Ed 
I am a realtor. I have been one for 19 years. And last week I spent seven days pretending I was a first-time buyer relocating to a city I did not know, with no agent, and only ChatGPT as my assistant. I wanted to see what would happen.
The short version. ChatGPT was genuinely useful for about 40% of what a buyer needs. For the other 60%, it lied to my face with the confidence of someone who absolutely knew what they were talking about. The scary part is, if I had not been a realtor, I would have believed almost all of it.
Let me walk you through what actually happened. The good, the bad, and the moment I sat at my desk and said out loud “oh no, people are going to lose money over this.”
This post is part of a series I have been writing about how to use AI to buy a house (and the sister post on selling with AI if you are on the other side of the table). This one is the experiment. The “what happens if you actually try it” post.
Where ChatGPT Genuinely Helped
Lets start with the wins, because they are real. I went into this thinking ChatGPT was going to be mostly hype. I came out of it thinking it is one of the best research partners a buyer has ever had access to. Just not for the reasons people think.
The first thing I asked was about neighborhoods. I picked a market I do not work in (Charlotte, North Carolina) and asked ChatGPT to compare three neighborhoods I had heard of. Not “which is best”, because that is a useless question, but “tell me the personality of each, what kind of person lives there, what the school philosophies are like, what the walkability situation is.” The answers were genuinely good. It described the vibe. It talked about which areas were more walkable and which were car-dependent. It noted the school philosophy differences in a way that sounded like it had actually read the district websites. For neighborhood research at the vibe level, ChatGPT is excellent.
Mortgage language was the next win. I asked it to explain PMI, points, DTI, and what “locking the rate” actually means in plain English. It nailed all of it. Better than most loan officers, honestly. (I know that sounds harsh but if you have ever sat across from a buyer trying to explain why their DTI is 43% and they cannot qualify, you know what I mean.) ChatGPT is a translator. It takes industry jargon and turns it into English. That is genuinely valuable.
I had it draft me a showing checklist. Questions to ask the listing agent. Things to look for in the attic, the panel, the water heater. It was solid. Not as good as walking the house with someone who has seen 2,000 houses, but better than walking in with no plan at all.
The one that surprised me most was contracts. I pasted in a standard purchase contract (with the names redacted) and asked ChatGPT to explain each section. It did a great job. It flagged the option period. It explained the financing contingency. It told me what an HOA addendum was. For someone reading a contract for the first time, this is enormous. I have had buyers sign contracts they did not fully understand for 19 years. ChatGPT closes that gap.
And then there was the must-have vs nice-to-have conversation. I told it about my fake situation (relocating, two kids, work from home half the week) and it walked me through which features actually mattered for my use case versus which were emotional purchases. It was thoughtful. It was patient. It was the kind of conversation I have with my buyers, except ChatGPT has unlimited time and I do not.
So yes. For research, education, translation, and thinking partner work, ChatGPT is a real tool. I am not anti-AI. I built an MLS connector for Claude because I think this technology is the future of real estate. But here is where it gets ugly.
Where ChatGPT Lied to Me with Full Confidence
The wheels came off the minute I asked anything about actual listings, actual prices, or actual current data.
First test. “What homes are for sale right now in the 28210 zip code under $500,000?” ChatGPT gave me five addresses. They came with bedroom counts, bathroom counts, square footage, and prices. They looked like real listings. I went to check them on the MLS. None of them existed. Not one. Two of the addresses were not even real street addresses. One was a property that had sold in 2019 for a completely different price than what ChatGPT quoted. It was not a search result. It was fiction, dressed up to look like data.
This is not new. A 2024 study found that more than half of ChatGPT’s references in academic queries were fabricated or contained errors, and the underlying problem is the same. Large language models are pattern matchers, not databases. When you ask one for current MLS data, it does not say “I do not have that.” It generates something that looks like the answer.
Second test. I picked a real house, a real address I knew was on the market, and asked ChatGPT for comparable sales. It gave me six comps. Confident, specific, dollar amounts to the thousand. I checked them against the MLS. Three did not exist. Two existed but had sold for different prices than what ChatGPT reported. One was real and roughly accurate. So out of six, I got one usable data point. The rest was hallucination wearing a suit.
Third test, and this is the one that got me. I asked “is this house a good deal at the asking price?” ChatGPT, working from its hallucinated comps, told me confidently that the house was overpriced by about $40,000 and I should offer below list. The problem? I know that market. I know that house. It is actually slightly underpriced based on real comps. If I had been a real buyer trusting that answer, I would have offered low, lost the house to a smarter buyer, and never known why.
Fourth test. HOA fees and short-term rental rules. I asked about a specific neighborhood. The answer was confident, detailed, and wrong. The HOA fee was off by about $200 a month. The STR rules ChatGPT described did not match what the HOA actually allows. ChatGPT had absorbed information from somewhere on the internet, probably a forum post from a few years ago, and was serving it back to me as current fact.
