BRENNEN

Industries / Retail, hospitality and personal services / Hospitality and tourism

Planning a trip

This work helps travelers decide where to go, what to see and how to fit it together, through guidebooks, package tours, travel agents, review sites, maps and now AI planners.

6

Milestones achieved, from 1836 to 2025.

2

To watch for: announced, not yet confirmed outside.

2

Open deliverables. Nobody has shipped them.

The road

AchievedWatch forOpen

Achieved

  1. 1836

    Guidebooks rate the sights

    In 1836 John Murray of London launched Murray's Handbooks for Travellers, which turned the writer Mariana Starke's exclamation-mark ratings into star ratings, rated sights by importance and laid out practical information along travel routes. In Germany Karl Baedeker began a series around the same time and had adopted Murray's star system by 1844, by The National Archives' account; the historian Sarah Schaefer Walton writes that Baedeker at first "borrowed heavily and explicitly from Murray" for its English editions.

    nationalarchives.gov.ukbranchcollective.org

    Credit

    John Murray III, son of the publisher John Murray II; Mariana Starke; Karl Baedeker

  2. 1841

    The package tour

    On 5 July 1841 Thomas Cook, a cabinet maker in the temperance movement, chartered a train to take 485 temperance society members from Leicester to Loughborough for one shilling each. Cook ran a first profit-making trip in 1845, later sold multi-stage tours of the Continent, and by 1851 had arranged travel for 165,000 people to London's Great Exhibition, by The National Archives' account.

    nationalarchives.gov.uk

    Credit

    Thomas Cook; later Thomas Cook & Son

  3. 2012

    Review sites tested on trust

    On 1 February 2012 the UK Advertising Standards Authority ruled on TripAdvisor's claims of "Reviews you can trust" and "More than 50 million honest travel reviews", after a complaint from the firm KwikChex and two hotels, and told it not to claim or imply that all its reviews were from real travelers or were "honest, real or trusted". In August 2024 the US Federal Trade Commission adopted a rule banning the sale and purchase of fake reviews, including AI-generated ones. Tripadvisor says it blocked 2.7 million fraudulent reviews in 2024, out of 31.1 million reviews submitted.

    LessonA review site could not show that every review was genuine, and fake reviews became a matter for advertising and trade regulators.

    asa.org.ukftc.govtripadvisor.mediaroom.com

    Credit

    Advertising Standards Authority; US Federal Trade Commission

  4. 2019

    A tour giant collapses

    On 23 September 2019 Thomas Cook UK said 26 of its UK companies had stopped trading, leaving about 150,000 customers abroad in 18 countries. The government told the Civil Aviation Authority to bring all of them home, including those without ATOL protection, the UK scheme set up in 1973 that protects holidays that include a flight. Between 23 September and 7 October more than 140,000 were brought back.

    LessonThe protection scheme covered only ATOL-protected customers, so the Department for Transport paid to repatriate the rest.

    nao.org.uk

    Credit

    UK Civil Aviation Authority and Department for Transport; reported by the National Audit Office

  5. 2024

    Planning tests for AI agents

    TravelPlanner, a benchmark of 1,225 trip-planning tasks with a sandbox of nearly four million records, found that even GPT-4 reached a success rate of only 0.6%, a success being a plan that meets every constraint. In July 2026 the TREK benchmark, 800 tasks scored by fixed rules, found the best of 15 AI agents, GPT-5.6, produced a fully feasible plan for 46.2% of solvable tasks, with a median of 6.6%.

    LessonTREK's authors found that meeting travelers' unstated needs is "the universal bottleneck".

    arxiv.orgarxiv.org

    Credit

    Jian Xie, Yu Su and colleagues (TravelPlanner, ICML 2024); Jinhu Qi, Irwin King and colleagues (TREK)

  6. 2025

    AI sends tourists astray

    In May 2024 the marketing agency SEO Travel asked ChatGPT for 100 two-day itineraries across 10 cities and found at least one error in 90%: 24% recommended a closed place and 52% a visit outside opening hours, Travolution reported. In September 2025 the OECD's AI Incidents Monitor recorded travelers sent by ChatGPT and similar models toward a "Sacred Canyon of Humantay" in Peru that does not exist. In January 2026 visitors were reported looking for Weldborough Hot Springs in Tasmania, which have never existed, after an AI-written article on a tour company's website.

    LessonThe Tasmanian company's owner said it had outsourced its marketing to a firm that used AI and that some posts went live without review.

    travolution.comoecd.aieuronews.com

    Credit

    SEO Travel; OECD AI Incidents Monitor; Euronews and Australia's ABC, which reported the Tasmanian case

Watch for

  1. Plan in chat, book elsewhere

    In October 2025 ChatGPT began hosting apps from Expedia and Booking.com. A traveler asks in plain words, such as for family hotels in Orlando this weekend, sees live listings with rates, photos and availability inside the chat, and clicks through to finish the booking on Expedia or Booking.com.

