think the biggest reason is, yes, I asked for 10 results in AI. But the way Yext as a company runs these searches, we ran it in four AI models, but we also run that same exact search in the same exact place on traditional search. So I'm gonna also ask for what are the top 10, higher changing businesses around me on Google. And I can tell you that there are not seven. There are not five. There are not 10 potential results. There's often close to 20 or 30 or 40 of these in any given market in traditional search. So if I look here, this is the number of businesses matching, the same keyword that we searched in AI, but matching that on a Google Maps pack. So this explains a lot of why they're in disagreement. So if you look at retail and rural, there's only 17 on average. There's only 17 options to even have in your word bank, whereas for food and very high dense areas, we almost have a full two pages worth of results. So it's not surprising that if the word bank is a bit bigger, then these different models are going to find a different different combination of five or six or seven or eight different businesses to recommend. Then recommend, and if they have a higher word bank, they're going to disagree a bit more often. So how did they define best and who do they consider? This is fascinating to me because, historically, when we talk about Google search, we always talk about, we always talk about best blank near me. And best blank near me historically in Google has had, a very defined bar. Right? We say in the review space that if you're not at least a four star in 99% of searches, if you're not a four star or above, then if somebody searches for the best blank near me, they won't even consider you. And what we saw is we were able to, just like I mentioned before, we were able to try to take these AI results and take our Google results and try to find the businesses named in AI and find those same businesses in traditional search so we could do our own research. We can say, if this HVAC company was named in AI and in a traditional search, we found that same name, then let's use the traditional search to look at their Google listing, to look at their website, to use the scout data. And, yes, we're gonna talk about citations, but let's do our own research. Right? Can we see physically what that Google profile looked like in time in terms of reviews, in terms of their photos, in terms of their completeness, in terms of who they are? Because if we can see in the in the actual data how that business is run, that might help us get an idea for what AI was looking at. And this, two numbers, incredibly intuitive, one number incredibly surprising. So if we look at that same business that AI is recommending, number one, and we look at its traditional stats, I can tell you that the number one recommended business, its median Google review and a best x near me AI search was 4.8 stars, which is incredible. I could also tell you that about half of the number one picks were 4.8 or above, which is incredible. Those two things are exactly what you would expect if your criteria was when somebody asked for the best to just give them the best rated. But what I thought was fascinating is that if we look at the number one recommended business in AI, 11% of the time, it's not even four stars. So the first two numbers, intuitive. Right? Very high rated business are at least half the time in that number one spot, and the number one spot is full of people that are almost a perfect five stars. But 11% of the time, the number one spot isn't even four stars. So what I what I took out of this is that that four star floor that exists in traditional search is not it does not exist in in AI. The other thing that I found fascinating was that 40% of the time, the number one ranked business in these AI best searches was the top rated business on the list. And I wanna say that, like, six more times because it's it's a bit of a brain buster to say that each model, about half the time, is putting its top rated business by reviews in the number one spot. And if they're putting their top if each model follows that same logic, they're all putting their top rated business in the number one spot. How can they only agree 4% of the time? Mostly, it's a math issue. Right? If they're doing it 40% of the time and they're doing it 40% of the time and 4040%, add them all together, it is gonna be a small number that they all picked the same 40%. But the best way I like to think about it is, the word bank example that I talked about. So let's use an example where, it's not who's the best, but who's the tallest. If if I ask everybody on this webinar to find the tallest person in your office, we're all working off the same rules. I'm picking the tallest person in my office, and you're picking the tallest person in your office, and they're picking the tallest person in their office. We're all following the same exact rules, but we're all in different offices. So it's not about what gets you from number five on the AI's list to get you number one on the AI's list. 40% of the time, it's just ratings. But the different question is what even gets you into this office, or what gets you into the room, or what gets you into that five? Because like we said, it's it's it's not looking