Historically, online shopping has been constrained by fixed interfaces, static catalogues, and predefined navigation patterns. Consumers could only explore within the boundaries established by the platform.
At the Data Engineering Summit (DES) 2026, organised by AIM, Arvind Jayaprakash, Senior Vice President of Technology at Glance, says generative AI is fundamentally changing how consumers discover products online.
For years, recommendation systems were built around a simple idea of matching users to items.
Whether it was an e-commerce platform suggesting a dress, a music app curating a playlist, or a travel site recommending a destination, the underlying logic remained largely unchanged. The goal was to predict which items a user was most likely to engage with.
“Classically, a recommendation system is that you have users, you have items, and you’re trying to do user-item matching,” Jayaprakash explains.
Success was measured in a fairly straightforward way. Recommendation systems predicted the probability of a positive action, such as a click or an add-to-cart, and optimised for that signal. It was a clean, well-understood problem, and the industry had become very good at solving it.
But generative AI is dismantling that model.
When Items are No Longer Finite
The most fundamental shift, according to Jayaprakash, is not about algorithms or infrastructure. It is about the nature of items themselves.
“What shifts is that the items are no longer finite. The items themselves can be generated,” he says.
According to him, the idea is that the inventory user browsing is no longer fixed and can instead be generated on the fly, thereby changing everything about how recommendation systems need to be designed.
Traditionally, products were displayed on neutral white backgrounds. The reason, he notes, was the need for mathematical and behavioural neutrality.
But consumers rarely shop in a purely logical manner. They imagine themselves wearing the product, using it in a certain place, or fitting into a certain lifestyle.
Generative AI allows platforms to bridge that gap. Instead of showing the dress isolated on a white background, the system can render the clothing on a user, inside a realistic environment, with different moods, poses, or scenarios.
The New Problem
In older recommendation systems, if a user ignored a product, the explanation was relatively simple. The recommendation was probably wrong.
But with generative AI, failure becomes far more ambiguous. Jayaprakash illustrates this with an example involving a user shopping for a sweater. The recommendation itself may have been correct, but the generated setting could have looked odd. The AI may have chosen the wrong context, or the rendering system may have distorted the user’s appearance.
“Any of these three failure modes could have happened, which could result in the user not positively engaging,” he adds.
This makes optimisation significantly harder because platforms are not evaluating a single prediction. They are evaluating the output of a layered decision-making process that involves recommendation, contextual generation, and multimodal rendering.
“How do you build a recommendation system now on a composite output?” he asks. That, Jayaprakash highlights, becomes the new challenge for AI-powered commerce systems.
From Similar Products to Infinite Exploration
One of the most significant changes discussed during the session was how generative AI changes the idea of ‘similar products’. Jayaprakash demonstrates this with a scenario where a consumer sees an outfit and simply types, “Make it blue”.
At that point, the platform’s predefined recommendation graph no longer constrains discovery. “The user can explore now in any direction,” he says.
The AI-generated modification may correspond to an entirely different item in inventory—possibly from another seller, another brand, or another pricing category. That means the recommendation engine must continuously re-evaluate the catalogue based on the user’s evolving prompts.
Jayaprakash explains that the process begins to resemble the experience of shopping inside a physical retail store.
Recommendation and Generation are Becoming Indistinguishable
Jayaprakash emphasises that recommendation and generation are increasingly blending.
The recommendation engine chooses the initial product. As the generative system determines the presentation, the consumer modifies the output through prompts. As a result, the recommendation engine re-enters the loop to match updated preferences with inventory.
This creates a hybrid discovery architecture where generation and recommendation continuously interact.
“You’d actually see recommendation and generation going hand-in-glove, perhaps in a very indistinguishable way in consumer flows,” he says.
The way consumers interact with e-commerce platforms could look very different over the next few years. Consumers may describe what they want conversationally, visualise these products, modify attributes in real time, and move fluidly between recommendation and creation.
Soon, discovery will shift from ranking products to collaboratively shaping intent.
Your reaction
Discussion
What's your take on this story?
No comments yet — be the first to share your view.