NextFin News - David Beitel, Zillow's chief technology officer and a founding member of the company, says artificial intelligence is finally doing what two decades of search technology could not: guiding home buyers through the judgment calls that begin after they find a listing. Speaking on a technology-focused interview program on Oct. 1, Beitel cited early data from Zillow's new AI Mode showing users spending three times as long on the platform, running three times as many searches, viewing twice as many homes, and connecting with agents at three times the rate. The claim reframes AI as a conversion engine rather than a search upgrade - and it lands as Zillow's investors are being asked to price a much larger thesis: that the company can turn a traffic monopoly into a transaction franchise without repeating the $881 million collapse of its iBuying venture.
The Moment Search Stopped Being Enough
Zillow turned 20 this year, and the anniversary marks a quiet inflection point. For its first two decades, the company's job was to make homes findable: digitize inventory, publish the Zestimate, and become the front door to the biggest purchase most Americans will ever make. The problem, as Beitel put it in an April essay on the company's own blog, is that "finding a home has never really been the hard part. Knowing what to do next is."
AI Mode, which Zillow began rolling out in beta in late March, replaces rows of filter toggles with a conversational interface. Shoppers can ask whether they can afford a neighborhood, why a price just dropped, what a fair offer looks like, or how one home compares with another - questions that previously required stitching together calculators, listings, and conversations. Beitel said the feature is now being extended across mortgages and financing, agent tooling, and rich-media listings, with fair-housing safeguards built into the guidance layer.
"We're connecting the entire housing journey with AI in a way that hasn't been possible before," Beitel said.
The engagement data he offered is striking because it maps cleanly onto Zillow's revenue funnel. More time on site leads to more searches, which lead to more home views, which lead to the metric that actually matters: connections with Premier Agent partners. If those multipliers hold as the feature scales beyond early adopters, AI Mode is not a product feature - it is a lever on the company's core lead-generation business. That is the crux of the matter, because lead generation is where Zillow makes most of its money, and any technology that moves that needle moves the entire valuation.
Beitel's own description of the product philosophy is worth reading closely. He draws a line between AI layered on top of existing workflows - a chatbot answering questions, a tool that summarizes documents - and AI that changes the workflow itself. "That version of AI is useful," he wrote in April. "It is not, by itself, transformative." Real estate, in his telling, is different because the stakes are different: a home purchase is local, highly regulated, emotionally charged, and coordinated across a web of professionals - agents, lenders, property managers, title companies - each playing a specific role. A generic large-language model trained on the open internet cannot schedule a showing, read a local market's pricing signals, or guarantee that its guidance follows Fair Housing requirements. Housing-native intelligence, he argues, requires proprietary data, embedded workflows, and accountability to real outcomes rather than engagement metrics alone.
Why This AI Bet Is Different From the $881 Million One
The ghost in Zillow's AI story has a name: Zillow Offers. The iBuying program used algorithms to price homes the company then bought, renovated, and resold. It wound down in 2021 after losing roughly $881 million, a failure that was part modeling hubris and part market timing - the models could not handle a turning housing market, and the balance sheet absorbed every error. The episode became a cautionary case study in what happens when an algorithm meets the physical and financial friction of actual houses.
The current AI stack takes the opposite risk profile. It is an information-and-matching bet with no inventory ownership and no balance-sheet exposure. The pieces are the Zestimate neural valuation model, which reports a nationwide median error rate near 2.4% for on-market homes; AI-enhanced listing media including drone-captured SkyTour imagery and virtual staging; ShowingTime+ for tour scheduling; Premier Agent Flex for routing buyer leads; and back-office tools such as Follow Up Boss for client management and dotloop for transaction paperwork. None of these requires Zillow to own a single roof.
