Context
In May 2025 Expedia Group relaunched the Hotels.com app with an AI-powered chatbot trip planner as the headline feature, part of a broader brand revitalisation following loyalty-programme backlash.
The chatbot is built on Expedia Group's underlying AI travel assistant Romie, announced at Expedia's 2024 Spring Release on 16 May 2024 by CTO Rathi Murthy as an alpha experiment on EG Labs, framed by Murthy "with hyper personalization in mind." The model and provider underneath Romie/Hotels.com has not been publicly disclosed.
What the AI actually does
Chatbot trip planning inside the consumer Hotels.com app: query → recommendations for properties, restaurants, and attractions; intended to assemble itineraries for guests.
Skift hands-on testing (20 May 2025) found:
- Wrong-borough recommendations. Skift surfaced a Manhattan property in response to a Brooklyn search. - Activity recommendations not bookable. The chatbot suggested restaurants and attractions but provided no booking links and no integration into a saved itinerary. - No itinerary-building functionality. Despite being called a "trip planner," it did not actually build itineraries. - No flight assistance (out of scope by design, but presented to users in a way Skift considered confusingly limited).
Skift's overall verdict: "frequent errors, inaccurate recommendations, and limited usefulness, particularly for activities and itinerary building", flagged as a reduction in user trust.
Measurable outcomes
- No retention or trust metrics published, Skift's review is qualitative
- No Expedia Group earnings-call mention specifically calling out the Hotels.com chatbot in 2025 calls
- No public redesign or pull-back announced as of April 2026, the chatbot continues to ship
What to copy
The negative pattern: do not ship a marquee AI feature without testing it against the basic accuracy bar of the non-AI feature it replaces.
The launch standard for AI features must be at least as high as the launch standard for non-AI features. A Brooklyn search returning a Manhattan property would not have shipped from a non-AI search team.
What doesn't transfer
Expedia Group has serious AI assets, Romie, the EG Labs alpha, the OpenAI ChatGPT Apps partnership announced October 2025. Capability is not the issue.
What shipped on Hotels.com was a marquee feature in a brand-relaunch press cycle that had not been tested against accuracy fundamentals. The lesson is staging discipline, not capability.
Open questions before buying
- What QA bar did the team apply pre-launch? Did anyone test 'wrong borough' or 'wrong city' as a basic case?
- Why has no public redesign or pull-back been announced 11+ months after Skift's testing?
- Consumer-facing hallucination doesn't need to cause a lawsuit to matter. Moffatt v Air Canada (Feb 2024) is the precedent: BC Civil Resolution Tribunal awarded CAN$812.02 against Air Canada for chatbot misrepresentation. Who owns the output for your AI feature?