Meesho

Objective

Get Indian consumers to buy online who do not know how to, or do not want to, type a query, often in a local language and for their first purchase: the feed guesses the intent for them, and voice replaces the search bar for those who prefer to speak.

The deployment

Meesho, an Indian marketplace that claims 264 million annual buyers at the end of March 2026, starts from an observation set out in its shareholder letter of May 6, 2026: most of its customers arrive without a specific product in mind and do not type a query. Its recommendation system PRISM (Personalised Ranking and Intent Signal Module) therefore builds most of the journey from a personalized discovery feed rather than from the search bar. According to the letter, PRISM combines more than 100 ranking models (up to 300 million parameters), reads the user's long-term profile and in-session behavior, and runs more than 6 trillion inferences per day. One module, Trendpulse, uses LLMs to spot regional search spikes and derive cultural trends to push in the feeds. Meesho quantifies three effects: more than 75% of orders come from these personalized feeds, conversion rose about 15% year on year thanks to improvements to PRISM, and the time it takes a new seller listing to find its first buyers fell by about 27%. The second system, Vaani, is a voice shopping agent launched in the fourth quarter of fiscal 2026 (January to March 2026). The user describes what they are looking for in their own language, asks questions, and the agent chains catalog search, customer reviews when the user hesitates, payment and confirmation; speech recognition runs on the phone to cope with slow networks. Meesho reports 1.5 million users in the first month, conversion up 22% among those who adopted it, and 79% of users saying voice makes shopping simpler. On the analyst call, CEO Vidit Aatrey adds that Vaani helped convert more rural customers and lowered acquisition cost, without a figure. Two caveats: the +15% and +22% are internal measurements whose method (control group, period) is not published, and the +22% concerns voluntary adopters, a population that probably already converts better than average.

Results Proof A

The figures come from the Q4 fiscal 2026 shareholder letter filed by Meesho Limited with the NSE on May 6, 2026, are repeated in the earnings release filed the same day and discussed by the CEO in the official transcript of the analyst call; Outlook Business reported them the same day. Financial communication from a listed company, several consistent sources.

The internal detail is not public. Here is a proven approach that leads to the same result, to adapt to your stack.

1. 1
   Signal collection Platform

   Every product view, add to cart, order, rating and return feeds the user's profile and the training data.

2. 2
   Feed ranking AI

   PRISM ranks the products to show in the home feed based on the profile and the intent estimated in the session.

3. 3
   Trend detection AI

   Trendpulse reads regional search spikes with LLMs and derives trends to push in the feeds before demand is expressed.

4. 4
   Voice journey AI and customer

   For users who choose Vaani, the agent understands the request in their language, searches the catalog, shows reviews when the user hesitates and guides them through to payment.

5. 5
   Measurement and experimentation Product and data teams

   The teams track the share of orders coming from the feeds, conversion and the ramp-up of new listings; Meesho says it doubled the number of experiments on the platform in Q4 compared with the previous year.