The moat Beitel is betting on is embeddedness. About 80% of U.S. residential transactions involve an agent who uses at least one Zillow-owned product, a figure the company has published. That gives Zillow something a generic large-language model cannot replicate: proprietary transaction data and a position inside the workflows where deals actually close. As Beitel wrote in April, housing-native intelligence "requires proprietary data built up over years" and "being embedded inside real estate agent and loan officer workflows where transactions happen - not summarizing them from the outside." The strategic pivot is explicit: after the iBuying loss, Zillow moved from balance-sheet AI bets to workflow AI bets - matching and productivity rather than house-flipping.
This is where the cyclical-versus-structural call comes into focus, and getting it wrong flips the conclusion. The housing downturn itself is cyclical: existing-home sales in 2025 hit their lowest level since 1995, mortgage rates sit at roughly double their pandemic-era lows, and inventory remains tight. Those conditions will mean-revert when rates fall, and three historical-cycle comparisons make the point - sales troughs in the early 1990s, after the 2008 crisis, and during the 2020 lockdowns were all followed by recoveries once financing conditions eased. But the shift from search to guided workflow is structural. It changes how buyers and agents interact regardless of the rate cycle, and it does not self-correct when rates drop. Zillow's own results are evidence of the separation: revenue of $2.6 billion in 2025, up 16% year over year, and second-quarter 2026 revenue of $772 million, up about 18% from a year earlier and above the $758.25 million consensus. Growth has persisted through one of the worst housing cycles in a generation, suggesting the lead-generation and workflow model is more resilient than the transaction volume it sits on top of.
The second-order implication is less comfortable. AI Mode raises engagement, which raises lead volume, which raises Premier Agent willingness to pay. But it also raises the possibility that AI handles the early conversation so well that the agent's role shrinks toward closing, squeezing the perceived value of the lead itself. Beitel's answer is to embed AI inside agent tools - Follow Up Boss, ShowingTime+ - so the agent captures the AI-generated lead rather than being disintermediated by it. Whether that containment works determines whether AI expands Zillow's take rate or cannibalizes it.
The Competition Zillow Did Not Lead With
Zillow's AI narrative assumes it starts the race with the largest installed base. It does - the platform draws roughly 207 million average monthly unique users, and the Zestimate covers more than 100 million homes. But the competitive map has changed faster than the story acknowledges, and three challengers attack different layers of the stack.
CoStar Group's Homes.com reached 115 million average monthly unique visitors in the third quarter of 2025 and launched its own AI search layer in February 2026. CoStar's advantage is symmetrical to Zillow's: it owns professional listing and data tools used by agents and brokers, the same "embedded workflow" logic Zillow claims as its own. If CoStar can convert its professional-data dominance into consumer traffic at scale, Zillow's data moat is no longer unique - it is contested.
Redfin, meanwhile, reports that users of its conversational search view nearly twice as many listings and are 47% more likely to request a tour than filter-based searchers - a comparable engagement story, but one that converts inside Redfin's own brokerage funnel rather than through an advertising model. That difference matters: Redfin captures the full commission, while Zillow sells the lead. In a market where commission economics are being rewritten, the integrated model has its own logic.
Then there is Google, the threat that never leaves. If AI Overviews begin answering "what can I afford in this neighborhood" directly on the search-results page, Zillow's top-of-funnel traffic is exposed in a way that no feature launch fully insulates. The National Association of Realtors settlement that reshaped commission economics in 2025 adds another variable: AI that explains buyer-agent value could strengthen the case for commissions, or it could arm buyers to negotiate them down. The direction is not yet knowable, and that uncertainty is priced into the wide dispersion of analyst views.
The Counter-Thesis: Engagement Is Not Earnings
The strongest argument against Zillow's AI bull case comes from Zillow's own recent numbers. Second-quarter 2026 revenue beat expectations at $772 million, but GAAP diluted earnings came in as a loss of $0.02 per share versus a profit of $0.01 a year earlier. Gross margin compressed by about two percentage points, and operating cash flow fell roughly 87% year over year to $11 million. Management's third-quarter revenue guidance of $745 million to $760 million landed below the $772.3 million consensus. The market is currently rewarding top-line growth while profitability deteriorates - a combination that has punished high-multiple platforms before, and one that should give pause to anyone reading the engagement multipliers as a straight line to earnings.
