I watched an episode by Steven Thomas, the economic forecaster we follow closely and have mentioned here before, and he made a point that has stuck with me ever since. Artificial intelligence is genuinely good at valuing housing at the macro level and genuinely bad at it at the micro level.
He is right, and it explains almost every conversation we have with homeowners about their Zestimate.
The big picture is where it works
At that level, the technology is impressive. Ask it what 92024 did year over year, whether inventory is building, and which direction rates are pushing demand, and it will tell you accurately. Zillow’s own published median error rate for homes actively listed for sale is 1.83%.
There is a reason that number looks so good. Zillow says the model takes in the list price and recent listing activity once a home hits the market. So a lot of that accuracy is the algorithm reading a price an agent already set.
The number nobody quotes
The number nobody quotes is the one for homes that are not listed. In other words, your home, sitting there while you check the estimate out of curiosity on a Sunday night. No list price to work from, and the published median error rate jumps to 7.01%. On a $2,500,000 home in Encinitas, that is just over $175,000. And a median error rate of 7% does not mean every estimate lands within 7%. It means half of them are worse.
Same floor plan, different value
Here is why. Take Encinitas Ranch. There were 4 different builders in Encinitas Ranch. Those different tracts are distinct. AI (and others) are often lumping all Encinitas Ranch properties together. But even within each builder’s project, you’ll have the same builder, same tract, same floor plan, three streets apart. One has been taken down to the studs with a real kitchen, new windows, and a backyard someone actually spent money on. The other is original from the 1990s with the builder tile still in the entry. The algorithm sees identical square footage, identical bed and bath count, identical lot size, and splits the difference. Those two homes are nowhere near the same value, and anyone who has walked both knows it in about fifteen minutes.
Same story with a view over the ridgeline versus a view of the neighbor’s roofline, the side of the street that backs to open space versus the one that backs to traffic noise, the cul-de-sac versus the through street. None of that lives in a public record, so none of it lives in the model.
What no model can price
And the biggest one of all. Your home is not worth what a database says. It is worth what the two most motivated buyers will pay in the specific window when it is available, and that changes week to week.
Use the estimate the way you use a weather app before a sailing trip. It tells you the general conditions. It does not tell you what the water looks like once you get out there.
The full Steven Thomas episode is worth watching. You can find it here: AI Thinks Your Home Is Worth What?!
If you want a real read on your place, give us a call at (619) 807-3423, and we will come walk it. No pressure attached.
