DirectBooker x ChatGPT

A new surface for an old problem

When OpenAI launched its Apps SDK and Model Context Protocol in October 2025, it opened a new distribution channel — one where users arrive with active, open-ended intent. Travel planning is a natural fit: most people begin with vague questions ("Where should I stay in Kyoto in May?") rather than rigid filter inputs.

DirectBooker, as a platform that focused on restoring power to direct hotel bookings, saw an opportunity. OTAs dominate hotel discovery online — they obscure direct prices, hide loyalty tiers, and capture margin that could go back to travelers. The ChatGPT canvas offered a way to intercept users before they defaulted to Booking.com or Expedia.

How do you design a product that lives entirely within conversation — no persistent UI, no page navigation, just text and intent?

Problem & Goals

OTA algorithms are designed to maximize their own revenue, not the traveler's value. Direct booking benefits — better cancellation terms, loyalty points, room upgrades, breakfast perks — are invisible on aggregator pages. Most users don't know what they're missing.

Goals

• Drive efficient hotel discovery through conversational AI

• Surface direct prices, loyalty benefits, and exclusive perks at high-intent moments

• Increase trust and conversion for direct bookings

Research - Understanding how travelers actually think

Before designing any flows, I ran a discovery sprint — competitive analysis, user interviews, and behavioral walkthroughs — to understand where OTA discovery breaks down and what conversational AI could uniquely address.

Key Findings

1. Direct prices are invisible

   Users didn't know direct prices are often equal or lower. They assumed OTAs had the best deals by default.

2. Intent-led discovery

   Many users begin with intent-based queries (‘near xxx Park,’ ‘pet-friendly,’ ‘good for remote work’) instead of predefined filters.

3. Loyalty benefits are invisible

   Users who belonged to hotel loyalty programs frequently didn't know how to apply benefits — especially at unfamiliar properties.

From vague intent to confident booking

The core design challenge was mapping the messy, nonlinear reality of travel planning onto a reliable, sequential conversational flow — without making it feel like a form disguised as a chat.

In a chat-native product, the words are the interface. I spent as much time on response structure, tone, and information hierarchy as I would on visual components in a traditional product.

Design system: adapting for a new surface

The core challenge was determining what to carry over: the full color palette was retained for brand recognition, while typography and components were rebuilt to align with ChatGPT's visual language. The result is a mini design system that keeps the branding of DirectBooker while feeling native to the platform.

What we tried, what we cut

Not every approach made it to the final product. Here are three directions we explored and the rationale behind the decisions.

Direction C is what we finally shipped — a modular adaptive response. Instead of one fixed layout, the system assembles a response from varieties of modules based on what the user actually asked and prompt.

Multi-Module Design: built for variable LLM output

The system is composed of multiple intent-driven modules, each surfacing in response to a distinct user need. When users want to evaluate options side by side — comparing pricing, cancellation terms, and inclusions — the comparison module activates. Location-based queries trigger the map module. As users approach a booking decision and seek deeper property detail, a dedicated hotel profile module takes over.

The product we shipped

The final DirectBooker app in ChatGPT is an opinionated conversational assistant — it asks fewer questions, surfaces more relevant context, and makes direct booking feel like the obviously smarter choice.