How did a 6.9% Zestimate error turn into Zillow's $500M iBuying loss?

In the third quarter of 2021, Zillow's iBuying arm was buying houses far faster than it could sell them. Zillow said it bought 9,680 homes in the quarter and sold only 3,032. At the end of October a KeyBanc analyst completed an analysis of 650 homes in Zillow's inventory, or about one-fifth of the homes owned, and found that 66% are currently listed below the purchase price at an average discount of 4.5%. In San Diego the figure was 94.3%, Phoenix 93.4% and Mesa, Ariz. 92.6%. On November 2, 2021, the Board of Directors of the Company made the determination to wind down Zillow Offers operations, which meant cutting about a quarter of the workforce.

The usual version of this story is "the AI failed." That is too loose, and the looseness is risky for anyone wiring valuation models into pricing decisions in 2026. The Zestimate was reasonably good at its original job, which was telling a homeowner roughly what a house is worth. Zillow lost the money because it turned that estimate into a binding bid, judged it on a metric that didn't match the new job, and took out the human brake just as the market turned. The thesis of this case is simple: a valuation model is safe to put money behind only when its error on the deals you actually win, plus the price drift over your holding period, fits inside your margin. Zillow's didn't, and nobody inside was measuring it that way.

How Zillow Offers turned the Zestimate into a binding cash offer

A cautious launch, then a $20 billion target

Zillow Offers started small. During 2018, Zillow Offers purchased a total of 686 homes and sold just 177, generating $52 million in revenue and losing $27.2 million. According to a case study in the Journal of Information Systems Education, Zillow Offers' initial approach to pricing decisions for home purchases also reflected a level of caution. Strategy changed when Rich Barton took over as Zillow's CEO in early 2019 and the company set a goal for Zillow Offers to reachreachThe number of unique people exposed to your message in a given period. Unlike impressions, reach counts each person once, no matter how often they see it.View full definition → $20 billion in revenue in three-to-five years. The same authors argue that Zillow Offers' focus on hypergrowth over profitability led to operational changes that failed to balance estimated price predictions and operational purchase price decisions.

Project Ketchup took the brake off

The volume push had a name. Bloomberg reported that an initiative called Project Ketchup added "overlays" to the algorithm's offers. Staff could raise offers by up to about 7% to win more sellers and meet volume goals, and pricing experts were told to stop questioning the algorithm's valuations. Under the same programme Zillow began using the Zestimate directly as its cash offer on qualifying homes.

The warning signs showed up before the write-down. One post-mortem notes that in summer 2021, local teams started to notice mismatches between the model's output and what homes were selling for in reality. Initially, they were treated as execution or collaboration issues. Meanwhile, operations couldn't keep up. Third-quarter Homes revenue came in at $1.2 billion, below the company's Q3 outlook of $1.45 billion at the midpoint of the range, due primarily to renovation and resale capacity constraints. Every week a house sat waiting for a contractor was another week of exposure to falling prices.

The accuracy metric answered a different question

Zillow published honest error figures. In June 2021 Zillow said the model had improved: its launch announcement gave a median error of 6.9% for off-market homes across more than 104 million properties. For listed homes the figure is about 1.9%, and the on-market figure is small partly because the model can see the list price. Look closely at how this is measured, because it's the same issue you face when reviewinghow automated valuation models are scored. Zillow calculates their publicly presented price-prediction error metrics by comparing the actual price of a home sold with the Zestimate made for that home on the date of sale. That tells you how well the model reads today's price. An iBuyer needs to know the resale price after renovation, months later.

There is also a selection problem. Sellers accept the offers that beat their own estimate, so accepted offers lean high. The published median covers every home. The losses came from the subset where Zillow overestimated, because those were the sellers who signed. Barton's own diagnosis on the earnings call: "We've been unable to accurately forecast future home prices at different times in both directions by much more than we modeled as possible."

How much did Zillow Offers actually lose?

The answer is more than $500 million on Zillow's first framing. The audited annual write-down came out lower, and the total cost is higher once restructuring is counted. The headline number combined two figures: a write-down of inventory of approximately $304 million within the Homes segment as a result of purchasing homes in Q3 at higher prices than the company's current estimates of future selling prices, plus an additional $240 million to $265 million of losses to be recognized in Q4 primarily on homes it expects to purchase in Q4.

The 10-KKThe average number of new users each existing user generates through referrals. Above 1.0, growth compounds on itself and becomes exponential.View full definition → tells a slightly different story. For the year ended December 31, 2021, we recorded a write-down to inventory totaling $407.9 million, and on top of that $71.2 million in impairment and restructuring costs within our Homes segment associated with our wind down of Zillow Offers operations. The costs carried into 2022: in the first quarter alone Zillow booked a write-down to inventory totaling $5 million, $30 million in restructuring costs and $30 million of charges associated with winding down Zillow Offers financing facilities. So if you quote one number, say what it includes. Inventory write-downs, losses on sale, severance and debt costs land in different line items, and published totals vary depending on which ones get added up.

The human cost was explicit. Zillow said it would scrap the service altogether and cut 2,000 workers. By November 1 the stock was still down 51.1% since closing at a record $199.90 on Feb. 16.

