What an AVM Actually Is — The Engine Under the Hood

An automated valuation model is a program that estimates property value by finding statistical patterns in historical sales data. The foundational logic is hedonic pricing: if a home sold for $500,000 with four bedrooms, two baths, a two-car garage, and 2,000 square feet in zip code 78701, then a similar home in that zip code should sell for a similar price, adjusted for measurable differences.

Early AVMs, deployed by mortgage servicers in the early 2000s, relied on linear regression — straightforward equations that weighted each characteristic additively. Modern AVMs are fundamentally different. Zillow’s Neural Zestimate, introduced in 2021, uses deep learning: the model does not simply compute a weighted average of comparable sales but learns complex, nonlinear relationships across hundreds of input variables. Gradient boosting and random forest models are also common in institutional lender tools. A 2025 systematic review found tree-based models to be the most frequently deployed machine-learning approaches across real estate pricing platforms globally.

The inputs vary by provider, but most models draw on some combination of the following:

• Public records. Tax assessments, deed records, permit history, and prior sale prices going back decades.
• MLS transaction data. Recent sale prices, list prices, days on market, and price-cut ratios in the subject property’s neighborhood.
• Structural characteristics. Square footage, bedroom and bathroom count, garage, lot size, year built, and property type.
• Neighborhood signals. Proximity to schools, parks, transit, and commercial activity; walkability scores; flood zone designations.
• Aerial and street-level imagery. Some advanced models infer exterior condition, roof age, and landscaping quality from satellite data and street view photos.

What most consumer-facing AVMs cannot see is the interior. Unless a trained property data collector has physically visited — as in Fannie Mae’s “Value Acceptance + Property Data” program — the model has no information about the kitchen renovation completed last year, the deferred HVAC replacement, the cracked foundation, or the fresh paint applied to mask a water-damaged ceiling. Condition is one of the primary drivers of actual market price, and it is essentially invisible to every consumer AVM.

The Major Players — Who Builds These Models and Why

The AVM landscape splits into two tiers: consumer-facing estimates, which prioritize coverage and freshness, and institutional models, which prioritize accuracy within a narrower confidence interval for financial decision-making.

Zillow

The Zestimate is the most-recognized AVM in the country. Zillow publishes its own accuracy benchmarks — unusual for the industry and worth noting as a transparency standard — reporting an on-market median error of 1.94% and an off-market median error of 7.06% as of Q1 2026. The Neural Zestimate improved off-market coverage meaningfully, but the fundamental challenge remains: off-market homes generate far less fresh transaction signal than listed ones.

Redfin

The Redfin Estimate integrates live MLS feeds more aggressively than Zillow, updating within hours of price changes and new listing data. This gives it faster reaction time in active markets. Published accuracy: 1.98% on-market, 7.72% off-market (Redfin, 2025). Marginally less accurate than the Zestimate on average, though the gap narrows in markets with deep, current MLS data.

Cotality (formerly CoreLogic)

The institutional heavyweight. CoreLogic rebranded to Cotality in March 2025. The company maintains a database of more than 5.5 billion records covering 99.9% of U.S. properties, updated daily, spanning more than 50 years of transaction history. Cotality’s AVM is not consumer-facing; it is licensed to mortgage lenders, servicers, insurance companies, and the government-sponsored enterprises. Because it is calibrated for financial decision-making, the model typically carries tighter confidence intervals than consumer tools — but Cotality does not publish median error rates.

Opendoor

The iBuyer’s internal pricing engine is effectively a purpose-built AVM optimized to minimize the risk of buying homes Opendoor cannot resell at a profit. It factors in estimated repair costs, local demand signals, and resale timelines — variables a pure price-estimating AVM typically ignores. Opendoor’s experience during 2021–2022, when rapid Federal Reserve rate increases erased margin assumptions across its portfolio, demonstrated how even sophisticated proprietary models can fail during structural market shifts that no historical dataset had captured.

Where AVMs Fail — The Systematic Blind Spots

The published error rates are averages. Underneath those averages are categories of property where AVM accuracy degrades substantially — and some of these categories are common enough to affect a significant share of buyers and sellers.

