Lily AI

§ 01 / Who they are nowThe translator nobody sees

Walk through the e-commerce sites of Macy's, Bloomingdale's, Gap or thredUP and you won't find Lily AI's name anywhere. That's the point. Lily AI is plumbing - the kind of plumbing that decides whether a search for "quiet luxury blazer" returns six relevant items or six hundred irrelevant ones.

The company has roughly 130 employees, around $59.5 million in cumulative funding, and a customer list that reads like a department-store wing of the S&P 500. Its product, at its core, is a giant taxonomy: more than 15,000 attributes that describe clothing, beauty and home goods the way humans actually shop for them, then a model that applies those attributes - automatically, at catalog scale - to product feeds, search indexes and ad placements.

§ 02 / The problem they sawCatalogs that don't speak human

Most retail catalogs were built by buyers, for buyers. They describe products in the language of the merchant: fabric weight, SKU code, vendor name, color family. Useful, sort of, for inventory. Less useful for the person typing "going-out top, not too clubby" into a search bar at 11 p.m.

The gap is small in any single transaction and enormous in aggregate. Retailers spend billions on Google Shopping ads, recommendation engines, site search and SEO - all of which rely on product metadata that was, for the most part, written by humans who think in spreadsheets.

Lily AI's bet was that fixing the metadata, not building yet another shiny consumer-facing AI gimmick, was the highest-leverage thing you could do in retail.

§ 03 / The founders' betA behavioral researcher and a systems engineer

Purva Gupta arrived at the problem from the wrong end of the building. Before Lily AI she worked at UNICEF and Eko India on behavioral communication - figuring out why messages land or don't land with the people they're meant for. Sowmiya Chocka Narayanan, her co-founder, came from the opposite direction: engineering at Texas Instruments, Yahoo and Box, the kind of CV that knows what a 10-billion-row table looks like before lunch.

Their first version of Lily was, briefly, a consumer app. The pivot - to selling the underlying AI to retailers as infrastructure - is the move that turned it into a business. It also explains why the company sounds more like a Cisco than a Casetify when you read its decks.

§ 04 / The product15,000 attributes and a generative quill

The platform has three jobs. The first is recognition: computer vision and NLP read every product image and copy block in a catalog and tag it against Lily's proprietary taxonomy. The second is enrichment: those tags flow into search indexes, recommendation engines, paid feeds and PIMs. The third, added more recently, is generation: a brand-aligned LLM that writes the product description itself, optimized for both human shoppers and the Google Shopping algorithm staring back at them.

It's a quietly ambitious stack. Computer vision does the looking. NLP does the listening. A large language model does the speaking. All of it pointed not at a chatbot but at the most boring page on the internet - the product detail page - because that is where money actually changes hands.

§ 05 / The proofNumbers that move buyers

Retail leaders are professionally skeptical of AI vendors, which is a polite way of saying they have been burned before. Lily AI's pitch survives the meeting because it lands in language CFOs understand: conversion rate, sell-through, cost per click. thredUP, the resale marketplace, reported a 15% lift in sell-through after rolling Lily AI's deep tagging across its categories. Conversion improved 2% for repeat shoppers and 2.8% on iOS. Favorites per user climbed too.

None of those numbers will trend on Twitter. All of them, applied to a multi-billion-dollar GMV base, are the difference between a good quarter and a board call nobody wants to take.

§ 06 / The missionCustomer language, end to end

Ask Purva Gupta what Lily AI is building and the answer is unfashionably long. It is not just better search. It is the same enriched layer of customer language - moods, occasions, trends, intents - flowing into search, recommendations, SEO, paid media, demand forecasting, even merchandise planning. One taxonomy. Many surfaces. Fewer translation errors at every step.

If that sounds like infrastructure, that's because it is. The most useful AI in retail this decade is probably not the chatbot at the bottom right of the page. It's the layer that decides which sweater shows up when somebody, somewhere, types "coastal grandma cardigan."

§ 07 / Why it matters tomorrowThe catalog as a language

Generative AI made it possible to write a billion product descriptions overnight. It did not make any of them good. The next phase of retail AI is less about generating more content and more about generating the right content - the words a real customer would search for, in the moment they're searching for them, in the channel they happen to be in.

That is a taxonomy problem before it is a model problem. Whoever owns the taxonomy - the layer that connects merchant data to customer language - owns a quiet, durable position in how the internet shops. Lily AI is wagering it can be that layer.

The bet is unsexy. It is also, increasingly, the one retailers say they wish they had made themselves.

§ CodaBack to the search bar

A customer types "cottagecore midi for a backyard wedding." Somewhere downstream, an enriched product feed has already learned that the dress on aisle seven is a midi, is floral, reads as cottagecore, photographs well on grass, and tends to convert on warm-weather Tuesdays.

The customer sees six options. She buys one. The retailer ships it. Nobody mentions the AI in the middle. That is exactly how Lily AI prefers it.