#296 How YPulse Built an AI Application for Market Research with Dan Coates, President at YPulse

Hi, I'm Maggie, VP of Curriculum at datacamp, and I know firsthand how overwhelming it can be to keep up with the rapid changes in data and AI. But staying ahead doesn't have to be hard. That's why our curriculum team has curated the best learning content to help you and your team stay ahead, covering everything from data and AI literacy for busy executives to cutting edge generative AI for developers, and everything in between. Whether you're upskilling yourself or your team, we make learning practical, engaging and flexible. It just works. Power your data and AI transformation journey on datacamp.com.

I've been running this company now for 17 years and I used to weep over the data that we'd collected in any given year that wasn't consumed by enough customers to sort of make it worthwhile. It's collecting data longitudinally is a real chore and I used to think about it like perishable produce or fruit or vegetables, right? That basically once I'd ask those questions that year, if there wasn't a customer at the receiving end of that, that it was all for naught, it was all gone. And now I don't look at it that way. Now I look at it like we've created a catalog or a music library. Everybody that I talked to when we were going down this path said, are you out of your mind? You're going to do a customer facing AI implementation first? I guess it makes some sense if you're going to stink off Broadway, right? What I'm really excited about is that unlocking that ability to kind of take a client into future directions and give them new business opportunities gives us the ability to kind of grow our relevance within organizations, change the sort of seat that we have at the table.

Welcome to Data Framed.

This is Ritchie.

There's a lot of pressure in many organizations to do more with AI and to squeeze AI into all of your.

Products wherever it might fit.

There are many wonderful opportunities, but sometimes it can be overwhelming. So today we're going to hear an AI success story. Our guest runs a market research company and when ChatGPT was first released, he realized that the way his customers were going to consume market research information was about to completely change. That meant his company's platform was now obsolete and he had to embark on a high stakes rewrite of the company's flagship product. There are many great lessons from understanding our guests journey, both from an AI transformation perspective and from the point of view of how we consume data. I'm sure we'll pick up some insights into market Research as well. Our guest is Dan Coates, the president and co founder at Ypulse. Ypulse helps brands like Apple, Netflix and Xbox understand and communicate with consumers aged 13 to 39, using data and insights from hundreds of thousands of interviews. Dan has been leading market research teams since the early 90s. This included a stint as VP of Sales and Marketing at SPSS, being an early innovator in online sampling at Polymetrics, and founding several startups. I don't want to say he knows everything about data driven market research, but he's certainly one of the world's leading thinkers in the space. Let's find out how he reinvented his company to use generative AI.

Hi Dan, welcome to the show.

Thanks for having me.

Brilliant. So, just to kick off, tell us, why did you decide you needed to build a chatbot for market research?

We do this annual customer engagement survey and in late 2023, we did our annual customer engagement survey. We found that a new problem was emerging for our customers. Corporate America has been downsized and people are expected to do a lot more with a lot less. What we found our customers were telling us were, you're producing too much information, we can't keep up with it. We ourselves did some time trials and found that it took 10 to 12 hours a week to keep up with all of our stuff. And so really we hit this point where we couldn't ask for any more time amongst our customers. We had to start saving them time. Meanwhile, we wanted to continue to be prolific and to cover as many possible topics that we could within our content. So that's when the idea of we could solve this problem with an AI chatbot sort of arose. The previous use case was in order to get the value from YPulse, you'd have to read everything that we produced. And the new model is find what's most relevant to you and just read that.

That's absolutely fascinating that the problem is just that people don't have time to read all the content that you're putting out and that's why we need AI. Obviously this is something I struggle with as well. I have a lot of things to read as part of my job, but then also sometimes like, do I need to read the whole thing or can I sort of skim it and maybe AI is going to be useful. So do you have a sense of when your customers want to use AI to consume content and when they just want to read the original document?

There's kind of two use cases now. The old one was we were kind of using breadcrumb like techniques to sort of lead people to the content that was relevant to them. So every day we produce a newsletter and contained in that is a brief description with daily article. So the purpose of the newsletter is to get you to read the article. The purpose of the article was to get you to be more interested in reading the report. The purpose of the report was to get you to want to dig into the data itself. And so we'd always sort of had that breadcrumb trail, but we found that some of our customers were telling us that they were worried that they were missing things. They knew that there was value in what we were creating. They just didn't want to miss anything and so let the content sort of wash over. You Take a look at the headlines for the daily newsletter and if nothing there is interesting, you ignore it. If the daily article is interesting to you, great. Read it. If not, ignore it. So we kind of have this. On the one hand, ignore the stuff that doesn't seem relevant to you based on the breadcrumbs that we're building. On the other hand, come use an AI power tool to find out exactly the answer to the question that you have on your mind at that point in time.

