From Social App to Data Giant: Foursquare's Evolution | Gary Little from Foursquare

The company was launched famously at South by Southwest by the team who was really playing with this notion of: we all have these smartphones in our pocket, and we now want to understand the world and interface with our friends through that digital medium, and evolved the business a bunch over time. We started as a consumer app, right? And with the rise of big platforms like Facebook and others in the social media space, it crowded out the notion of a location-based social application being the primary mode by which you engage with your friends in the world.

And since that time, I've been thinking about what is the value of what we've built over time. And so, Matt, we were talking just a little bit before, I got involved in the business six years ago, raising capital for the prior iteration of the business, and really developed this thesis around location and geospatial as a broad category, and the need for software to really unlock the value that's inherent in these very large, very complex, very valuable datasets that, for most customers, are unwieldy. It's very hard to get to value.

Today, and so when we look at what Foursquare had built over time, it was the ability to understand where devices move in the real world in real time, and that technical architecture lends itself very well to all these use cases that are basically exploding in the world. And so, as you think about where we've shifted the business, it's really around taking those component pieces and building a location and geospatial platform for our customers to really not only use the data that we have, but use the data that they have to build great products, to do great data analytics, which in practice is a much harder problem than folks would have thought 10 years ago with the rise of mobile compute.

I think one of the great things about the legacy that we get to build off of in the social app context is we were crowdsourcing, effectively, a map of the world. And so we operate in 190 countries globally, with 200 million global POIs, or places of interest, including a level of specificity of knowing the difference between Foursquare's headquarters and being in Fat Denny's cafeteria, and having that telemetry of what's in the world, and then building mechanisms that keep that fresh. The world's obviously super dynamic in terms of how places change over time, and so the starting foundation of what we do is really understanding places in the world at scale on a global basis, and then attaching, in real time, how devices move in relation to those places.

So what does that mean for all sorts of pattern development, for understanding foot traffic? If you're a retailer, you can think about applications in places like logistics or real estate, understanding neighborhood patterns or so forth. And so that's the sort of nature of creating this very dynamic asset, which is constantly changing, and building the mechanisms to be able to have what we think is unique, other than basically Google in the world, which is human confirmation. We have elements of humans telling the machines that we're right when we're making informed decisions around where a device is.

And you can imagine in a place like New York City, that's a highly complex technical problem because, again, the GPS probably assumes all of us are on 6th Avenue right now. And so how do you get that specificity, that you have a cluster of devices in places like this? And it's understanding that from a machine perspective using ML techniques, but also then having human verification that's telling our machines all the time, we know when we're correct. We also know when we're completely using math to guess where we are.

And that's a big differentiator in the space.

I think what can be confusing sometimes from the outside is, to your point, Matt, we have a lot of products that we've developed. For those that don't know, we've done a lot of acquisitions, so we've acquired end-product capabilities and, therefore, revenue streams and customers and so forth. But if you look from the inside out, everything that we do is based on sort of three data primitive layers, right? Places in the world, users, and movement of those users in relation to place.

And so every product that we build, when we think about the platform, everything starts from an understanding of those data primitives and more. We're now starting to add things like transaction data into that, but it's really that corpus of place and movement data that allows us to build every product that we have, and then more and more as we expose that for customers to build based on those same primitives. So when we think about what is the platform strategy, we talk a lot about it.

One analogy I like to use for the team internally is what we've been up to over the last three years is really similar in some sense to Amazon building AWS many moons ago, right? Realizing they had built up a bunch of infrastructure. Hey, a bunch of customers might want this infrastructure. How do we then productize that and make it readily available for others to use? We're in the phase of building, re-engineering our platform for that scale effort, where more and more customers will come to us, not just for the data, but our ability to actually process and systematize a bunch of these big data assets in real time, which is very, very difficult and very costly in the current infrastructure that exists in the world.

The platform layer for us is being able to use our data, our systems, our tools to manage those data workflows and get to value much faster. Right now, we have so many customers that will buy data from us over very long periods of time, and often when they churn now, it's because they can't get to value. They can't build whatever they were trying to build. And a lot of that is because they don't have the relevant software to get there, and they don't have specialization in this otherwise very complex space.

And so when we think about platform, a lot of what we're trying to do is take everything we've learned from these very big distributed products that we have and get to the common components of the infrastructure. How do you think about place, device, and user movement, and how do you orchestrate that? How do you build in things like privacy and security the right way? And so that's what we are really on this journey to do. This year will be the first year that we really start to release pure platform products as opposed to end-to-end either data assets or packaged products.

Yeah, across our portfolio today, we do everything from giving mobile developers APIs to build personalization into their mobile apps. We use data analytics to do things like measuring ROI and attribution. In some sense, we'll do some version of almost everything that we do today. I think the nuanced difference around how we are approaching problems going forward is more and more we're exposing the components that we've built to power a product like Attribution, where we take data from the world, location data, we take impression data, we do a bunch of statistical analysis on lift relative to your exposure, and we give back to the customer a report that says, here's your metrics.

