For the last of its six-part series produced in partnership with OpenAI, INMA’s latest Webinar focused on one of the most urgent questions facing publishers today: how to use AI to personalise news experiences in ways that deepen engagement and support sustainable growth.

The session was moderated by Sonali Verma, lead of INMA’s Newsroom Transformation Initiative, and featured Erica Greene, director of engineering for machine learning at Yahoo News, and Ritvvij Parrikh, senior director of product management/AI at Times Internet.

They shared their personalisation experiences to offer insight into the opportunities and obstacles news organisations face as they integrate AI‑driven personalisation into core strategies.

Inside Yahoo News’ personalisation strategy

As director of engineering for personalisation at Yahoo News, Greene oversees the systems that shape what 190 million users see across one of the world’s largest news and information platforms.

Yahoo News has evolved over time and now combines news, entertainment, lifestyle, games, weather, and more. Today, the company is a global content aggregator with thousands of publishing partners, making personalisation not just a feature but the backbone of the user experience.

At the centre of Yahoo’s strategy is its continuously updating, multimodal homepage feed: “There’s an infinite scroll feed of content, sort of like Twitter, but this is all video and article content from our publishing partners, and it’s personalised,” she explained. “So if you go and you read some things and you come back five minutes later, it will have been updated, and you will see content that is personalised to you.”

Working behind the scenes is a sophisticated ranking system that evaluates up to 80,000 potential content items at any moment and narrows them to roughly 200 that appear on the homepage.

The goal, she said, is to “build a best‑in‑class recommendation system to enable effective products and drive growth.”

Personalisation at Yahoo is multidimensional. The system considers behavioural signals, such as what users click, how long they stay, and what they return to, but also contextual factors such as time of day, day of week, and location. Users can also explicitly tell Yahoo what they like, a direction Greene said her team is working to expand.

Of course, editorial judgment remains part of the equation as well, and the editorial team curates stories based on “what is important in the world.” This ensures personalisation doesn’t overshadow news value or public‑interest priorities.

Powered by data

To power all of this, Yahoo relies heavily on structured data — metadata, taxonomies, embeddings, entity extraction, and other layers of information that help machine‑learning models understand content at scale. This structured data is essential for building not only recommendation systems but also new consumer features, Greene said.

“When content comes in, we extract a bunch of information about that content to make it easier and better and more accurate to recommend to people,” she said. “And we can build a lot of other products, not just the recommendation system, but further products on top of that structured data.”

This enables products like the recently launched “cheat sheet,” which highlights the key people or topics in an article and lets readers explore related coverage. It also underpins the Yahoo 100, a new feature that tracks trending topics in real time by combining clusters of similar content with editorial ranking and summarisation.

Greene framed this work within a broader shift in the industry: the move from unstructured to structured data, accelerated by large language models. Tasks that once required months of engineering — such as parsing recipe ingredients or tagging content — can now be done “in five minutes,” she said, though she cautioned that LLMs introduce new challenges around accuracy, reproducibility, cost, and model deprecation.

Still, the opportunity is enormous: “Basically, the world is your oyster if you’re sitting on the data,” she said.

Yahoo doesn’t view personalisation as simply a way to increase engagement; it is an essential tool for sustaining a healthy ecosystem of publishers. Because Yahoo monetises through advertising and revenue‑sharing, better recommendations mean better outcomes for partners.

“If we grow and are successful, then our publishing partners also grow and are successful,” Greene said.

Her message to the industry was both pragmatic and optimistic: Personalisation is hard, but AI is making it faster, cheaper, and more powerful than ever.

“Being able to ... build reproducible pipelines in this world is going to be the core product problem of our time,” she said. “We are in an industry that needs some kind of killer products, and I think it's going to be really fun to see all of you build stuff on top of this.”

How Times Internet is reimagining personalisation

Times Internet has faced a very different challenge: building algorithmic sophistication within a traditional news organisation that competes with the speed, scale, and data advantages of the world’s biggest tech platforms.

“In India, historically, news companies have relied extensively on Google and YouTube for their traffic, and that traffic is extremely volatile,” Parrikh said.

Since the entry of LLMs, these publishers have been watching the traffic drop.  

The challenge is nothing new, he said. Media — be it movies, music, newspapers, or digital news — has moved progressively toward personalisation. Movies went from single screens to multiplexes to streaming services; music went from the radio to personal listening devices to streaming services.

He framed the challenge for news media companies simply: “We are operating a Netflix‑category product with the tools of a multiplex.”

In other words, audiences expect the fluid, predictive, deeply personalised experiences of modern digital platforms, but most newsrooms are still working with fragmented data, limited compute, and legacy workflows.

Within the broader Internet economy, Parrikh explained, news products compete not only with other publishers but with quick‑commerce apps, ride‑hailing services, e‑commerce giants, and social platforms — all of which command far richer user data and far higher engagement.

“All of us compete for the same advertisement and the same subscription dollars. But if you look at news products, typically they tend to have lower login rates, they have little to non-first party data, they have very little intent gathering.”

Someone reading a yoga article, for example, does not necessarily want to buy a yoga mat. This structural disadvantage makes personalisation both more difficult and more essential.

Changing optimisation

To thrive, news organisations must shift from rule‑based segmentation to machine learning‑driven optimisation, Parrikh said:

“A lot of news media rely on analytics and if/then business rules or segmentation. Analytics is essentially looking in the rear-view mirror and driving your car. It is past data. It is data from a week ago, a month ago, etc.”

Machine learning, by contrast, continuously tests, adapts, and corrects itself, allowing the system to “stick very closely to reality.”

He shared a chart that showed there is no such thing as a rich country with low electricity consumption and compared it to the current reality of algorithms:

“Basically, there is no rich country without a lot of use of electricity. Similarly, I think there’ll be no rich media company without heavy use of algorithms. That is where everything is headed.”

At Times Internet, feedback loops refresh every 15 to 30 minutes, enabling the algorithm to respond to breaking news, shifting interests, and emerging trends.

To balance personalisation with editorial judgment, Times Internet built a suite of AI agents. The Editorial Judgement Agent evaluates importance, relevance, geography, and public‑interest value, helping filter out stories that are “too niche, too irrelevant.”

A “shelf‑life prediction” model estimates how quickly a story will become stale or common knowledge. And an AI Irreducibility Agent identifies stories that should not be summarised because doing so would strip away essential nuance or emotional depth.

Despite these innovations, Parrikh stressed the company remains conservative in production. Every model is evaluated with labelled datasets, editors have override authority, and the team optimises for false negatives:

“That we are OK to let go of good stories not coming in the feed, but we do not want to amplify bad stories,” he said. “Yes, it reduces the outcome that the app achieves, but this is something that we are comfortable with as a company.”

. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .