Can AI Accurately Price a Home? What Agents Must Know
By Emily Terrell — Top Coach and Speaker at Tom Ferry International. Licensed since 2016. Closing 70+ deals/year while coaching agents nationwide.
AI can price a home within roughly 2% when that home is already listed — and roughly 7% when it isn’t. Zillow publishes both numbers, and the gap between them is the whole story. This guide breaks down why the off-market estimate is the one your seller is looking at, and how to run that conversation.
Key Takeaways
- Zillow reports a median error rate of 1.9% for on-market homes and 7.0% for off-market homes — and every pre-listing seller is looking at the off-market number.
- Accuracy is measured against the most recent estimate before sale, not the estimate your seller saw six months ago.
- In Texas, TREC rules bar license holders from using the word “value” for their own analysis and require a verbatim disclaimer on any estimated sale price.
- The winning move at the listing table isn’t arguing with the algorithm. It explains it better than the seller can.
What is an automated valuation model?
An automated valuation model, or AVM, is a computer program that estimates a property’s likely sale price from public records, tax data, prior sales, and market trends — without anyone physically inspecting the home. The Zestimate, the Redfin Estimate, and the institutional models lenders use are all AVMs. A Zestimate is not an official appraisal, and Zillow says so directly on its own accuracy page.
What every AVM has in common is pattern recognition. The model looks at what similar properties sold for, then extrapolates. It does not value your listing. It runs a match against history.
Why this matters for real estate agents
Here’s the thing nobody wants to tell you: the accuracy number your seller quotes and the accuracy number that applies to their house are two different numbers.
Zillow reports a median error rate of 1.9% for U.S. homes currently on the market, and 7.0% for homes that are not on the market. (Zillow, “How Accurate Is My Zestimate?”) That second figure is the one that matters, because a seller checking their estimate before they call you is — by definition — off-market.
Run the math on a $600,000 Stone Oak home. A 7% median error means the estimate could reasonably land anywhere from $558,000 to $642,000. That’s an $84,000 spread. If your seller anchors to the top of it, you’re negotiating against a number the algorithm itself isn’t confident in.
And “median” is doing heavy lifting. A 7% median error doesn’t mean estimates are within 7%. It means half are closer and half are further off.
“The seller isn’t wrong to look at the Zestimate. They’re just looking at the least accurate version of it. My job in the first ten minutes is to show them which number they’re holding — and then show them the one that actually applies.” — Emily Terrell, Tom Ferry Coach
How is AVM accuracy actually measured?
This is the part that reframes the entire conversation, and almost no agent knows it.
Accuracy is scored against the most recent estimate, not the original
Appraiser Ryan Lundquist has documented that Zillow computes accuracy by comparing the final sale price to the Zestimate on or before the sale date — not the estimate from before the home was listed. (Sacramento Appraisal Blog)
Why that matters: a home could list at $380,000, drop through a series of reductions, and sell at $350,000 — 8% below the original estimate — while the most recent estimate had already slid to $353,000, letting the reported error come in at 1%. (Sacramento Appraisal Blog)
The estimate moves toward the market once the market gives it something to move toward. That’s not a scandal. It’s a measurement choice — and it’s why the on-market number looks so much better than the off-market number.
What the model can see, and what it can’t
| What the AVM can see | What the AVM cannot see |
| Recorded sale price and date of comps | Seller concessions not captured in the recorded price |
| Public record square footage and room count | Unpermitted additions, conversions, or renovations |
| Assessed property tax information | Interior condition, finish quality, deferred maintenance |
| Geographic proximity to recorded comps | Micro-location factors — views, street noise, privacy |
| Historical price trend for the area | Market shifts inside the last 30 to 90 days |
| School district and basic neighborhood data | HOA restrictions or pending special assessments |
| Listed amenities from public fields | Staging, curb appeal, and buyer emotional response |
Every row on the right is a conversation you can have. Every one is a reason your judgment is the layer the algorithm was never built to supply.
Where AVMs break down worst
Thin comp pools are the common thread. Luxury tiers with three or four sales a year, custom homes with no true match, rural and low-turnover markets where the model reaches back eighteen months for a comparable, and markets that moved sharply in the last quarter. In all four cases the model is doing its best math on data that doesn’t support the question.
What are the Texas rules on quoting a home’s value?
If you practice in Texas, this section is not optional reading.
Under TREC rules, a real estate license holder may not perform an appraisal of, or provide an opinion of value for, real property unless licensed or certified as an appraiser. (22 TAC §535.17, via Cornell Law School) You can give an estimated sale price. You cannot call it a value.