at every business. It's already whittled it down to four or five or six or seven. So how do you get on the list in the first place? Well, I I mentioned this already, but we're only naming seven. But because of the traditional search, I can tell you that there's 37 on average businesses that could be considered. And now we have to talk about, what I like to think of as the these models diets or citations. So we talk about this quarterly, but when you get your list of five or six or seven businesses at the bottom, you'll get not every source that they've looked at because these AI models are grounded in some types of sources too, but every every source that these models, admitted to citing. So we get first party sites like your website. We get local websites, which are one level deeper than just a local landing page. We get directories. We get industry specific ones. We get reviews platforms. We get social. And you need to understand that each one of these models in these subjective unbranded questions have a different diet. So Gemini, for example, is almost, most of the time, it's looking at your website. Claude, out of everybody, is most often looking at these reviews platforms. Perplexity is is pretty even across the board, and ChatGPT is the highest share of these third party directories and aggregators. So this is important to realize because they're all, yes, once you get onto the list, they're super concerned with your ratings, but getting on the list is, is is different per model. Right? So, unfortunately, you kinda have to eat your broccoli here. There isn't a if you're joining the webinar to find out what the silver bullet is, on Gemini, I'd tell you to focus on this. And on perplexity, I'd tell you to focus on this. And on chat GPT, I'd tell you to focus on this. So the answer is just you have to be you have to kinda cover all your bases because you're not positive exactly which model somebody's gonna be using and what that model is gonna care about when when it's summoned. So getting on the list is about having really good readable content, schema, your pages, your profile, your descriptions, your attributes. It's it's about having the information that's even gonna put you up there and then winning, like we said, they're all we're using a lot of the same rules to get you to the one spot, which is rating, review volume, recent activity, and responses. We'll we'll drill into this a bit more, but it's different. Right? Because if I if all I did was tell you that you need to focus on the review platforms because that's what gets you to the number one spot, you might be the world's best business that isn't even in the word bank. Right? Because if you got there, if you had an at bat, you could wally pip your way through a a hall of fame career. If you got there, you'd crush it, but it's getting there in the first place. So it's it's just like traditional search in terms of needing more and more just machine readable content. So I looked at because we have that traditional search information, I looked at every attribute about every business whether or not they were ranked or whether or not they were even named. And I could tell you three time, if you have a location page, you're three times more likely to get on that list. If you have business description, you're two times more likely to even be on the list. If you've claimed your profile, if you have schema markup, if you have recent posts on Google or otherwise, if you have at least 10 attributes, these are all things that you could tell more often than not from that big word bank of traditional search. More often than not, these are the ones that get pulled out and named in the list of five, six, or seven businesses that AI is even considering. And, unfortunately, we find that it's different for everybody. Right? If I if I limit this to just retail and food, I could tell you that it was almost four times more often to get you on a list if you had a good local page with good schema. In financial services and health care, it was more about those third party directories and aggregators, things like claiming a profile and having more attributes that was more likely to get you on the list by three times. If it was local services, one of the biggest things that matter from even getting on the list on local services was the frequency and recency of posts, Google posts and otherwise. So not only are the the different areas you could be in using a different criteria and the different models that you could be named on using a different criteria, but even the different industries are using different criterias, which if you remember last webinar, I I talked about the the importance of knowing exactly who you are and making sure you're benchmarked to the right things. I did find it incredibly interesting to dig a little deeper into when they agree on number one. Is there anything in common that they cited? Right? Is there anything that when they have agreement, did they all look at the same thing? And what I found interesting was that nine over 90% of the time that there was an agreement amongst models that the number one source that was cited was that number one ranked business's personal website, or local landing page. And that was interesting because getting you on the