An automated valuation model estimates what a property would sell for today without anyone visiting it. Lenders use AVMs to screen and monitor collateral, insurers to size coverage, investors to mark portfolios and consumer sites to show a number next to a photo. Appraisals still exist because regulation requires them for many decisions, but the AVM is what runs at scale.
What the model learns from
The training signal is recorded sales: the price on the deed, the date it recorded and the property it describes. Around that sits everything the model can know about the property and its market. Assessor characteristics give it size, age and type. Prior sales give it the property's own history. Nearby recent sales give it the neighborhood's level and trend. Active listings give it the current asking prices and how long they have sat. Permits tell it something changed, like a new roof or an addition, that the assessor roll will not reflect for a year.
Modern AVMs are ensembles of tree-based models, sometimes with a hedonic regression underneath for interpretability. The specific algorithm matters less than the features and the freshness of the sales the model is trained on. A model retrained weekly on last week's recordings will track a moving market; a model retrained quarterly will lag it.
Why accuracy varies by market
Every provider's error is lowest in dense suburban markets with homogeneous housing stock and frequent sales, and highest in rural areas, in neighborhoods with unusual properties and in markets that are moving fast. This is not a flaw in any one model. It is the shape of the problem: the model is interpolating from comparable sales, and where there are few comparables, or the property is unlike its neighbors, the interpolation is weak.
A national accuracy figure therefore tells you little. Ask for the error distribution by metro, and for the share of properties the model declines to value or values with low confidence. A provider who quotes one number for the country is either hiding the tail or has not measured it.
Reading the confidence score
A good AVM returns three things: the estimate, a range and a confidence score. The range is the interval within which the model expects the actual sale to fall. The confidence score summarizes how much the model trusts itself, and it should be driven by observable things: how many comparable sales were found, how recent and how similar they are, how much the ensemble's members agreed, and how typical the property is for its area.
Use the confidence score as a routing rule. High confidence can go straight into an automated decision. Low confidence should trigger a broader comparables search, a desk review or an appraisal. Treating every estimate as equally reliable is how AVMs get blamed for errors they flagged.
The comparable sales are the evidence
An estimate that arrives without its comparables is an opinion. An estimate that arrives with the five sales it rests on, their distances, their dates and the adjustments made for size and age is an argument a human can check. This is why we return the comps with every value and also offer the comparable sales on their own. Some teams want to apply their own adjustment logic, and some want to show the evidence to a customer.
How to run a fair test
Take a set of properties whose sale prices you know and the model does not. The cleanest way is to score sales that recorded after the model's training cutoff, which is how we validate internally. Send the addresses without the prices, receive the estimates, and compute the median absolute error and the share within ten percent, by metro. Compare providers on the same set. Then look at the low-confidence tail separately, because how a model behaves when it is unsure matters more than its median.
Ask the provider for the same test on their side and compare answers. A provider who will not run a blind test on your portfolio is telling you something.
What an AVM is not
It is not an appraisal, it is not a listing price recommendation and it is not a commercial valuation. Commercial property is valued on income, and public record rarely carries the rent roll, so a commercial "AVM" is usually an assessor value dressed up. It is also not a substitute for knowing the condition of the property, which is why joining the estimate to permit history improves both the number and the story behind it.
Questions
How accurate are AVMs?
It depends on the market. Dense suburban markets with frequent sales produce low error; rural and unusual properties produce high error. Ask for the error by metro and for the low-confidence share, not a national average.
Can an AVM replace an appraisal?
For screening, monitoring, marketing and consumer display, yes. For decisions where regulation requires an appraisal, no. Most lenders use both, routed by confidence.
What does an AVM cost?
A lookup costs a small fraction of an appraisal and returns instantly. Bulk valuations for a portfolio are priced per property or as a flat file.