Off-market homes

The 7% off-market error is the single most important number for homeowners to internalize. The reason is structural: when a home hits the market, live transaction signals flood in — buyer interest metrics, days on market, price reductions — and the AVM updates toward the market price. Before that listing, the model works from historical comparables and structural characteristics only. In markets with low transaction velocity, those comparables can be years old.

Rural and thin-market properties

AVMs are trained on transaction density. Areas with few recent sales provide limited comparable data, and the model’s confidence intervals widen accordingly. The pronounced regional price divergence visible in the current national market — with some Midwest markets outperforming and Sun Belt markets correcting — creates genuine challenges for models trained heavily on pre-2022 data. The algorithm may apply price assumptions appropriate to a 2021 suburban landscape that no longer exist.

Unique and complex properties

A standard three-bedroom suburban colonial maps cleanly to comparables. A 1920s craftsman with an in-law suite, a barn, and a non-conforming lot is harder. Condos present particular difficulties: HOA financial health, special assessments, building reserves, and unit-specific floor premiums are largely invisible to most AVMs. Commercial-use zoning overlays, multi-family structures, historic designations, and properties with significant acreage add layers that most consumer AVMs do not handle well.

Model failure during structural shifts

The 2022 Federal Reserve rate cycle exposed the core vulnerability of historically trained models: they cannot anticipate regime changes. AVMs trained heavily on 2020–2021 data — a period of record-low rates and near-universal price appreciation — overestimated values by an estimated 5–8% during the 2022 correction period, according to industry observers. A model that learned “prices rise when rates are low” had no template for a 300-basis-point rate shock arriving in nine months.

The Racial Bias Problem — An Unresolved Technical and Civil Rights Issue

Research from the Urban Institute has documented a consistent pattern in AVM performance: models produce greater percentage error in majority-Black neighborhoods than in majority-white neighborhoods, even when controlling for structural property characteristics. The Urban Institute is careful to distinguish between two different problems — systematic undervaluation bias and elevated error magnitude — and finds stronger evidence for the latter. But greater error has practical consequences either way: valuations that run high can harm sellers who rely on them to price accurately, while estimates that run low affect refinancing eligibility and home equity calculations.

The root cause is data quality, which in turn reflects historical disinvestment. Majority-Black neighborhoods historically faced redlining, lending discrimination, and lower transaction velocity, all of which produce thinner, less reliable training data for machine learning models. When the algorithm has fewer high-quality comparables to learn from, its predictions carry greater uncertainty — and that uncertainty compounds over generations.

A 2024 year-end review by Veros, one of the major AVM providers, found that this error disparity persists even as models grow more sophisticated. The finding underscores a core limitation of data-driven systems: a model trained on historically biased data will reproduce patterns embedded in that data, absent explicit corrective measures. The new federal AVM quality control rule, discussed below, makes nondiscrimination compliance a legally mandated requirement — not a soft goal — for any AVM used in a mortgage credit decision.

The Regulatory Turn — Washington Moves to Govern AI Appraisals

For much of its history, the AVM industry operated without sector-specific federal oversight. That changed in July 2024, when six federal agencies — the OCC, Federal Reserve Board, FDIC, NCUA, CFPB, and FHFA — jointly issued a final rule on quality control standards for automated valuation models. Mandatory compliance took effect October 1, 2025, making this the first substantive federal regulation directly governing AI-based property valuation in the mortgage market. The same agencies had already moved to govern AI-driven loan processing and underwriting; the AVM rule extends that framework to collateral valuation.

The rule applies to mortgage originators and secondary market issuers that use AVMs in credit decisions or securitization determinations on loans secured by a consumer’s principal dwelling. Covered institutions must implement policies ensuring their AVMs meet five quality control factors:

• High confidence in estimates. Models must meet defined accuracy thresholds with ongoing validation against actual sale prices.
• Data integrity. Systems must protect against manipulation of input data, including fraudulent comparables inserted to inflate or depress valuations.
• No conflicts of interest. Institutions cannot use AVM providers in which they hold a financial stake that could compromise the output.
• Random sample testing and review. AVM outputs must be tested against actual sale prices through random sampling, with documented remediation when accuracy falls short.
• Nondiscrimination compliance. AVMs must comply with ECOA, the Fair Housing Act, and other applicable nondiscrimination laws, with explicit bias testing required.