list means that you had good really good machine readable content, and getting unanimously on the list was all about that first party content. The other thing that was interesting that, that was really consistent with unanimity, was not just being on every single third party, but being on these canonical industry specific ones. So for example, yes, we we we have plenty of research about being on as much of the publisher network as you possibly can. We call that a rest of network research, and there's a new version of it coming out because Google cares where you are. But as far as these AI models, having a consistent and agreed upon number one recommendation, what we found was that way more often in in a specific industry, if there was this consistent number one, it was it was often because they cited a very industry specific source. The examples here I use are CARFAX and GasBuddy. If there was consistency with number one, you were probably seeing one of these niche publishers in citations because maybe they all found these this niche publisher. They all saw you recommended on it, and then they said, wow. This is this is this is, illuminating. So you couple that with having a good website that might be the second citation. And now that, you have this machine readable content, you're not just on the list, but now everybody agrees that you belong there. And the last thing I wanna look at was we have a a wealth of this, this traditional data to go along with the AI one. So I I hate to keep saying that you just kinda have to eat your broccoli, but I can tell you some of these attributes for a typical business within this dataset. They've got 19 reviews. They're getting two reviews a month. They've got 80 photos. They have a decently populated profile on Google business profile, and their star rating is under four. When models agree three out of four or four out of four, those numbers are higher. And when they agree unanimously, those numbers are in a different stratosphere. So I I'd love to tell you that there's exactly one thing to focus on to get on the list and to get up the list, but what I can tell you is that these AI models are just prioritizing good businesses that are run well and run consistently. So, if you remember my my webinar from last time about just being the best version of you that you can based on everybody around you and having the right benchmarks, it's it's just as true in AI as it was last quarter in in traditional. It's, there's there's no way to cheat the system and get up and get on. You just you just have to have as much information as many places as possible because you don't you don't know for sure who's gonna be looking where. So I I thought this was fascinating because I I wanted to recontextualize a lot of this that it's not as much of a science as it is an ARC here. These these are not all knowing systems. These are these are prediction models. So I'm gonna stop sharing here. I feel like I took a little bit more of the time than I was allotted, but, Javi, let's have a conversation. Hey, Anthony. Can you hear me? Sorry. I'm having Mhmm. some audio issue. Okay. Cool. Awesome. Hey. Thanks so much for that. There are some questions in the chat, and I also had some here to ask you in the follow-up also on some questions that were presubmitted. So let's start with the first question. So for AI engines, the same question, half million times. They agree on the best business, on the best business 4% of the time. If I own that business, should I... Should that scare me? Should that excite me? Like, what should I do? I yes. Both. I think that it is it's just the number is just grounding. I I think it's just, like, it's a number that everybody should have in the back of their mind, and and I don't think anybody ever thought that SEO was solved. And I kinda want everybody to just pump their brakes that GEO isn't solved either. Yeah. That just I want everybody to think back that killing two birds with one stone and remember that this isn't at least right now, they're all they're all marching to the beat of their own drum, and and it's not yet solvable. So the best thing you can do is just be the best version for wherever they're looking, however they're looking. So I I think the note like I said, I it was a perfect research number selfishly because it was it was a sexy number that is is both surprising and unsurprising. So would I be excited or scared? Maybe medium rare, maybe a little bit more scared than excited, but it's nice to know that number so you're not judging yourself on an unrealistic. Yeah. For sure. Alright. Let's see. The next one was, how will the continued rise of AI with Google being Chad GPT affect organic search traffic for websites? I suspect it could reduce it at least initially. Cool. Yeah. I I I suspect you're right. There was a bunch of articles that came out the other week. I think I touched upon one of them on LinkedIn about Google going from a a it's not a walled garden, but it's I I like in New York, I like to think of this Gramercy Park where, like, you can see into the garden. You're just probably not gonna go there. I think it's important to to still focus. We talked about this before, but still focus on the metrics that matter because you're right. I don't think we're