Okay, so that's interesting. I like the idea that you've gradually eased people into reading, like getting in depth into the content. So you can start off with just an overview and decide whether you like it or not.

Okay.

So I'm curious as to what constitutes success for this. Like, how do you know whether the chatbot is working or not?

One of the first things that we noticed when we baited over the course of 2024, we had the chatbot in a free open beta for about three months across two different beta opportunities. And we noticed that during those beta opportunities, engagement with our content went up, not down. So we feel like AI is another tool. From the engagement stats, we feel like AI is another tool that leads people into our content as opposed to drives people away. Other metrics that we look at, sales metrics, have to admit we made a healthy six figure investment last year in AI. So one of the metrics was could we sell a healthy six figure amount of AI? So for our current installed base, we introduced AI as an upgrade. So it was really rewarding when we were able to sell more than we'd spent in the first year. So I'd say it's a pretty positive roi. All the usage reporting, report downloads, everything that we do in Pendo kind of indicated that this was getting people more interested in the traditional products that we sell. Internal feedback. We're the number one user of our own tool. Our writers use our tool to find previous content they can refer to in today's article, for instance. That helps us understand whether the tool is working. And they're the harshest critic, by the way. My own internal team beats us up all the time in terms of what we're doing and how we're doing it. And then the last point is just customer feedback. Multiple points along the way through the deployment, we would ask people questions about post beta. We would ask them what they liked, what they didn't like, whether they would pay for it, how much they would pay for it. So all of our decisions were sort of grounded in their feedback. For instance, they said that they would pay for the AI upgrade, but they thought it represented about a 30% increase in value. And so we didn't feel that shy about a 20% surcharge to be able to have AI tacked on. That went so well last year that at this point, new customers don't have the option of, you know, would you like fries with that? Starting on January 1, 2025, all new customers had AI just baked into their product. We increased the price significantly, but we made sure that there wasn't an opt out on the AI. And we're feeling so bullish now that our current customers won't have the option as of January 1, 2026. We may have to leave a few customers behind if they really don't want to have AI. But as of the first day of next year, it won't be an option.

So that's a pretty radical shift then, if this is coming to all your customers. But I thought some of those statistics were quite interesting. The fact that you said it cost you a sort of six figure dollar sum to create this thing, but you're getting that money back in terms of revenue after, what was it, less than two years since it's been released?

We've been live now for about 10 months, so we broke even in under a year. Oh, wow.

Less than a year. That's pretty impressive.

Yeah.

Okay, cool. And you said like, the main thing you're sort of looking at is like whether your customers like it. It's like, are they engaging with the AI? Like, how much time are they spending chatting with this thing?

Yeah, there's two parts to it. The behavioral side, everything that Pendo can observe, all the clicks, all the downloads. And then the other side is the attitudinal, which is what do you think? What do you like? What do you hate?

Okay, yeah, obviously like ASP wasn't a point, people spending a lot of time with the bot if they don't actually like what they're doing. So yeah, I like these sort of tracking some sort of customer satisfaction metrics there as well.

Interestingly, everybody that I talked to when we were going down this path said, are you out of your mind? You're going to do a customer facing AI implementation first? I guess it makes some sense if you're going to stink, stink off Broadway. Right. But we really felt like by engaging our customers and bringing them into the process and the sort of healthy beta periods that we put into it, if we had a dog, we'd know about it pretty soon. Our customers are definitely vocal and express their opinions freely. But the other thing is we wanted to show them that we were trying. And we do a lot of studies on innovation amongst young people and young people certainly feel like it's okay to fail. It's not okay to not try. So we just wanted to try. We wanted to bring them along in the process, explain what we were doing along the way. And I think we were rewarded for that. Very much a product centric approach, right? Not technology for technology's sake, but we're trying to solve a problem that you've expressed to us that you have. And so here's our attempt to solve that problem. What do you think? We solved it okay.

