Direct Answer: Local AVM Accuracy Depends More on the Market and Property Than the Brand
A local Automated Valuation Model, or AVM, is a statistical tool that estimates a home’s market value using facts such as comparable sales, lot size, living area, condition, location, and sometimes images or text. It is not an appraisal performed by a licensed appraiser, and it should not be treated as an offer to buy. The most useful question is not simply whether an AVM is accurate, but what “accurate” means in a particular ZIP code, property class, and price segment.
Also worth reading: Which AI Home Valuation Metrics Are Most Accurate in 2026? · How Accurate Is AI Property Search, and How Should Buyers Verify Its Results? · What Does Verified Property Data Mean for Buyers, Sellers, and AI-Powered Real Estate Tools in 2026?
As of September 27, 2026, there is no single national accuracy percentage that applies to every local AVM. A model can perform well on ordinary, well-tracked sales in established neighborhoods while performing poorly on unique homes, recent construction, distressed properties, or markets with few transactions. Buyers and sellers should therefore ask sellers, agents, lenders, and model providers for a documented local median error, a typical error range, and the number and recency of comparable sales.
A practical starting point is to treat a local AVM with a median percentage error below 5% as useful for broad screening, 5% to 10% as moderately useful but uncertain for a transaction, and above 10% as a prompt to obtain stronger evidence. Those are analytical thresholds, not regulatory guarantees. The final price can still differ materially because condition, renovations, views, floor-plan changes, seller concessions, and competition among buyers are not always represented correctly.
What Local AVM Accuracy Metrics Actually Measure
The standard metric is the median absolute percentage error, often abbreviated MAPE, although the technical version divides the average of absolute percentage errors while “median absolute percentage error” uses the median. In consumer discussions, many people use the terms loosely, so the calculation must be checked. A median error of 4% means that half of the tested estimates fell within 4% of the eventual sale or benchmark price, subject to the dataset and methodology.
A mean error can be misleading because one badly mispriced home can distort the average. Median error is usually better for describing the typical result, while the 80th or 90th percentile error shows how bad the weaker results become. Buyers should also distinguish directional accuracy from pricing accuracy: a model may correctly predict that a house is worth more than another house while assigning it the wrong dollar amount.
Coverage matters as much as the headline percentage. An AVM evaluated on 20 recent sales is less persuasive than one evaluated on 2,000 sales, especially if the smaller sample represents only one neighborhood or home type. Quality control should include the evaluation period, repeat-sales or holdout method, source of “actual value,” and treatment of missing data. Metrics should be refreshed monthly or quarterly because a model that was accurate two years ago may no longer reflect current market behavior.
| Feature | Strong local evidence | Weak local evidence | Why it matters |
|---|---|---|---|
| Median error | 3%–5% | 10% or more | Describes typical pricing error, not just model direction |
| 80th percentile error | Below 8% | Above 15% | Reveals how poorly the model performs in difficult cases |
| Validation sample | At least several hundred recent sales | Fewer than 30 relevant sales | Larger, current samples are more dependable |
| Coverage | High for local property types | Many missing or excluded sales | A narrow test may hide neighborhood variation |
| Benchmark | Closed-sale price or credible appraisal | List-price-only comparison | List prices reflect asking strategy, not completed value |
| Frequency | Updated within 90 days | Annual or older | Markets and property values change quickly |
Local AVMs are shaped by the economics of the immediate market. In a tract with many nearly identical homes sold close together in the last 90 days, the model can identify a central value with reasonable confidence. In a rural area where comparable sales may be 20 or 40 miles away, local housing features change rapidly with distance, road access, wells, septic systems, terrain, and school zones. The model may report a ZIP-code-wide number that conceals this instability.
Property similarity creates another problem. A model trained or tested on standard three-bedroom tract homes may not predict a renovated kitchen, a converted garage, a new septic system, a lot with an accessory dwelling unit, or a house with serious foundation issues. Zillow’s Zestimate is a widely recognized consumer AVM, but its accuracy is not a promise that every estimate falls within a fixed margin. National brand visibility does not eliminate the need to test performance in the specific city and property class.
