Ahmadreza Samadi / Home

Analysis · Real estate AI

What AI home-value estimates miss

An automated estimate can be a useful starting point. It is still an estimate built from the property and market information the model has—not a direct inspection of everything that matters about a home.

By · · Updated

At a glance

An automated estimate depends on the property and market information available to the model. Before relying on it, check current property facts, suitable comparable sales and any missing information or uncertainty. Zillow’s Zestimate, discussed below, is not an appraisal.

What goes into the number?

The US Consumer Financial Protection Bureau describes an automated valuation model, or AVM, as a computer-generated estimate that compares property facts such as bedrooms, bathrooms and floor area with recent sales and other local market information. Different valuations can use different comparable sales, dates or purposes. A precise-looking figure does not make those inputs complete.

Zillow's Zestimate explanation gives a concrete product example: its model uses public records, listing-service data and details supplied by users, as well as market trends. Zillow says an unreported renovation may not be reflected and that incorrect or incomplete home facts can affect accuracy. That describes Zillow's product; other AVMs may use different data and methods.

What the model may not see

Imagine two illustrative homes with the same recorded size and bedroom count. One has a recently renovated kitchen; the other needs major roof work. If those conditions are absent from the available records, the shared facts alone cannot tell the full story. A weak set of comparable sales can add another source of error. This is why I would check the underlying facts and comparisons before acting on an estimate.

A single point estimate also compresses uncertainty. Zillow publishes an estimated sales range alongside its Zestimate and says the final sale price may vary with negotiation and market conditions. Its own explanation explicitly says the Zestimate is not an appraisal. Treat the range and the evidence behind the number as part of the answer.

Errors can differ across communities

There is a broader quality question beyond one house. A 2025 Urban Institute study of US AVMs found that valuation errors and undervaluation can differ across homeowner groups in its studied markets, even after accounting for property and neighborhood characteristics. That result is not a measured claim about every model, country or property. It is a reason to test error by relevant groups and locations, not only report one overall accuracy score.

How this connects to my work

At Appraiva, my technical work connects property imagery analysis with property identity, data enrichment and search for investment research. The Appraiva case study describes that workflow: connecting an image observation to the property it describes and other available information.

Three checks before relying on an estimate

  1. Are the property facts current? Confirm size, rooms, renovations and condition against reliable records or inspection.
  2. Which sales are genuinely comparable? Consider location, timing, property type and important differences, not just a nearby sale price.
  3. What is missing, and how uncertain is the result? Look for a range, review the evidence, and seek professional valuation when the decision warrants it.

The $500,000 estimate in my visual posts is hypothetical, not a valuation of a real home or a result from Appraiva. This article explains how to evaluate an AVM output; it is not property-specific financial or appraisal advice. For the related question of where a person should review AI-supported real estate work, see where AI should stop in real estate.

See the visual explanation

My Instagram carousel and LinkedIn document walk through the illustrative homes and the three checks.

Sources