The pattern here is not subtle. Anything that requires current, local, verified data is a danger zone. ChatGPT does not know what is on the MLS today. It does not know what just sold. It does not know what the HOA voted on last month. But it will tell you anyway, because it is designed to give you an answer.
The Moment That Made Me Glad I’m a Realtor
The “is this a good deal” test is the one that haunted me. I want to slow down on it for a second because I think it is the whole game.
Here is what ChatGPT told me, based on its hallucinated comps. “This home appears to be overpriced relative to recent sales in the neighborhood. Similar 4-bedroom homes in the area have sold between $445,000 and $475,000 in the past six months. The asking price of $525,000 is approximately $50,000 above the typical sale price. I would recommend offering between $460,000 and $475,000.”
Sounds professional right. Sounds like advice you would pay for. The numbers are specific. The reasoning is logical. The recommendation is concrete.
Here is what was actually true. Recent sales in that neighborhood for that floor plan were $515,000, $530,000, and $545,000. The asking price of $525,000 was below the middle comp. A reasonable offer was full price or maybe $5,000 below. An offer of $465,000 would have been ignored. The house would have sold to someone else within a week. And the buyer who trusted ChatGPT would have spent another three months looking, watching prices keep going up, and eventually overpaying for a worse house in a worse location.
That is the cost of confident wrong information in real estate. It is not “you got a worse outcome.” It is “you got no outcome and the market moved away from you while you were arguing with the listing agent.” (Kahneman’s whole thing about overconfidence applies here. The confident voice on the screen is the most dangerous part. If ChatGPT had said “I am not sure, you should verify this with a local agent,” I would not be writing this post. But it does not say that. It just tells you.)
What This Means If You’re Actually Buying
Lets get practical. Here is what I would tell my own buyers about using ChatGPT during their search.
Use it for the stuff that does not change day to day. Neighborhood personality. Mortgage terminology. Contract language. Inspection checklists. School philosophies. The 30,000-foot stuff. ChatGPT is great at that and you should absolutely lean on it.
Do not use it for anything time-sensitive. Current listings. Current prices. Current HOA rules. Current property tax rates. Current STR ordinances. Current comps. If the answer changed in the last 12 months, ChatGPT is probably wrong, and worse, it will not tell you it is wrong. It will give you an answer that sounds right.
Ground your AI in real data when you can. If you paste in actual MLS data, actual tax records, actual HOA documents, ChatGPT can analyze them beautifully. The problem is not analysis. The problem is that ChatGPT does not have access to current data unless you give it to the model directly. Garbage in, garbage out becomes nothing in, fiction out.
And honestly, work with a realtor. Not because I am a realtor and I want your business (although I do, and you can reach out here if you are in Central Texas). Work with one because the entire job of a buyer’s agent is to be the verified data layer that ChatGPT cannot be. We know what is on the market. We know what just sold. We know what the HOA does. We pay for the systems that have the real numbers. ChatGPT is a research assistant. It is not a data source.
The Fix I Built
I hit this wall hard during the experiment, and I am not the only one. So I built the Austin MLS MCP. It is a server that connects Claude (and through Claude, ChatGPT and other AI tools) directly to live MLS data in Central Texas. When you ask the AI about listings or comps, it stops guessing and starts pulling from the actual database. No hallucinations. No fake addresses. Real numbers.
The Austin MLS MCP page has the technical details if you want to dig in. I have a follow-up post coming about why this matters for buyers, but the short version is, this is the only way AI becomes useful for actual home shopping instead of just sounding useful. The data has to be live.
My Realtor’s Takeaway
A week of pretending to be a buyer using ChatGPT taught me something I did not expect. AI is going to make smart buyers much smarter. It is also going to make naive buyers much more confidently wrong. Those are not the same thing, and the gap between them is going to widen.
The buyers who win with AI are the ones who use it for what it is good at (research, education, translation) and who get human verification for what it is bad at (anything current, local, or transactional). The buyers who lose with AI are the ones who trust the confident voice on the screen without checking. The technology does not warn you which mode you are in. That is on you.
If you are a buyer, the takeaway is simple. ChatGPT is a research partner. It is not a source of truth. Use it accordingly and you will be miles ahead of the buyer who does not use it at all. Treat it like a Wikipedia article that occasionally makes things up, and you will be fine.
I am still bullish on AI in real estate. I would not have spent the last year building MCP connectors and writing posts like this if I was not. But the version of AI that helps buyers is the one that knows what it does not know. Right now, ChatGPT does not. That is the gap. And until that gap closes, you still need someone like me sitting next to you when the numbers actually matter.
Frequently Asked Questions
Want a Realtor Who Actually Uses These Tools?
If you are buying in Central Texas and you want to work with someone who has actually tested every AI tool in the market and knows what works, reach out to Neuhaus Realty Group. I built the Austin MLS MCP because I believe AI is the future of how buyers and sellers work with their agents. But the future only works if the data is real. That is what we do. Real data, real analysis, real outcomes. Selling instead of buying? Check the ChatGPT home value experiment sister post.
