Flipkart Product Discovery

A catalogue of well over a hundred million products does not have a search problem; it has a discovery problem, and they are not the same thing. Search assumes the user can name what they want. Most of the interesting shopping, and most of the abandoned sessions, happens when they cannot: they know the job, the budget, the occasion, or a picture in their head, and the platform asks them for a keyword.

The specific failures

Search overload: a generic query returns too much, and the ranking has to guess an intent the query never expressed. Category ambiguity: the same product lives in several places, so browsing by category is a coin toss. Personalisation gaps: recommendations that ignore region, price sensitivity and the context of the visit feel generic in a market where those three things decide almost everything. Mobile-first reality: patterns inherited from desktop, with its wide filters and long result lists, do not survive a five-inch screen and a patchy connection. Language: a large share of users do not think in English product names, and both text and voice search punish them for it.

What I would build

An intent layer above search: semantic understanding of the query, visual search from a photo, voice tuned for Indian accents and product vocabulary, and parsing of compound queries that carry several constraints at once. A discovery engine that treats context as a first-class input (location, time, device, budget signals, what similar households nearby actually buy), so that the home surface is a shelf arranged for this person rather than a leaderboard. Progressive navigation: categories that adapt to behaviour, filters with visual previews, one-tap presets for the common jobs, and comparison that does not require opening five tabs. And discovery moments (trending, deals, inspiration, social proof) as editorial surfaces with a point of view, not a carousel of whatever converts.

Design principles

Intent over keywords. Progressive disclosure: the right few options first, more on demand. Contextual intelligence: adapt to where the user is and what they are doing, not to who they were last month. And reduce the distance between finding and buying at every step, because each step is where the session dies.

What I would measure

Share of sessions that end in a purchase when they began without a specific product in mind. Search abandonment after the second reformulation: the moment the user has told you the query is not working. Pages viewed before the first add-to-cart. Conversion by language and by region, because an average hides exactly the users this is for. And the lift, if any, from each discovery surface measured against a proper holdout rather than against last week.

The learning

Scale creates its own problems. At this size, discovery is not a feature of the search box; it is the product, and it has to be designed as a system, with models, information architecture and editorial judgement working together, or the catalogue's size becomes the reason people leave.