BRENNEN

Industries / Construction and property / Real estate / Short-term rentals

Listings and search

This is getting a home onto the websites where travellers look for one, and those websites' work of showing each traveller the homes most likely to suit them and to be free on their dates.

6

Milestones achieved, from 1995 to 2022.

2

To watch for: announced, not yet confirmed outside.

2

Open deliverables. Nobody has shipped them.

The road

AchievedWatch forOpen

Achieved

  1. 1995

    Owners list their own homes

    David Clouse, then an intranet webmaster at United Airlines, built a database-driven website in 1995 whose first listing was his family's ski condominium in Breckenridge, Colorado, and bought the name VRBO, for Vacation Rentals By Owner, in 1996. Owners paid a yearly fee to advertise, as in a classified ad, and dealt with guests themselves: the first 250 listings were free and later ones cost $72 a year, by his account.

    skift.com

    Credit

    David Clouse

  2. 2005

    Listing sites merge

    Brian Sharples and Carl Shepherd, backed by the firm Austin Ventures, bought CyberRentals, GreatRentals, Rent 101 and Holiday Rentals, launched HomeAway in the summer of 2005, and bought VRBO in late 2006. Several of these sites had already shared one calendar and rates tool and their reviews through a group called Rentors.org; under HomeAway one company held the main owner-listing sites, and Jon Gray, later its chief revenue officer, called the VRBO purchase the most important moment in its history.

    skift.com

    Credit

    Brian Sharples and Carl Shepherd

  3. 2015

    Classified ads become bookings

    On 4 November 2015 Expedia agreed to buy HomeAway for about $3.9 billion and said it would speed HomeAway's "shift from a classified marketplace to an online, transactional model" across its 1.2 million properties. In the new model guests book and pay on the website, which takes a fee, in place of owners paying to advertise and handling the booking themselves.

    LessonA HomeAway team that studied Airbnb after a 2010 trade conference concluded, in John Banczak's words, "We need to quickly move to online bookings, to a consumer fee, get out of the subscription model".

    sec.govskift.com

    Credit

    Expedia (Dara Khosrowshahi); HomeAway (Brian Sharples)

  4. 2017

    Hosts turn many guests away

    Using Airbnb data for one large US city in 2013 and 2014, the economist Andrey Fradkin found that hosts rejected 36% of first contacts and 42% of all booking inquiries, partly because calendars showed homes as free that their hosts would not let. His model estimated that without availability tracking and filters, searches ending in an accepted inquiry would fall by 68%, and that ranking homes by how likely the host is to accept would raise the share of accepted searchers by 10%.

    LessonWhen every home is one of a kind and the owner can say no, a search engine has to predict the owner's answer as well as the guest's taste.

    ide.mit.edu

    Credit

    Andrey Fradkin, MIT Sloan

  5. 2018

    Search learns from clicks

    Airbnb's Mihajlo Grbovic and Haibin Cheng described listing "embeddings", lists of numbers learned from guests' browsing sessions that place similar homes close together, which Airbnb used to personalize search ranking and similar-home suggestions in real time. Their paper won the best paper award of the applied data science track at KDD 2018, a leading data-mining conference; the same year Malay Haldar and colleagues described Airbnb's move of search ranking from decision-tree models to neural networks.

    kdd.orgkdd.orgarxiv.org

    Credit

    Mihajlo Grbovic, Haibin Cheng, Malay Haldar and colleagues, Airbnb

  6. 2022

    Verified photos fill more nights

    A study of 7,423 Airbnb properties over 16 months, published in Management Science in August 2022, found that homes with photos taken by Airbnb's photographers had 8.98% higher occupancy than homes with photos taken by their hosts. The authors also measured 12 features of a good listing photo, covering composition, colour and how the subject stands out from the background.

    doi.org

    Credit

    Shunyuan Zhang, Dokyun Lee, Param Vir Singh and Kannan Srinivasan

Watch for

  1. Search by describing the trip

    Airbnb launched a search box in the US on 30 September 2026 that takes a request in plain words or by voice, suggests filters from keywords (a search mentioning a baby brings up cribs and playgrounds), shows an Airbnb-written summary of each home in the results, and compares saved homes side by side. In August 2026 its chief executive, Brian Chesky, told investors that for guests who switch it on "it's going to convert very, very well". Expedia's Vrbo already offers plain-language search.

