When the Algorithm Misses: What New Research Reveals About AI Home Valuation Blind Spots
Automated valuation models have quietly become one of the most trusted starting points in residential real estate. Type an address, get a number, and suddenly a homeowner has a mental anchor for what their property is worth. But a number without context is just a signal waiting to be interpreted, and new research out of San Antonio is a useful reminder of that.
Michael Marelli, a licensed Texas broker and founder of Waymark Real Estate, published a study examining where automated home value estimates tend to diverge from actual market outcomes. Rather than asking whether AI can price a home, which the data already answers well for most properties, the research narrows in on the exceptions. That framing matters. In data science, the interesting patterns usually live at the edges of a model, not in its average performance.
The study identifies two recurring failure modes using two documented Texas transactions. The first is a data representation problem: features like privacy, views, or lot condition relative to neighboring properties simply are not captured in the structured fields most automated models ingest. If a variable is not in the training data, it cannot show up in the output. The second is a comparable selection problem, where a model pulls comps from an adjacent area with its own distinct value driver, a historic district or a different construction tier, that does not actually reflect the subject property’s competitive set. Both point to the same underlying truth about AVMs: they are only as good as the geographic and structural assumptions baked into how they cluster properties.
Context matters here too. The research cites Zillow’s own published median error rate of 1.9 percent for active listings, which is genuinely strong performance at scale. The two case studies are not presented as a statistical measurement of how often these failures occur across the market, and the paper is transparent about that limitation. Texas is a non-disclosure state, meaning sale prices are not public record, which constrains any researcher’s ability to run a large-scale accuracy study specific to the state. That is a data availability problem as much as a modeling one, and it is worth flagging for anyone building or evaluating property intelligence tools in similar markets.
The question isn’t whether AI can price a home. For most homes, it can. The real question is whether your home is one of the exceptions.
What this really underscores is a principle worth repeating: automated intelligence is a layer, not a verdict. A well-built model can compress enormous amounts of market data into a fast, useful estimate, but the sellers best served by that estimate are the ones who understand what it cannot see. Structured data has edges, and homes with atypical characteristics or unusual geographic positioning sit right at them. Pairing an automated estimate with a targeted, human gut check on those two specific failure points is a small step that meaningfully improves the odds of pricing a home correctly from the start.