Yeah, certainly. I like that. Yeah, you might as well try even if you're going to fail. On the other hand, it does seem quite risky going straight from we've not worked with AI before to we're going to go straight into this is our flagship product going for all our customers. So I want to know how you pulled this off. Let's get into the weeds a bit on like how you built this thing. So I guess to begin with, like, what sort of tools, technologies did you.

Use as part of the AI for about six months prior to when we started, we formally kicked off pretty much a year ago, late February, with a vendor. Prior to that, we were inundated with all of these quote unquote experts that had said that they could help us. And we, we had this sort of standard response. We responded to all of them. We just said, can you give us an example of a production version of something that you've created? And that winnowed it down dramatically. But basically we managed to find one that was, I don't know if you know Professor Scott galloway, he's an NYU Stern professor, pretty prolific. He created something called Profg AI about 15 months ago. And he's prolific, we're prolific. Basically querying his body of content, that's the same thing that we wanted. So in About November of 2023, we tracked down the people that had created Prof. G AI and that led us to Haystack, which was what they used, sort of an open source platform. And then that led us to Deepset, which is a company that will help you deploy. So in late February of 2024, we had our first kickoff meeting with the Deepset team. And basically. So Deepset is kind of one of those intermediate layers. They create your vectorized database from your content. They allow you to have sort of a modularized approach so we could toggle between various LLMs at will. And just sort of that intermediary approach suited us well. They had a lot of expertise and quite a few webinars actually. So we watched a few webinars before we engaged with them even. They had a white paper on how to structure our team in order to successfully deploy. So we were so naive. We just thought, these people seem to know what they're doing because they've done it a bunch of times. They could show us multiple production examples of what they've created. So let's just sort of follow their lead. And I have to admit, that paid off for us. We probably could have figured it out ourselves, but instead on their side, there were a couple of engineers that were always involved, some account handling kind of people, but their core engineers got us up to speed really, really quickly. So within the Deepset platform, we use ChatGPT4 omni as our LLM. And now we're using Sonnet 3.5 on text to SQL. So the very first version of the product just looked into our Repository of reports, 4000 reports and articles that we produced. The second version now taps into 15 billion data points that we've collected. Survey answers, responses from young people across seven countries over the last decade. And that's all sitting in Snowflake. And so we're using text to SQL techniques to tap into those snowflake databases to get even more value.

Okay, that's very cool. And it sounds like to begin with, you didn't have that many skills in house. Like you were looking to partner with someone who'd sort of built something before. Can you talk me through, like, what did you have internally? Like, where were you starting from? Did anyone know anything about AI at all?

We have a product team, a couple of developers, product manager and then some data scientists that administer our Snowflake database. And basically we all sort of knew about our data really well. We knew about our customers really well, really didn't know much about AI at all. And of course the developer types thought, well, let me look into this. But we felt like we could spend a long time wandering around even the debate of which LLM to use, right? I mean, people tend to get sort of stuck in these moments where in the overall scheme of things it really doesn't matter which LLM. And we were really happy to be able to toggle between LLMs because that's when we discovered it really didn't matter much which LLM. But the important thing was to get started and to move quickly. We asked our customers what their pain points were. They told us we didn't want to wait another year and a half or two years to show them something. We wanted to get it done in a six month period. That was when I made the strategic decision of we're going to get some help here. We're not going to argue with them, we're just going to follow them until we get the first version of the product out into the world. And then based on its success or failure, we'll either ditch those people or keep going with them. So when we got really positive feedback from the first version, the one that was just focused on our documents only, we knew that we could stick with Deepset and get to the finish line on the second part.

I'm right there with you. That there's so many different tools now. The AI tooling is just grown like exponentially in the last couple of years. So it's very easy to get that paralysis by analysis and not be able to get anywhere. So I do like these with a simple approach for, okay, well, we'll get some people to know what they're doing and then build something and see what happens. And I'm sure if they're listening and they're breathing a sigh of relief that you're not going to ditch them after V1.

I mean, it's funny, if they're listening, let them take a shot across the bow based on the fact that they used Haystack, which is an open source platform. Platform, basically, Deepset is the organization that created that. When we talked about it internally, we were like, listen. And once we know what we're doing, we can always revert back to the open source version away from the full featured corporate version and do it ourselves. And listen, I would say that that's always a future Possibility. But what we've learned is after each new iteration of the product, there's just so much more that we can do and so much faster than we can move that the idea of retrenching back to something that's fully internalized as opposed to relying on an external partner isn't really that appealing to us yet.