Local market conditions also change the error rate. Rising interest rates reduce transactions and make fewer homes available as comparables. Inventory shortages can produce abrupt price changes, while a cluster of distressed sales can pull local values downward without representing normal owner-occupant demand. A useful report separates owner sales, foreclosures, short sales, and cash purchases, and it states whether the model was tested during the same economic period buyers will face when making an offer.
Unique features do not merely add uncertainty; they can make the concept of a single value less informative. A property with a major addition, unusual lot, view, or legal restriction may have a wide reasonable-value range. In such cases, an AVM should be treated as a screening estimate, and a broker price opinion, comparable-sales analysis, or full appraisal should carry more weight.
How to Test an AVM for a Specific Home
Start with a single most recent estimate, but do not compare different AVMs as if they were independent appraisals. Several models may reuse the same public tax records, deed transfers, permits, and listing feeds, so agreement among them can create false confidence. Independent evidence comes from a licensed appraisal, a well-supported comparable-sales analysis, or a broker’s documented knowledge of recent local conditions.
Next, check three to five genuinely comparable sales rather than the first results returned. The strongest comparables should have a similar location, sale date, size, lot, bedroom and bathroom count, condition, and property type. A sale from six months earlier may be less useful than a recent sale farther away if the older property was in materially different condition. Adjustments should be explicit: every 1,000 square feet, garage, bedroom, or major renovation can have a different dollar effect by neighborhood.
Use a range rather than defending one number. If the adjusted comparable evidence supports values from $410,000 to $445,000, that is a usable negotiating interval, not an administrative error. A close-in AVM estimate near $427,000 may help locate that range, but it should not be the sole basis for a purchase, refinance, tax appeal, divorce, or estate decision. For an ordinary home in a liquid market, a local error below 5% is encouraging; for a complex property, even 5% may be too small to guide a transaction.
Buyers can request a written AVM or comparative market analysis from a real-estate professional and ask what happens if the first three comparable homes sell away. Sellers should test the estimate against the likely net proceeds, commissions, transfer taxes, and repair costs. The estimated gross value does not equal the amount the owner will receive after a transaction.
Comparison of AVMs, Broker Analyses, and Full Appraisals
An AVM is fast, inexpensive, and useful before a listing or during early property discovery. It can update daily and provide an immediate rough range without visiting the property. Its weakness is that it may not account for craftsmanship, odors, deferred maintenance, a precise floor plan, neighborhood noise, or the quality of recent renovation work. It is therefore well suited to triage, search, and rough affordability screening.
A comparative market analysis, or CMA, is prepared by an agent using selected sales and market knowledge. It can account for differences that a model misses, but its quality depends on the agent’s data selection and adjustments. A CMA is valuable for setting a listing range and evaluating offers, though it remains an opinion rather than a certified appraisal. Asking the agent for the underlying comparables and adjustment reasoning is essential.
A full appraisal is commissioned through an appraisal-management process and is commonly required for a mortgage. It is more expensive and slower, yet it is the strongest standard tool for a lender’s risk decision. The appraiser observes the property, verifies relevant data, and applies professional methods. A primary residence purchased with a conventional mortgage typically requires an appraisal under the ECOA and Uniform Closing Disclosure rules, although appraisal waivers and other policies can apply in eligible transactions.
| Feature | Automated valuation model | Broker comparative market analysis | Licensed appraisal | Best use |
|---|---|---|---|---|
| Turnaround | Seconds to days | Often within 24–48 hours | Days to several weeks | Rapid screening |
| Typical cost | Often $0 to $30 per query | Commonly included as agent service | Commonly $300–$700 or more | Mortgage or legal decision |
| Property inspection | Rarely | Sometimes | Yes | Trustworthy full valuation |
| Comparable-sales detail | Model-dependent | Usually shown | Yes | Verify local pricing |
| Major limitation | Hidden data and weak feature capture | Agent judgment | Timing, access, and comparable availability | Cross-check other tools |
Common Mistakes That Distort Local AVM Conclusions
The most common mistake is treating a point estimate as a guaranteed price. A 4% error on a $500,000 home equals $20,000, while a 10% error equals $50,000; the dollar consequence may matter more than the percentage. Another mistake is comparing an AVM with the listing price. A seller usually prices to create competition, and a reduction or increase does not necessarily show that the model failed.
