Here is the uncomfortable answer: nobody at the appraisal district sat down and decided the second house was worth more. A model did, using records that were probably already stale when the numbers went live.
That gap between what a computer spits out and what a buyer would actually pay is where most valuation disputes are born. So let's pull the machine apart.
What an Automated Valuation Model Actually Is
An automated valuation model, usually shortened to AVM, is software that estimates a property's market value from data instead of from a person walking the lot. Lenders use them to check collateral before approving a refinance. County appraisal districts use them to mass-appraise tens of thousands of parcels in a single cycle. Zillow's Zestimate is the version most homeowners meet first, usually at 11 p.m. on a phone.
They exist because the alternative doesn't fit. Appraisal districts in large metros have to produce a value for every parcel under their jurisdiction before notices go out each spring, and sending a licensed appraiser to every single door is physically impossible. The model is the only tool that fits the calendar.
That's the whole design brief, by the way. Not accuracy. Coverage. A model that's 92% accurate for every parcel beats a precise appraisal for 3% of them, at least from the district's side of the desk.
The Four Inputs Feeding Every AVM
Strip away the branding and every AVM runs on four things. Some systems weight them differently. None of them escape these categories.
- Recorded sale prices of nearby comparable properties, pulled from county deed records and multiple listing data.
- Physical attributes held in the assessor's file: square footage, bedroom and bath count, lot size, year built, pool or no pool.
- Location adjustments for school attendance zones, floodplain maps, traffic exposure, and proximity to commercial corridors.
- Time trending, which nudges older sale prices forward or backward to reflect where the market has moved since the comparable sold.
Of the four, recorded sale prices carry the heaviest weight in almost every system I've seen discussed. Which is exactly why the model's weak spots cluster around neighborhoods where sales are rare, unusual, or nobody's reporting them accurately.
Where Models Get It Wrong
Every one of those four inputs assumes the county's record is correct and current. Spend an afternoon looking at property records in older Houston neighborhoods and you'll learn how generous that assumption is.
Ranch-style houses that got a second-story addition in 2009 often still show their original square footage. Renovated kitchens and finished garages exist only in the owner's memory and the contractor's invoice. Meanwhile, a tired duplex that sold cheap because of a mid-probate estate pulls down every model value around it for the next eighteen months.
There's a structural version of this problem, too. Housing in the United States has gotten measurably bigger over the decades, a long-term trend the U.S. Census Bureau has tracked for years in its housing data. Any model leaning on older comparable sales from a neighborhood with a mix of eras is quietly comparing two different kinds of house.
And the vendor databases feeding real estate models miss plenty. Property valuation models built on national data sources carry known gaps in coverage, which is a recurring theme in the research published by HUD USER, the Department of Housing and Urban Development's research arm.
A Five-Minute Record Audit You Can Run Today
Before you argue about a value, confirm the value was built on facts. Pull your parcel record from the county appraisal district website and check these against reality:
- Square footage under air, measured, not eyeballed from the listing.
- Bedroom and bathroom count. Recount them.
- Year built and year of any major renovation. Check for duplicate entries.
- Lot dimensions. Compare against your survey, not the plat map thumbnail.
- Exemptions on file. Homestead, over-65, disability, whatever applies to your household.
Photo-document anything that's wrong. A tape measure and a phone camera solve more disputes than a paragraph of complaint ever will.
When the Model Outpaces the Market
Values usually get set in January based on sales that closed the previous autumn, then sit there until notices mail in spring. If your neighborhood cooled off in February, you're arguing against a snapshot of a market that no longer exists. That timing lag is baked into the process and it's not going away.
This is where a lot of owners get stuck, because the correction path runs through a formal protest with deadlines and evidence rules most people have never touched. Plenty of homeowners hand the filing to a firm that does the math daily, and you can compare how a houston property tax protest gets handled by a consultant who deals with the district year round. Doing it yourself is completely viable. You just need your numbers tight and your documentation tighter.
One deadline note worth internalizing now: protest windows in Texas are short and they do not move for you.
What Makes Some Models Better Than Others
I'd rank the four inputs in order of how much they matter to you, not to the model.
Sale recency beats everything. A comparable that closed in the last six months in your own subdivision tells you more than any adjustment the software applies for school ratings or a cul-de-sac premium.
Physical accuracy is second, and it's the one you can actually fix. Everything else is the district's problem.
What nobody enjoys admitting: a model's precision sounds impressive until you ask how often it's ever audited against reality. Probably close to never.
Knowing the Machine Is the Whole Game
Your value isn't a verdict. It's an estimate built from records, comps, and timing, and two of those three things are things you can verify or correct. Fix the record. Capture the comps that closed near you. Watch the calendar. When a model's output protects the district's deadline more than your equity, you're allowed to push back.
Pull your parcel record this week, tape measure in hand. What do you find that the database got wrong?