Okay. Yeah, I think there's definitely a trade off between using a lot of free open source tools, stitching things together yourself, and then cheap software, but more time spent versus you've got a sort of prepackaged corporate solution which is going to be faster, but you've got to pay for it upfront.

Yeah, I can only imagine the war for talent right now. We couldn't hire the people on the deepset side for any amount of money. They wouldn't come work for an organization like us. So at some point you have to understand who you are and what your value proposition is. I don't think that we would be the kind of organization that somebody who's really proficient in AI would want to join. So best that we rent those people instead of trying to buy them.

That makes a lot of sense. Yeah. I think certainly the top end sort of AI engineers, they're making bank at the moment. So, yeah, difficult to get hold of for sure. And so how long did this process take? You said you started building this, what, a year ago and then it's been live 10 months. Did you really build this thing in.

Two months then or so, the first version? So we had a kickoff meeting at the end of February and we launched V1 on June 15th. So that was 120 days. Actually we launched on June 15th and had our first paying customer by day 120. So that was a good sign. In that 120 days, we had a kickoff meeting. We moved to prototyping really quickly and by that we just loaded a subset, not even all 4000 documents, a subset of our documents, into their sort of drag and drop vectorized database. And I was able to show board members, I went on tour with a number of significant clients and just sort of say like, this is the direction that we're heading. One of our clients is Apple. And it was really nice. They said, listen, we're Apple and we're never going to use your AI. But we. What we would say from our point of view is you're doing the right things, you're not doing the wrong things, and so keep going in that direction. Even though you count on us wanting to use our own LLM. And by the Way, that vectorized database that you just built, it'd be great if you could deliver that to our doorstep so we could include it. We could ingest it within our LLM, which of course the answer to that was yes. But basically within 120 days we had the kickoff meeting. We prototyped, we toured it with clients, got some feedback. I was able to show my board what we were doing. Even internally, a lot of our internal team members were nervous about us doing this. So the prototype really helped me show them you don't have anything to fear. We felt like with the prototype we had about 70% accuracy. And so with absolutely no prompt engineering, no tweaking, no custom nodes built by the Deepset team, that we were at 70% just by dragging and dropping some documents into a bucket seemed to be a good start. We then went through the process of developing it, and through that process, what was really helpful was we had a set of benchmark questions and then we made our own editorial team answer those benchmark questions. So let's say 20 to 25 questions and we had our editorial team say, how should an AI answer this? So we had this framework, this reference framework, and then as we were going through the prompt engineering, the custom node development process, we would just be seeing how much closer it got to that benchmark answer that we had. So then we finally got the development, launched a beta in late May and gave customers about a month to interact with it, took their feedback through the beta process, re engineered the product so that V1 was 120 days. Then you have to understand Deepset's based in Berlin. July and August, one of the Deepset team members had a baby. We, we really didn't get onto version two until late August, and at that point, again, another 120 days connected everything to the Snowflake database. Many more custom nodes developed. And so now what happens is the customer asks a question, it runs off and gets an answer from our repository of documents, runs off, gets an answer from our Snowflake database, and then takes and then combines those two. So the chart of the different components of the software that we look at Deepset Studio is kind of complex and things are kind of getting rounded back and forth and here and there across different LLMs. But at the same time, you know, again, within 120 days, we were in beta mode. We decided to stay in beta mode for a long time. We spent a month and a half till the end of the year. But at this point we've incorporated all that feedback. We're doing A slight kind of UX touch up. We feel like we could do some interesting things from an interface design perspective that would not only signal that this is a new version and a finished completed version, but we learned a few things. Things like people kind of like to see what processes are going on. They feel a little bit nervous. The more black box it is, the harder it is for them to sort of trust in the results. And so even giving statuses to what we're doing through the course of the minute or so that it takes for the AI to answer the question kind of gives them that comfort of like, okay, they're explaining the process, they're making it a little bit less black box. And I can understand the answer that much better as a result.

Yeah, that does certainly seem to be an important thing, is trying to find ways to gain trust with your users. I think there's a lot of people, especially when just like, well, AI gave an answer, not sure why your users are going to be like, well, I'm not sure whether this is the correct answer. I suppose since a lot of your work seems to be on retrieving facts, there's a real danger around hallucinations here. So talk me through, how do you make sure or that you reduce the number of hallucinations there?