getting as many page views as we've ever gotten before because AI overviews are not just coming. They're here. And even the people who don't love an AI overview are are tempted to take its word a lot of the time. So we're not we're not building our I don't know when we're gonna get more and more bot traffic and whether or not we're gonna get more information on when these these AI overviews cite you, in a volume way. Right now we can do it on a one off way. But I think that you're absolutely right. Like, these websites not aren't necessarily for people anymore. They're for machines. And if they're for machines, then, you know, we're getting these the the rise of zero click like we've never seen it before. Right. No. Absolutely. And then, one more... Or another question here. I... And and for those on the webinar, I did answer in a thread. So if you scroll through the comments, you should be able to see it. But I figured I'd ask it again just, as well. But, when it comes to ratings, right, we talked about four stars, you know, and all of that earlier. Does business rating play does business rating play into the results? Does business rating plan to the results? Yeah. Absolutely. Like I said, I what here okay. This is fascinating. I didn't say this out loud in the webinar, but I think it's the undertone of couple of these slides. If you saw the diet slide that talked about the citations, one thing I think is. fascinating is that all of these businesses agree on the number one ranked business based on their rating. However, the highest that any one of these these models is citing a review site explicitly is, like, 20% of the time. I think it was ChatGPT or Gemini or one of those. So I think it's fascinating that reviews seem to be, like, the underbacking of what gets you to number one, but these models are explicitly citing review specific sites less often than I would have ever expected. So I'm trying to do further research on, does that mean that these these models are grounded in GBP maps more than we thought that they were and they just have that data within them and they don't need to explicitly cite it, does that mean that everything else about your business is just coincidentally leading them to to put the number one rated business number one, even if they're not looking at business ratings? I think it's fascinating because I would have expected that if the number one business is ranked highest 40% of the time, then they'd all be looking at these review sites, but they're not. So I think having your rating readable is incredibly important too. Because if they're looking at your first party site, there is a schema markup for business ratings. There's a way to get these ratings first party onto your site. There's a way to get these ratings in more machine readable places than than just these review sites. So I think there's too much research to ignore that rating seems to matter, But I'd actually say, you know, let's actually get let's get that rating in more places so that it can be read even if they're not looking for it. No. Absolutely. Alright. Next question. What are the top three to four directories to focus on in helping your GEO for AI platform presence? Historically, I would have said just be everywhere, and I'd still if you have the time, Right. if you have the energy, still be everywhere. There's there's a ton of research that we and other companies have done around that network lifting, all chips. But coming out of this, I think I was I was surprised to see how big of a role these canonical industry specific ones matter. So, yes, you definitely be in as many places that you can name. Right? Be in the big ones. Get to as many of the rest of them as you can. But if you're gonna cherry pick some of those, you should probably cherry pick some of your industry specific ones. Absolutely. And instruction data matters. Alright. Next one. Would like some tips on measurement reporting. Trying to figure out new ways of reporting, that now leads to... Or or... Sorry. I'm trying to read this question. Trying to figure out new ways of reporting now that leads and page views have been deteriorating. So tips to measure... Tips on measuring and reporting. Lead it's interesting because leads and page views are come from a different two different worlds. Leads are top of funnel. Sorry. Leads are bottom of the funnel. The page views are top of the funnel. So maybe this question is halfway there already because I I think for years, even before AI models became as prevalent as they were, impressions and real top of the funnel stuff has been falling. They've been falling because of algorithm shifts. They've been falling because of just the way that Google has changed counting of map pins on on maps.google. Now they're following because of AI. So my biggest recommendation would be to try to try to look as far down the funnel as you can, and that's way easier said than done. But on GBP stuff, that means not tracking impressions, but instead looking at the the CTAs, phone calls, directions, website clicks. We've always struggled as a company because a lot of our businesses are brick and mortar, and the closest thing to a transaction is a GBP directions click, and we're not gonna follow that person into a store. So there's always been