A lot of our customers want to build some of those capabilities natively, and many have tried, and most have failed because it's a very, very hard technical problem. So when you think about how we evolve the product, you can either buy the product from us with reporting and UI and customer service and all of that, or there's a bunch of the component pieces that we will make available for developers to build their own versions of these things, where you can take any event catalog and say, I exposed a customer to X.

Did that drive more people into my store? Did it lead to more incremental sales? And so, in some sense, most of the products that we have today, we've been deprecating a lot of products over the last 3 years as we get to the core of what we think the biggest opportunities of value are. And we will manage those products going forward, but more and more open them up for developers to build their own versions of it with their own datasets, right?

As opposed to moving data back and forth constantly.

Yeah, look, I think there's kind of two ways I think about that. One is very specific to location and geospatial, which is, as fascinating as this world is to us, we don't see enough gravity in the world for the specialization to be deep enough across all of our customer cohorts. So what we've seen, the prevailing business model for a long time is companies like mine existed, in the past, to create data, give that data to a customer, and then the customer does stuff.

And more and more, we found our customers don't have the sophistication to build these very specific either product categories or some of the data analytics they're trying to do because of the specialization. And if you think about that from the outside, this was our general observation when I was in your seat, Matt, and investing in Foursquare before I came to run the company, which is, that's a software problem as we defined it. We understand all of the specialization between the raw data and the customer value point, and there are no tools and systems that exist specific to what we do that allow you to do that very easily.

So you have to be a highly specialized geospatial data person on the other side, generally speaking, to get to value. And so we had two ways to think about that problem. We can either package up every product and try to solve every use case that exists in the world, or we can step back and say, okay, most of the things that we do are predicated on a similar set of problems within understanding place and movement data and users. And if you could actually build workflows for that and expose those, you can get the customer much closer to the value proposition that they're trying to build for.

That gave rise to the start of our platform strategy, and then in came a bunch of privacy regulation around location data. And with the rise of that, what it basically constrains is the ability to freely share data across walls. And so more and more, you think about most of our customers that are operating in the real world have huge datasets in location and geospatial. They're coming to us to buy data generally because it's not just the raw data; we do some things like normalization on top of it.

They're not even using, in most cases when we talk to our customers, data that they've created inside their own walls. Because it's just too complicated. So again, the beauty of taking a more developer-centric and platform-centric opportunity is it also brings to life the ability for the customer to use their own data. And given privacy constraints, we see that as the more optimal path, where I can package up things like the tooling and infrastructure, but also the models that we've trained on debiased data across the world that allow you to then apply those techniques to your dataset.

And get to the outcomes that you're trying to get that you might not even be able to build the data science in-house anyway. And so when you think about the packaging of basically data with spatial functions and compute, you have something really interesting that gets your customer much closer to value.

Yeah, I think there's really been kind of two things I would sort of say. If you go back to the founding of the company, I give Dennis and the team a lot of credit from the founding for understanding that you're kind of working with superpowers when you understand a person's location. And so from the beginning, the way that the systems at Foursquare had been built was really this thing we've all gotten used to now because Apple and Google have caught on, and at the operating system level, we have to opt into location tracking, and it's very, very clear to the user.

In the early days, that was not the case. Foursquare, from the very beginning, whether it was tech we built or when we would allow folks to use our SDKs and otherwise, we always had that location-tracking opt-in, and that's been very key to the business going forward. What's changed, Matt, is ironically, with the Dobbs decision in the Supreme Court overruling Roe v. Wade, you kind of set off this interstate sort of blue-state, red-state dynamic, where blue states are very concerned that red states can subpoena data around women's health, which is a very interesting thing for a location enterprise software company to realize, okay, this is the thing that I would not have predicted three, four, or five years ago.

And what that's done is it's put parameters on a really important set of data. We call it sensitive locations. It just means, is there something about a location that gives you data you shouldn't otherwise have about a person? So female health clinics or religious affiliations or things like that that you wouldn't want freely exchanged in the world. And so we've joined trade groups to really—we think the industry was acting very blasé about privacy before. We're glad to see that there's a lot more gravity toward a better set of rules around how do we track that, how do we anonymize data, and then how do we make sure we don't keep things that we don't need in our systems that are personal identifiers that actually harm can come from people understanding, right?

And so I think from our perspective, the biggest shift and most important shift, and one that we're the most happy about, is we're moving from this very personalized ID, so mobile identifiers and so forth, to a more anonymized world where people are looking at these datasets in aggregate, which has a lot of privacy protection built into it from how we freely exchange, how do we connect datasets, and so forth. And I think that's been one of the more exciting things for us to see, is that there's now institutional push against that, because competitively, it was hard for us always to rationalize being on the side of privacy always when you're competing in a market where others aren't exactly behaving that way.

If you're looking at datasets of hundreds of millions of places with 70 rows of attributes, with a bunch of movement data and a bunch of user data, the real question from a data science perspective is: how do you understand those relationships in the world to answer really basic questions? Like, what are the most important intersections in New York City if I'm Diageo and I'm trying to make sure my alcohol is in every major bar in New York that's most trafficked during happy hour, and things like that?

