The Zestimate is the most-cited home valuation in the United States. It shapes seller expectations, anchors buyer offers, and shows up in mortgage conversations and divorce filings. Despite that, very few people who quote it can tell you when it launched, how it actually works, what its published error rate is, or what Zillow itself says about it. This article is a factual walk through all four — sourced to Zillow's own disclosures and to public reporting — and an honest explanation of why Zeego's property estimate is built differently.
A note on tone before we start: nothing here is a claim that Zillow has done anything wrong. Zillow has been more transparent about its model's accuracy than almost any other AVM provider. The point of this article is that even with that transparency, a single AVM is a single estimator — and single estimators are, mathematically, beaten by ensembles. That's the case Zeego makes.
A short history of the Zestimate
Zillow launched in beta on February 8, 2006, founded by former Expedia executives Rich Barton and Lloyd Frink. The product's defining feature on day one was the Zestimate: a free, automated valuation for roughly 67 million U.S. homes, displayed on a map alongside aerial imagery. Nothing like it existed for consumers. Real estate agents had access to MLS comparables; appraisers had their proprietary tools; banks had internal AVMs. Putting a number on every house in America, for free, was a genuinely new product.
The model has been revised many times. Zillow exited beta and shipped its first major Zestimate accuracy improvement in early 2008. In 2017, the company launched the Zillow Prize on Kaggle — a $1 million public competition to improve the Zestimate's algorithm. The winning team, announced in January 2019 after nearly two years and roughly 3,800 competing teams from 91 countries, beat Zillow's benchmark by about 13%, and Zillow incorporated those gains into the production model. In 2021, Zillow rolled out a 'neural Zestimate' — a deep-learning rebuild that the company said improved the model's responsiveness to fast-moving markets and reduced median error.
The Zestimate, in other words, is not a static algorithm. It has been actively improved for two decades. That's worth crediting.
What the Zestimate actually is
Per Zillow's own published documentation, the Zestimate is an automated valuation model that uses public records, MLS data when available, prior sale and listing history, tax assessments, and user-submitted home facts. It is not an appraisal, and Zillow says so explicitly on every property page and in its help center. It produces a point estimate and, on most homes, a range.
The model is statistical, not physical. It does not see the inside of your house. It does not know whether your kitchen has been remodeled in the last five years or in the last fifty. It infers condition from data — when a home last sold, how it was described in past listings, what comparable homes in the same area sold for. On homes where those signals are strong, the Zestimate is genuinely useful. On homes where they're weak or contradictory, it can be far off.
How accurate is it, really
Zillow publishes its accuracy figures, and the figures are the right place to start any honest discussion. The two numbers Zillow reports are median error rate and the share of estimates within 5%, 10%, and 20% of the eventual sale price. Both are reported separately for on-market homes (currently listed) and off-market homes (everything else).
Per Zillow's published data, the nationwide median error rate is roughly 2% on on-market homes and roughly 7% on off-market homes. The share of Zestimates within 5% of sale price is meaningfully higher on listed homes than on unlisted ones. Those numbers move over time and by metro, so the version you see on Zillow's site today is the one to trust — but the structural pattern holds: the Zestimate is much more accurate on a home that's actively for sale than on one that isn't.
Why? Because once a home is listed, the model gets a powerful new input: the list price. That price reflects the listing agent's read on the local market — current inventory, buyer demand, recent neighborhood sales — that an AVM can't fully reconstruct from public records alone. The Zestimate is best, in a sense, when it has the least work to do.
Translated into dollars on a $1.2 million home, a 2% median error is $24,000. A 7% median error is $84,000. Median means half of estimates miss by more than that. These are not rounding errors on the buyer's side of the table.
The Zillow Offers episode
The most public stress test of any AVM in the last decade was Zillow Offers, the company's iBuying business. From 2018 through late 2021, Zillow used algorithmic valuations (closely related to, but not identical to, the consumer Zestimate) to make cash offers on homes, then resold them. In November 2021, Zillow shut Zillow Offers down and wrote down roughly $304 million on inventory in Q3 alone, ultimately announcing thousands of layoffs and a wind-down of the homes it had bought.
In Zillow's own communications and in subsequent reporting, the company attributed the failure to its model's inability to forecast home-price appreciation accurately enough in a fast-moving market. It is a fair, and important, distinction that Zillow Offers used a different pricing model than the public Zestimate. But the broader lesson is the same one the academic literature on AVMs has been making for years: any single algorithm, no matter how well-funded, has blind spots. Zillow had more data, more engineers, and more financial incentive to get pricing right than any other AVM operator on earth, and the model still couldn't price homes accurately enough to run a profitable iBuying book.
If a single AVM wasn't enough for Zillow's own checkbook, it isn't enough for yours.
How the Zestimate moves the market
The Zestimate's influence isn't theoretical. Surveys consistently find that a majority of home shoppers consult Zillow during their search, and a substantial share of sellers cite the Zestimate when forming their pricing expectations. That creates a quiet feedback loop: the Zestimate predicts what a home will sell for, sellers anchor on it, buyers anchor on it, and the eventual sale price is then pulled toward the prediction. This is not a controlled experiment, but the anchoring effect is well-documented in behavioral economics, and there is no plausible reason real estate would be immune.
