Industries / Construction and property / Real estate / PropTech
Search and listings
This work puts homes for sale and rent in front of the people looking for them, through maps, photos, tours, ads and search boxes.
5
Milestones achieved, from 2005 to 2025.
1
To watch for: announced, not yet confirmed outside.
2
Open deliverables. Nobody has shipped them.
The road
AchievedWatch forOpen
Achieved
2005
Listings land on a map
Within weeks of Google Maps' launch in February 2005, Paul Rademacher, a computer animation engineer in Santa Clara, California, built HousingMaps.com, which placed Craigslist rental ads as pins on Google's map, without help from anyone at Google. Google released an official Maps API (a public interface that lets other sites use its maps) in June 2005, and by July 2007 it counted more than 50,000 sites that combined its maps with other data, a form that became known as a "mashup".
Credit
Paul Rademacher; Google
2015
Agents launch their own portal
In Britain, two websites, Rightmove and Zoopla, carried most homes for sale. On 26 January 2015 Agents' Mutual, a company owned by estate agents, launched a rival portal, OnTheMarket, and required members to list their homes on no more than one other portal. When a member agent also listed on Zoopla, the dispute went to court; the Competition Appeal Tribunal found on 5 July 2017 that the rule was lawful and served "pro-competitive" aims, and the Court of Appeal dismissed the appeal on 24 January 2019. By then OnTheMarket had become a listed company and agents could sign listing agreements without the rule.
LessonA rule that limits where listings may appear was judged lawful when its aim was to break a two-portal market, yet the portal dropped it as it grew.
Credit
Agents' Mutual (OnTheMarket)
2018
Virtual tours, measured
A virtual tour lets a buyer move through a home on screen. Zixuan Meng, Yong Tan and Zhenbin Yan studied nearly 43,000 homes listed on a large Chinese property platform in 2018 and 2019, where tours were viewed on an ordinary screen without a headset; with a VR tour, average time on the market fell from 34 days to 19, with no effect on selling price (Information Systems Research, 2025). Isamar Troncoso of Harvard Business School and Mengxia Zhang of Ivey Business School studied 75,178 homes sold through Redfin in greater Los Angeles from March 2019 to March 2021, about 22% of them with virtual tours: tours shortened selling time during the pandemic, and once photo and listing quality were accounted for, their effect on price was not significant.
LessonTours speed up sales but have not been shown to raise prices.
Credit
Meng, Tan and Yan; Troncoso and Zhang
2022
Housing ad delivery reworked
On 21 June 2022 the US Justice Department sued Meta, Facebook's owner, alleging that the algorithms deciding which users saw housing ads relied in part on characteristics protected by the Fair Housing Act, such as race and sex. Under the settlement Meta stopped using its Special Ad Audience tool for housing ads, paid the maximum civil penalty of $115,054, and built a Variance Reduction System, which keeps measuring who is shown each housing ad and adjusts delivery so that the audience matches the people eligible to see it. It launched for US housing ads in January 2023, and an outside reviewer, Guidehouse, has reported on its compliance since June 2023.
LessonAn ad system can discriminate even when the advertiser chooses no protected group.
Credit
US Department of Housing and Urban Development and Department of Justice; Meta
2025
Climate scores pulled from listings
In September 2024 Zillow, the largest US home-search site, added climate risk scores from First Street, a company that models flood, fire and heat risk, to its listings. In November 2025 it removed them from more than a million listings after the California Regional Multiple Listing Service (CRMLS), a large listing service, objected; Art Carter, its chief executive, said that showing the chance of a home flooding this year or within five years can greatly affect how desirable it seems. Zillow left a link to First Street's records in their place, and Realtor.com, Redfin and Homes.com kept showing the scores.
LessonThe listing services that supply homes to a portal can shape what the portal shows about risk.
