Adjina Dekidjiev, a real estate broker at Coldwell Banker Warburg, put it simply to MoneyWise: automated valuation models "don't have information about renovations, property improvements, cosmetic improvements, and additions that significantly affect property value." She was explaining why Zillow's Zestimate sometimes misses by tens of thousands of dollars. She could have been describing every major AVM on the market.
A midrange kitchen remodel in the United States costs $82,793 on average, according to Zonda's 2025 Cost vs. Value Report. It returns $42,130 at resale, a 51% recoupment. A minor kitchen refresh costs $28,458 and returns $32,141, a 113% recoupment that makes it the fifth-best renovation for resale value in the country. Either way, the improvement matters to the home's price.
But to the algorithm? Neither renovation exists until the home sells or a human updates the record.
What Algorithms See, and What They Miss
Zillow publishes its Zestimate's median error rate at 1.76% for on-market homes and 7.22% for off-market homes. On a $600,000 house, 7.22% translates to a $43,320 margin of error. Jeff Lichtenstein, CEO of Echo Fine Properties, captured the absurdity of this to MoneyWise: "I always ask a homeowner, 'When Zillow came over, what did they think of the kitchen?'"
Zillow never came over. Neither did Redfin, which reports a 2.1% on-market error rate. Neither did CoreLogic, which claims 1.8%. None of these systems have ever looked inside your house. They process tax assessments, comparable sales, square footage, lot size, and listing data. Interior condition, the difference between laminate countertops and quartzite, between original 1990s cabinets and a soft-close custom build, is invisible to every major consumer AVM operating in the United States today.
Zillow's own guidance acknowledges the problem. Their help page instructs homeowners to "claim your home and edit your home facts," manually entering renovation details to improve accuracy. That is a remarkable admission for a $14 billion company: the algorithm works better when you do its homework.
Researchers Proved It Could Work
In 2026, a team from the Norwegian University of Science and Technology published a study in The Journal of Real Estate Finance and Economics titled "Giving Eyes to Automated Valuation Models." Working with 15,702 condominium transactions in Oslo, they trained convolutional neural networks to classify room types and grade interior condition from standard listing photographs, the same images already uploaded to every real estate platform in the world.
Results were striking. Room classification hit 95% accuracy. Kitchen and bathroom condition grading reached 66-73% accuracy on a five-point scale, with misclassifications rarely exceeding one level. SHAP analysis confirmed the models were relying on economically plausible features: surfaces, fixtures, visible wear, not random background objects.
When these computer vision scores were folded into automated valuation models, prediction error dropped. Mean absolute percentage error fell from 8.0% to 7.7% in an XGBoost model, and from 12.8% to 12.4% in a hedonic regression. A one-level improvement in overall room condition correlated with an 8.2% to 8.7% price increase, holding all other variables constant.
Crucially, the computer vision assessments performed comparably to human-assessed condition labels. Not identically, but close enough that the researchers concluded automated condition grading could substitute for costly manual inspections at scale.
One result was conspicuously weak: living room condition grading hit only 43% accuracy. Living rooms present less standardized visual vocabulary than kitchens or bathrooms, where fixtures, appliances, and surface materials signal quality in ways a neural network can parse. A living room's condition depends on flooring, paint quality, trim detail, and proportions that humans read effortlessly but algorithms struggle to codify.
A Rule That Requires Accuracy but Not Vision
On October 1, 2025, a federal AVM quality control rule took effect, issued jointly by six agencies including the CFPB, FDIC, and FHFA. Mandated by Section 1125 of the Dodd-Frank Act, the rule requires mortgage originators and secondary market issuers to adopt policies ensuring AVMs produce estimates with "a high level of confidence," protect against data manipulation, avoid conflicts of interest, submit to random sample testing, and comply with nondiscrimination laws.
Read that list again. High confidence. Nondiscrimination. Random testing. All reasonable requirements. None of them mandate that an AVM actually observe the property it is valuing.
