Why AI Doesn’t Know What Just Sold in Your Neighborhood (And What You Need Instead)

Try this right now. Open ChatGPT and ask it, “what houses sold on my street in the last 90 days?” Watch what happens. It will confidently invent transactions. Wrong addresses, wrong prices, wrong dates, sometimes even wrong buyer types (“a cash investor purchased your neighbor’s home for $812,000 in March”), for sales that never happened. The numbers look real. The houses don’t exist.

For a buyer this is annoying. For a seller about to set a list price, it’s expensive. I’m Ed Neuhaus, I’m a broker here in Austin and I’ve been doing this for 19 years. Lets talk about why AI cannot see what just sold in your neighborhood, and what you actually need if you’re going to use AI to help you price your house.

Why This Matters for Pricing

Pricing a house is three things stacked on top of each other. Comps (what actually sold), market velocity (how fast things are moving right now), and buyer psychology (what the next buyer through the door is willing to pay). Get the comps wrong and the other two don’t matter, because you’re already starting from a number that has nothing to do with reality.

And comps are not “what is for sale on Zillow right now.” Comps are what closed. Past tense. A house listed at 850k that’s sat on the market for 70 days tells you almost nothing. The house that listed at 799k and closed at 785k in 11 days, that tells you everything.

If your comps are off by 10% on a 700,000 dollar house, you’re talking about 70,000 dollars. Price too high and you sit. The market punishes overpriced listings by ignoring them, then by forcing price drops, and price drops show up in the listing history forever. Price too low and you literally hand money to the buyer. Both mistakes start the same way, with bad comp data.

Why AI Gets Comps Catastrophically Wrong

Here is the part nobody at OpenAI or Anthropic will tell you on the home page. The AI cannot see the MLS. Period.

A few reasons for this. First, training cutoffs. Even the freshest LLM is months behind the actual day. By the time a model “knows” the world, three quarters of the comp window you care about has already passed. Second, no live MLS feed. Closed sale data lives in the MLS, behind a broker membership, behind rules, behind a license. Public sites like Zillow and Redfin show you some of it, but they show you what they’re allowed to show you, and the AI models were trained on whatever happened to be sitting on the public internet when somebody scraped it.

Third, and this is the worst one, a lot of what AI has absorbed about home values came from public automated estimates like Zestimates. Zestimates are Zillow’s automated guess at what a house is worth, and Zillow publishes its own median error rate (in the low single-digit percent range for on-market homes, higher for off-market). So you’ve got an AI trained on guesses, generating guesses about other guesses. That’s not pricing, that’s a game of telephone.

Fourth, and this one is sneaky, the AI cannot tell you it doesn’t know. It just generates something. Ask it “what closed at 1234 Oak Street last month” and it will produce an answer with a price and a date and a tone of total confidence. None of it has to be true. Daniel Kahneman called this the “what you see is all there is” problem (his whole point in Thinking Fast and Slow was that confident answers feel true even when the underlying data is garbage). LLMs do that for a living.

What Real Comp Data Looks Like

When I pull a real comp set for one of my sellers, here’s what the MLS actually gives me on each closed sale:

• Address, square footage, beds, baths, lot size, year built
• Original list price and any price changes along the way
• List date, contract date, close date, days on market
• Close price, the actual number the title company recorded
• Seller concessions, repair credits, closing cost help
• Type of buyer financing (cash, conventional, FHA, VA, jumbo)
• Property condition notes, updates, features that did or didn’t drive value

That’s a real comp. That’s what I pull in about 30 seconds from the MLS. That is what AI is making up out of thin air when you ask it the same question.

And the buyer financing piece matters more than people think. A cash close at 750k and an FHA close at 750k look the same on the outside, but the cash buyer probably negotiated harder and got a real number. The FHA buyer might have had 15,000 dollars in seller concessions baked in, so the “real” net to the seller was 735k. AI does not know any of this. The MLS does.

What I Built. Live Comps for AI.

So here’s where I went a little nerdy. I’m a realtor with broker-level access to the Austin MLS. I sit on a license that lets me query the real data every day. And I got tired of watching AI invent comps, so I built something.

It’s called the Austin MLS MCP. MCP stands for Model Context Protocol, it’s just plumbing that lets an AI ask my MLS database a real question and get a real answer back. The same data we already display on NeuhausRE.com for the public, delivered programmatically so an AI can use it. Not a data subscription, not a feed for sale. Same MLS data, just wired up so Claude or another model can actually see it instead of guessing.

