A $949,000 Listing Doesn't Mean the House Is Worth $949,000
By The HomeHubAI Team · · 6 min read
There's a house we keep coming back to: 1573 19th Ave, San Francisco, in the Sunset. According to public records, five bedrooms, two bathrooms, 1,881 square feet, built in 1926. Same family since 1994. It listed on May 26 at $949,000. The listing came down on June 10, still at $949,000. Nobody bought it.
We first valued the house in August, two months later. Our algorithm looked at nearby sales and came back at $2.43 million. That's a wild gap with the ask, so we also ran it through an AI model we keep as a second opinion. The AI, looking at the same comps, said $2.35 million. Close enough that we weren't worried. Neither one had really dealt with the $949,000 list price that had already been on the market.
Both of those numbers are wrong. So is $949,000. The useful part is why.
The house was already on the market
The listing was not new information. It had been up in May and pulled in June. Our first pass in August just didn't use it. When we ran the house again and actually fed in the $949,000 ask, the two systems split. Similar homes nearby were still selling around $1,250 to $1,320 a square foot. Plug that into 1,881 square feet and you get something in the low $2 millions. The algorithm did that math and landed at $2.21 million.
The $949,000 list price was so far below those sales that the old code mostly shrugged. It stretched the range down to $920,000 and knocked the confidence score. The point estimate barely moved. A range from $920,000 to $2.43 million is another way of saying we don't know.
The AI went the other way. We'd given it a rule: don't go more than 15% above an active list price without a good reason. It treated $949,000 as the seller knowing something the public record doesn't. Original condition, deferred maintenance, a house that hasn't been opened up in 30 years. So it parked the estimate on the asking price, in a tight band from $880,000 to $1.02 million, and it was more confident than the algorithm was. That would have been a stretch even on a live listing. This one had already been pulled.
What neither of them had was how list prices actually behave in that neighborhood.
List price is a starting bid
In a lot of San Francisco, the asking price is a marketing floor. Sellers price low to start a fight. We checked 21 recent sales from the same pool as this house. The median sale closed at 1.5 times the list price. Only 3 of 27 sold within 5% of what they asked.
Apply that to $949,000 and you get something closer to $1.43 million. Which is a very different number from either of the ones we had, and a better read than treating an ask that already came off the market as the value.
The AI never saw that, because we never showed it the list prices of the comps. It saw what they sold for. It saw what this house was asking. It had no way to know that Sunset sellers routinely list low on purpose.
That's what we missed, and we would have kept missing it if the two systems had stayed close. A $1.25 million gap on the same house is hard to ignore. We went and counted.
What we changed
We now measure the sale-to-list ratio on the cleaned comp pool and move the estimate partway toward list price times that ratio. The ask still matters. We just convert it into what similar homes actually sold for. A listing that hasn't had a status change in months gets less weight, because a stale ask is weaker evidence. This one had been down since June.
We also apply a modest discount when a home hasn't had a real sale in 15 years or more and we have no other condition data. This one has been in the family for 32 years. The comps around it are often renovated. Treating them as interchangeable is how you get to $2.2 million on a house that may not have been touched since the nineties.
On this house, that lands here:
- Algorithm, before the fix: $2.21 million, range $920,000 to $2.43 million
- AI second opinion: $949,000
- Algorithm, after the fix: $1.85 million, range $1.68 million to $2.02 million
Confidence went up too, because the range stopped pretending the house might sell for either the ask or a renovated comp. You can read the current report.
This is still an estimate
It is not an appraisal. Someone who walks through the house will know things we don't. If the interior is a wreck, $949,000 might turn out to be the right number. It might also be a classic Sunset underprice and sell for well over a million.
We're also not done. The algorithm now knows that a Sunset list price is a starting bid. The AI second opinion still doesn't, because those per-comp list prices are not in its prompt yet. That's the next fix. The point of keeping the two systems side by side is that one of them will eventually say something the other can't defend.
If you're pricing a home, especially in a market where people list low on purpose, look at what nearby houses actually closed for relative to their asks. The list price is a starting point. It is not the value.
Want to see what we think your house is worth? Run a free valuation.