The seller turns their phone around at the kitchen table and shows you a number. It is higher than yours. Their whole posture says they have already decided which of the two is real, and it is not the one you spent four hours building.
Most agents lose this conversation in the first ten seconds, and they lose it the same way: by attacking the number. Those things are never right. The seller hears you dismissing the only independent-looking opinion in the room, and now you are the salesperson arguing down their equity.
There is a better move, and it costs nothing. Zestimate accuracy is not a mystery you have to argue about from memory. Zillow publishes it. Two numbers, on their own website, and one of them settles the conversation without you having to attack anything.
The two numbers Zillow publishes
| Home type | Median error rate, nationwide |
|---|---|
| On-market homes | 1.83% |
| Off-market homes | 7.01% |
Nearly four times the error, and almost nobody notices there are two figures. The one that gets quoted in headlines and repeated by sellers is the first one, because it is the flattering one.
But the seller sitting in front of you has a house that is not for sale yet. Their house is off-market. The number that applies to them is 7.01%, and they have almost certainly never seen it.
Why the gap is so wide is not complicated. Once a house is listed, the model can see the asking price, the photos, the description and how the market responds to it. Before it is listed, the model is working from public records and whatever the owner typed in. It is guessing about condition, finishes, layout and every improvement nobody filed a permit for.
The word doing the heavy lifting is “median”
This is the part to slow down on, because it is where the real force of the argument is and most agents skip past it.
A median error of 7.01% does not mean every home is off by about 7%. It means half of homes are off by less than that, and half are off by more. Sometimes much more. There is no ceiling on that top half.
On a $450,000 house, 7.01% is roughly $31,500 in either direction, and that is the midpoint, not the worst case. A coin flip decides whether this specific house is one of the better estimates or one of the worse ones. Nothing on the screen tells the owner which.
That framing matters because it is not an attack. You are not saying the tool is bad. You are saying it is a statistical estimate behaving exactly as advertised, and a coin flip is a poor basis for the largest financial decision most people ever make.
Zillow already agrees with you
The single most useful thing in this conversation is that you do not have to be the one making the case. Zillow’s own wording does it. On their site, the Zestimate is described as “a useful starting point”, and they state plainly that it “is not an appraisal and should not be used as a substitute for one”.
Nobody at Zillow is claiming this is a valuation. They built a starting point and they say so. The only people treating it as a final answer are the ones reading it on a phone.
What the model cannot see
If a seller pushes back — why would it be off by that much? — the answer is concrete, and it is more persuasive than any statistic.
An automated estimate is built from what is written down: tax records, prior sales, square footage, bed and bath counts, lot size, and whatever the owner has typed into the listing record. That is a real dataset and it gets you most of the way on a street of identical tract houses.
It gets you nowhere on everything that decides what a buyer actually pays:
- Condition. Two houses with identical records, one with a kitchen from 1994 and one renovated last spring. Same inputs, same estimate, forty thousand dollars apart.
- Work done without a permit. Common, invisible to public records, and often the most valuable thing in the house.
- Layout and light. Square footage does not distinguish a good floor plan from a chopped-up one, or a south-facing living room from a dark one.
- What is next door. The commercial lot, the road noise, the neighbor who parks a boat on the lawn. None of it is in the file.
- What actually happened in the last sale. A recorded price of $460,000 with $15,000 in seller concessions is not the same trade, and the model usually cannot tell.
Say it as a question rather than a lecture: has anyone from Zillow been inside this house? The seller answers it themselves, and the point lands without you having criticized anything.
What to actually say
Do not argue. Agree, then add the information they are missing.
Good, I’m glad you looked. That’s a real starting point and I use them too.
One thing worth knowing, and it’s on Zillow’s own site: they publish two accuracy numbers. For houses already listed, they’re within about 1.8%. For houses not on the market yet — like yours today — it’s 7%. And that’s the middle, so half of homes are off by more than that.
On a house at this price, 7% is about thirty thousand dollars in either direction. So the question isn’t whether their number or my number is right. It’s that theirs is a range and mine is a specific house, and here’s what mine is built from.
Then show your work. Not a printout of comps: the reasoning. Which three sales you chose and why, what you added and subtracted for, and where you decided to land in the range. If your CMA cannot survive being explained out loud, the problem is not the Zestimate. Building that narrative is the one part of a valuation these tools are genuinely good at, which is worked through in the piece on where AI helps with a CMA and where it does not.
When the Zestimate is lower than your price
It happens as often as the other way round, and agents handle it badly in the opposite direction: they get quiet and hope nobody mentions it.
Raise it yourself, before the buyer’s agent does. The same 7% applies in both directions, and a seller who hears you explain it while it is working in their favor will believe you later when it is not. It also inoculates them for the moment a buyer arrives quoting that low number as justification for an offer.
The thing to do before the appointment
- Look the house up yourself. If you walk in without knowing what the seller has already seen, you are reacting instead of leading.
- Check the home facts on the listing record. Wrong bedroom count, wrong square footage, a finished basement that is not recorded. Owners can update these, and a genuinely wrong fact is the fastest way to show the estimate is working from bad inputs — without insulting anything.
- Know your local number. National figures are averages of wildly different markets. Zillow publishes accuracy by area; unusual housing stock and thin comparable sales make it worse.
- Have the sentence ready. “They publish two numbers and yours is the second one” only lands if it comes out cleanly. Rehearse it once in the car.
That last one generalizes. The moments that decide a listing appointment are four or five specific sentences, and the piece on prompts for the listing presentation covers rehearsing the rest of them.
Common questions
Is the Zestimate just bad?
No, and saying so damages your credibility. Within about 1.8% on listed homes is genuinely good for an automated model with no access to the inside of the house. The problem is not the tool, it is the gap between what it is and what people assume it is.
Can a seller change their Zestimate?
They can claim the home and correct the recorded facts, and if those facts were wrong the estimate can move. It is worth doing when the record is genuinely inaccurate. It is not worth doing as a way to manufacture a higher number to justify a price, and that motive is obvious to everyone including the buyer’s agent.
Should I show sellers the accuracy figures?
Yes, on Zillow’s own page rather than in a slide you made. The credibility comes from it not being your document. Pull it up on your phone, let them read it themselves, and say very little while they do.
Does an AI valuation tool do better?
Not in a way that helps you here. Any model working from public records has the same blind spot about condition. And a general chatbot is worse, not better, because it cannot see sold data at all — the reasons are in the piece on whether ChatGPT can reach MLS data.
What if the seller insists on the higher number?
Then you have a pricing conversation, not a technology conversation, and the honest version of it is that overpricing costs them the first three weeks of buyer attention they will never get back. Whether you take the listing anyway is a business decision. Just do not let an automated estimate be the reason you took it.
Sources
- Zillow, published Zestimate median error rates: 1.83% for on-market homes and 7.01% for off-market homes nationwide, and Zillow’s own description of the Zestimate as a starting point that is not an appraisal.
Checked August 2026. Zillow updates these figures; check the current numbers and your own state before quoting them to a client.