Analyst price targets reflect the disagreement. Estimates range from roughly $30 to $100 per share, with medians clustering between $50 and $76 depending on the source - dispersion wide enough to signal that Wall Street has not decided whether AI monetizes or merely costs money. That split is itself a data point: when a stock's valuation depends on a thesis about user behavior that has not yet been proven at scale, the market prices both the promise and the doubt.
Beitel's answer, in effect, is that margin pressure reflects AI investment ramping ahead of revenue recognition, and that the threefold increase in agent connections is the leading indicator that earnings will follow. That logic is coherent but unproven at scale. AI Mode was live for only about 5% of Zillow's user base as of the first-quarter earnings call, and early-adopter behavior is notoriously skewed toward the most motivated shoppers. The engagement multipliers could compress as the feature reaches a mainstream audience that asks harder questions and converts less eagerly. The history of product rollouts is littered with early metrics that looked like laws of nature and turned out to be selection effects.
There is also a regulatory overhang that Beitel acknowledged and that investors should not treat as a footnote. An AI assistant that steers buyers toward or away from neighborhoods can run afoul of the Fair Housing Act if it learns from biased historical patterns. Zillow has built safeguards into the feature, but the legal standard for algorithmic steering is still being written, and enforcement risk rises with adoption. This is not a speculative concern - fair-housing advocates have already challenged algorithmic advertising practices in real estate, and the scrutiny will only intensify as AI makes more of the recommendations that shape where people live.
What To Watch, and What Would Prove It Wrong
The short-term read is straightforward: expect volatility. Guidance already disappointed, and the next two quarters will show whether revenue re-accelerates toward consensus while margins stabilize as AI compute costs normalize. Investors should watch the cadence of AI Mode's rollout, the trajectory of gross margin, and whether operating cash flow recovers from its second-quarter trough.
Over the next 12 to 24 months, the decisive questions are whether Premier Agent pricing power improves as lead quality rises, and whether AI Mode can expand from roughly 5% to a quarter or more of monthly users while holding its engagement multipliers. If both happen, Zillow's traffic advantage compounds into a workflow advantage that CoStar and Redfin cannot easily match.
Structurally, the winner in real estate AI is whoever owns the housing-native intelligence layer: proprietary data, embedded agent and lender workflows, and demonstrable Fair Housing compliance. Zillow enters that race with the strongest hand on data and distribution. Execution is the variable, and execution is where the $881 million lesson lives.
Three scenarios frame the path. In the base case, AI Mode reaches about a quarter of users by the end of 2027, engagement multipliers settle at 1.5 to 2 times, Premier Agent revenue grows in the mid-teens, and margins stabilize - a modest rerating. In the upside case, AI Mode becomes the default search interface, the threefold agent-connection rate sustains, Zillow takes share from Homes.com and Redfin, and margins expand as AI replaces manual operations - multiple expansion follows. In the downside case, engagement proves to be early-adopter skew, margins keep compressing, Google or CoStar wins the AI search layer, and Fair Housing scrutiny produces an enforcement action. That last path would mark the return of the $881 million ghost in a new form: overconfidence in models that cannot capture the full reality of the market.
The falsifying signal is specific: if AI Mode penetration exceeds 25% of monthly users and Premier Agent revenue per lead is flat or down year over year, the thesis that AI converts engagement into monetization is wrong. At that point, the multipliers would be revealed as a product novelty rather than a business-model shift.
Zillow's AI bet is no longer that machines will buy houses - it tried that and lost $881 million. The bet now is quieter and harder: that AI can make people decide faster, and that the company sitting between the question and the agent gets paid for it.
更多独家洞察尽在 nextfin.ai.
