AI Can’t Tell You What a Home Is Worth
At an open house last weekend, a buyer pulled up two screens before he had even walked into the backyard. One was a Zillow Zestimate. The other was a summary from ChatGPT, laying out what he should offer. He held up his phone like it settled the question. I had to ask him a simple thing: where did those numbers actually come from?
That question is worth sitting with. An estimate is supposed to be a starting point, something you check and refine. But the word gets used like it's already settled, and I think that's part of why people stop asking where the number actually came from.
This is happening across California, not just in the Tri-Valley. More buyers are opening a search engine or an AI chat window before they open a conversation with their agent. I understand why. These tools are fast, free, and available at eleven at night when you can't stop thinking about a house you saw that afternoon. But there is a real gap between what these tools can tell you and what you actually need to know before you write a number on an offer.
What These Tools Actually Do Well
I want to be fair here, because these tools are useful for plenty of things. They're good at narrowing a search, learning the basics of a neighborhood, pulling together public information quickly, and helping a buyer get comfortable with a market before they start touring homes. I use some of them myself for research.
Where it gets risky is when a number pulled from public data starts getting treated the same as a professional's read on a specific home, on a specific street, this week.
What an “Estimate” Can't See
Zillow publishes its own accuracy data, and it's worth knowing. For homes currently listed for sale, the company reports a median error of roughly 2 percent. That sounds tight. But for homes that are not actively listed, which is the situation most buyers are actually estimating, the median error jumps to roughly 7 percent.
On a home in Pleasanton, Livermore, Dublin, San Ramon, or Danville, where price points regularly run well above the national median, 7 percent is not a rounding error. It's tens of thousands of dollars. And that's the median, meaning half of all estimates miss by more than that.
An algorithm works from public records, tax data, and recent comparable sales. It doesn't walk the property. It doesn't know that the kitchen was remodeled two years ago, or that the seller is motivated because of a job relocation, or that three other buyers are already circling. And critically, it doesn't always know which comps it should be ignoring.
The $100,000 Gap
Back to that buyer at the open house. He showed me his ChatGPT summary as though it were settled fact, a clean number with no hedging. When I compared it to what I actually knew about the home and the surrounding sales, the gap was about $100,000.
I asked him a simple question: what comps did that summary use? He didn't know, because the tool doesn't explain its reasoning in a way that lets you check its work. So we looked together. One of the comparable sales it had pulled in was from a neighborhood with known foundation issues tied to ground movement, a fact that gets passed around among local agents and longtime residents but doesn't live in any public data field an algorithm can read. That single comp was quietly pulling the estimate in the wrong direction, and there was no way to know it without knowing the street.
A local agent isn't just pulling comps, we're deciding which ones don't belong in the conversation. That's the part no algorithm is doing yet.
What Buyers Should Actually Do With an AI-Generated Number
I'm not telling anyone to stop using these tools for research. I'm telling you where the line is.
● Treat it as a starting point, not an offer number.
● Ask what comparable sales the estimate is based on, and whether any of them have condition, location, or history issues a public record wouldn't capture.
● Walk the block, not just the listing. A neighborhood tells you things a data table can't.
● Bring the number to your agent before you write it into an offer, not after.
If you can't explain where a number came from, it's probably not ready to go on an offer yet.
What This Means for Buyers Right Now
In a competitive Tri-Valley market, offer strategy is where deals get won or lost, sometimes over a few thousand dollars, sometimes over whether the listing agent even calls you back. If you can't defend your number because you don't know where it came from, that's a real risk to take into a negotiation. A number built on real comps and an actual read on the seller's situation puts you in a much stronger position.
FREQUENTLY ASKED QUESTIONS
Are Zestimates and AI summaries always wrong?
No. They are often reasonably close, especially for homes that are actively listed. The issue isn't that they're always wrong, it's that you can't tell, from the number alone, whether this particular one is close or off by six figures.
Should I stop using these tools altogether?
Not necessarily. They're a fine starting point for general research. Just don't let a starting point become your final offer number without a local, human review of the comps behind it.
How do I know if an online estimate missed something local?
Ask your agent to review the comps behind it. A local agent will know about issues like soil movement, HOA disputes, busy streets, or pending sales that haven't hit public data yet, the kind of details that don't show up in an algorithm's inputs.
Does this apply to sellers too?
It does, and it's worth its own conversation. But for buyers specifically, an inflated or deflated estimate can mean overpaying for a home with real issues, or losing a home entirely because the offer came in under what the market actually supported.
Have questions about what a home is actually worth in the Tri-Valley, beyond what an app can tell you? Let's talk before you write the number down. If this was useful, share it with a buyer who's about to trust a screenshot over a conversation.
Sources: Zillow Zestimate Accuracy Methodology
Patty Recupero
REALTOR® | ELATION REAL ESTATE | CA DRE 02112275