Well, retrieval augmented generation, generally speaking should be fairly straightforward. Go check the repository for the most relevant document and when you get that, feed that through an LLM to kind of synthesize and summarize a series or chunks of various reports that get fed in on a zero day retention basis through an LLM. If you can't find anything relevant within our results, it doesn't give the LLM a chance to make something up. It basically returns, sorry, we can't find anything related to your question in our corpus. So retrieval augmented generation really helps in that if there's nothing there to work with, it says. So instead of pretending and going through, we did have in the early days with Sonnet, we had a couple of times in the text SQL where it would hallucinate a little bit on the SQL queries, but we managed to figure out where and how that was happening and stamp that out. But again, retrieval augmented generation really helps because there's a fixed corpus. We're trying to stay true to that. I don't know if you've ever heard of a groundedness score, but our groundedness score is in the high 80s, which means basically nearly 90% of the time whatever it returns from the LLM matches identically what our corpus said. We Find that that extra 10% is interesting. We haven't stamped that out completely because sometimes the LLM will make suggestions. One of the great use cases is can you suggest some campaign ideas based on the information that you've just seen? Can you make some campaign recommendations? And it'll be really creative in that regard, different ways to sort of engage young people. You can also ask the question, can you raise a creative brief for my ad agency? And there's no creative briefs in our corpus. It's pulling that from the LLM. What is a creative brief? What needs to be in a creative brief? And so we've left that last 10% to really open up some workflow and productivity enhancements to taking our information and creating new documents out of it.

That's good that the Retrieval augmented generation is working as it should be. And it's interesting that on top of that sort of fact based retrieval stuff, you're then letting the LLM just come up with the creative ideas. You can do sort of, I guess, writing, copy writing briefs, that sort of thing.

Yeah, we did have one issue where we had this debate internally. It was recency versus relevance. Right. So with youth information, probably something that's 10 years old isn't quite as good as something from yesterday. So we found that with Retrieval Augmented generation, every once in a while we'd have this recency versus relevance tug of war. There would be some content that was perfectly more relevant than something that was newer. So really trying to get the retrieval augmented generation model to favor recency over relevance was a bit of a tug of war, but I think we managed to get there in the end.

Yeah, I can certainly see how youth information dates pretty quickly then. Okay, so what's the solution to that then? Is it just you're prioritizing reports that are more recent then over something that might have a higher factual or higher match score? Is that the case?

Basically Retrieval augmented generation drives off of the match score. How relevant is this? Based to the query and just training it that like years matter and something from last year is more relevant than something from 10 years ago. We every once in a while it's funny because the underlying LLMs, they're not static. In fact, you might even still be on the same version of ChatGPT4 omni and all of a sudden they change something back at the ranch at OpenAI and all of a sudden you're getting a different result, which is where those benchmark questions have really come in handy. In fact, every Monday morning we have a staff member that re asks those 25 questions just to see what the results are and to see if they're changing at all, and if so, how? I'll give you an example. When we first got started, we would get answers back from the LLM. They were kind of unstructured, and so we asked for more bullet points. We wanted things to be punchier and more takeaways, and so we asked for bullet points. And then at some point through the first version, it seemed like the LLM decided that it liked bullet points as well. So we were getting bullet points on our bullet points. So then we had to back out of the prompt engineering the request to make it all bullet pointed, because they'd somehow figured that that would be a good thing to do just on a standard basis. So we find that the prompt engineering isn't always. It never ends. You're constantly prompt engineering, tweaking it, checking the benchmark questions, seeing what answers it produces, seeing how that's tilting in one direction versus another, and then redoing the prompts to account for that.

I think we've all seen that PowerPoint presentation where someone's got like, bullet points nested in bullet points nested in bullet points forever on the slide. But yeah, that's interesting that your testing process then, if I've understood this correctly, you've got a fixed set of questions, but then when you run them, a human has to look at the answers and judge whether or not these are good answers or not. So there is still a human in the loop during the testing process.

Yeah, indeed. And we've even started archiving these things so that we kind of see how the answers have changed over time. It's fascinating. The LLMs aren't in any way static, even version to version.

Is this something that you think you'll automate eventually, or do you think you always want a human checking this out?

