Related Recommendations as a Service

Problem Description

Our existing related content recommendation system lacked the flexibility and editorial control needed across our multisite publishing network. Editors often found the automated recommendations irrelevant or too limiting, which led them to manually curate related articles, a time consuming and inconsistent process.

One of the major pain points was the inability to select which brand(s) within the network content should be recommended from. Editors wanted to blend content from multiple sites, or restrict it to the current one, depending on the article’s context and audience. Additionally, they wanted the ability to manually override recommendations by entering a specific keyword or phrase to surface more relevant matches.

Control over how many related items appeared and where they were placed on the page was also essential, as different content types and layouts had varying needs. Without these capabilities, the editorial team struggled to deliver a cohesive and personalized content experience to readers, and the user journey suffered as a result.

Stakeholders

Editorial Teams

The primary users of the system, editors needed an intuitive and reliable way to surface relevant content for their readers. They wanted more control over recommendation sources, manual override options, and flexibility in placement and quantity of recommendations. The lack of these features led to increased manual work and frustration.

Product Management

Product stakeholders were focused on improving user engagement, time-on-site, and content discoverability. They needed a system that could scale across multiple brands while supporting consistent user experience and editorial freedom.

Engineering Teams

Engineering was responsible for implementing and maintaining the recommendation engine across multiple websites built on different stacks, including WordPress. They required a centralized, easy-to-integrate solution that could support both custom applications and CMS-based platforms without additional overhead.

Commercial / Revenue Teams

While not day-to-day users, commercial stakeholders were invested in increasing page views per session and ad impressions by keeping users engaged. They supported the project to drive measurable ROI through improved content recirculation.

Business Value & Commercial Impact

Replacing the third-party recommendation system eliminated an annual subscription cost. The new in-house solution increased editorial efficiency by reducing manual content curation time by 90%, allowing teams to focus on higher-impact work.

Its flexible architecture enabled integration across multiple brands and platforms with minimal development effort, cutting implementation time for new brands.

• Eliminated vendor subscription cost 

• Increased user engagement : pages per session grew by 10–15% due to better content relevance

• Boosted ad impressions and revenue, longer sessions and more page views directly increased ad-based revenue

• Reduced editorial time ,less manual work freed up ~5–10 hours per editor per month, reducing cost and improving output

• Faster onboarding for new brands , reduced integration time from 3–4 weeks to 2–3 days, accelerating time-to-value

• Improved UX , consistent, smarter recommendations led to lower bounce rates and higher return visits

My Role

• Designed the system integration between Elasticsearch and the cloud function

• Created the query structure and handled index configuration

• Developed and deployed the Cloud Function to dynamically return valid RSS XML

• Collaborated with product and marketing teams to define content types and metadata

• Collaborated with product managers to define requirements that maximize business impact

• Presented solution benefits to sales teams, enabling better positioning and faster adoption

• Provided technical demos and training to business users and external clients to facilitate onboarding

Technical Overview
The related recommendation system was built as a serverless API using Google Cloud Functions, which allowed us to decouple recommendation logic from the frontend and make it accessible to both WordPress sites and custom applications.We used Terraform to provision and manage the infrastructure, ensuring consistency, repeatability, and ease of deployment across environments.

Key Technologies Used:

• Google Cloud Functions — Stateless backend, scalable on-demand

• Elasticsearch — Content indexing and similarity scoring

• Terraform – For infrastructure provisioning and configuration, including Cloud Functions, IAM roles, and environment variables

• Custom WordPress Plugin (PHP) – Handled metadata collection, API requests, admin settings, and output rendering within the site theme

• Cloud Logging & Monitoring

• OpenTelemetry

How It Works (Flow):

1. On article load, the frontend sends a request to the recommendation API with article metadata (e.g., tags, brand, content type).

2. The Cloud Function parses the request, builds an Elasticsearch query with optional filters (brand, tag boosting, etc.).

3. Elasticsearch returns a list of relevant content items ranked by relevance.

4. The service returns a clean JSON response.

5. The widget renders the results in the site’s UI.
   
   

Additional Features:

• Manual keyword override support

• Cross-brand and multi-brand filtering

• Configurable number of recommendations and placement

• Fallback rules (e.g., show recent popular content)

Impact / Results

• Deployed across 24+ brands with different tech stacks

• Reduced engineering duplication across platforms

• Average 15–30% increase in engagement (e.g., time on page, CTR)

• Added flexibility for editors to control relevance logic (e.g., curated tags, filtering by recency)