The fifth requirement is the most technically demanding. Testing an AVM for fair lending compliance requires running the model across protected class groups, comparing error distributions by neighborhood demographic composition, and documenting what happens when disparities are found. For companies that have historically treated their models as proprietary black boxes, this is a material operational change.

One important scope note: the rule does not apply to consumer-facing tools like the Zestimate or Redfin Estimate, which are used for informational purposes rather than credit decisions. It directly governs the institutional AVMs embedded in lender systems and in Fannie Mae’s and Freddie Mac’s collateral programs — the models that actually determine whether loans are made.

Value Acceptance — When the Algorithm Replaces the Human Appraiser

The place where AI home valuations carry the most direct financial consequence for borrowers is not in the consumer apps but in Fannie Mae’s and Freddie Mac’s collateral programs, which determine whether a purchase transaction requires a full human appraisal at all.

Fannie Mae retired the term “appraisal waiver” on September 3, 2025, replacing it with “value acceptance.” The terminology shift reflects a broader repositioning: these are not simply waivers of appraisal requirements but a positive determination that the agency’s proprietary AVM is sufficiently confident in the property’s value and the transaction’s risk profile to proceed without independent human review.

In Q1 2025, Fannie Mae expanded eligibility substantially. For purchase loans on primary residences and second homes, the maximum loan-to-value ratio eligible for value acceptance increased from 80% to 90%. For the higher-tier “Value Acceptance + Property Data” program — which adds an interior inspection by a trained data collector — the LTV ceiling moved from 80% to 97%. The practical effect is that many more purchase transactions now qualify for the algorithmic shortcut, including higher-leverage transactions that were previously excluded.

The appeal is straightforward. Traditional appraisals cost $400 to $700, take one to three weeks, and can become the bottleneck in a time-sensitive closing. In competitive markets where the gap between offer acceptance and closing matters, eliminating the appraisal contingency has real value. For borrowers with strong loan profiles on standard properties in data-rich markets, value acceptance can legitimately accelerate a transaction without meaningful additional risk.

The tradeoff is a reduction in independent oversight. A traditional appraisal is partly designed to protect the lender, but it also gives the buyer an independent third-party assessment of what the property is worth relative to what they are paying. In a value acceptance transaction, that check simply does not occur. Borrowers who pay above an appraised value in a traditional transaction make a visible, explicit choice with that information in hand. In a value acceptance transaction, no comparable data point is generated — the AVM has implicitly certified the price without anyone physically inspecting the property. The rate environment of 2026, with purchase prices remaining elevated in many markets, makes this distinction particularly relevant for buyers who are already stretching.

For context, actual usage of value acceptance remains modest as a share of all purchase originations — estimated at roughly 1 to 2 percent of loans in early 2025. Lenders continue to conduct full appraisals on the majority of loans, particularly jumbo, non-conforming, and complex property transactions that fall outside agency guidelines.

What AI Valuations Mean for Buyers and Sellers Right Now

The practical implications differ depending on which side of a transaction you are on.

If you are a seller

The Zestimate on your home is a starting point, not a price. It is almost certainly based on exterior characteristics and historical comparables only. It does not know about the improvements you have made, the systems that need replacement, or the micro-market dynamics that an active agent working your specific neighborhood would recognize immediately. Use it as a rough orientation, then request a comparative market analysis from a licensed professional who has physically walked the property. The real market’s verdict comes from buyers who see the home in person — the AVM is a first draft of that verdict, not the final one. Understanding how to price and prepare your home effectively in today’s market requires more than any algorithm can currently provide.

One specific risk for sellers: if your Zestimate is significantly higher than what comparable sold properties would support, you may set your list price above market and spend weeks learning that the AVM was wrong — at the cost of days-on-market stigma that depresses your eventual sale price.

