You've probably seen the story by now. A Florida homeowner used ChatGPT to sell his house — skipped the agents, ran the whole process through a chatbot, and closed for $100,000 more than local agents had estimated. It went viral for good reason. It's the kind of headline that makes anyone wonder if they still need a Realtor.
Here's the part that didn't make the headline. Once the offers started coming in, real estate analysts who looked at the deal walked away skeptical that the AI-guided price was actually the win it looked like — some put the number the seller left on the table anywhere from $75,000 to $225,000, from not fully working a bidding war. Pricing a home is one skill. Reading a room full of competing offers in real time is a different one entirely — and no chatbot is doing that for you.
So I wanted to see how this plays out here in New Mexico — a non-disclosure state, meaning actual sale prices aren't public record the way they are in most of the country. That's a real blind spot for any AI tool pulling from public data. I wanted to know exactly how big.
I picked two real properties — one in Rio Rancho, one in Albuquerque — and tried to price them the way an AI (or a curious seller typing an address into ChatGPT) would: using only what's publicly available. No MLS. No client-only data. Just what anyone can find with a search bar.
Property 1: Rio Rancho
Public automated valuation models put this home at roughly $610,000. The one comparable I could find nearby was a similar home listed — not sold — at a price that, per square foot, implied a value closer to $330,000.
Automated estimate
$610,000
Nearby comp (list price, per sq ft) implied roughly $330,000 — the two public numbers disagree by nearly $280,000.
Built from real comparable sales data
$660,000
Meaningfully higher than either public number — for reasons no amount of address-only searching could have surfaced.
Two "AI-accessible" numbers on the same street, landing roughly $280,000 apart. Neither one is a sold price, because in New Mexico, sold prices simply aren't published. My actual CMA on this property landed at $660,000.
Property 2: Albuquerque
This one was worse. Every public source I found — half a dozen real estate sites — showed this property as an active listing, asking $415,000 to $425,000, with a specific square footage. Two problems: this home has never been listed and has never sold. Not now, not ever.
This home has square footage AI can't see. Additions were made to it over the years — work that pushed it well past 3,000 square feet — and none of it ever made its way into the public records or listing sites an AI pulls from. The internet didn't have stale data here. It never had a shot at the right number to begin with.
My real CMA valued this property at $525,000 — about $100,000 above what a chatbot pulling from those same sites would have confidently reported, for a listing status that doesn't even exist and a square footage that was never publicly correct to begin with.
What This Actually Proves
"AI lowballs everything" is the easy takeaway here, and it's wrong. The two misses had nothing in common. One was a flawed algorithm working from real, if incomplete, data. The other was a structural blind spot — an addition no consumer site, and often no AI, will ever see unless someone tells it. Type in an address and you can't tell which kind of wrong you're getting, or whether you're getting any at all.
Neither test ran the way the viral Florida story did. Both New Mexico properties came in low, not high — and for two completely different reasons. The size and direction of the miss isn't predictable, and that's the actual problem. You can't correct for an error you don't know is there.
If You're Going to Ask AI About Your Home's Value, Ask It This
You don't have to avoid these tools — I use them constantly. But if you're going to ask ChatGPT, or anything else, what your home is worth, don't stop at the first answer.
Watch what happens. In my experience, the tool often walks back its own confidence once you push on it — which tells you everything about how much weight that first number deserved.
Where AI Actually Helps, and Where It Doesn't
AI is genuinely useful for organizing information, summarizing comps, drafting the narrative around a pricing strategy. What it can't do — and what neither test property proved otherwise — is know which listing is stale, which square footage is wrong, which comp actually closed, or how to read a room once five offers are sitting on the table.
That last part is still, stubbornly, a people problem. The closing is earned, not calculated.
Curious what AI would say about your home — and what it would miss?
I'll run both numbers for you: the address-only guess, and the real one. No obligation, just a straight comparison.