that attribution gap a little bit. And that same thing is on a website, but a little bit easier because once you get to the site, which might not be happening by a human as often as it was, but when you get to the site, you can at least start to track that most important thing. So I this is scary. But I guess my advice would be track the thing as far down as you can, and how we connect those dots is gonna get harder and harder because we don't have perfect referral sources. for each one of these models. But you kinda have to understand that top of the funnel is just you more not useless, but less useful than it's ever been because they might still be doing things, just doing them in a different way. And we we need to catch up to that, but I would recommend you mentioned page views and refer page views and leads. Focus on the leads. Don't stop focusing on the page views. Awesome. Alright. And let's get one more in here so we can get Arjun on. But, this one was interesting also from the chat. How do we correlate the earlier review rating stats with the citation stats, lower percentage attributed to reviews? Oh, cool. That's yeah. I think I think I accidentally just mentioned the answer that's a minute ago. I I'm asking the same thing. Yeah. Always review so important when the review sites are cited so infrequently. 20% was the highest. I think 23% is what I saw. It's one of two answers. One is either they're getting these review ratings from nontraditional review places. Right? Because, for example, we would include GBP in directories and not a review site. So, again, if they're getting review ratings from GBP, that's not necessarily the bucket that you would expect on that chart. It's not a review site. It just has reviews. And the same thing goes for basically every one of those buckets. Can you inject where are they getting reviews from? They could be getting it from the scheme on your page. They could be getting it from first party reviews. They could be getting it from a directory that happens to have reviews. I just would would look at those what I keep calling is the diet charts. What what are they eating? I I would just keep looking at those with a understanding that they're not all exclusive buckets. Right? Reviews can be gotten from somewhere other than the reviews sites. Right. The other thing is I'm trying to do more and more research onto grounding, right, because I do think a lot of these models the same way I don't think they went out to the website to find the capital of Albany. I think they just they know it because of their historical grounding. I suspect that there's a lot to learn from citations that aren't happening. Right? What types of questions do they feel comfortable answering without a citation, and what where's that information coming from? So I suspect a lot of historical ratings already exist, and they're just pumping them in from from the new readable places constantly. Awesome. Thank you so much, Anthony. Alright. Let's go ahead and pass over to Arjun and go to his part of the webinar. Hello. Alright. I'm gonna share my screen. Thanks for the intro, Havi. Sweet. So I'm gonna be going a little bit deeper than what Anthony talked about when we're talking about, what information Scout gets for us. So if you think about Anthony's framing of his research topic, it was looking at which brands are mentioned by AI. I'm actually gonna be talking today about which facts or which business attributes about a brand are mentioned and whether or not those are relevant to what people are actually searching for. So I'm gonna keep going back to these three pillars that I'm gonna go over these slides, but I wanna show you what facts businesses are filling in, and that's on their Google profiles or whether through Yext or not. And then what people are searching for and what AI actually volunteers in terms of the information it displays. And I'm gonna show you how the three of them don't always necessarily agree. So for this research, we went through about 1,800,000, 1,760,000 answers, exactly. We went through about 400,000 locations, a 178 facts, and about six models that we scraped information from. So I like to frame it with a study that Samuelson, an economist, did in 1938. If you saw any of Christian's recent blog posts, he talked about the the dichotomy between what a person states and what they reveal. So if you think about the, the phrase put your money where your mouth is, what someone says might not always be equivalent to necessarily to what they actually buy or purchase. And so what I wanted to do was think of a similar framing when it comes to businesses and AI. And so you're gonna hear me say these two words a lot today, and I'm gonna, you know, have an academic slide a little bit later to talk about exactly what they mean. But I'm gonna talk about stated and mentioned. So stated is what the business says about itself. So that's through your GBP profile or, again, through the x knowledge graph or through anything else. And then what the model repeats. So what does a what factor attributes out of the subset that businesses could possibly feed into your GBP is the AI or the different AI is actually mentioning when you ask a certain type of question. And again, we ask this question 1,700,000 times