Credit
Zillow; First Street
Watch for
Home search inside chat assistants
Zillow says its connector for Claude, Anthropic's AI assistant, lets a user search homes for sale or rent, see maps, home value estimates and affordability sums, and request a tour or an agent inside the conversation, with the same listing display and attribution rules as Zillow.com; it has offered an app inside ChatGPT since 6 October 2025 and, for rentals, a similar service in Gemini. It gives no figures on use or accuracy.
Achieved whenAn outside party reports how many people search this way, or tests how often the answers match the live listings.
Claimed by
Zillow, October 2026
Open
01
A fair-housing wording test
A public set of about 2,000 short texts of the kind listing software now writes, such as ad sentences and chat replies, each labeled by two people trained in US fair-housing rules as lawful or as a likely violation, with the protected group involved. Include lawful sentences that look risky and unlawful ones that look harmless. Then score the tools that claim to catch discriminatory wording.
The briefCredit · unclaimed02
Steering test for home-search assistants
A paired test, on the model of fair-housing "secret shopper" testing, of the home-search assistants that listing sites now connect to chat apps. Send the same request from fictional users who differ only in a signal of race, family status or national origin, record every neighborhood and listing suggested, and measure the differences, without booking tours or contacting agents.
The briefCredit · unclaimed
Open 01 · the brief
A fair-housing wording test
The first to publish is credited on this line, and the deliverable moves from Open to Achieved.
What to build
A public set of about 2,000 short texts of the kind listing software now writes, such as ad sentences and chat replies, each labeled by two people trained in US fair-housing rules as lawful or as a likely violation, with the protected group involved. Include lawful sentences that look risky and unlawful ones that look harmless. Then score the tools that claim to catch discriminatory wording.
Why it is still open
Zillow released its Fair Housing Guardrail, a word list plus a fine-tuned text classifier, as open source in 2024, but its labeled dataset and trained model are available only on request and no accuracy figures were reported; the National Fair Housing Alliance invited others to test it. A search of Hugging Face found no fair-housing dataset.
Built with what exists
Zillow's fair-housing-guardrail code on GitHub; AI models used as checkers, such as Claude Fable 5.2 (see the Note "Three working devices") and the open-weight Mistral Large 4 (see the Note "A trillion open parameters"); Cohen's kappa to report agreement between the two labelers; Hugging Face Datasets to publish.
First step
Write 200 texts, half of them traps, have two people label them separately, and run the Zillow guardrail and one AI model over them.
Done looks like
2,000 labeled texts with the labelers' agreement rate, and published precision and recall, by protected group, for at least three checkers.
Open 02 · the brief
Steering test for home-search assistants
The first to publish is credited on this line, and the deliverable moves from Open to Achieved.
What to build
A paired test, on the model of fair-housing "secret shopper" testing, of the home-search assistants that listing sites now connect to chat apps. Send the same request from fictional users who differ only in a signal of race, family status or national origin, record every neighborhood and listing suggested, and measure the differences, without booking tours or contacting agents.
Why it is still open
Researchers at MIT ran 168,000 prompts through GPT-4 and found racial steering, including steering of minority home-seekers toward neighborhoods with lower opportunity scores (EAAMO 2024). A 2026 audit of seven AI models in four US cities found steering that changed with the city and the way a preference was phrased. Both tested bare AI models, not the assistants that read live listings, as did a September 2026 audit that gave three models a fixed pool of New York listings. A search in November 2026 found no outside test of the listing sites' own chat apps.
Built with what exists
The assistants' ordinary chat interfaces; a script of paired requests; American Community Survey data for each suggested census tract's makeup; an opportunity index of the kind the MIT study used; Python for the statistics.
First step
Write 50 paired requests for one city and run them by hand through one assistant, logging the ZIP code of every suggested home.
Done looks like
At least 2,000 paired requests across five metro areas and two assistants, with the gap between paired users in the racial makeup and opportunity score of suggested neighborhoods, its confidence interval, and all transcripts published.