Simultaneously, the Government-Sponsored Enterprises are implementing Uniform Appraisal Dataset 3.6, shifting appraisal reporting from narrative text to structured XML. An $11.3 billion industry is rebuilding its data infrastructure. Computer vision could slot directly into this structured format, populating condition fields automatically from listing images. Instead, those fields will be filled by human appraisers, when human appraisers are involved at all. For the growing number of transactions where AVMs replace human appraisals entirely, the fields stay empty.
Who Loses
Lower-priced homes bear the worst of this blind spot. Researchers in Oslo found that condition effects were strongest in less expensive properties. An $80,000 kitchen in a $1.2 million house sits among comparable renovated homes that calibrate the algorithm. An $80,000 kitchen in a $400,000 house, where renovations are rarer and comps are sparser, creates a valuation gap the AVM cannot close.
Consider what that means for a homeowner refinancing. If the algorithm undervalues a renovated $400,000 home by 8%, that homeowner loses access to $32,000 in equity. A HELOC application rejected. A debt consolidation derailed. A down payment on a second property that never materializes. The renovation happened. The algorithm did not notice.
Sellers face a different version of the same problem. Buyers walk into negotiations armed with a Zestimate. When that number ignores a $28,000 minor remodel that added $32,000 in value, the opening offer anchors below fair market. Agents work around this constantly, but the asymmetry persists: the algorithm's number carries an authority that exceeds its information.
Strongest Case Against
AVM providers would argue, reasonably, that comparable sales already capture renovation value indirectly. If renovated homes in your neighborhood sell for more, the algorithm learns from those transactions and lifts your estimate. This is true. It is also slow, circular, and geographically contingent. Your renovation improves the comps only after someone else's similar renovation sells, and even then, the model cannot attribute the premium to the kitchen versus the new roof versus the school district's rising scores.
Providers would also note that deploying computer vision at scale introduces new risks. Staging and photography quality vary. A professionally photographed laminate kitchen can outshine a poorly lit quartzite build. Models trained on listing images inherit the biases of real estate photography, which is itself an industry designed to make properties look better than they are. Deploying condition grading without solving the staging problem could create new inaccuracies, not fewer.
Both points have merit. Neither explains why the technology sits in academic papers while the regulation demanding accuracy sits in the Federal Register.
What to Do Right Now
If you have renovated your home and don't plan to sell immediately, three steps protect you from the AVM blind spot.
First, claim your home on Zillow and update the home facts. Add the kitchen remodel, the bathroom update, the new flooring. This is manual labor the algorithm should do for itself, but until it does, the data gap is yours to close.
Second, contact your county tax assessor and report the improvements. Tax records feed AVMs. If the assessor's file still shows your 1990s kitchen, every algorithm reading that file inherits the error. Yes, this may increase your property tax assessment. It also increases every AVM's estimate of your home's value, which matters the moment you refinance, borrow against equity, or sell.
Third, when selling, ensure your listing photos clearly document the interior condition of every room. Until AVMs deploy computer vision, the images serve human appraisers and buyers. But the Norwegian study used standard listing photos, not specialized imagery. When AVM providers eventually integrate vision models, the training data will be the photos already in the MLS. High-quality images of your renovation become a long-term asset.
Limitations of This Analysis
The primary research cited here studied 15,702 condominiums in Oslo, Norway, a market with different transaction mechanics, housing stock, and data availability than the United States. Norwegian homes sell through ascending-bid auctions with near-universal agent involvement, producing cleaner pricing data. Whether computer vision condition grading transfers to the American market at similar accuracy levels has not been independently verified.
Living room condition accuracy at 43% represents a genuine constraint. Bathrooms and kitchens have standardized visual vocabularies, porcelain, chrome, tile, that neural networks parse well. Living rooms, bedrooms, and other flexible spaces are harder. Full-home condition assessment will require either better models or supplementary data sources.
Finally, AVM providers have proprietary data and methods not disclosed publicly. It is possible that internal condition proxies, such as permit records, renovation-related text in listing descriptions, or satellite imagery of exterior changes, capture more of the renovation signal than public-facing error rates suggest. Without auditable model documentation, this cannot be confirmed or denied.