Overview of this campaign

In a digital newsroom, stories evolve constantly, and so does audience interest. Our Insight and Analytics team has access to huge amounts of performance data, but that data often arrives in a form that’s hard to act on: article-level metrics can spike or collapse for reasons unrelated to medium-term trend performance (such as breaking news or virality), and traditional topic detection struggles to keep up with the pace of change.

The aim of this initiative was to use generative AI to create a more practical form of editorial intelligence: an automated system that detects emerging themes across the Telegraph’s output and translates them into segments that analysts can actually use.

The approach combined three elements:

• LLM embeddings generated from full article text, capturing semantic meaning at scale
  

• Grouping of articles on the same topic using a sparse autoencoder (CompresSAE) 
  

• LLM-based segment naming, where generative AI produces concise, interpretable labels based on article metadata, making segments understandable without technical knowledge
  
  

The system refreshes regularly, producing an up-to-date picture of the news landscape and surfacing it through dashboards. Our Insights and Analytics team can monitor emerging topics, compare engagement and conversion across themes, and spot “hidden” segments that are performing strongly but might otherwise be overlooked.

Success meant delivering something fast enough for actionable use, explainable enough to build trust, and scalable enough to run continuously, using GenAI to strengthen the team's ability to report on audience behaviour in real time.

Results for this campaign

This initiative gave The Telegraph a new layer of real-time visibility: instead of analysing performance story by story, the analytics team can now see how audience interest is moving across broader themes, and they can see it in language that makes sense immediately.

The system refreshes every 15 minutes, meaning analysts can track emerging segments almost as quickly as they appear. This supports faster reporting around promotion and commissioning, and helps enrich reporting to explain the difference between short-lived spikes and longer-running narratives that steadily accumulate attention.

By analysing at the segment level, we gain more stable insight than individual article metrics can offer. Dashboards make it easy to compare readership, click-through and conversion across topics, inspect time-series behaviour, and surface outliers, including segments that sit in top-performing percentiles and may indicate under-exploited themes worth further coverage.

By automatically naming segments with clear labels, the system removes much of the interpretation burden. Analysts don’t need to decipher clusters or model outputs, they can focus on the editorial implications.

Contact

To contact a company representative about this campaign, click here for the INMA Member Directory