If you are a buyer

When you encounter a Zestimate or Redfin Estimate on a home you are considering, pay attention to the surrounding context. How many recent comparable sales exist in the immediate area? Is the local market fast-moving or stagnant? Does the AVM have access to data from the past 30 to 60 days, or are the nearest comparables from 12 to 18 months ago? In a market with deep, current transaction data, the estimate is more reliable. In a thin or unusual market, it should carry less weight in your offer calculus. Tracking national and regional market analysis alongside any individual AVM estimate helps calibrate what the algorithm is — and is not — capturing about local conditions. Resources on buyer strategy and offer negotiation are also worth reviewing before using any AVM as a basis for bid decisions.

If you are refinancing

Check whether your lender intends to use an AVM-based evaluation rather than a full appraisal. If your property has significant improvements or unique features that do not appear in public records — a finished basement, a remodeled kitchen, an addition — the algorithmic estimate may understate your equity. A full appraisal costs a few hundred dollars and could meaningfully change your loan-to-value ratio, affecting your rate, your PMI status, or your eligibility for a cash-out refinance. The cost is frequently worth it if there is meaningful uncertainty in the AVM estimate.

The Next Frontier — LLMs, Computer Vision, and Narrative Appraisals

The AVM space is evolving quickly enough that the tools available at the end of 2026 will look meaningfully different from those of 2024, and the direction is toward both greater accuracy and greater transparency.

Several companies are integrating large language models with property and spatial data to produce what might be called narrative valuations — AI-generated reports that describe not just a number but the reasoning behind it, citing specific comparable sales, neighborhood characteristics, and risk factors in natural language. Academic research published in 2025 explored incorporating LLM-extracted features from property listing descriptions into valuation models, finding improvements in accuracy on data-sparse properties where structured inputs are thin.

Computer vision is advancing in parallel. Models that assess roof age and exterior condition from aerial imagery are commercially deployed today. Interior condition assessment from listing photographs is an active research area, with companies working on systems that can identify indicators of deferred maintenance, dated finishes, or above-average renovation quality from the photos that sellers upload to the MLS. If interior condition data becomes computable from photographs at scale, it would close the largest remaining gap in AVM accuracy.

The federal regulatory framework also creates pressure toward explainability. A model that must demonstrate nondiscrimination compliance to regulators cannot remain a pure black box indefinitely. The combination of regulatory requirements and competitive differentiation is pushing the industry toward AVM outputs that describe their confidence level, identify which comparable sales drove the estimate, and flag when the property’s characteristics fall outside the model’s reliable range.

None of these advances eliminate the fundamental limitation: a trained professional who has examined a property in person, understands its neighborhood context, and can exercise judgment about variables that have no clean data representation will remain valuable for complex valuation questions. But the gap between what the algorithm knows and what the appraiser knows is narrowing, and the pace of that narrowing is accelerating.

The Bottom Line — A Decision Framework for Using AI Valuations

The algorithms powering home valuations in 2026 are better than they have ever been, and they will keep improving. But “better than 2019” and “accurate enough to replace professional judgment” are very different claims. The practical framework for any buyer, seller, or owner is knowing which side of that line your property falls on.

• Trust the AVM more if: the property is a standard single-family home in a suburban market with high transaction volume; the estimate has updated within the past 30 days; multiple recent comparables exist within a quarter-mile; the property type is among the most common in its zip code.
• Trust the AVM less if: the home is off-market or has not sold in more than five years; it is a condo, multifamily, or property with unusual characteristics; the local market has experienced rapid price movement in the last 12 months; you are in a rural or low-transaction-density area; or your neighborhood is in a historically underinvested area where the model’s training data may be thin.
• Seek an independent appraisal if: you are making or accepting an offer significantly above or below the AVM estimate; you are refinancing and want an accurate basis for equity calculations; the property has known improvements or deficiencies that are not visible in public records; or the stakes of the transaction make a few hundred dollars of additional certainty worthwhile.

The federal AVM quality control rule, in effect since October 2025, improves the integrity of the models used in lending decisions. But it does not make any individual estimate correct. The Zestimate on your specific home, on this specific day, could be off by 10% or more — and you would have no way of knowing until a buyer or an appraiser tells you otherwise. Running your own numbers alongside any automated estimate is always a sound discipline. You can use our mortgage payment, affordability, and buy-vs-rent tools to model how different price points translate to real monthly costs, independent of what any algorithm says the property is worth.