AVM vs Appraisal vs Zestimate: Which Home Value Should You Trust in 2026?

What an AVM actually computes

An AVM is a regression model. It takes a target property's attributes (beds, baths, square footage, lot size, year built, location) and finds the N nearest comparable sales — typically 50–200 — within a geographic and time-window radius. It then runs one or several of these regressions:

1. Linear regression on comparables. The simplest AVM. It fits a line through recent sale prices vs property attributes, then plugs in the target's attributes. Accurate when the comparable set is homogeneous (a tract subdivision); inaccurate in mixed-stock neighbourhoods.
2. Hedonic regression. A property-attribute model that prices each feature independently — an extra bath is worth $X, an extra 1000 ft² is worth $Y. Better at extrapolating to atypical homes.
3. Gradient-Boosted Trees (XGBoost / LightGBM). A non-linear ensemble that can capture interaction effects (a pool is worth more in Miami than in Buffalo). State-of-the-art for most modern AVMs including Zestimate's recent revisions.
4. Vision-based condition adjustment. New in 2024–2026. The model inspects listing photos and adjusts the value down for "fair" or "poor" condition. Twellie uses Claude-Sonnet-4 + Gemini-2.5-Flash for this layer; CoreLogic and Quantarium use proprietary vision models.

The output is a single number plus a confidence interval ("$487k ± $24k, 80% confidence"). The confidence interval is the part most people ignore, and it's the part that matters most.

What a formal appraisal computes

A formal appraisal is a USPAP-compliant report from a licensed appraiser. USPAP (Uniform Standards of Professional Appraisal Practice) requires the appraiser to consider three approaches:

1. Sales-comparison approach. Same idea as an AVM, but with 3–6 hand-picked comparables and explicit, line-by-line adjustments ("subject has 1 fewer bath: -$8,000; subject has updated kitchen: +$15,000"). This is the dominant approach for residential.
2. Cost approach. What would it cost to rebuild this house from scratch (land + improvements – depreciation)? Only used as a sanity check on residential, primary on new construction.
3. Income approach. What is the property worth based on the rent it generates? Primary on rental properties, ignored for owner-occupied.

The appraiser physically visits the property — measures rooms, inspects condition, photographs every room. They produce a 30–50 page report ($400–$700, 5–10 days turnaround) that the lender uses to decide whether to fund the loan.

Accuracy benchmarks

Median Absolute Error (MdAE) is the standard metric — the median of the absolute percentage error across a benchmark set of recent sales. Lower is better.

Two warnings:

1. Active-listing accuracy is much better than off-market accuracy for free AVMs. Zillow's reported 1.9% MdAE is for active listings — they have the listing photos, description, and recent comp data right in front of them. Off-market is the harder problem and the number that matters.
2. National averages hide huge variance. All AVMs are dramatically less accurate in markets with thin sale volume, atypical homes, or sparse public records. Rural Vermont is harder than tract Phoenix.

When to trust which

The mistake people make: treating Zestimate as ground truth for an offer. Zestimate is good enough for a starting point, but a +/- $30k confidence band on a $400k home means the difference between a winning offer and a losing one. That's where a paid AVM with the deeper analysis (Twellie, HouseCanary) earns its fee.

Why Twellie's report is different from a single AVM number

The Zestimate-style number tells you what the value might be. A report tells you why and what to do about it:

• Comparable sales with line-item adjustments, so you can verify the math instead of trusting a black box.
• Photo condition grading that adjusts for the listing's actual state (Zestimate doesn't see photos for off-market).
• Confidence band, prominently shown, so you know how much wiggle room exists in the offer.
• Cost-of-ownership model that puts the price in context against taxes, insurance, maintenance, and 5-year equity.
• Negotiation strategy that translates the number into "offer $X, expect counter at $Y, walk-away at $Z".

If you only need a quick number, Zestimate is free and good enough. If you're about to write a $400,000 offer, a $50 deeper analysis is the cheapest insurance you can buy. The full Twellie methodology and benchmark data is at /methodology, and you can browse a complete sample report at /mockup/report before you spend a dollar.

Common pitfalls

1. Reading one number, ignoring the band. A $487k AVM with a $40k confidence band tells you the real market range is $447k–$527k. Always read the band.
2. Comparing different AVMs as if they agree. Zestimate at $487k and Redfin at $501k aren't disagreeing by $14k — the 95% CI on each is probably ±$30k. They're agreeing within noise.
3. Ignoring photo condition. A 2400 ft² home with deferred maintenance is not worth what a 2400 ft² fully-renovated home is worth. Free AVMs miss this entirely.
4. Trusting the AVM when public-records data is stale. If the county hasn't updated the assessment in 8 years, every AVM trained on that data inherits the staleness.

What to do next

Pull a Twellie report on the next address you're serious about — the report explicitly shows you the AVM number, the comparable adjustments, the photo grades, and the recommended offer range. Then read the methodology page (/methodology) to verify the math is honest.