K.

Back

K.

Back

K.

AI Powered Recommendations

Defining an AI-powered recommendations system to improve content discovery and eliminate dead-ends across Radio Times & Good Food.

Project details

Team

Kanza Leghari

Snr Product Designer

Magda G.

Product Designer

Software Engineers

Alex M.

Lead App Developer

Platforms & Softwares

UserTesting

User research & Card sorting

Figma

Design & Prototyping

Figma + Claude

Claude & FAISS

Vector embeddings

Summary

Users across Radio Times and Good Food frequently reached dead-ends in their browsing journeys, with limited support for meaningful next steps after viewing content. Existing related content relied on basic rules or keyword matching, often resulting in generic or irrelevant recommendations.

I led the research on defining a new approach to content discovery by identifying the most meaningful “next best actions” for users and shaping an AI-powered recommendation system to support them. Through journey mapping, user research and card sorting, we established the types of relationships users expect between content. These insights informed the development of a model that generates more contextually relevant recommendations based on deeper content relationships.

Users across Radio Times and Good Food frequently reached dead-ends in their browsing journeys, with limited support for meaningful next steps after viewing content. Existing related content relied on basic rules or keyword matching, often resulting in generic or irrelevant recommendations.

I led the research on defining a new approach to content discovery by identifying the most meaningful “next best actions” for users and shaping an AI-powered recommendation system to support them. Through journey mapping, user research and card sorting, we established the types of relationships users expect between content. These insights informed the development of a model that generates more contextually relevant recommendations based on deeper content relationships.

Expected increase in stickiness

Expected uplift in session durations

Continue reading for the detailed version.

Research &
Discovery

Flow mappings

To understand where discovery was breaking down, we mapped key user journeys across both Radio Times and Good Food, focusing on how users navigate from one piece of content to the next. This allowed us to identify moments where users reached dead-ends or lacked clear next steps after viewing content.

The mapping revealed that while users could successfully arrive at content, the onward journey was often unsupported, with limited or low-quality related content options. This highlighted a gap in the discovery experience, where users were not being guided towards meaningful next actions, reducing opportunities for deeper engagement.

Card sorting and user research

To define what meaningful “next best actions” look like, we conducted user research and card sorting to explore how users expect content to relate to one another.

Across both brands, users consistently preferred recommendations that felt contextually relevant rather than generic. For entertainment, this included relationships such as franchise, genre, sub-genre, tone and cast. For food, users valued connections based on ingredients, dietary needs and meal type.

These insights highlighted that effective recommendations require a deeper understanding of content relationships beyond simple metadata or keyword matching.

From rules-based to AI-powered

To enable more meaningful recommendations, the team explored an alternative to building increasingly complex search queries. Based on our research, users were looking for content that felt similar in tone, mood or narrative, something that traditional rules-based systems struggled to capture.

Instead, we moved towards an AI-driven approach that represents each piece of content using multiple signals, including titles, genres and descriptions. By placing greater emphasis on descriptive context, the system can identify deeper relationships between content, such as shared tone or themes, rather than relying on keyword overlap alone.

This allows recommendations to be generated based on overall similarity, surfacing content that feels genuinely related in experience. For example, rather than matching on a single keyword, the system can identify films or shows that share a similar mood, storyline or genre blend, resulting in more relevant and intuitive recommendations.

Design & Experimentation

Phase 2

In-Context related content

Based on the insights gathered and empowered by FAISS vectorised embeddings, we already have contextual related content live on the RT app to get data on usage and impact on onwards journeys. Next step is to get it live on BBC Good Food recipes in the app.

Tailored discovery enabled by AI

Research showed that users don’t think in rigid categories, but look for content that feels similar in tone, mood or context.

Based on this, we defined key signals such as genre, sub-genre, tone, cast, quality and recency to inform an AI-powered recommendation model. This allows content to be matched on deeper relationships, rather than keyword overlap alone.

This enables a more flexible discovery experience, where users can explore content through combinations of attributes like mood, themes or editorial groupings, and refine results based on relevance.

In-app UI is in-progress

Outcomes &
Learnings

The AI-powered recommendations are currently being implemented across show pages, improving the relevance of related content and reducing dead-end journeys in discovery. Early iterations demonstrate a clearer alignment between recommendations and how users naturally perceive content relationships.

The second phase of the work, focused on enabling more expressive, user-driven discovery within the app, is still in progress, with UI and interaction design currently being explored.

This project reinforced the importance of using AI to support genuine user needs rather than forcing its application. The value came from defining the right signals and relationships upfront, ensuring the system reflects how users think about similarity, rather than introducing AI as a feature in itself.

The AI-powered recommendations are currently being implemented across show pages, improving the relevance of related content and reducing dead-end journeys in discovery. Early iterations demonstrate a clearer alignment between recommendations and how users naturally perceive content relationships.

The second phase of the work, focused on enabling more expressive, user-driven discovery within the app, is still in progress, with UI and interaction design currently being explored.

This project reinforced the importance of using AI to support genuine user needs rather than forcing its application. The value came from defining the right signals and relationships upfront, ensuring the system reflects how users think about similarity, rather than introducing AI as a feature in itself.