    Achieved whenAn outside party reports how many trips are planned and booked this way, or an independent test finds that the prices and availability shown in the chat match the agency's own site.

    hospitality.today

    Claimed by

    OpenAI, Expedia and Booking.com, October 2025

  2. Itineraries built after booking

    Expedia says its conversational Activity Planner, for US users, turns a traveler's saved favorites into activity, restaurant and event suggestions and a daily itinerary, one of more than 30 new or improved features it calls AI-powered. It gives no usage or accuracy figures.

    Achieved whenAn outside audit checks a sample of its itineraries for closed places, wrong hours and impossible timings and publishes the error rate.

    hospitalitynet.org

    Claimed by

    Expedia Group, October 2026

Open

  1. 01

    Check AI itineraries against reality

    A repeated public audit of the AI trip planners people actually use. Give at least four planners, such as ChatGPT, Gemini, Mindtrip and Expedia's planner, the same 50 requests across 10 cities, then check every place each answer names: whether it exists, whether it is open on that day and hour, and whether the traveler can get between stops in the time allowed. Ground truth comes from official websites, OpenStreetMap opening hours and public transit timetables, with a person settling every doubtful case. Repeat every three months.

    The brief
    Credit · unclaimed
  2. 02

    Accessible rooms, measured and compared

    For one city, and with each hotel's permission, measure the accessible guest rooms of at least 30 hotels: door widths, bed height, shower type and threshold, grab bars and turning space. Publish the measurements under an open license in A11yJSON, an open format for describing accessibility, and compare each room with what the hotel's website and two booking sites say about it.

    The brief
    Credit · unclaimed

Open 01 · the brief

Check AI itineraries against reality

Credit · unclaimed

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

What to build

A repeated public audit of the AI trip planners people actually use. Give at least four planners, such as ChatGPT, Gemini, Mindtrip and Expedia's planner, the same 50 requests across 10 cities, then check every place each answer names: whether it exists, whether it is open on that day and hour, and whether the traveler can get between stops in the time allowed. Ground truth comes from official websites, OpenStreetMap opening hours and public transit timetables, with a person settling every doubtful case. Repeat every three months.

Why it is still open

The best-known audit, SEO Travel's in 2024, tested one tool once, was run by a marketing agency, and this research found no replication covering other planners. Academic benchmarks such as TravelPlanner and TREK score plans against fixed sandbox data, not against what is true on the ground on the day, while incidents such as Peru's nonexistent canyon keep being recorded.

travolution.comarxiv.orgoecd.ai

Built with what exists

The planners' public apps; OpenStreetMap and its opening-hours tags; public transit timetables in the GTFS format with the open routing engine OpenTripPlanner for travel times; a low-cost AI model built for jobs such as extraction to pull place names and times from each answer, such as Meta's Muse Ember (see the Note "The jobs agents repeat"), with a person checking.

  1. First step

    Write 10 requests for one city you know well, run them on two planners, and check every place by hand.

  2. Done looks like

    200 itineraries (50 requests on four planners) with every named place checked, published with a table for each planner giving the share of itineraries with at least one nonexistent place, one closed place, one visit outside opening hours and one impossible connection, then the same audit rerun three months later.

Open 02 · the brief

Accessible rooms, measured and compared

Credit · unclaimed

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

What to build

For one city, and with each hotel's permission, measure the accessible guest rooms of at least 30 hotels: door widths, bed height, shower type and threshold, grab bars and turning space. Publish the measurements under an open license in A11yJSON, an open format for describing accessibility, and compare each room with what the hotel's website and two booking sites say about it.

Why it is still open

US law requires hotels to describe accessible features in enough detail for a guest to judge independently whether a room meets their needs. Yet NPR, which interviewed 50 wheelchair users and heard from more than 200 people in December 2025, found accurate information hard to find, with beds now often 25 to 30 inches high and shower doors too narrow for a wheelchair. Measured guides exist but are not open data: AccessAble sends surveyors to venues in the UK and Ireland, is paid by partners who commission its guides, and offers them free to view; Wheel the World reports more than 5,000 verified listings on its own platform. This research found no openly licensed, room-level dataset and no published tally of how often online descriptions miss or contradict the measured room.

ecfr.govopb.orgaccessable.co.ukeugenecascadescoast.org

Built with what exists

A tape measure and a laser distance meter; a phone camera; A11yJSON, MIT-licensed and maintained by Sozialhelden, the Berlin nonprofit behind Wheelmap, whose code already defines doors, showers and restrooms; OpenStreetMap to place each hotel; a low-cost model such as Meta's Muse Ember (see the Note "The jobs agents repeat") to list the claims in each online description, with a person checking.

  1. First step

    Ask three hotels in your city for permission, measure one accessible room in each, and write the results into A11yJSON.

  2. Done looks like

    At least 60 rooms in 30 hotels measured and released under CC BY, a published table of how often each measured feature is missing from or contradicted by the hotel's site and each booking site, and a second person re-measuring 10% of rooms with every measurement within 1 cm.