I think what I would like to do long term is have this move out of the product team's hands and into our editorial team's hands, just because I'd like them to feel more involved so that I would like the human in the loop to be the creator of the content as opposed to a product team, a technical team. And the reason why is that we found that our editorial team has been a little skittish around the AI. They're not so sure that it's a good thing for them or a bad thing for them long term. And I understand their hesitancy completely. But I think to give them the reins of the beast and let them ride it around would be smarter and because they'd feel a little bit more in control.

Yeah, I can certainly see how editorial people might feel like they're competing with the AI in some sense. So you got the conflict, actually. How do you manage that as a manager? What happens when your employees are freaked out by the AI?

At first, of course, we said what everybody else says, which is, this isn't meant to replace anybody or displace anybody. This is meant to be something that allows us to just do more. I think that we've been true to our word on that front. Nobody's been let go. And I think that it's at a core level, AI and particularly retrieval augmented generation versions, they're good at repurposing some original idea, but they're not really good at the original idea. So the thing that we told our editorial team is we need you to keep doing what you do, which is you're the ones that are coming up with the original idea. This just allows that original idea to take more and more shapes and forms over time.

Yeah, okay. That's actually one of the sort of common things I've had from guests on the show. It's like when you ask, well, what's happening? Is AI going to take people's jobs? The standard answer is no, it's just going to make people more productive. But you see how well, yeah, in some cases it is going to take people's jobs. So I like the idea of just saying, well, humans should be focused on the creative side of things and then use the AI for repurposing stuff. And that works pretty well. So I'm curious, like, one of the biggest, big advantage of AI is you can personalize content. You can make the responses different for different audiences. Do you have to do that with your AI or market research as a sort of homogenous crowd, are there like a single Persona that you have to deal with?

We do have multiple Personas, for sure, but we've sort of aligned our products around those Personas. So we have three SKUs, as it were. One is Daily Intelligence. And that's meant for the casual user who just wants to keep up to date. We find that a lot of executives, they just want to not sound stupid and stay culturally relevant or with it. And that tool is really meant for people at the top of their organization that either manage processes that involve younger people and don't want to appear to be lame. That's a lightweight tool. You know, $299 a month, charge it to a credit card and you're done. And of course, what we do with that one is we restrict it just to our articles. So it's these daily articles that talk about young people in a certain way, of course, informed by all of our data and reports, but just in and of itself, it's an article. So that's the one Persona. The second Persona is more marketers. And marketers, they don't want to dig into the data, they don't want to slice and dice things, they just want the big idea so they can go off and kind of implement that. And that aligns with our reports a little bit more than an article, but not nearly something that you want to sort of dig in from a data perspective. And then the last product that we have is Pro and that allows not only dashboard data dashboards that we deliver up via tableau, but then also the ability to kind of go in and manipulate data. I don't care about 13 to 39, I really care about 18 to 24 and by the way, 18 to 24 year old females. So they're digging in and they're, generally speaking, they're the market researchers now implementing AI. Basically what we thought was, okay, we'll have the AI for daily intelligence, just use the articles as the input and we'll have the AI for the prime to just use the reports and then AI for Pro to then have additionally access to data. So that's worked out pretty well in that our products align with our Personas. But trust me, a bunch of clever folks in the marketing department have decided to cross the streams which we've resisted to date. But why shouldn't executives have access to the core data through AI? It's really fascinating. Again, we were curious as to whether deploying AI would mean people would read the reports less and people would dig into the data less, but it turned out to just be the opposite. If you give them a really fast path to answering a question, then of course they're going to want to fact check and dig in a little deeper and then go in and slice and dice the data. So so far it's working well. Our Persona is aligned with our product SKUs. Our product skews limit what's being fed into the AI and we've resisted the temptation across the streams. And I think I'm going to stay there for a little bit longer strategically before I let executives tap into data.

So I'm curious as to whether the existence of this AI has changed how you create content. Do you create these reports now knowing that they're going to be fed into AI. Does that change how you make them or the structure of them?

