An automated valuation model is software that estimates a property's value using public records, tax assessments, recent comparable sales, property characteristics, and machine learning trained on millions of transactions. The output looks like a single clean number, but under the hood it is a probability distribution with a confidence score.

How accurate is Zillow's Zestimate?

As of 2025, Zillow's published nationwide median error is roughly 1.8–2.4% for on-market homes and 7.0–7.2% for off-market homes. On a $500,000 off-market home that is a typical miss of about $35,000–$36,000, and because it is a median, half of all estimates are off by more.

Why did Zillow's iBuying program fail?

Zillow shut down Zillow Offers in November 2021 with estimated losses of more than $880 million. The algorithm did not break mathematically, it failed from concept drift as the 2021 market shifted, seller adverse selection that concentrated its portfolio in mispriced homes, unmodeled features like condition and neighborhood feel, and operational realities like renovation costs.

Should I rely on an AVM when buying property abroad?

Use AVMs for direction, not decisions. AVMs perform worst exactly where cross-border investors put their money, unique and luxury assets, thin markets, and emerging markets with fragmented registries. Trust the AVM to filter listings, then pay for a human appraiser on anything over roughly $250,000.

Will AI Kill Real Estate Valuation? AVMs, Zestimate, and the $880M Lesson Every Investor Should Know

AI-powered valuation tools promise instant property appraisals. But Zestimate's 7% off-market error rate and Zillow's $880M iBuying collapse reveal the limits. Here's what cross-border investors need to know in 2026.

Published on: April 23, 2026

Quick answer: AI is not killing real estate valuation, it is bifurcating it. Automated valuation models now cover more than 116 million U.S. homes and are genuinely useful for screening, portfolio monitoring, and loan origination at scale, but Zillow's Zestimate still carries a median error of roughly 1.8–2.4% on-market and 7.0–7.2% off-market, and Zillow lost more than $880 million when it bet its iBuying program on its own model. For unique, luxury, thin, and cross-border assets, most of what serious international investors buy, a human appraiser remains irreplaceable. The winning approach is hybrid: AI for speed and screening, human expertise for the deciding input.

Ask any proptech founder in 2026 and you'll hear the same pitch: artificial intelligence is about to do to real estate appraisal what Uber did to taxi medallions. Algorithms process millions of data points in seconds. They don't miss comparables. They don't get tired, biased, or paid by the deal. And crucially, they cost a fraction of a licensed appraiser.

The numbers sound convincing. Industry data suggests firms using machine learning in commercial real estate have lifted net operating income by up to 10%, and by mid-2025 more than 90% of commercial real estate companies had launched AI pilots. Residential AVMs, automated valuation models, now cover more than 116 million U.S. homes, generating values in milliseconds.

So the question is no longer whether AI can value property. It can. The question is whether it should be the primary input for a six- or seven-figure investment decision. And the answer, if you look past the marketing, is more interesting than either the boosters or the skeptics want to admit.

What an AVM actually is (and what it isn't)

An automated valuation model is a piece of software that estimates a property's value using a combination of public records, tax assessments, recent comparable sales, property characteristics, and, increasingly, machine learning models trained on millions of transactions.

Most modern AVMs stack two or more techniques. A hedonic model assigns values to individual property attributes (an extra bedroom is worth X, proximity to a top-ranked school is worth Y). A comparable-sales model operates the way an appraiser would: it finds similar recently sold properties nearby, adjusts for differences, and triangulates a value. More sophisticated systems layer in a repeat-sales index, tracking how specific properties have appreciated over time, and weight each method by a confidence score.

The output looks like a single clean number. Under the hood it's a probability distribution with a standard deviation, a confidence score, and a quiet assumption: that the data it trained on still describes the market today.

This is where things get interesting.

The Zestimate: useful, but don't confuse it with an appraisal

Zillow's Zestimate is the most publicly tracked AVM in the world, which makes it an unusually honest case study. The company publishes its own accuracy data, and the picture is more nuanced than the headline suggests.

As of 2025, Zillow's published nationwide median error rate is roughly 1.8–2.4% for on-market homes and 7.0–7.2% for off-market homes. Here's what that actually means in dollar terms:

$500,000 home, on-market: typical miss of ~$9,000–