A search that returns nothing the shopper clicks is normally filed as a failed session. Flipkart has started treating it as a mailing list. Wiring a product-research agent into CRM converts retail search from a channel that waits for intent into one that schedules its own next occasion, and it moves discovery spending off the media plan and onto infrastructure the retailer already owns.

The pipeline starts by hunting for a specific kind of failure. Mandar Kulkarni, Pooja A. and Samir Shah of Flipkart describe a rule-based filter, in a paper posted to arXiv on 19 August, that runs over the company’s search logs and keeps only queries satisfying four conditions. The search drew no clicks; the user carries a high-affluence flag; the wording contains a subjective qualifier such as best, latest, top or under; the product sits in the mobile phones vertical. What survives is a shopper who asked something the search box could not answer and then left. Over 23 days the pipeline turned those into 15,061 WhatsApp messages, each carrying two or three products behind a shortened, tracked link.

The decisive move is where the architecture puts the research. A Discovery Agent goes out to Google, review blogs and video, issues several reworded versions of the query, and keeps only products that multiple independent sources name. That is the trip the shopper was about to make alone, taken by the retailer instead and returned with a price attached. A Flipkart Search Agent grounds each candidate to a live catalogue entry, checking stock against the delivery pincode and pricing it against the customer’s tier and current bank offers. A Review Agent then reconciles every specification claim in the drafted message against catalogue metadata before anything is sent. The off-platform research step that used to end the session has been absorbed into the message that restarts it.

Flipkart reports a click-through rate roughly 285% above its historical WhatsApp campaigns. The number needs an asterisk the paper supplies itself: read rates rose only about 8%, because a read can be attributed only to the original recipient. Clicks were counted for everyone, since the tracked link kept working after the message was forwarded. The authors credit both relevance and forwarding for the lift without separating them, so the headline figure measures reach and quality at once. The comparison is against historical campaigns rather than a holdout, and the GMV claim arrives without a figure.

The forwarding is the part worth copying. The campaign generated 37,258 visits, about two and a half for every message sent, and on several days it drew more clicks than notifications. Nobody forwards a promotional push. What travelled was a shortlist with its reasoning attached, moving through the channel where a shopper would otherwise ask a friend which phone to buy. Flipkart did not buy that distribution.

The retailer, not the shopper, now sets the date of the second conversation.

The cost line is what makes this a fashion problem rather than an Indian electronics story. Flipkart runs the pipeline on Gemini 2.5 Flash and puts the inference at two to three cents a query, which is why it can afford to filter for high-affluence users instead of serving everyone. That filter has an older name in this industry. Clienteling is the practice of a salesperson who knows a good customer and writes when something arrives, and we covered Printemps in July putting a personal shopper’s telephone number on screen and counting relationships opened rather than baskets closed. The practice has never scaled because it costs a salary. What Flipkart added is not the idea. It is the price.

Most current writing on agentic commerce points the agent the other way. The assumed threat is a general assistant that sits between retailer and shopper, intercepts the query and demotes the retailer to a fulfilment endpoint, which is why so much retail budget now goes on being legible to somebody else’s model. RTB House’s June and July survey of 1,840 shoppers, reported by PPC Land, found 42% of US respondents saying AI tools lengthen their purchase decisions, a figure usually filed under bad news for conversion. Read against this architecture it inverts: a longer research phase leaves more unanswered queries in the retailer’s own logs, timestamped and attributable to a known customer. The defensive posture treats the gap between question and purchase as territory being lost. Flipkart treats it as a queue.

The condition under which this does not transfer is worth naming precisely. Smartphones have ground truth, since a specification is an integer and a launch date is a fact, which is why the Review Agent scored 99.1% on specification accuracy when a human team checked 2,218 of its recommendations. The Discovery Agent’s method depends on independent sources converging, and technology blogs converge because there is something to converge on. Fashion has no agreed ranking of the best coat for a mild autumn, no launch date to re-verify, no external corpus that settles. If verification is what makes an unsolicited message safe to send, and fashion cannot verify, this architecture stops at the electronics vertical.

The objection misreads what the Review Agent does. It does not certify that a recommendation is good; it certifies that nothing in the message contradicts the catalogue. Fashion catalogues carry checkable fields of their own: fabric composition, the price after this customer’s discounts, whether the shopper’s size is in stock, the delivery date, the return window. Those are the claims that cost trust when they turn out to be wrong, and they are the ones outbound fashion messaging gets wrong most often. On the missing corpus, Amazon researchers built an agent for subjective gift-selection needs at WSDM 2025, reading customer reviews rather than expert rankings on the grounds that taste lives in reviews and not in catalogue data. The same logic extends to fashion, where retailers hold more of that material than any technology blog holds about phones.

If agentic search keeps stretching the distance between a shopper’s first question and their purchase, the value of a retailer’s own search log rises with it. That log is the one record of what a customer wanted that no platform can sell back to the brand at auction. Turning it into scheduled contact reclassifies discovery from a media buy, rented and repriced every quarter, into an operating asset with a marginal cost measured in cents. Retailers who make that move stop bidding for the second visit and start issuing it. The ones who do not will go on paying a platform to reintroduce them to people who already asked. Every retailer already holds a record of the questions it failed to answer, and the only decision left is whether that record stays a diagnostic or becomes a channel.