So far, again, in order to keep the sensitivities of the editorial team in check, we've used the same process to create the original content. But to the editorial team's credit, once it's created, they're doing some really interesting things beyond that. So one of the things that really vexed them or seemed kind of low value was all of the sorting out of SEO keywords for every article, for every report. We're using AI to do that now, which they love that. Take the drudge out of it. I love the quote, which is, I didn't want AI to write poetry and music for me so that I could do the laundry. I want AI to do the laundry so that I could write music and poetry. So what can we do to take the drudge work out? So developing the SEO keywords out of each article, that was drudge work, AI is doing a great job of that now. We also have these mass repurposing for variation. So you write an article. So now give me that article as a LinkedIn post. Give me that article as a link listicle that we could put a carousel on LinkedIn. Give me that article as a series of 20 second videos that we can put out into the world. And that's been really helped by AI. So so far we've had this notion of, hey, the original content, the original creation process, we're not going to touch that. Then everything that follows thereafter, we're going to see how AI can help. Certainly AI has been embraced by our marketing team, which they used to have to either have somebody beg somebody on the editorial team to repurpose something for marketing purposes. And they're like kids in a candy store nowadays and they're using AI to take and repurpose stuff for marketing reasons. So they don't have to really worry about marketing content anymore. They all sort of go to the watershed and develop it from there.

I do like the idea of just using the AI to create all the sort of secondary content. So all those sort of social posts, the marketing copy, all that kind of stuff. Because it's not super, super fun content to create where you start taking an existing report, right? Well, okay, but it's got to turn this into a LinkedIn post. I mean, I'm sure there are people who enjoy doing that, but it just seemed like AI can help you a lot.

Say it a third way, say it a fourth way, give it to me a fifth way. I mean, I'd Imagine in the world of copywriting and editing, these poor guys, Don Draper sitting at a typewriter, coming up with yet another tagline, Right? That was kind of what it felt like to them.

Cool.

All right.

So having built this thing, are there any things you wish you'd done differently during the process?

I've been running this company now for 17 years, and I used to weep over the data that we'd collected in any given year that wasn't consumed by enough customers to sort of make it worthwhile. It's collecting data longitudinally is a real chore. And I used to think about it like perishable produce or fruit or vegetables. Right. That basically, once I asked those questions that year, if there wasn't a customer at the receiving end of that, that it was all for naught. It was all sort of gone. And now I don't look at it that way. Now I look at it like, way we've created a catalog or a music library, and we can tap into that, that there is residual value in the data that we collected five years ago, 10 years ago, if for no other reason than to kind of create these trend lines of which direction things are moving. And to be honest, we could have done it in the past without AI I would have had to sit some data analyst in a room and make them sit there for five weeks combing through past data reports, taking the numbers from this question last year, the year before that, the year before that, drawing that trend line. And I like our data analysts too much to do that to them. Now we've got AI which is perfectly happy to comb through historical data at record rates of speed and give us that trend line over time. And it's really valuable, seeing the directionality of things. Yesterday's information's great. Yesterday's information, as the latest data, point to 10 previous years of data is even better. So if I had a time machine, I'd go back and I'd say, listen, don't worry about that. Just collect as much data as you can. You'll thank yourself for it later. And even now, we find ourselves accelerating the types of data that we're collecting so that 10 years from now, we'll have something really interesting to show people.

That's cool. I like that the use of AI has now unlocked new insights because it makes the data analysis so much easier. Particularly like the stuff that you thought maybe wasn't so exciting, the older data, the stuff about they're no longer teens, they're now in their 30s. It's like, what's changed. So, yeah, I like that it's unlocking new use cases. Have you got a favorite insight you've got from the AI then?

So one of our customers is a paint company, so they make paint for a living. Pretty exciting stuff, right? So we'd loaded all of this stuff into AI and I was just giving them a demo. Right. And I still, even as president of the company, love to give our clients demos of AI and sort of chat with them about the results. But we basically, we were asking the AI about trends in home decorating, especially amongst Gen Z, so younger half of the age ranges we deal with. And it came up with something really fascinating. It said, well, young people tend to be renters, and renters have certain needs for temporary design that homeowners don't. So paint is kind of. It's a tough thing for a renter because if you paint it, there's a chance you have to repaint it or that you're going to get charged money because you painted something. And again, this is where the LLM kicked in. It was talking about removable wall decals. Decals where you could basically, if you think about those fathead sports ones, things like that, you could dress up a room without applying anything that was permanent, any paint. It even talked about temporary flooring. And you can just sort of see the lights going off on the other end with the client. It's like, wait, I totally get it. We think about paint as being something that everybody wants to do, but a renter may not want to do that. But a renter does want to make their home look better or more individualized or more personalized. So what can we do? That's paint that you can put on and then immediately take off or at some point apply some other chemical to it that isn't toxic. That basically removes the paint that you use to sort of make that space your own.

