By Michael Marelli, Licensed Texas Real Estate Broker, TREC #639078
Published: July 31, 2026 · Reading time: approximately 12 minutes

In this article:

Research Snapshot

Question: Can an automated valuation model accurately price a home, and when should the result be trusted?

Why it matters: For many homeowners, an automated home value estimate is one of the first tools they consult before deciding whether, or how, to sell.

Methodology: Analysis of direct brokerage experience across two documented case studies, published accuracy data from Zillow, and published industry research on AVM design.

Key takeaway: Automated valuation models are an effective starting point for pricing a home, but two identifiable categories of limitation, information gaps and market-boundary mismatches, can materially affect their accuracy in specific situations.

Introduction

Before listing a home for sale, most homeowners check an automated value estimate. It is fast, free, and available with a single search. For many properties, the number that comes back is a reasonably close approximation of what the home will eventually sell for.

For some properties, it is not. The question this article set out to answer is not whether automated valuation models (AVMs) work. They clearly do, and they work well for the majority of homes they are asked to price. The question is more specific: in the cases where an automated estimate turns out to be materially wrong, is that unpredictable, or is it possible to identify, in advance, the kinds of situations where an estimate deserves a closer look?

This article examines two documented cases in which an automated home value estimate diverged meaningfully from a property's actual market value, one understating it, one overstating it. Each case is used to isolate a specific, identifiable mechanism behind the divergence, rather than to argue that automated pricing tools are broadly unreliable. The goal is to build a practical framework that a homeowner can apply to their own property before relying on an automated estimate for an important decision.

Methodology

This research was conducted to answer a specific question: when can an automated home value estimate be trusted, and when does it warrant closer scrutiny? It was not designed as a statistical study measuring how often AVMs are wrong across a large sample of homes.

Three sources of evidence support the findings presented in this article.

Direct brokerage experience. The two case studies are drawn from real transactions handled directly by a licensed Texas real estate broker. They were selected because each demonstrates a distinct, identifiable mechanism behind a mispriced automated estimate, rather than because they represent a comprehensive sample. Property and neighborhood details have been described using anonymized composites rather than specific addresses or identifying information, since the purpose of each case study is to illustrate a pricing mechanism, not to document a particular transaction. Where dollar figures are described as approximate, they reflect the broker's recollection of the transaction rather than figures pulled directly from closing records.

Published accuracy data from Zillow. Zillow publishes its own median error rate for its Zestimate tool: 1.9 percent for homes currently listed for sale, and 7.0 percent for homes not currently on the market. Zillow's own published guidance also notes that Zestimate accuracy depends on the volume and completeness of data available in a given geographic area, a factor directly relevant to both categories of limitation described in this article.

Published industry research on AVM design. General information about how AVMs select comparable properties, including the role of geographic proximity as a primary filter, was drawn from published industry sources, including research from the Mortgage Bankers Association on automated valuation model methodology.

Limitations of this research. This article does not claim to measure how frequently either category of limitation occurs across the broader housing market. The two case studies represent specific, observed instances of each mechanism, not a statistical frequency. In addition, Texas is a non-disclosure state, meaning sale prices are not part of the public record in the way they are in many other states. This limits the ability to conduct a large-scale, data-driven study of AVM accuracy specific to Texas, and it is one reason this research relies on direct brokerage experience rather than a broader dataset. Readers should treat the two categories of limitation identified here as a starting framework, not an exhaustive account of every way an automated estimate can be inaccurate.

Case Study One: Pricing Beyond Structured Data

A homeowner in the San Antonio area checked an automated home value estimate before preparing to sell. The estimate placed the home's value at approximately $330,000.

The estimate was based on a set of comparable properties, recently sold nearby homes that shared similar recorded characteristics. Most of those homes needed work: dated kitchens, worn systems, lots that backed directly onto major roads.

The subject property differed from those comparable properties in several meaningful ways. It had been well maintained and recently updated. Its lot was larger than those of many of the selected comparable properties and backed onto several acres of undeveloped land, providing privacy and open views from a covered back porch.

An AVM estimates value using structured property data such as square footage, bedroom and bathroom counts, and recent comparable sales. It has no structured field for characteristics such as privacy, open views, or the qualitative appeal of a property's setting.

A review of the comparable properties suggested the estimate was drawing from homes that did not fully reflect the subject property's condition, lot characteristics, or buyer appeal. A subsequent property walkthrough confirmed that the home's larger lot, privacy, and open-land views distinguished it from the comparable properties used in the estimate.

