In this talk, the speaker recounts how Zillow tackled the classic marketplace cold‑start dilemma by inventing the Zestimate in 2005‑2006. He explains that while existing real‑estate sites focused on “what’s for sale,” Zillow identified the more fundamental consumer question—“what is my house worth?” and decided to provide a valuation for every home, not just listed ones. To achieve this, they recruited analytics lead Stan Humphreys from Expedia, raised a $6 million pre‑seed round (largely funded by the founders’ own proceeds from Hotwire), bought extensive property data, and built valuation algorithms. When launched in February 2006, Zillow offered 40 million Zestimates with a 14 % average error and attracted a million users on day one and four million in the first month, a launch speed surpassing many later tech products. Over the next 15 years, continuous improvements using AI, machine learning, and computer vision reduced the error to about 4 % overall (2 % for homes on sale) and expanded coverage to 100 million homes. The speaker reflects that while the initial launch created a massive traffic spike, maintaining that level took two years of sustained effort.
Chapters
You know, in the very beginning as marketplaces, we all face the kind of the chicken and the egg, right? Or the kind of cold start problem. So, you know, how did you kind of solve for that? We solved for that with the Zestimate. And the Zestimate was this idea that we had. Well, let me back up. So we're like, okay, we should build a real estate product, a real estate website of some sort. It'll provide information transparency. It'll empower consumers with access to this information, which is hidden behind secret databases at the county courthouse. in the MLS, you know, in consumers' or real estate professionals' brains, and we'll provide information transparency. And then we said, well, okay, what's the most important question that everybody asks about real estate? And the way most websites at the time, this was 2005, there were plenty of real estate websites at the time. The question they were all answering was, what's for sale? Every single one, realtor.com, every, I mean, every website at the time was answering what's for sale. And we're like, well, that's interesting. But the question that most people ask themselves is not actually what's for sale, it's what's my house worth? And there were lead gen sites at the time that would say, what's my house worth? In fact, House Values was a publicly traded company that would have TV ads that would say, go to housevalues.com and we'll tell you what your house is worth. You go to housevalues.com, you'd enter your address and hit submit, and then it would say a real estate agent will call you to tell you what your house is worth. So it was lead gen to real estate agents to tell you what your house was worth. And so, you know, I still remember we were in a high-rise in downtown Seattle. We were looking across at Queen Anne, which is a residential neighborhood kind of across the lake. And we were like, imagine if you had goggles, like superpowers, like godlike goggles. There was no VR, AR, Oculus, Snap Spectacles, none of that existed yet. We're like, what if you could put on these glasses and you could like see a price on every rooftop and you could see all this Like you could look into the houses and see on every single house, not just the ones that are for sale, but every house. That would be really cool and voyeuristic and, you know, incredible and illuminating and traffic generating. And so we said, okay, to do that, we need to figure out what every house is worth. That seems really hard. And so we called this guy Stan Humphreys, who ran analytics at Expedia and worked for us at Expedia. And we're like, hey, we finally figured out what we're going to do. It's been 3 months. We've been sitting in this conference room, you know, editing family photos. And like, we're going to try to figure out what every house in the country is worth. You should, it's time to leave Expedia and join us. And he is like, I don't know anything about real estate. And we said, well, neither do we. Like, this isn't about real estate. This is a math problem. This is a simple, well, not so simple. It's a pretty complicated math problem, but it's a math problem. It's not a real estate problem. And we're going to figure out what every house is worth. And he said, okay, fine. So he joined and then we did a $6 million pre-seed round from the founders, from us. Remember that down round I mentioned at Hotwire? When we sold Hotwire for $700 million, I made $1 million and I was the co-founder. And with that $1 million, I bought a $700,000 house or $650,000 house or $650,000 down payment on an $800,000 house. And with the other $350,000 that I made from Hotwire, I put it into the pre-seed round at Zillow. So I just You know, put it all on, you know, all out, you know, bet it all on black. And actually, that's a bad metaphor because that would make it sound like there's a 48% chance of success. I bet it all on like double zero and put the $350,000 into the pre-seed round at Zillow. And we raised $6 million from the rest of us. My co-founders each did $2.5 million. They had more money than I did from their Expedia experience. And then the other couple employees did the rest. And so we did $6 million pre-seed. We spent a couple months buying data and building algorithms to try to value every home in the country. And when we launched February 2006, we had 40 million Zestimates. So that was about half the country with a 14% error rate, meaning that every time a home is sold, the Zestimate was about 14% off on average. So 40 million Zestimates, 14% error. Fast forward 15 years later, today Zillow has 100 million Zestimates. So basically every home in the country at a 4% error rate. And on homes that are for sale, it's about a 2% error rate. So the Zestimate went from 14% wrong to 2 to 4% wrong over 15, the 15 years that I was CEO. And the way that it did that was through AI and machine learning and computer vision. And we can talk about that if you want, but, you know, 15 years of just the algorithm improving and us improving it and it improving itself. But to answer your question briefly, Mike, the answer on the cold start problem was the Zestimate. When we launched, we had a million users on day one. We had 4 million users in month one. That's still the fastest launch of any product, including ChatGPT. They didn't have a million users on day one. Faster than Snap, faster than Instagram, faster than WhatsApp. And, you know, so, and by the way, it took 2 years to get back to that first month traffic. So huge spike, 4 million users in month one, which is only a 3-week month. Yeah. And then, and then it took 2 years of the grind to get us back to where we were on month 1. So I guess we didn't really solve the cold start problem. We had a very hot start and then it became lukewarm and then it took 2 years to get it back to where it was.