over and over again for different industries. So if you think about the three pillars, you have the person, what they're asking for or what they're searching for. And I'm gonna talk about this demand side a little bit at the end. But I'm gonna be focusing more on the supply. I'm gonna be talking more about what are businesses stating and what what are the models mentioning. And for the study, we actually wanted to focus on just one type of question. So what I'm focusing on is objective branded questions. It's essentially a pretty generic or vague prompt where I'm just I I I gave an example over here saying, tell me all that you can about Home two Suites by Hilton Galveston. I'm not saying, do they have this? Do they have that? No. I'm just asking about a specific place, and that's why it's a branded question. But I'm asking a very objective question about everything that it could possibly have. And so I'm giving AI the the the the AI models the ability to freestyle as Anthony said earlier on. And in this exact example, you can see that when I when I asked the question, it talked about the address, the phone number, beach access, attractions, accommodations, breakfast, and whether or not it has a pool and fitness center. So these are the kinds of attributes that businesses can feed in, and these are the kind and on the right hand side, you can see what it's actually showing, on the Google AI overview. So some quick examples. One thing that we saw that was stated a very high proportion amount of the time, but almost never mentioned was wheelchair accessibility. And on the flip side, something that again, this is only relevant, to hospitality, and I'll dive into the different industries later, was that spas were only mentioned seven stated 7% of the time, but they were mentioned 64% of the time. Now when when I say mentioned, it doesn't necessarily imply that a business has a spa. It's just that the AI overview could say, does it have a spa or not? So whether you have it or not, this is the kind of information that the AI models are volunteering. So I wanna go back to the wheelchair example because the number there was 0.2% of the time is it actually mentioned. And I if you if you think about the percentage of the population that probably needs something that is wheelchair accessible, that makes sense. But I don't want to, like, deprioritize how important that is because the two words that come to mind here are context and memory. I'm gonna give you an example here where I was talking to chat GPT. None of this information is true, but I had never spoken to chat GPT about my grandfather before. And they said in the first prompt, my grandfather has a walking stick. Do you have any recommendations for a new one? And it gives me a bunch of different options. And then I open a new chat and ask it something completely unrelated. I say, now my grandfather and I are going to get lunch near near me in New York. That's where I live. Do you have any suggestions? And within the second line, it's mentioning somewhere that's comfortable and easy to get into, I e accessibility. So again, just a quick overview on the terminology I'm gonna be using today. A factor and attribute is something that you feed to the GBP. So we have a 178 that we're tracking, 47 of which are native to all five industries, and then you have a 131 which are category specific. So they could be relevant to anywhere between one or one to four of the industries that I'm gonna be going over today. And then stated is how often the business filled that field in and mentioned is how often the AI answer actually mentions that thing. And given free range, when you allow it to freestyle, these are the things that you talk that they talk about. So, again, if relevant to hospitality, you have things like fitness centers and pools. Relevant to food and beverage, you have, like, breakfast and bar, and then you have things that are more generic like free Wi Fi. The thing that we the the overarching theme that I'm gonna be talking about today is that, actually, what you see is that the more you state something or the more businesses state something does not necessarily imply that AI is gonna volunteer that information more often. So if you look at the examples on the left, you have wheelchair accessibility, debit cards, credit cards, and restrooms. Things that you expect a lot of places to have or things that by default you expect a lot of locations to fill in. And those are actually mentioned or volunteered by AI a very small percentage of the time. You can see the how the arrows go down into the right. And on the flip side, you have things like fitness centers, breakfast, spa, and indoor pool, which are mentioned varying degrees amount of time anywhere between zero to 75% of the time. And yet they're mentioned a lot more than other things. And another way to visualize this is on a on a scatter plot. So again, you would expect here you have on the x axis, what percentage of locations offer the field, and on the y axis, what percentage of AI answers that mention it. You would expect it to go up into the right if it was directly correlated. But what we're seeing is that that's not actually true. Facts that are stated more are not necessarily mentioned more. And when we looked