An Algorithm Won't Fight for Your Offer

Aug 13

Every week I get some version of the same message: "Zillow says the house is worth $1.12M and they're asking $1.19M. Should we offer $1.1M?"

Lately there's a second version: "I asked ChatGPT and it said in this market we should offer 3% under list with a 17-day inspection contingency."

It's a fair question. And the number isn't crazy. But it's the wrong question, because it assumes the transaction is a math problem. In San Diego right now, it almost never is.

An automated valuation model is very good at one thing: telling you what similar homes recently sold for. That's genuinely useful. What it cannot do is get on the phone with a listing agent at 8:45pm on a Sunday night and find out that the seller's real problem isn't price at all. No amount of training data fixes that. It's that she's closing on a place in Mission Hills on the 20th and every offer she's holding wants a 45-day escrow.

That call is where houses are won.

What the model knows, and what it doesn't

An algorithm sees square footage, bed/bath count, lot size, and sold comps. Here's a partial list of things it does not see, drawn from deals I've worked in the last year:

The comp it leaned on hardest was a probate sale that closed 11% under market because the heirs lived out of state and wanted out.
The house is on the flat side of the canyon, so the "view lot" premium the model applied to the whole street doesn't apply here.
The listing has been sitting 38 days not because it's overpriced, but because the first photos were shot at 4pm facing west and the living room looked like a cave. New photos went up Tuesday.
Two of the three comps are in a Mello-Roos district and this one isn't, which is a real monthly difference the model flattens into nothing.
The seller's agent and I have closed four deals together and she knows my pre-approvals don't fall apart.

That last one has no field in any database. It has decided more than one of my clients' offers.

The ChatGPT version of this problem

The valuation sites have been around long enough that most buyers know to take them with a grain of salt. The newer thing is clients showing up with a full negotiating strategy from ChatGPT or a similar AI assistant. Honestly, some of it is decent. It'll explain contingencies clearly. It'll walk you through how an appraisal gap works. It's a patient tutor at midnight when you're too embarrassed to text your agent again.

Where it falls down is that it's answering from general knowledge about real estate, not from this market on this week. Ask it about San Diego and you'll get advice that's directionally reasonable and locally wrong: a 17-day inspection contingency because that's the default, when the winning offers on that street have been coming in at seven. Or a recommendation to "offer below asking in a slowing market," which might be right in Escondido and get you laughed off a listing in South Park.

It also can't verify anything. It doesn't know the listing agent. It hasn't seen the property. It can't tell you the seller already lost one escrow, and it can't call and ask.

So bring it. Use it to get fluent in the vocabulary and to pressure-test what I tell you. Just don't confuse a confident paragraph for a read on the actual situation.

The part nobody puts in a spreadsheet

Last spring I had buyers on a three-bedroom in Clairemont. Conventional financing, ten percent down, clean file, but ten percent down all the same. Six offers came in. Two were cash.

On paper they lose that every time, and any model would have told them so. What actually happened: I called the listing agent before we submitted and asked what the seller actually needed. Turned out she'd already had one escrow fall apart on that house six weeks earlier. The buyer got cold feet during the inspection period and walked with the deposit intact. She wasn't chasing the top number anymore. She was terrified of going back on market a second time.

So we didn't just send paperwork. We sent a ten-day inspection window instead of seventeen, a deposit large enough to have teeth, appraisal gap coverage up to a stated figure, and a lender letter with a direct cell number and a note saying call him tonight if you want. She did call him. My buyers came in under one of the cash offers and got the house.

No algorithm makes that call. No algorithm knows there was a call to make.

What "fighting for your offer" actually means

It's not shouting. It's mostly unglamorous work done before anyone sees a contract:

Finding out what the seller actually wants. Sometimes it's the highest number. Often it's a rent-back, a specific closing date, a shorter contingency period, or simply the buyer least likely to blow up escrow. You only learn this by asking a human being.

Structuring the offer around that. Price is one lever out of eight or nine. Deposit size, contingency timelines, appraisal gap language, who pays what. That's where a well-built offer separates itself from a well-priced one.

Defending the number after we're in contract. When an appraisal comes in low, somebody has to build the rebuttal: the comps the appraiser missed, the permitted work, the corrected square footage. Then somebody has to push it through. That's a person making a case, not a query returning a result.

Knowing when to walk. A model will happily tell you a house is a good buy. It won't tell you the seller has already refused two reasonable repair requests and is going to be a problem for the next 30 days.

I'm not anti-technology

I use the tools constantly. Automated estimates are a great sanity check. Mapping tools, permit records, price-history data, listing alerts that hit my phone faster than I could ever refresh a page. All of it makes me better at my job, and I'd be foolish to pretend otherwise. I use AI myself for the unglamorous parts: summarizing a 90-page HOA document, drafting a first pass at a listing description, organizing my notes after a long day of showings.

But there's a difference between a tool that informs a decision and a system that makes one. The tools are excellent at the first thing. San Diego is a market where inventory is tight, the coastal-to-inland spread is enormous, and a lot of the real information never touches the MLS. In a market like that, the second thing still runs on relationships and phone calls.

The estimate tells you where the conversation starts. Somebody still has to have the conversation.

The bottom line

If you're buying here, use the data if you want it. Pull the estimates, look at the price history, question my recommendations. And if you'd rather not do any of that, no problem. Some of my best clients showed up with a spreadsheet and a shortlist. Others showed up with a neighborhood they liked and a free Saturday. Sorting through the information is my job, not the price of admission.

Then understand what the data can't do for you. It can't read a seller. It can't build a relationship with the agent on the other side. It can't sit across from you at 9pm and tell you honestly that the house you love has a foundation issue you're going to regret.

And when your offer is one of seven on a Tuesday afternoon, it won't be the algorithm picking up the phone to make your case.

That's still my job.