When a license holder provides a broker price opinion, comparative market analysis, or estimated worth or sale price, they must also provide a written statement reading: “This represents an estimated sale price for this property. It is not the same as the opinion of value in an appraisal developed by a licensed appraiser under the Uniform Standards of Professional Appraisal Practice.” That statement must be part of the written analysis and reproduced verbatim in at least 12-point font. (22 TAC §535.17, via Cornell Law School)
It reaches your marketing too. TREC’s guidance on Rule 535.155 lists as a potentially misleading advertisement one including the value of a property, unless it is based on a disclosed appraisal readily available on request, or given in compliance with §535.17. (TREC, “TREC’s Advertising Rules — What You Need To Know”) So the “what’s your home worth” graphic you were about to post has a compliance dimension.
This is general information, not legal advice. Confirm your own advertising and disclosure practices with your broker or an attorney.
The five-step AVM conversation
Knowing this is worth nothing if you can’t deploy it under pressure. Here’s the system.
- Bring it up first. Don’t wait for the phone to come out. Pull the estimate yourself, early. That reads as confidence.
- Name which number they’re holding. On-market versus off-market. This is the single highest-leverage sentence in the appointment.
- Pull the comp set. Most platforms show at least partial comps. Walk through which ones don’t fit and why.
- Point at two or three specific gaps. The unpermitted casita. The greenbelt lot. The roof was replaced last spring.
- Position your analysis as the completing layer. Not a replacement. The part the model was never built to do.
Five steps, under three minutes, same every time. That’s not talent. That’s a system — the same way the rest of your business should run. (More on building repeatable systems around your listing data)
Where AI actually fits
AI tools like Claude and ChatGPT are not valuation tools. They’re research and communication tools. Use them to draft your client-facing AVM explainer, build a side-by-side comparison sheet, or script the five-step conversation in your own words. Do not use them to produce a price. A properly built comparative market analysis anchored in real local data is still the work.
Common mistakes
- Getting defensive. The second you sound like you’re attacking Zillow, you’ve lost the room.
- Quoting the on-market accuracy figure to a pre-listing seller. It’s the wrong number for their situation and a sharp seller will catch it.
- Skipping the comp set. Abstract explanations of methodology don’t move anyone. Their actual comps do.
- Using the word “value” in Texas. It’s a rules problem, not a style preference.
- Treating this as knowledge instead of a script. Knowing it and delivering it under pressure are different skills.
Frequently Asked Questions
How accurate is the Zestimate?
Zillow reports a median error rate of 1.9% for homes currently on the market and 7.0% for off-market homes. (Zillow) Median means half of estimates fall inside that range and half fall outside. Accuracy also depends heavily on how much public data exists in a given area, so thin or unusual markets perform worse than the national figure suggests.
Why is the off-market number so much worse than the on-market number?
Once a home lists, the model gains a major new input — the listing price and current market activity — and its estimate adjusts toward it. Accuracy is then scored against that adjusted figure rather than the original. Zillow computes accuracy by comparing the final sale price to the estimate on or before the sale date. (Sacramento Appraisal Blog)
Can AI replace a comparative market analysis?
No. AI can draft your explanation, format your comparison, and script your presentation. It cannot inspect a property, assess condition, or read a buyer pool. Current large language models are not valuation engines, and the AVMs that do produce estimates work from public data that misses condition, unpermitted work, and micro-location entirely.
Can a Texas agent tell a client what their home is worth?
Not in those words. A Texas license holder may not provide an opinion of value unless licensed as an appraiser, though they may give an estimated sale price accompanied by a verbatim written disclaimer in at least 12-point font. (22 TAC §535.17) This is general information, not legal advice — confirm your practice with your broker.
What should I say when a seller quotes their Zestimate?
Ask which tool they used, then show them whether they’re looking at the on-market or off-market estimate. Explain that the off-market figure carries a materially higher published error rate, walk them through the specific comps the model chose, and identify two or three property-specific factors it cannot see. Position your analysis as the completing layer, not a correction.
Do AVM estimates differ between platforms?
Yes, often by tens of thousands of dollars on the same property. Different providers use different data sources, comp selection logic, and calibration. Pulling two or three estimates side by side in a listing appointment is one of the most effective demonstrations available, because it shows the seller directly that these are modeled opinions rather than facts.
Bring this to your team or event
Emily Terrell speaks at brokerage events, real estate conferences, and team trainings on AI, systems, and social media — the exact playbook in this post, delivered live to your audience. As a Top Coach and Speaker at Tom Ferry International and an active agent closing 70+ transactions a year, Emily speaks from the stage about what’s working right now, not theory. Recent stages include NAHREP and eXp Con.
Book Emily to speak at your next event:
Email: eterrell@yourcoach.com
Phone: (210) 400-9191
Web: coachemilyterrell.com
For real estate agents who want to implement this: Get the weekly real estate prompt library at weeklyrealestateprompts.com or follow @coachemilyterrell on Instagram for daily systems and AI breakdowns.