at this across the six different models in our study, we saw that there was not much variation here either. Between five to 7% of all answers that mentioned, that, mentioned had a field filled specific to that. And this was pretty much consistent across all the models. So before I get to the industry breakdowns and talk about the specific fields and trends that we saw, I wanna say two things. First of all, this is a prioritization exercise. This is probably the best way to look at it. This is not me telling you that fields at the bottom don't matter at all and fields at the top matter a whole lot more. Because, the fields at the bottom, for example, like wheelchair accessibility, are the while they might not be searched for us off as often, on the on the rare occasion where someone is searching for something and they they need wheelchair accessibility, you want your brand to be the one showing up with relevant information and facts and attributes about it to encourage consumers to go to your brand or your location. And on the flip side, we have past studies, and you can go to the new Yext Research web, landing page that prove that an or, the overarching theme that higher completion rates, these are fill rates. It's essentially how we calculate what percentage of fields, on the Google business profile that you've completed is highly correlated to the number of impressions and clicks that your brand, and your locations actually get. And so even even if things at the bottom might not seem as important, I think the best way to look at this is sort of if you have a few hours a week or if your team has a few hours a week to update certain fields or to maintain, you know, some fields over over the other, this is probably best used as a prioritization exercise rather than just a, you know, I'm gonna do this and not do that. So let's start with food and drink. The most interesting one here I saw was that breakfast is only mentioned, is only stated by brands about 45% of the time, and yet it's mentioned about 68% of the time. And same thing with delivery. It's mentioned about 25% and mention and stated about 25% of the time and mentioned about 55% of the times. So these are some opportunities for businesses to be filling out fields that AI often volunteers and helps her brand stand out. And then with health care, I was quite surprised with this. You see that with accept walk ins, it's almost never stated by a lot of businesses in our study, and yet it's mentioned 36% of the time. And, again, when I say it's mentioned 36% of the time, that's not me saying that 36% of businesses have or accept walk ins. It's saying that the AI model says whether or not it accepts walk ins or not, which is implying what the model sees as an important attribute for your business in health care. And then if you look at retail, delivery is mentioned in about 34% of answers, but it's only stated just over 20% of the time. So, again, something that I would have thought more businesses would be filling out, in their GBP profiles more consistently actually seems not to be the case. If you look at hospitality, we can see an interesting example in the inverse. So here's where you start to see the more generic fields. If you look at free Wi Fi, it's mentioned almost a 100% of the time. Let's call it 95% of the time. And yet, it's only mentioned 55% of the time in AI answers. And I imagine the reasoning for this is that there's a lot of information, like I gave restroom before or free Wi Fi here that are so unanimous and, like, you expect almost all businesses to offer it. So it's not necessarily gonna be a differentiator when it comes to you filling out those fields as important as it is overall. And then the same thing with financial services. You look at whether or not they accept credit or debit cards. Obviously, nowadays, you would accept most, most places to accept both. And while over 50% of all the brands that we were tracking do accept them or and they filled it out their their profiles out accordingly, it's only mentioned in the AI answers or AI overviews about four or five percent of the time. So, again, if you look if you go back to the three pillars I was talking about, there's how often people search for something. And I just wanted to pull up a few examples here. This is using Semrush data who we we have we have access to their data. And then we have the business and how often businesses actually fill in a certain attribute or that piece of information. And then there's the AI model in between and how often they actually volunteer that information. And what you can see here is there's not really that high of a correlation between any of them. If anything, some of the fields that are filled in the most often are not necessarily gonna be mentioned by the AI model all the time. And so again, to go back to how to look at this, you it it it helps to know which fields in your industry differentiate your brand and your the the locations that you have. But at the end of the day, the ones at the bottom still matter for two reasons. One, context and memory, because you do still wanna serve that person in a wheelchair when it comes to him who is searching for your brand. And second of all, because we know that higher completion rates are directly correlated