Zillow tried to out-price every agent in America and lost $881m

In November 2021 Zillow announced it was shutting Zillow Offers, the division that used the company's own pricing algorithm to buy houses straight from sellers and resell them at a margin. By the time the books closed on the year, that one business unit had lost $881 million. Roughly a quarter of Zillow's workforce had been cut, and the company was sitting on thousands of homes it had paid more for than the market would give back. The Zestimate, the tool that made Zillow a household name, had nothing to do with the failure. What failed was the bet that a good enough algorithm could stand in for the person who actually knows the street.

A valuation for every home, whether you asked for one or not

The Zestimate launched on 8 February 2006, putting automated value estimates on tens of millions of US homes at a time when finding out what your house was worth meant phoning an agent or waiting for the place two doors down to sell. Coverage grew from roughly 40 million homes with monthly-updated values to more than 110 million homes today, refreshed several times a week. In June 2021 Zillow replaced the original model with a deep-learning rebuild it calls the "Neural Zestimate", trained on far more granular data.

What the Zestimate actually earned Zillow is traffic. 221 million average monthly unique users across its apps and sites in the fourth quarter of 2025 alone. That is the payoff of a free, instant valuation that is always on. Buyers and sellers start with Zillow because Zillow already has an answer, before any agent has heard from them.

Now look at where the tool is accurate. Zillow's own published figures put the median error rate at 1.9% for homes currently on the market and roughly 7.0% for homes not listed at all. Those two numbers carry the whole story. Once a home is listed, the model has an agent's asking price to anchor to, so it is marking its own homework against a number a human already produced. Take the anchor away and the error nearly quadruples. The algorithm performs best exactly where a professional has already done the hard part.

Then Zillow became the buyer

In April 2018 the company launched what became Zillow Offers, starting in Phoenix. Rather than telling you what your house might be worth, Zillow would buy it from you at a price its own model generated, do light renovations, and resell it. The programme grew to 25 cities across 12 states. This was a different bet altogether. A wrong Zestimate cost Zillow nothing. Zillow Offers put the company's own balance sheet behind every number the model produced, at portfolio scale, mostly on homes no one had physically walked through for Zillow before the offer went out.

The wind-down

On 2 November 2021 Zillow announced the wind-down. CEO Rich Barton put it plainly: "We've determined the unpredictability in forecasting home prices far exceeds what we anticipated, and continuing to scale Zillow Offers would result in too much earnings and balance-sheet volatility." About 2,000 people, roughly a quarter of the staff, lost their jobs in the restructuring that followed.

The final numbers landed on 10 February 2022. Zillow's Homes segment, the unit built around Zillow Offers, posted a pre-tax loss of $881.5 million for the full 2021 year. The rest of the business, the advertising and leads operation the Zestimate feeds, stayed profitable enough to hold the group's total net loss to $527.8 million. In its final quarter of active buying and selling, Zillow lost an average of roughly $25,000 on every home it sold.

Why the algorithm broke

Researchers at Stanford Graduate School of Business who studied the collapse point to a specific failure. Zillow's model had learned its sense of normal from a relatively stable pre-pandemic market, and it kept bidding as though the 2021 price surge would hold while conditions shifted underneath it. More conservative iBuyers pulled back sooner. The statistics were sound as far as they went. What they could not do was price one specific house in a market that had stopped behaving. That is local, current judgement, and a model bidding at national portfolio scale is the wrong shape of tool for it, no matter how much data it has seen.

What this means for South Africa

Part of Zillow's problem was the ground it stood on. The United States has no single national record of who paid what for which house. Price data comes from roughly 489 separate regional multiple listing services, on the Real Estate Standards Organization's own count, each with its own rules and coverage, stitched together with county assessor records of very uneven quality and update speed. Even a company with Zillow's resources had to model its way across that patchwork. When a once-in-a-generation price surge arrived, there was no clean, current ground truth to check the model against.

South Africa starts from different ground. Every transfer of ownership here passes through one national deeds registry. A South African pricing model would still run into plenty of what sank Zillow Offers, because condition, position and buyer appetite need a person who has stood in the road outside the house. What the registry changes is the job the technology should be doing. Zillow built a guessing machine because no reliable national price record existed. A tool built here can start from an actual record of what was paid, and put that record in front of the person who has to turn it into a number a client will trust: the agent.

There is a second, more local echo. In 2023 South Africa's Competition Commission found that Property24 and Private Property were charging agencies rate-card prices with non-cost-justified gaps of more than 300% between comparable tiers, and ordered remedies. An American company modelling prices at national scale and a local portal pricing access to listings raise the same question: who controls the price data in a property market, and what makes them use that position responsibly. Zillow's failure was technical. The Commission's finding was about conduct. Both come back to concentrated control of property price data.

The findproperly angle

The two halves of the Zillow story point the same way. The Zestimate worked because it made checking a price a normal, expected thing for ordinary people to do. Zillow Offers died because it asked an algorithm to do the agent's job. findproperly is built on the first half. It puts sourced deeds registry sale prices, ownership history and comparable transactions in an agent's hands, so the number they bring to a client rests on an actual record rather than a private guess. The agent still sets the price. They just never have to argue it without the record in front of them.