In the spring of 2024, I sat at my kitchen table in Richmond, Virginia, refreshing the Zillow app for the fourth time that morning. My three-bedroom colonial in the Museum District had a Zestimate of $431,000. I'd purchased it in 2019 for $385,000, and I'd spent the better part of three years renovating — new kitchen, refinished hardwoods, a bathroom gut job that cost more than my first car.
I was preparing to list the house. My agent, Dana, had suggested $485,000 based on recent comps. But the Zestimate haunted me. What if buyers see that number and think she's overpriced? What if $431,000 is what the market actually says? I didn't realize it then, but I was about to learn something that would change how I think about property valuation entirely.
Here's what was actually happening inside Zillow's algorithm — and why it matters for anyone buying, selling, or refinancing a home. Zillow publishes its own accuracy data, and the numbers are more sobering than most homeowners realize. As of 2025, the nationwide median error rate for Zestimates on on-market homes is 2.4%. That sounds reasonable until you understand what it means: half of all estimates are off by more than 2.4%.
For off-market homes — the ones you're browsing on a Sunday night, the ones you're considering making an offer on — the median error jumps to 7.5% nationally. In some states, it's dramatically worse. In Nevada, the off-market median error has exceeded 14%. In rural markets across the Mountain West and Deep South, errors regularly hit 10-15% (Zillow, "Zestimate Accuracy," 2025).
On a $430,000 home, a 7.5% error means the Zestimate could be off by more than $32,000 in either direction. That's not a rounding error. That's a down payment on a rental property. That's three years of property taxes. That's the difference between an offer that gets accepted and one that doesn't.
When we listed at $485,000, something strange happened. Within 48 hours of the MLS listing going live, the Zestimate on my home jumped from $431,000 to $497,000 — a $66,000 swing based on nothing more than the fact that my house was now listed for sale. No renovation had been done. No new comps had closed. The algorithm simply absorbed the listing price and recalibrated.
We received three offers within a week. The highest was $505,000. But the buyer's lender ordered an appraisal, and it came back at $478,000. That's a $27,000 gap between the Zestimate and a licensed appraiser's opinion of value — on a home that was actively on the market. The deal nearly fell apart. We negotiated. We settled at $488,000.
What happened to my home illustrates a fundamental problem with Automated Valuation Models, or AVMs. A Zestimate is generated by a proprietary machine-learning algorithm that weighs five primary inputs: public property records (tax assessments, lot size, square footage, year built), prior and current transaction data, comparable sales from the MLS, local market trends including seasonal patterns, and any information the homeowner has manually updated on Zillow's platform.
What the algorithm cannot evaluate is essentially everything that actually determines a home's value to a human buyer: interior condition, quality of renovations, curb appeal, natural light, noise levels, school reputation at the neighborhood level, walkability to amenities, and the subjective "feel" of a property. A 2023 study published in the Journal of Housing Economics found that AVMs systematically underperform on properties with non-standard features — additions, lot premiums, view premiums, and custom finishes — because these attributes aren't captured in the structured data the models consume (Bokhari & Geltner, 2023).
Research from the Federal Housing Finance Agency confirms this pattern. Their 2022 analysis of AVM performance found that valuation accuracy degrades significantly for homes in the top and bottom quartiles of their local market — the unique, the renovated, the distressed. The models work best on tract homes in homogeneous subdivisions where every property is structurally similar. The more your home deviates from the statistical norm, the less reliable the estimate becomes (FHFA Working Paper 22-01).
Stories like this, plus the math behind real estate decisions. Weekly from Maren.
After closing, I went back and studied the full Zestimate history on my home. From 2019 to 2024, while I was actively renovating — adding real, appraisable value — the Zestimate tracked roughly $30,000 to $50,000 below what the comps supported. The algorithm had no way of knowing that the kitchen was no longer the original 1993 builder grade. It saw square footage and tax records. It didn't see the $42,000 I'd spent making the house worth more.
Then came the listing spike. Then the appraisal gap. Then the negotiated close. At every stage, the Zestimate was present but unreliable — a number that felt authoritative because it was displayed prominently on the most visited real estate website in America, but that was ultimately just a statistical guess with a known error rate that most users never investigate.
The deeper issue is one of anchoring bias. Behavioral economists have documented for decades that people's judgments are disproportionately influenced by the first number they encounter — even when that number is arbitrary (Tversky & Kahneman, 1974). When a buyer sees a Zestimate of $431,000, that number becomes their mental anchor. A listing at $485,000 feels inflated. A seller who sees $497,000 after listing feels validated. Neither response is based on a reliable valuation — both are based on an algorithm's output that Zillow itself acknowledges has a median error that can exceed 10%.
A 2021 study from the Real Estate Economics journal found that homes with Zestimates below their listing price took an average of 12% longer to sell than comparable homes where the Zestimate was at or above the list price — even when the listing price was justified by comps (Bucchianeri & Minson, 2021). The Zestimate wasn't reflecting the market. It was shaping it.
I think about this now every time someone sends me a Zillow link and says, "Look what that house is worth." They're not looking at what it's worth. They're looking at what a machine-learning model — trained on incomplete data, blind to interior conditions, and known to err by 7 to 15 percent — estimates it might sell for. There's a difference. And on a $400,000 transaction, that difference can be the price of a new car.
My home sold for $488,000. The final Zestimate, updated after closing, settled at $486,200. Close enough to look credible. Wrong enough to have cost me leverage at every stage of the process.
If you're using a Zestimate — or any AVM, including Redfin's Estimate, Realtor.com's valuation, or Chase's home value tool — here's what the research says about how to calibrate your expectations. These tools are starting points, not conclusions. They work reasonably well for cookie-cutter homes in data-rich markets. They fail quietly and expensively for everything else.
Before listing your home, get a Comparative Market Analysis from a local agent who has physically walked through properties in your neighborhood. Before making an offer, understand that the Zestimate may be anchored to data that's months or years old. Before refinancing, know that your lender will order an appraisal regardless of what Zillow says — and the two numbers may not be in the same zip code. The FHFA's own research recommends treating any AVM estimate as having a confidence range of ±10% for off-market properties. On a $400,000 home, that's a $40,000 window. Plan accordingly.