Zillow is upfront that the Zestimate is not an appraisal and should not be used as one. Lenders don't underwrite from it. Appraisers don't pull comps from it. But buyers and sellers, in practice, treat it as a starting point — and starting points matter. An offer written off a high Zestimate overpays. An offer written off a low one loses to a competing bid.
Can a homeowner influence their own Zestimate?
Yes, partially, and Zillow publishes how. The official paths are documented in Zillow's help center and consumer guides, and they're worth knowing — both as a homeowner and as a buyer who should expect that the seller across the table has used them.
- Claim the home on Zillow and update home facts. Bedrooms, bathrooms, square footage, lot size, year built, and recent renovations can all be edited by a verified owner. Corrections are reflected in the model's inputs.
- Add or update photos. Zillow's model uses listing imagery as a signal; better and more recent photos affect how the home is read.
- List the home for sale. Once a home is listed, the model incorporates list price and shifts toward the on-market accuracy band.
- Make sure the public record is correct. Tax-roll errors — wrong square footage, wrong bedroom count, missing permits on additions — flow into the Zestimate.
What an owner cannot do, despite a long-running internet myth, is simply 'set' their Zestimate to a number. Edits to home facts feed the model; they don't override it. But the practical effect of cleaning up a stale record on a renovated home can still be material — sometimes tens of thousands of dollars on the displayed estimate.
For buyers, the relevant takeaway is the inverse: the Zestimate you see on a listed home reflects whatever the seller has chosen to share or not share. That isn't dishonesty — it's how the product works — but it's a reason not to treat the number as an objective valuation.
Legal and regulatory context
The Zestimate has faced legal challenges over the years, including a 2017 suit filed in Illinois that argued the estimate functioned as an unlicensed appraisal. That case was dismissed at the federal-court level, and Zillow's position — that the Zestimate is an estimate, clearly disclosed as such, not an appraisal — has held up in court. The product continues to ship today on roughly 100 million U.S. homes with the same disclaimer language. Nothing in this article disputes any of that.
Why one AVM is the wrong tool, even at its best
Set Zillow aside for a moment. The structural problem is the same for every single-source AVM — Redfin's Estimate, CoreLogic, HouseCanary, Quantarium, ATTOM, anyone else. Each is a model. Each model has training data, feature weights, and assumptions. Each disagrees with the others on most homes, and not at random — the disagreements track real differences in how each model handles condition, comp selection, and market velocity.
The mathematical literature on ensemble forecasting is clear and well-replicated: a simple average of several independent estimators almost always outperforms the best individual estimator over time, because each model's idiosyncratic errors partially cancel. This is true in weather forecasting, in election polling aggregation, in clinical risk scoring, and there is no reason to expect it to be false in residential valuation.
That is the entire thesis behind Zeego's property estimate.
How Zeego's estimate is built
Zeego's estimate is not a competing AVM. We don't train our own model and ask you to trust it. We pull four independent valuations on every property and present the blended average — alongside the underlying sources — so you can see both the consensus and the disagreement.
- Zestimate — Zillow's model. Strong on listed homes with deep comp volume.
- Redfin Estimate — different training data and a different feature set than Zillow; often diverges meaningfully on the same home.
- CoreLogic — the AVM most widely used by lenders and institutional investors under the hood.
- Quantarium — a computer-vision model that scores condition from listing and street imagery, which is the variable the other three struggle with most.
When the four converge in a tight band, the estimate is high-confidence. When they spread wide — say, $150,000 between the lowest and highest — that spread is itself a signal that the home is hard to value and you should weight your own diligence (comps, disclosures, condition, hazard exposure) more heavily. A single AVM can't tell you which situation you're in. Four can.
What this means for a buyer
- Treat the Zestimate the way Zillow itself describes it: an estimate, not an appraisal, useful as one input among several.
- Expect higher accuracy on currently listed, tract-style homes with strong comp volume; expect materially lower accuracy on off-market, custom, remodeled, or unique properties.
- Assume the seller has, legitimately, optimized what they can — claimed the home, updated facts, refreshed photos. That's not bad faith; it's how the product is designed to work.
- Don't anchor your offer on any single AVM. Look at the spread between sources and at the comps the estimate is actually built on.
- When the spread is wide, the answer isn't to pick the source you like best. It's to do more work on comps and disclosures.
The bottom line
The Zestimate is a useful, transparent, well-documented product that has been continuously improved for nearly twenty years. It is also a single algorithm with structural limits that Zillow has acknowledged in its own published accuracy data and that became visible at scale during the wind-down of Zillow Offers in 2021. None of that is a scandal. It is a reason to treat any one AVM — Zillow's, Redfin's, anyone's — as one estimator among several, not as the answer.
Zeego's property report shows the four-source blend, the spread, the underlying comps, and the disclosures together, in about a minute, at no cost. That's the version of the number a buyer should price an offer off of — not because Zillow is wrong, but because no single estimator should carry that much weight on a six- or seven-figure decision.