What an AVM actually does
An AVM ingests two streams: the target property's attributes (address, beds, baths, square footage, lot size, year built) and a comparable-sales pool drawn from the surrounding geography and recent time window. It then runs one or several regressions to produce a single point estimate plus a confidence band.
The four regression families you'll see referenced in vendor methodology docs:
- Linear regression on comps. Fits a line through recent sale prices vs property attributes. Cheap, transparent, accurate in homogeneous tract subdivisions; falls apart in mixed-stock neighborhoods.
- Hedonic regression. Prices each feature independently — an extra bath is worth $X, an extra 1,000 ft² is worth $Y. Better at atypical homes than plain linear.
- Gradient-Boosted Trees (XGBoost / LightGBM). Non-linear ensembles that capture interactions (a pool is worth more in Miami than Buffalo). State of the art for Zestimate, HouseCanary, Quantarium, CoreLogic.
- Vision-based condition adjustment. New since 2024. The model looks at listing photos and adjusts down for "fair" or "poor" condition. Twellie's stack uses Claude + Gemini for this layer.
Each architecture has its own failure mode. The honest vendors publish their median absolute error (MdAE) — the median of the absolute percentage error across a benchmark set of recent sales.
Why AVMs are not appraisals
A formal appraisal is a USPAP-compliant report from a licensed human appraiser who physically visits the property, applies the three USPAP approaches (sales-comparison, cost, income), and produces a defensible 30–50 page document. AVMs are statistical, not human-judgment, which is why a lender funding a mortgage requires the appraisal — not the AVM — for the loan file.
The two are complementary, not interchangeable. An AVM is fast, cheap, and good enough for screening properties before you visit. An appraisal is slow, costs $400–$700, and is the legal benchmark for the loan and any tax-appeal or estate settlement.
For the long version of this comparison see AVM vs Appraisal vs Zestimate.
Typical AVM accuracy
| AVM | Median Absolute Error | Tier |
|---|---|---|
| Quantarium / HouseCanary | 3–5% | Lender-grade |
| CoreLogic | 4–6% | Lender-grade |
| Redfin Estimate (off-market) | ~6.7% | Free consumer |
| Zillow Zestimate (off-market) | ~7.5% | Free consumer |
| Twellie | ~7% | Paid consumer + photo + report |
| Formal appraisal | 1–3% | Legal benchmark |
Two things to remember when reading these numbers:
- Active-listing accuracy is much better than off-market. Zillow's headline 1.9% MdAE is for active listings, where the model has the listing photos and recent comp data. Off-market — a home that hasn't sold in 8 years — is the harder problem and the number that matters when you're shopping.
- National averages hide huge variance. Every AVM is dramatically less accurate in markets with thin sale volume, atypical homes, or sparse public-records data. Rural Vermont is harder than tract Phoenix, and the confidence band reflects that.
When an AVM is the right tool
- Screening listings before you visit — is this in budget?
- Sanity-checking a list price against the comparable set
- Tax-appeal preparation, paired with a formal appraisal
- Pre-offer pricing, paired with the confidence band
When it isn't
- Loan funding. Lender requires a USPAP appraisal.
- Legal proceedings (divorce, estate, eminent-domain). Need an appraisal.
- Atypical property. A custom log home on 40 acres has no comps the AVM trusts. The confidence band will be wide; treat the number as a placeholder, not a value.
- Stale public records. If the county assessment is 8 years old and the home has been renovated since, every AVM trained on that record will inherit the staleness.
The right read on any AVM number is "point estimate plus a band". A $487,500 AVM with a $40k confidence band is telling you the real market value lives somewhere in the $447k–$527k window. That window is your negotiating range — not the headline number.