    Achieved whenAn outside test shows that plain-language requests return homes that meet them at least as often as filter search does, or Airbnb publishes its test results with the method.

    skift.comfool.com

    Claimed by

    Airbnb (Brian Chesky), August and September 2026

  2. Made-up queries train search

    Airbnb researchers described CASTLE, a method that has an AI model write synthetic search requests from listing data and builds relevance labels from pairs of homes seen in booking sessions, used to train Airbnb's plain-language search before real requests existed. They report that the synthetic requests resemble real users' requests far more closely than other methods (a divergence score of 1.01 against 9.33 for the best alternative) and checked the labels with people on 200 samples.

    Achieved whenThe queries and labels are released, or an outside team reproduces the result on its own data.

    arxiv.org

    Claimed by

    Wendy Ran Wei and colleagues, Airbnb, May 2026

Open

  1. 01

    A public home-search test

    A public test set for searching holiday homes: about 300 plain-language requests ("two-bedroom flat with step-free entry, ten minutes' walk from a metro station, under €150 a night") run against a frozen copy of the listings in three cities, with each of the top 20 homes for each request graded by two people as a good, partial or poor match. Any search system, from keyword search to an AI model, can then be scored on it.

    The brief
    Credit · unclaimed
  2. 02

    Check AI-written home summaries

    An outside check of the summaries Airbnb now writes for homes in US search results: collect 500 of them and test every claim in each against the host's own description, amenity list, photos and recent reviews, marking it supported, unsupported or contradicted.

    The brief
    Credit · unclaimed

Open 01 · the brief

A public home-search test

Credit · unclaimed

The first to publish is credited on this line, and the deliverable moves from Open to Achieved.

What to build

A public test set for searching holiday homes: about 300 plain-language requests ("two-bedroom flat with step-free entry, ten minutes' walk from a metro station, under €150 a night") run against a frozen copy of the listings in three cities, with each of the top 20 homes for each request graded by two people as a good, partial or poor match. Any search system, from keyword search to an AI model, can then be scored on it.

Why it is still open

Airbnb built its own synthetic requests and labels for plain-language search and has not released them. The nearest public test, PriceBench (2026), measures which of 179 New York hotels 28 AI models choose to book, not whether a search finds homes that fit a request; searches of Hugging Face datasets and GitHub for a holiday-rental search relevance set found none.

arxiv.orgdoi.org

Built with what exists

Inside Airbnb's listing files, free to reuse under a Creative Commons Attribution licence; an open-source keyword search engine such as OpenSearch; open embedding models; AI models such as the open-weight Mistral Large 4 (see the Note "A trillion open parameters") or Claude Fable 5.2 (see the Note "Three working devices") as rankers to compare; the trec_eval scorer used in search research.

  1. First step

    Download one city's Inside Airbnb listings, write 50 requests, and grade the top 20 results of a plain keyword search for each.

  2. Done looks like

    300 requests and at least 6,000 graded request-home pairs published with the listing snapshot, the agreement rate between graders, and scores for at least three search methods.

Open 02 · the brief

Check AI-written home summaries

Credit · unclaimed

The first to publish is credited on this line, and the deliverable moves from Open to Achieved.

What to build

An outside check of the summaries Airbnb now writes for homes in US search results: collect 500 of them and test every claim in each against the host's own description, amenity list, photos and recent reviews, marking it supported, unsupported or contradicted.

Why it is still open

Airbnb launched the summaries in the US on 30 September 2026 and gave no accuracy figures. The closest check is the consumer group Which?'s July 2026 investigation of Tripadvisor's AI hotel summaries, which found summaries calling hotels "spotless" and service "friendly" while recent reviews reported illness and harassment; it did not say how many hotels it checked.

rentalscaleup.comwhich.co.uk

Built with what exists

A web browser and screenshots; an AI model that reads images, such as Claude Fable 5.2 (see the Note "Three working devices"), to split each summary into claims and match them to the listing, with two people confirming every flagged claim; open object-detection code that finds amenities in listing photos, such as Daniel Bourke's public replication of Airbnb's amenity detection (github.com/mrdbourke/airbnb-amenity-detection).

  1. First step

    Collect 50 summaries for one US city by hand and label their claims against the listing pages.

  2. Done looks like

    500 summaries from at least five US cities, each claim labelled by two people, and a published rate of unsupported and contradicted claims with its margin of error and all screenshots.