Workflow · Production

trivago experiments with generative AI to summarize hotel information for users

The problem

trivago had no prior experience with AI product development and faced challenges around output reliability, non-deterministic results, and how to assess the quality of AI-generated content for their hotel search use case.

First attempt

During testing the AI produced inaccurate and misleading outputs: a recommended restaurant was placed in the wrong city, neighborhoods were placed on the wrong continent, and the model sometimes returned only apologies with no useful answer.

Workflow diagram · grounded in source
1
Use case selection via design sprint
trigger
“We generated ideas in design sprints and ideation sessions with the core product team and stakeholders from across the company”
2
AI hotel review summarization
ai_action
“we want to use AI to summarize hotel information and reviews to provide concise summaries to our users”
3
Prompt engineering iteration
validation
“Evolving from this point towards a reliable output machine for our users requires a lot of 'prompt engineering'”
4
Live testing and feedback
feedback_loop
“Features mentioned in this article are currently undergoing live testing and may not be available to all users”
Reported outcome

trivago reached MVP release for the hotel summarization feature, now in live testing; the team found that simple prompt engineering was more effective than complex model training at improving output quality.

Reported metrics
AI output accuracynot 100% accurate and is non-deterministic
Prompt engineering effectivenessremarkably effective in enhancing the model's output
Source
https://tech.trivago.com/post/2023-09-15-experimenting-with-ai-to-enhance-our-product-firsthand-experience-from-our-product-managers
Read source ↗

Frequently asked questions

What did this team achieve with this AI workflow?

trivago reached MVP release for the hotel summarization feature, now in live testing; the team found that simple prompt engineering was more effective than complex model training at improving output quality.

What results were reported?

AI output accuracy: not 100% accurate and is non-deterministic; Prompt engineering effectiveness: remarkably effective in enhancing the model's output (source-reported, not independently verified).

What failed first in this deployment?

During testing the AI produced inaccurate and misleading outputs: a recommended restaurant was placed in the wrong city, neighborhoods were placed on the wrong continent, and the model sometimes returned only apologie…

How is this workflow AI workflow structured?

Use case selection via design sprint → AI hotel review summarization → Prompt engineering iteration → Live testing and feedback.