What that means in practice. I can now ask Claude “show me all 4-bedroom closed sales within half a mile of 1234 Oak Street, in the last 90 days, between 2,200 and 2,800 square feet” and Claude pulls a real list. With real addresses, real close prices, real days on market. The AI stops hallucinating because it stopped having to guess.

If you want the longer story of how this got built, I wrote that up over here. We just connected Claude to the live Austin MLS.

A Side by Side. AI Without vs With Live Comps.

Picture a 4-bed, 3-bath, 2,400 square foot house in Bee Cave. Owner is thinking about selling. They want a quick gut-check on price before they call an agent.

AI without MCP. They ask ChatGPT. It produces three “comparable” sales. Two of the addresses don’t exist on Google Maps. One is real but it was a 5-bedroom on twice the lot. The model lands on a “fair market value” that just feels confident. The owner gets excited. They tell the spouse. They start mentally spending the money.

AI with MCP. The same question, run against the actual MLS. Claude pulls real recent closings within half a mile, filters down to true 4-bed comps in the right square footage range, and reports back the median close price, the median days on market, and which ones had seller concessions baked in. The number that comes out is a defensible range, not a vibe. With notes on what’s actually moving and what’s sitting.

The delta between those two answers, on a house in that price range, can easily be six figures. That’s not a rounding error. That’s a year of college, a year of someone’s retirement, a real number that vanishes if you trusted the first answer.

For Sellers: What to Do Instead of Trusting AI Alone

Ok so what should you actually do? Three things.

One. Get a real CMA from a real agent. Not a Zestimate, not a “what would you sell my house for” line from ChatGPT, not the number your neighbor’s cousin said over Thanksgiving. A real comparative market analysis, pulled from the MLS, by somebody who knows your zip code. That’s a free service most of us offer. Here’s ours.

Two. If you want to use AI tools, make sure they’re wired to live MLS data. Otherwise you’re paying for confidence, not accuracy. And confidence priced wrong is worse than no answer at all.

Three. Do both, and compare. Run your house through ChatGPT, then run it past a real CMA. If they’re within a few percent of each other, you’ve probably got a defensible range. If they’re 100k apart, the AI alone is wrong and you now know it. That’s a useful experiment. We have a whole post on this if you want to see the experiment play out. The ChatGPT home value test.

What I’d never do, and what I’m watching people do every week now, is set my list price based on what ChatGPT told me. That’s how houses sit. That’s how price drops happen. That’s how you end up trying to “chase the market down” three months in.

For Agents Reading This

If you’re a realtor and you’ve been watching all of this AI stuff from the sidelines, here’s where I’d start. The Austin MLS MCP works with Claude, ChatGPT Developer Mode, Cursor, Cline, Gemini CLI, and Perplexity. Different tools for different workflows. The setup is the same idea every time, you wire your AI to live data instead of guesses. I have step-by-step guides for each tool linked off the Austin MLS MCP page. Pick whichever tool fits how you already work.

The version of “AI in real estate” that gets sold to agents on Instagram is mostly fake. Generated listing descriptions, auto-replies, content factories. That’s not where the value is. The value is having an AI that can actually answer “what just sold” the right way, and that requires real data, not better prompts.

The Realtor’s Take

AI as your sole pricing source is an expensive mistake. Not because AI is bad, but because the AI you’re using right now is blind to the only data that matters. It cannot see the MLS. It will generate something anyway. And that something will sound right until you list at the wrong price and the market tells you otherwise.

AI with live MLS data plus a human who knows the market, that’s a different conversation. That is genuinely a better way to price a house than what most sellers got 10 years ago. Faster, more complete, less hand-wavy. The future of selling isn’t AI or a realtor, it’s both, working off the same real numbers.

If you’ve been using AI to figure out what your house is worth and the number it gave you felt off, it probably is. Pull a real CMA and find out.

Want a Real Price for Your House?

If you’re thinking about selling in Austin, Westlake, Lakeway, Bee Cave, Spicewood, or anywhere in the Hill Country and you want a real number (not a Zestimate, not a ChatGPT guess, an actual CMA pulled from live MLS data) we’d love to put one together for you. No pressure, just real numbers.

Get in touch with Neuhaus Realty Group, or if you want to nerd out on the AI side of how we built this, read about the Austin MLS MCP. If you’re earlier in the journey and just trying to learn how AI fits into buying or selling, here are the two pillars. How to buy a house with AI and how to sell a house with AI. For the buyer-side version of this same data problem, see why AI doesn’t know what’s actually for sale. And once you’re under contract, here’s how AI can help write your listing description and how it can help you read buyer offers.