The views expressed here reflect personal observations from decades in the field and are for informational purposes only. They do not constitute financial, investment, or professional advice. Every property and situation is unique, consult a qualified professional for guidance specific to your circumstances.
In November 2021, Zillow shut down its home-flipping division and laid off 2,000 employees. The company had lost approximately $881 million buying homes based on its own Zestimate algorithm. Rich Barton, Zillow's CEO, said the algorithm had created "an unacceptable level of variance" in their offers.
Let me translate that: Zillow trusted its own numbers, bought houses based on those numbers, and then discovered that the houses were worth significantly less than the algorithm predicted. The most sophisticated automated valuation model in the world, backed by the most comprehensive real estate dataset ever assembled, got it wrong badly enough to lose nearly a billion dollars.
And yet millions of Americans still check their Zestimate every month as if it were gospel.
What the algorithm actually does
Here's what most people don't understand about the Zestimate: it's a statistical model, not an appraisal. It looks at data points, square footage, lot size, bedrooms, bathrooms, year built, tax assessments, recent sales in the area, and runs them through a regression algorithm to produce a number.
It's good at identifying broad patterns. If three-bedroom homes in a particular zip code have been selling for around $500,000, Zillow will get you in that neighborhood. But "in the neighborhood" can mean a range of $100,000 or more. That's the difference between a fair deal and a terrible one.
"Zillow knows the price of everything and the value of nothing. Those aren't the same thing, and confusing them is how people lose money."
What the algorithm can't see
The Zestimate can't walk through the front door. It can't smell the mold in the basement. It doesn't know that the kitchen was remodeled by someone who cared versus someone who was flipping the house as cheaply as possible. It doesn't know that the neighbor runs an auto repair shop out of his garage. It doesn't know that the sellers are divorcing and need to close in 30 days.
All of these things affect value. None of them appear in the dataset.
This is the fundamental limitation of automated valuation: it can process data, but it can't exercise judgment. And judgment is exactly what real estate valuation requires, because every home is a unique combination of physical characteristics, location specifics, market timing, and human decisions that no algorithm can fully capture.
Why this matters to you
If you're buying a home and the seller is pointing to the Zestimate to justify their asking price, you should know that the algorithm has a median error rate of about 7% nationally. On a $500,000 home, that's $35,000. In some markets and for some property types, the error rate is significantly higher.
If you're selling and your agent is using the Zestimate as a pricing tool, you should find a different agent.
If you're refinancing and the bank is ordering an automated appraisal instead of a human one, you should understand that you might be leaving money on the table, especially if you've made improvements that don't show up in the public data.
The Zestimate is a tool. It's not a bad tool. But it's a rough tool, and treating it as precision is how people get hurt. The best protection against algorithmic error is understanding, knowing what the number represents, what it misses, and when to trust a human who's actually walked through the house instead.