Second, users often select a broad ZIP code when a model is tested at a city or state level. ZIP codes are postal delivery areas, not uniform real-estate markets. A performance figure for one ZIP code should not be transferred to a different ZIP code without evidence. Neighborhood boundaries, school districts, transit access, terrain, and development can produce large differences over short distances.
Third, people forget to inspect the time period. A metric calculated across transactions from 2019 may not describe a 2026 market, particularly when rates, inventory, remote work, or local zoning changed. Fourth, they assume that three AVMs provide three independent opinions. Many estimates are driven by common transaction feeds, so the models may converge without independently validating one another.
Finally, users ignore uncertainty around renovations and defects. An AVM can misread square footage, fail to recognize permits, or assign value based on an incomplete data set. A reported 2,200 square feet may include a basement, addition, or finished room of uncertain quality. In older housing stock, condition can shift value by tens of thousands of dollars, so local inspection evidence is more useful than a tiny change in algorithmic output.
When Buyers and Sellers Should Act—and When They Should Pause
A buyer should use a local AVM before touring homes to estimate whether the search is affordable and to flag properties that may be substantially above or below the neighborhood pattern. During offer preparation, the estimate should be updated and compared with at least three recent closed sales. If the estimated value and comparable evidence differ by more than about 7%–10%, pause and investigate rather than submitting an offer based on the estimate alone.
A seller can use an AVM to choose a listing range, but should avoid selecting only a favorable estimate. A realistic range should reflect likely buyer competition, the home’s condition, and the effect of repairs or concessions. If the seller is deciding whether to renovate, obtain a contractor’s scope and a professional’s value analysis; a machine estimate does not measure construction cost or return on investment.
Pause for an appraisal or specialist review when the property is unusual, the sale is legally sensitive, or the financial consequence is large. Examples include divorce, estate settlement, a tax appeal, a divorce-related valuation, a trust sale, a property with an accessory dwelling unit, or a home with major structural changes. A comparable-sale dataset that has fewer than three highly similar recent sales is also a reason to slow down.
The date of the data is decisive. As of September 27, 2026, an estimate supported by sales from the last 90 days is generally more decision-relevant than one based on a stale index, but no interval makes a poor comparison adequate. Ask the provider when the model was last trained, when it was last calibrated locally, and whether the reported accuracy uses closed sales that occurred after the prediction date.
Costs, Reliability, and a Decision Rule for 2026
Most consumer AVMs are free to view, although paid brokerage tools, desktop products, API calls, and premium reports can cost roughly $20 to several hundred dollars per month. Agent CMAs are often included in brokerage services, while full residential appraisals commonly fall near $300 to $700, with complex, remote, commercial, litigation, or extensive-inspection work costing more. These are market ranges rather than regulated all-in quotes.
The best 2026 decision rule is conditional. If the local median error is below 5%, the test has at least several hundred recent and relevant sales, and the estimate is within a few percent of adjusted comparable evidence, use it as a useful screening tool. If the median error is 5%–10%, use it to form a range and seek professional confirmation. If it is above 10%, the data is stale, coverage is weak, or the property is materially different from the tested group, do not use the estimate for pricing or financing decisions without stronger evidence.
No AVM should be judged solely by an accuracy number supplied in advertising. A credible evaluation reports the exact formula, benchmark, geography, property type, time window, sample size, and error distribution. A provider willing to show a 90th-percentile error and identify failure cases is giving a more honest picture than one showing only a favorable median.
For realtigence.com’s AI-driven matching and property-discovery use case, the responsible message is not that software can replace local valuation judgment. Software can narrow the search, surface records, compare estimates, and flag questions for a human expert. The user still needs a broker CMA or appraisal when the stakes demand one. The platform adds value by showing the evidence behind a match and explaining uncertainty, not by presenting an algorithmic number as certainty.