How home valuation tools work

How AVMs work, and where they break

9 min read

What an automated valuation model is, why it looks accurate on listed homes, and the places it quietly fails.

The short version

An AVM, an automated valuation model, is software that prices a home by running comparable sales and property data through a statistical model. It is the engine behind the Zillow Zestimate, the Redfin Estimate, and the home-value tools at Chase, Bank of America, and most lender sites. AVMs are fast, free, and reasonable where data is dense. They break where data is thin, or where the number depends on something the model cannot see. The gap between those two cases is large, and most homeowners check their value in exactly the case where the AVM is least reliable.

How an AVM builds a number

The inputs are public records and listing data. Deeds, tax assessments, and building permits establish the property's characteristics; recent comparable sales from public records and the MLS establish the local price level. A machine-learning model weighs those comparables and adjusts for differences in square footage, lot size, age, and location to produce a single estimate. Zillow's Zestimate runs this across more than 104 million homes using neural networks and ensemble methods; the Redfin Estimate uses more than 500 data points and live MLS data, retrained on actual closing prices. The methods differ in detail. The shape is the same: comparable sales, processed statistically, returned as one number.

Why the on-market number looks so good

Both companies publish their accuracy, and the headline figures are strong. Zillow reports a national median error near 2% for homes that are currently listed, and Redfin reports about the same. The catch is in why. Once a home is listed, the AVM gains the listing itself, fresh MLS data, professional photos, and the asking price, and the estimate drifts toward that asking price. As appraiser Ryan Lundquist has pointed out, the accuracy is measured against the most recent estimate before sale, not the one from six months earlier. The on-market number is impressive partly because, by then, the model is reading the agent's pricing back to you.

The number you check is the weak one

For off-market homes, the ones not currently for sale, the picture changes sharply. Zillow's national median error for off-market homes is about 7.5%, and Redfin's is roughly 6% to 7%. Median means half of all estimates miss by even more than that. On a $400,000 home, a 7.5% miss is about $30,000; on a $600,000 home, about $45,000. The homeowner casually checking their value is looking at the least accurate version of the number, the one with the fewest fresh inputs, while it reads as a precise figure down to the dollar.

At a glance, median error by listing status:

Tool	On-market	Off-market
Zillow Zestimate	about 2%	about 7.5%
Redfin Estimate	about 2%	about 6% to 7%

Where AVMs break

The failures cluster in predictable places:

Condition. The model cannot see inside. A gut renovation and a deferred-maintenance shell can carry the same public record and the same estimate.
Off-market homes. The number with the fewest inputs and the widest error, and the one most homeowners are looking at.
Thin-data markets. Rural and low-volume areas starve the comparables. Off-market median error runs near 13% in states like Vermont and Maine.
Luxury. Above roughly $2 million, error rates jump to 10% to 20%, per Cotality's 2026 housing data, because comparable sales are scarce and each home is more unique.
Unreported work. Renovations and additions that never reached the public record never reach the model.
New construction and unique homes. Too few comparables to anchor to.
Fast-moving markets. AVMs lag, because they are built on sales that already closed.

The cost of trusting the number

The most expensive AVM failure on record is Zillow's own. Its home-buying arm, which bought houses based on its estimates, lost roughly $880 million and shut down in 2021, laying off about a quarter of the company, because the model paid real money for what it could not see. The most quoted failure is smaller and sharper: Zillow's former chief executive sold his own home for nearly 40% below its Zestimate. If the number can miss that badly for the people who built it, the lesson for everyone else is to treat it as a starting point, not a verdict. Bank estimators are no different. Drawing on the same public data, they land in the same 7% off-market range.

What this means for an agent

The AVM is not the enemy and it is not going away. It is a fast, free first read with a known blind spot. That blind spot, condition, micro-location, demand, the renovation the record never captured, is Context Blindness, and it is exactly the territory where an agent's read carries information the model does not have. The agents who handle the Zestimate well bring it up themselves, explain how it was built and where it is reliable, and show where their own analysis sees more. A reasoned valuation that shows its comparables and an honest range, like the one on our home-value page, is built around the same idea: state the number, then state what the number cannot see. For the consumer-side explainer, see why the Zestimate is wrong on your house, and for how the AVM sits against a CMA and an appraisal, see AVM, CMA, or reasoned valuation.

Frequently asked questions

What is an AVM?

An AVM, or automated valuation model, is software that estimates a home's value by running comparable sales and property data through a statistical model. It powers the Zillow Zestimate, the Redfin Estimate, and most bank home-value tools. AVMs are fast and free, but they cannot inspect a property and are far less accurate for off-market homes.

How accurate is the Zestimate or Redfin Estimate?

For homes currently listed, Zillow and Redfin both report a median error near 2%. For off-market homes, Zillow's median error is about 7.5% and Redfin's is roughly 6% to 7%. Median means half of all estimates miss by even more. On a $600,000 home, 7.5% is about $45,000.

Why are AVMs more accurate for listed homes?

Once a home is listed, the model gains the listing, fresh MLS data, and the asking price, and the estimate drifts toward that asking price. The accuracy is measured just before sale, so the on-market number partly reflects the agent's pricing rather than an independent estimate.

Where do AVMs fail most?

On condition they cannot see, on off-market homes, in thin-data rural markets, and on luxury homes above about $2 million, where error rates reach 10% to 20%. They also miss unreported renovations and lag in fast-moving markets, because they are built on sales that already closed.

Should I trust an AVM to price my home?

Use it as a starting point, not a final price. Zillow's own former chief executive sold his home for nearly 40% below its Zestimate, and Zillow's home-buying arm lost about $880 million pricing homes by algorithm. An AVM cannot see condition, demand, or micro-location, which is where a local agent's analysis adds the most.

Sources: Zillow and Redfin published accuracy disclosures (on-market and off-market median error rates), appraiser Ryan Lundquist on how on-market accuracy is measured, Cotality 2026 housing data on AVM error at luxury price points, and public reporting on Zillow's home-buying losses and its former chief executive's home sale. Error rates are self-reported medians and vary widely by market and property type.

This article is general information and analysis, not financial, lending, or appraisal advice. Verify any home value with a licensed professional before acting.