Those questions sound very benign, but when you actually think about the data science, trying to put together SQL queries in such a manner to understand those interrelationships is virtually impossible and, honestly, was part of the reason that you don't see a lot of the value derived when you buy the data, because you have to spend so much time, energy, and effort just to try to make sense of it. And by the time you've done that, it's changed. Knowledge graphs are very interesting because you can understand the inference of how these things are related in a much more dynamic and real-time way, which is basically the core of our value from a business perspective.

And so we've re-architected our entire data architecture to be much more in the data-primitive knowledge graph construct. I wish we could say we did it because we knew LLMs were coming around as fast as they are. But now, when you combine it with natural language, you can actually ask very basic questions and allow the techniques around understanding hundreds of millions of points of inference that get you to the answer that you're actually looking for. Whereas you would not have been able to even remotely do that in the prior version of trying to string together these datasets.

You wouldn't necessarily even know what you should be looking for, per se. And by exposing it on a knowledge graph, we can actually do all sorts of very dynamic analytics on what's happening in the real world, because all of our questions are some nexus of a whole bunch of things happening in real time. And I want to know something like: am I going to have a traffic problem? Do I need a new road, a bridge, a toll? Do I need a new bar here?

Or whatever the question I'm trying to get to. And you can do that in much more real time at massive, massive efficiency gains in terms of your infrastructure costs, your compute costs to even ask the question to the machine. So it's foundational to basically the core of everything we do from a product stack going forward. Everything will be built off of the same data primitive layer, the same graph layer that allows us to really surface the data in real time.

Yeah, well, I mean, the first thing it means is we keep renaming things. So just when I think I know what's going on, we come up with a new name, right? No, in all seriousness, the way we think about a computational shift like this is, I actually was asked at the board level, I don't know, probably two board meetings ago, "Who's your AI person?" And I thought that was a really funny question. What does that mean? Who's my internet person?

I don't have an internet person either, right? All of the things that we're doing, these big data systems, are all informed by all of the techniques being applied. And I know, obviously, LLMs are like everybody's favorite bag today, and we'll certainly use them. But the way I think about the mission that we're driving at with the geospatial platform is, we believe that ultimately, by taking this approach, it democratizes the concept of location and geospatial being highly valuable to the world. And with AI, what that does is it just gets you there much faster, right?

Because again, if I think about what am I trying to solve for my customers, most of my customers can't do geospatial data joins at the level of efficiency that you need. And if we can actually take that away and just ask a basic question of a machine, like, "What are the most important intersections in New York City today at five o'clock?" that's just massively democratizing. So many business analysts who aren't technical need to ask these types of questions to run businesses more efficiently every single day.

And using these techniques just completely changes the aperture of who our customer is, right? So it's now not only every customer on the planet with a physical location, which is most businesses. Inside the business, it's now not the data science department. It's literally the finance department, the sales department, marketing department, everybody. And so when I think about how do we evolve the business, this is just giving tools and infrastructure to make that happen much more quickly. That said, it's really, really early for us.

From pure AI, when we think about what's coming, I think the most interesting thing for us really is when the cost of inference models becomes highly efficient. That's where, for us, that's probably the most valuable crossing-over of a new platform, because most of our questions are real-time in nature on massive datasets. And if we can do that at a cost-efficient structure for either ourselves or to our customers, it just unlocks everything that we talk about all the time wanting to do.

LLMs will have some application as well in it, but I think it's really interesting to watch the evolution of the core modeling that we're doing every day become more efficient, more capable, more dynamic.

It's like the supertrend before the AI supertrend took over everyone's consciousness, Matt, at least in terms of the way we think about it, which is, again, my journey here going from advising the company to investing in the company to running the company. The investment thesis, really, when you take a step back, is everything's connecting. And not only are they connecting, but devices actually function now with context in the real world—autonomous vehicles being a very simple example, right? And then COVID happened, and everyone heard the term smart supply chain for the first time in history, all this.

But if you think about what all these things mean, they basically mean that a computer has to actually understand where it is at any given time to provide the value that it's trying to give to a consumer. And so when we think about that, that's like a supercycle of growth for the wider lens of location and geospatial being an interesting place to build a platform and build tools for folks to be able to do all this stuff much more efficiently, because the capture of these datasets and how you use them and how you do that efficiently is a huge opportunity, because more and more things are connecting.

And I'm not smart enough to know if AR, VR—is it a thing? Did Apple break the market? I don't know. I don't see any goggles here tonight. But I know more and more things are going to talk to centrally located computers and give back value to us as end users in all sorts of really compelling ways. And to do that, you need a lot of the tools, software, and capabilities that we're building to do it efficiently.

And we think, again, that that's specialized. Most developers want to focus on whatever the Glass experience or whatever you're trying to build. They don't want to figure out—the analogy I like to use to the team all the time is, this is why Twilio became so big so fast. People didn't actually want to sit around and bang their head against trying to figure out how to text the customer, right? They're like, oh shit, there's an API for that. That's phenomenal.

So we think there's a very similar analog, and there's more and more things that are going to connect and drive value based on spatial compute. And we like that.