That's fantastic. They just unlocked an entirely new business idea just from having a little conversation with the AI and seeing what the market wanted. Okay, fantastic stuff.

Cool.

All right, so tell me, is there anything else you're particularly excited about in the world of market research as we.

Move from the what to the so what to the now what? It's really exciting for me to be able to tell clients, here's what you should do next. You should come up with a series of removable wall decals. You should come up with a flooring concept that you can put down, and it will stay in place until you unlock it and pull it back up again. So Those kinds of things are really interesting use cases. What I'm really excited about is that unlocking that ability to take a client into future directions and give them new business opportunities gives us the ability to kind of grow our relevance within organizations. The to ability change the sort of seat that we have at the table. Instead of folks that are out there to fetch you an answer to a question or kind of folks that you would want to have in your hip pocket as you were designing the future of your organization. That's pretty helpful. And in our industry, the big fish eat the little ones. So YPulse is the fifth company that I've created. What I'm super excited is how AI is kind of unlocking all this latent value in something that I could sell to a bigger competitor where we're small and nimble enough that we've made these sort of steps and moves with the scarce resources that we have. Much larger organizations than us haven't even gotten started on this front. So I could see us selling our organization into a bigger organization that has even more data than we have, that needs these techniques even more. And I'm following in the footsteps of some really great entrepreneurs that have created business value and then harvested it. Got an impactor Comer who sold his company Lucid to another company in our space called Sint for a billion dollars. Or Amy Pressman and Berger Hald, who sold their company Medallia to. Well, first they took it public and then they sold it for thoma Bravo for $6.5 billion. I don't know if you've heard of Ryan Smith or Qualtrics, but not only did he sell Qualtrics to SAP for 8 billion, but then they spun it out and took it public at a 21 billion valuation. Data and content are certainly valuable. These are companies that have unlocked tremendous value without the benefit of AI. Right now, in some dorm room at MIT or Harvard or Princeton or some other university, maybe Oberlin College, there's some young people that are going to use these tools, upend everything we've created, and lock even more value. So I'm kind of excited about the future of industry because the folks that are coming into it now will be kind of AI first thinkers, taking everything that we're doing and rethinking it.

That's very cool. I like the idea of having these business insights at scale. So it's not just the paint company you're helping, it's every single customer just getting these new business insights there. And yeah, I mean, you're talking Billion dollar companies here if you can help all these people at scale. So that's pretty exciting. So just to wrap up, I'm always interested in follow recommendations. Is there anyone whose insights you're particularly keen on at the moment?

So I thought about this from a few different angles. Let me recommend a book. A guy named Jack Skeeles wrote a book called Unmanaged and what that's taught me is how much value destruction occurs at the top of the organization by folks that figure they're so smart and strategic and they're going to decide what to do next. Anything that we've done that was good really kind of came from the bottom up as opposed to top down. That's been really helpful. Kyle Poler from Growth Unhinged has really helped me in terms of account based marketing and we've got a great product. What we really need to do is make sure that as many people as possible get a chance to sort of think about whether it could help them in the challenges they're facing today. So Kyle Poler is really helpful. And then the last thing that I really feel strongly about is especially as a CEO, you need a community and you can't just talk to your staff. It's really helpful to have a community of like minded people who are facing similar issues and you could sort of talk to them about how they're dealing with it. So I belong to an organization called SaasRise which is led by Ryan Alice who was the founder of Icontact and he's created. Now there's about 400 of us SaaS CEOs, we meet at various levels. So my level is the we've got a greater than $50,000 a year ACV. And so me and the other guys that have greater than $50,000 a year ACVs get into a corner and talk about what we're seeing. And I can't tell you how many problems I've solved just by talking to somebody who has a similar business to mine and how they've faced issues maybe a year or two ago and how I might leverage their experiences for my benefit.

Okay, that's great stuff. I've said that Unmanaged book sounds like a great gift for your boss, like.

Whoever you are, especially if your boss is an idiot. Nice.

And yeah, the other thing, like the group for Susan CEOs like in general, like whatever your job is, find people who are doing something similar to you. You have a work therapy session, gripe about whatever and yeah, solve some problems together. Yeah, it's a great idea. I love it. All right. Thank you so much for your time, Dan.

Okay, great. Well, thanks so much. I really enjoyed.