to more traffic. And with that, I will take some questions, Havi. Thanks, Arjun. Appreciate that as always. Let's see here. So question number one. Your research says AI loves to talk about the pool and the X-ray machine and almost never mentions the parking of which credit cards you take, for example. What does AI care about? What what... I I guess, why does AI care about the things it cares about? I mean, because it's trying to tell you something maybe you don't already know or that you think you would like to know with the context and memory piece. So, like, the examples I gave, like, every business offers, you know, probably credit card or debit card. They accept debit and card credit card fields. So saying that helps very few people choose. Whereas when it comes to hospitality or health care, like, things like the pool and the X-ray are probably the differentiators that help, that only specific places have, and those help you get mentioned about seven times more. Perfect. Thank you. Alright. Next question is, you showed three different opinions about what matters, Consumer search for pet friendly hotels, for example, 12,000 times in a month. Businesses fill in wheelchair accessible parking, which gets searched 40 times. And AI talks about something else entirely. So out of those three, who's right? Yeah. I don't I don't know if it's as straightforward as who's right, but. the fun the funny part is that AI isn't the odd one out from what I would say. Like, I would say the one that's actually the odd one out is probably businesses. If you look at the correlation between the three, like, there's not a strong correlation between any, but the weakest correlation like, the weakest correlation between any of the other two is probably with businesses. So models and searchers mostly mostly again, I say mostly, want the same stuff, pet friendly, beach access, and yet hardly any businesses fill those in. Meanwhile, all the businesses are filling in things like wheelchair parking and Wi Fi, which not many people are necessarily searching for because they assume that the look that what they're searching for already has that stuff. Perfect. Alright. So we have three minutes left, I believe. Let's do one more here. So shoot me straight here, Arjun. So I run marketing, for a 500 location brand. I've got, you know, one afternoon and a team of two. Where do I get started? What do I do first? I'd say go straight to the custom fields that only your industry or your category gets. So for example, hotel amenities or clinical services and not things like payment types because they don't get mentioned often anyways. But it doesn't mean that those aren't important at all. As I mentioned, the research has proven overall overall fill rates are highly correlated to more clicks, so it's still useful to fill in things like accessibility, which are unanimous and, you know, filled out pretty much universally. And not just for traffic, but because the one time you actually like, the one time the person who actually needs it asks, you really want that information to be there. But, yeah, definitely use, you know, you use the fields that help your brand stand out and that have the highest mention rates that are not, stated as often by businesses to help you differentiate yourself right at the start if you wanna, you know, prioritize. Yeah. Awesome. At at. at internally and with clients, we have we have the we have the metric that we call completion rate, which is how often are you filling out fields. Internally, we have started to to group these fields as core, secondary, or tertiary. Core fields are table stakes. Like, this this whole webinar is not about core fields. You're gonna fill out your name, your address, your phone number, things like that. So if you're a team of two, if you have more time, yeah, like Arjun saying, don't just start arbitrarily going alphabetically through the fields we have. Start going to just like those those industry canonical directories, think about think about your industry first. But, yeah, fill out everything if you can. Then if you can't, fill out the core stuff. If you have a little bit more time, let's let's go industry specific. And I guess one more thing I would say is don't be too cute because hospitality is a really fun example. We've done some research where we saw that businesses in hospitality that filled out hours were actually showing up a little less often than the ones that didn't. I'm not that's not something I'm gonna recommend to just stop filling out or delete as a hospitality brand, but the research shows that I wonder if that's sending a signal that shows certain searchers that you're not a twenty four hour check-in. There are certain things that we are seeing in the data, but I don't wanna get too cute yet. Again, fill up fill up generally as much as you can, and we'll we'll get further from there. Thanks, Anthony. Arjun, anything else? No. I think. we're good. Perfect. Alright. Well, gentlemen, thank you again for an insightful q three, webinar. And for all the attendees as well, thank you so much for joining us as always. We will be back here Wednesday, December 9 at 1PM. So hopefully, we'll see you back here then. Hope you all have a great rest of the day and a better week. Thank you.