The home was listed at $399,000, about 21 percent above the automated estimate. It sold for $390,000, still roughly 15 percent above where the algorithm had placed it, and drew three offers along the way. The seller believed the property was worth more than the eventual sale price. Because that opinion was not independently validated by the market, it is included only as an observation rather than evidence.

This case is not a story about the algorithm being careless. It is a story about what the algorithm was built to do. When a meaningful share of a property's value sits outside its structured data, in the land, the view, or the condition relative to what is actually available nearby, the estimate has no way to account for it. This does not necessarily represent a flaw in the algorithm itself. Rather, it reflects the limits of any valuation model that depends primarily on structured property and sales data.

This case illustrates an important category of limitation in AI-assisted home pricing: property characteristics that materially influence buyer demand but are difficult to represent in structured property data.

Failure Mode Identified

Category: Pricing Beyond Structured Data
Mechanism: Important buyer-valued characteristics were not fully represented in the structured property data used by the AVM.
Observed Outcome: The automated estimate materially understated the property's eventual market value.

Case Study Two: Pricing Across Market Boundaries

Before listing, the owner of a renovated home in the San Antonio area obtained an automated home value estimate. The estimate placed the property's value at approximately $300,000.

The estimate selected comparable properties from an adjacent neighborhood, approximately half a mile away and separated from the subject property by a major road. That neighborhood carried a recognized historic designation and a corresponding price premium not present in the subject property's own neighborhood.

AVMs typically prioritize geographic proximity and recent sales activity when selecting comparable properties. While additional variables may also influence model selection, geographic distance remains a foundational component of most automated valuation approaches. These models cannot reliably distinguish between adjacent areas that carry materially different value characteristics, such as historic designation, unless those differences are represented in the data available to the model.

A review of comparable properties within the subject property's own neighborhood indicated a market value closer to $200,000 to $225,000, reflecting the absence of the price premium associated with the neighboring historic district.

The property owner, who had recently completed renovations intended to support a resale, listed the home based on the automated estimate. The automated estimate materially overstated the property's market value. The property subsequently experienced an extended time on market and multiple price reductions before selling.

A related pattern was independently observed in a separate transaction, in which a property adjacent to a golf course community was priced using comparable properties drawn from a nearby neighborhood with different construction characteristics, including differences in exterior masonry coverage and roofing material. In both cases, the estimate incorporated comparable properties from a neighboring market area without accounting for a defining characteristic that separated the two markets.

An AVM's search radius typically expands when nearby sales data is limited, in order to produce an estimate. When that expansion crosses a meaningful market boundary, the resulting comparison set can no longer represent the competitive market in which the subject property would actually compete.

This case illustrates a second category of limitation in AI-assisted home pricing: automated comparable selection can cross recognizable market boundaries, causing the comparison set to no longer represent the market in which the subject property actually competes.

Unlike the first case study, where important property characteristics existed outside the structured data, this example demonstrates that even accurate property data can produce misleading estimates when the comparison set is drawn from a different competitive market.

Failure Mode Identified

Category: Micro-Market Boundary Selection
Mechanism: Comparable properties were selected from an adjacent market area with a distinct value driver that was not represented in the subject property's competitive market.
Observed Outcome: The automated estimate materially overstated the property's market value. The property subsequently experienced an extended marketing period and multiple price reductions before selling.

Analysis: What These Cases Reveal

The two case studies point to a broader conclusion about how automated home value estimates should be interpreted.

An AVM can only produce an estimate based on the information it can observe and the assumptions built into its comparison process.

In the first case study, important buyer-valued characteristics were not represented in structured property data. The estimate was working as designed, but the available data did not fully represent the factors that influenced buyer demand.

In the second case study, the structured data was largely accurate. The issue was not what the model knew about the property itself, but where it looked for comparison. The comparison set no longer represented the competitive market in which the property would actually compete.

Although the mechanisms differed, both case studies demonstrate the same underlying principle: when important information is unavailable or when the comparison set no longer reflects the property's true competitive market, confidence in the resulting estimate should decrease.

A Practical Framework for Interpreting AI Home Value Estimates

Together, these case studies identify two recurring situations in which home sellers should interpret AI-generated pricing estimates with additional care. The framework below summarizes the two observed limitations identified in this research.

Observed Limitation Underlying Mechanism Potential Impact
Pricing Beyond Structured Data Important buyer-valued characteristics could not be represented in structured property data. The estimate may not fully reflect buyer demand for unique property features.
Micro-Market Boundary Selection Comparable properties were selected from a different competitive market. The estimate may reflect pricing dynamics from a market in which the property does not actually compete.

