At India Today, an AI experiment asks whether audience behaviour can be predicted
From the outset, India Today was clear about what Audipulse was – and was not. “Audipulse was not meant to overcome any bottlenecks, “said Bal Krishna, who leads the Fact Check team at India Today. “It was meant to assist the editor on the basis of real data, rather than assumptions.”
Analytics dashboards tell editors what worked yesterday. India Today wanted to predict what might work tomorrow.
The result was Audipulse, an AI-powered audience prediction engine that analyses engagement data and generates predictive signals around story performance, publishing time, and content format.
To support the project, India Today joined the 2025 edition of Newsroom AI Catalyst, a WAN-IFRA accelerator programme in partnership with OpenAI, which supports publishers experimenting with AI initiatives in the newsroom.
Founded in 1975, India Today is one of India’s largest news media organisations, with operations spanning television, digital, print, and video. Its digital newsrooms publish across politics, entertainment, sports, business, and breaking news.
When analytics arrive too late
The project emerged from a familiar newsroom reality: editors had access to large volumes of engagement data, but turning that information into clear editorial decisions remained difficult.
Manual analysis was slow and subjective. Existing analytics systems were useful in measuring performance after publication, but less effective in helping editors make forward-looking decisions.
“In today’s digital space, where people do not actively choose the source of news, and instead become passive consumers of whatever the algorithm pushes, it is important to know what your audience expects from you to keep them loyal,” said Bal Krishna.
The project also reflected concerns around sending Google Analytics and Comscore data to external cloud environments. To address that, the system was deployed on-premises using local GPU infrastructure.
Predicting engagement before publication
Audipulse combines previous-day analytics with draft headlines prepared for publication the next day.
Using data from Chartbeat and Google Analytics, the system analyses clicks, engagement, time spent, story topics, and formats including text stories, videos, picture stories, and interactive formats.
Based on those inputs, the model predicts engagement outcomes and recommends publishing times and formats.
The system also retrains continuously by comparing predicted outcomes with actual performance data.
During a 15-day pilot, Audipulse achieved a prediction precision rate of 64 percent compared to a 52 percent editor baseline.
Internally, prediction accuracy was measured by evaluating how stories predicted to perform well ultimately performed the following day.
“AI is very efficient at analysing data and identifying trends,” Krishna said. “Even if ample data on audience behaviour is collected by an organisation, it is extremely difficult to reach a definitive conclusion without the help of AI.”
Why cricket improved the model
One of the clearest lessons from the pilot was that audience prediction required more than performance data alone.
The team found that adding contextual taxonomies such as elections, cricket, and Bollywood improved prediction precision by 11 percentage points.
At the same time, the experiment exposed the limitations of purely data-driven systems.
“The biggest concern was that while the data-driven approach can be good for predicting trends, it struggles to capture the deeper context associated with the stories and the topics,” Krishna said.
Improving the outputs required repeated monitoring, additional data, and manual refinement.
“To overcome this, it needs more data and monitoring the output and refining it, which in turn needs dedicated resources,” he said.
The pilot also revealed editorial scepticism around predictive recommendations early in the process, until side-by-side results were demonstrated during testing.
Operationally, the team managed cloud costs through overnight batch retraining.
Prediction still needs context
Audipulse remains in development and testing, with the newsroom continuing to refine the quality and reliability of predictions.
Future plans include extending the system to video thumbnails and push alerts, while also building an explainability layer designed to show the factors influencing individual predictions.
The team is also exploring a longer 30-day A/B test as part of the next phase of experimentation.
For now, however, the project remains focused on a narrower question: whether predictive systems can meaningfully support editorial decision-making in environments increasingly shaped by algorithms.