The rest of the sector split two ways. Redfin held on another year, then announced that it will eliminate 862 positions, or 13% of its workforce, and closed RedfinNow. CEO Glenn Kelman warned that even before its overhead expenses, the RedfinNow properties segment will likely lose $22, $26 million dollars in 2022. The company said maintaining a profit with rising interest rates would make our offers on homes insultingly low. Opendoor stayed in, and its 2026 filings describe a rebuild. These are the company's own figures and partly non-GAAPGAAPThe standard set of accounting rules companies follow to prepare consistent, comparable financial statements, dominant in US reporting.View full definition →, so read them as management's account of its turnaround. The CFO said contribution margin has climbed from a low of 1% three years ago to 5.8% in Q2, landing in the middle of our 5%-7% range, and in the first quarter aged inventory decreased to 10% from 51% in 3Q25. Opendoor is also growing Cash Now, More Later, our capital-light offering that generates attractive returns while placing substantially less capital at risk per home.

Why a single valuation miss can exceed an iBuyer's whole margin

Here is a hypothetical, labelled as such and kept to arithmetic, using the sourced percentages above on a $400,000 house:

• At Opendoor's reported 5.8% contribution margin, a well-run iBuyer clears about $23,200 per home in 2026.
• The Zestimate's 6.9% median off-market error is $27,600, and by definition half of homes miss by more.
• KeyBanc's 4.5% average listing discount below purchase price is $18,000.
• A Project Ketchup overlay at the top of the reported range, about 7%, adds $28,000 to the bid.

The numbers come from different companies and years, so this only shows scale. But scale is the point. The model's typical miss is bigger than the entire margin of a disciplined operator. And the winner's curse decides which way the misses land.

What property AI leaders should take from Zillow Offers

The spread test

Here's the decision rule. It fits on a whiteboard:

Error on deals you win + price drift over the hold < spread after costs. If it isn't, the model advises and a person or a fixed haircut decides.

Each term needs its own measurement:

• Accepted-offer error compares the model's value with the realised resale price, only for offers the counterparty accepted. Track it separately from the vendor's headline median.
• Hold-period drift is how far local prices can move between contract and exit. It depends on your renovation and resale capacity and onwhere the local cycle sits. Zillow's contractor backlog stretched this term just as appreciation slowed.
• Spread is what's left after renovation, carry, selling costs and financing. For an iBuyer it is single digits, as Opendoor's 5% to 7% target shows.

When the test fails, you have four options: cut the bid, shorten the hold, move risk off the balance sheet (as Opendoor's capital-light product tries to do), or bring back a reviewer who can veto. Zillow did the opposite of all four in 2021.

One more rule follows from Project Ketchup. A human override that only ever moves offers up, to hit a volume number, is itself an unvalidated pricing model. Log every override with its direction and the reason, and test it like any other model.

Where your context differs

If you originate mortgages, part of this is now law. Under the interagency rule, institutions using AVMs for certain credit decisions must adopt policies, practices, procedures, and control systems to ensure that AVMs used in these transactions to determine the value of mortgage collateral adhere to quality control standards covering confidence, data manipulation, conflicts of interest, random sample testing and nondiscrimination. This final rule is effective October 1, 2025. Random sample testing is the closest regulatory cousin to measuring accepted-offer error. The fifth factor has a fair housing edge: the agencies' rule goes beyond the four statutory quality control standards to include a fifth discretionary standard: compliance with "applicable nondiscrimination laws." An AVM whose error varies by neighbourhood is a lending-discrimination exposure as well as a pricing one. The preamble also recognises that AVMs are typically owned by vendors and not the mortgage originators subject to the rule, but the obligation still sits with you.

If you buy homes for cash, run single-family rentals or bid on build-to-rent portfolios, that rule probably doesn't cover you, so the discipline has to be internal. Brokerages that show AVM values in listing presentations carry advisory error. A miss there costs credibility and anchors the seller's expectations, but it doesn't put a loss on your balance sheet.

The 2026 twist is how easily an LLMLLMA Large Language Model is an AI system trained on vast text data to predict and generate language, enabling tasks like writing, summarizing, and answering questions.View full definition → agent can now chain a valuation call to an offer letter. Google Cloud now hosts OpenAI, Gemini and Claude models side by side, according to The Decoder. TechCrunch reports that Walmart is letting AI agentsAI agentsAgentic AI refers to AI systems that pursue goals autonomously by planning, taking actions through tools, and adapting based on results, with minimal step-by-step human direction.View full definition → make purchases for customers, with spending controls. Apply the same logic to property: an agent drafting offers needs hard caps tied to the spread test, not a prompt asking it to be careful. The Decoder also reported this month that OpenAI can't yet explain why its coding agents switch off misalignment monitoring. Project Ketchup was the human version of switching off your own monitoring.

What to do on monday

• Pull every decision from the last 12 months where a valuation model set or anchored a price you committed to. Compute error separately for accepted and rejected offers.
• Write the three numbers of the spread test for one product line on a single page, and get finance to sign off on the spread figure.
• Export your override log and sort by direction. If more than a small minority of overrides move price the same way, ask who set the volume target.
• Check whether the team that owns acquisition volume can also change pricing inputs. If it can, separate the two this quarter.
• Lenders: mapmapUsing software to automate repetitive marketing tasks and campaigns, enabling personalisation at scale across channels like email, web, and social.View full definition → each AVM vendor contract to the rule's five factors and ask for the vendor's latest random-sample and nondiscrimination test results.
• Set a weekly tripwire on the share of your inventory listed below purchase price, with a named owner who has authority to pause buying.

KeyBanc found Zillow's problem with public listings and a spreadsheet, by comparing what Zillow paid with what it was asking. Checking the model's bets against the market every week costs almost nothing. Zillow's mistake was telling the people who did that check to stop.

The full course on this sector:AI in Real Estate.