Last year Americans spent ten weeks on average looking for a home. The $1.6 trillion market that resulted still runs on a filter form.

Americans bought about 4.1 million existing homes in 2025. At the national median price, that is roughly $1.6 trillion changing hands, close to the size of Spain's entire economy, spent one house at a time. Almost every dollar of it moved through a search that now takes a median of ten weeks, and a majority of buyers say the hardest part of the whole thing is just finding the right home. The biggest purchase most people ever make still starts with the same four boxes: price, beds, baths, ZIP code. Those boxes are the problem.

Here is what the boxes miss. Picture a couple in Chicago with a $550,000 ceiling and a saved alert for three-bedroom listings in Lincoln Square. Six weeks in, they have toured seven homes and put an offer on a two-bedroom greystone in Logan Square. Their own filters would have hidden it twice over, on price and on bedroom count. Their filters described the house they thought they wanted. Their clicks described the house they bought. Everything expensive about the home search lives in that gap.

The ProblemThe search is built on what buyers say, not what they do

Every listing portal works the same way. A buyer types a maximum price, a bedroom count, and a ZIP code, and the system hands back everything that survives those hard filters. The model treats a stated preference as a fact. In real life it is closer to an opening bid, and the whole search inherits three faults from it.

The filter throws away the home that would have closed. Hard filters subtract. A $550,000 ceiling deletes the $565,000 listing the buyer would have loved; a three-bedroom rule hides the two-bedroom with a finished basement that fits the same family just fine. Because buyers routinely transact outside the box they drew, the filter discards its best candidates first.

The agent is working from a hunch. A typical tour list comes out of a fifteen-minute intake call and a folder of links the buyer happened to forward. The agent reads taste by instinct and only recalibrates after a wasted Saturday reveals that the buyer quietly hates galley kitchens. There is no structured read of what this person actually responds to.

Time leaks at every wrong showing. Ten weeks and a run of poorly matched tours are costly on both sides. The buyer burns evenings and momentum. The agent burns windshield hours that do not pay until something closes. Since the 2024 commission settlement, that unpaid time is harder to justify, because buyers now sign a written agreement and negotiate what their agent earns out in the open.

The strange part is that buyers are already generating the answer. Almost everyone searches online first, yet the system does almost nothing with the trail they leave.

A buyer's clicks are a more honest description of the home they will buy than the filters they set to find it.

The SignalEverything an agent needs is already in the browsing trail

The behavior a buyer leaves behind while browsing is a running, honest record of preference. Dwell time, saves, return visits, how deep they scroll into a photo gallery, the listings they dismiss in under a second. All of it carries information the filter form never captures. Read together, those signals decode into the attributes an agent would otherwise spend weeks guessing at.

Streaming worked this out a decade ago. Spotify never asks what music you like; it watches what you play, replay, and skip, then groups you with listeners who behave the same way and serves you the next song from inside that group. We built the same machinery for homes and call it the Housing Taste Graph: a model that learns a buyer's preferences from behavior, sorts homes into granular genres instead of bedroom counts, and matches buyers to listings the way a streaming service matches a listener to a track.

Why It WorksRecommendation engines already run the markets real estate hasn't touched

This is not a bet on an unproven idea. It is the last large consumer market still running on declared preference while everyone else moved to revealed preference years ago. The numbers from those other markets are blunt.

The engines pay for themselves. Netflix has put the value of its recommendation system at roughly $1 billion a year, almost all of it from keeping subscribers who would otherwise drift away. Salesforce found that shoppers who engage with a recommendation convert at about 4.5 times the rate of those who do not. McKinsey pegs the revenue lift from good personalization at 5 to 10 percent. Three things make the approach hold up when you move it to housing.

What buyers do beats what they say

Typed preferences are aspirational. Behavior is honest, and it is expensive to fake. Dwell time, saves, and re-views accrue continuously, so the profile sharpens with every session instead of going stale the moment an intake form is filed.

Genres are learned, not declared

Embeddings sort homes by latent style (the felt difference between a Logan Square greystone, a Streeterville high-rise, and a new-construction infill) rather than the beds-and-baths fields a buyer can name. This granular genre layer is where most of the matching value sits, and it is exactly what a filter cannot see.

Collaborative filtering solves the cold start

The hardest moment for any recommender is a brand-new user with five clicks. Clustering that buyer against established taste groups produces a useful shortlist before they have said much of anything, then refines it as their own behavior comes in. It is the same logic behind "listeners like you."

Coded to SolutionEach wall maps to a capability that removes it

The Housing Taste Graph answers the three faults in the conventional search one for one. Every capability exists to take down a specific wall.

The ImpactFewer wrong showings, and the close pulled forward

The point of all this is simple: get the right homes in front of the right buyer sooner, so the agent spends Saturdays on tours that convert. In the Chicago pilot we model the change as a funnel. More buyers reach a genuine fit, more reach an offer, and the search compresses.

The FlywheelWhy the lead widens once it starts

1. 1Every view, save, and skip trains the model.Each buyer's taste vector and the shared genre map both sharpen with use.
2. 2Sharper genres mean better shortlists.Tours convert at a higher rate and offers come sooner, which is the outcome agents actually get paid for.
3. 3Faster closes pull agents in.An agent who closes in six weeks instead of ten brings their next buyer onto the platform.
4. 4More buyers deepen the data.Cold-start recommendations get sharper for the next person who shows up with five clicks.
5. 5The genre map gets finer.New styles and micro-neighborhoods emerge, and the lead over filter-based search keeps growing.

The WindowThe conditions just converged

Three things have lined up at once. The clustering and embedding tools that once needed a research team are now cheap and fast. The browsing data to train them already exists, generated on every portal. Zillow alone treats the space around the transaction as a market worth roughly $187 billion. And the 2024 NAR settlement, born from a $1.78 billion jury verdict and settled for $418 million, has put agents under real pressure to prove their worth now that buyers sign for their services in writing. A tool that closes deals faster has never been easier to sell.

The money is moving the same way. The global PropTech market sits near $47 billion in 2025 and is on track for roughly $104 billion by 2030, growing about 16 percent a year, with residential the largest slice. More than 750 startups are already building AI tools for real estate. The buyer was always ready; the technology and the incentives finally caught up.

The home search is the last big market still asking people what they want instead of watching what they do. The first mover that reads the browsing trail, clusters it, and hands agents a shortlist that closes will set how this category works for the next decade.