These two categories are observations drawn from the current research. They are not presented as an exhaustive list of every way an automated pricing estimate can become unreliable. As additional cases are analyzed, this framework may expand to include other categories of limitation.

An Important Clarification

These observations should not be interpreted as evidence that AVMs are unreliable. On the contrary, they are among the most useful pricing tools available to home sellers. Their greatest strength is providing a fast, consistent, data-driven estimate. Like any analytical model, however, their usefulness depends on understanding the situations in which additional context becomes important.

Rather than asking whether an AI estimate is simply "right" or "wrong," a more useful question is whether the estimate is being applied in a situation where these observed limitations are likely to influence the result. Recognizing those situations allows sellers to use AI as it was intended: as a powerful decision-support tool rather than a substitute for judgment. The next section provides a practical framework for making that assessment before relying on an AI-generated home value estimate.

Practical Guidance for Sellers Using an Automated Home Value Estimate

An automated estimate is a reasonable starting point for most homes. The two limitations identified in this research point to specific situations where that estimate deserves a closer look before it is used to set a listing price or shape expectations about a sale.

Questions Related to Pricing Beyond Structured Data

  • Does the property have features that would be difficult to describe on a standard listing sheet, such as an unusual lot shape, uninterrupted views, or a border on open land, greenbelt, or park space?
  • Have significant updates or renovations been completed recently that might not yet be reflected in public property records?
  • Does the property's overall condition differ meaningfully, in either direction, from what would be typical for the immediate area?

If the answer to any of these is yes, the automated estimate is best treated as a starting point rather than a conclusion. A comparison built specifically from properties with similar characteristics, not just similar size and age, will generally provide a more complete picture.

Questions Related to Micro-Market Boundary Selection

  • Is the property located within a short distance, roughly half a mile or less, of a neighborhood with a distinct identity, such as a historic district, a gated section, a golf course community, or a notably different range of home values?
  • Are there relatively few recent sales within the property's own immediate neighborhood or subdivision?
  • Does the property's construction, including materials, age, or builder, differ from homes in a nearby area that could plausibly be included in an automated comparison?

If the answer to any of these is yes, the estimate may be shaped by sales activity outside the property's actual competitive market. A comparison built specifically from the immediate subdivision, rather than a wider surrounding area, can help confirm whether that is the case.

A General Verification Step

Regardless of the specific property, sellers can take one additional step before relying on an automated estimate for a major decision: review the individual comparable properties the estimate is based on, when that information is available, and compare them directly to the subject property on the two dimensions identified in this research, notable site or lot features, and proximity to a distinct neighboring market. A licensed real estate professional or a licensed appraiser can typically provide this kind of comparison.

The purpose of this research is not simply to identify where automated estimates can become less reliable. It is to help sellers recognize those situations before making important pricing decisions. This step does not replace the value of an automated estimate. Instead, it helps determine whether the estimate is being applied in a situation where additional review is warranted. In many cases, the estimate is likely to provide a reliable starting point. In others, these simple checks can identify circumstances where additional analysis may support a more informed pricing decision.

Conclusion

The two cases examined in this research point to a consistent finding: AVMs are a strong, reliable tool for pricing a home in most circumstances, and their accuracy depends on two things: the completeness of the structured data describing the property, and the accuracy of the comparison set used to price it. When a property has significant value characteristics that are hard to capture in structured data, or when it sits near a distinct micro-market that could plausibly be pulled into its comparison set, an automated estimate is more likely to require a second look.

Neither of these situations is common for every home, and neither means an automated estimate should be dismissed. They mean the estimate should be interpreted with the property's specific circumstances in mind, using the questions outlined in this article as a starting point.

The broader lesson extends beyond home pricing. An AVM is one example of a general-purpose tool applied to a specific, high-stakes decision. It performs a defined task well, using the data and assumptions it was built on. Recognizing where those assumptions hold, and where a property's circumstances fall outside them, is what allows a homeowner to use the tool effectively rather than either over-trusting or dismissing it.

An accurate home value estimate is available to most sellers with a single search. Knowing when to trust that number, and when to look closer, is what this research was designed to help answer. If you are weighing how to price your home without a listing agent in Texas, or wondering whether a tool like ChatGPT can help you sell a house, the same principle applies: AI is a strong starting point, and knowing where to look closer is what protects the outcome.

References

Related Reading:

See how Waymark applies this research to your property with an AI-guided pricing analysis reviewed by a licensed Texas broker.