Natural-language product discovery on a sportswear PDP

Twistag built a conversational product discovery layer for a global sportswear brand — natural-language search embedded directly in the existing product detail page, trained on a catalogue of 500-plus products, deployed and user-tested with real shoppers in under six weeks by three engineers and a designer. The brand's shoppers could describe what they wanted ("something bold and minimal for a city weekend") but could only express it through a colour dropdown. Twistag built a headless component that layers over the existing PDP without touching the e-commerce platform, using Pinecone for semantic product matching at sub-500ms latency and LangGraph for conversation state. The brand is not named publicly.

Product detail pages have looked the same since the early 2010s — hero image, tabs, size selector, filters in a sidebar — while the way people interact with software has changed completely. Shoppers ask questions. They describe what they want in natural language and expect the system to understand intent rather than parse keywords. A shopper looking for "something bold and minimal for a city weekend" had no way to express that through a colour dropdown. The commercial constraint made experimenting hard: product pages are the highest-converting pages on any e-commerce site, which makes them the riskiest place to change anything. The brand's digital team did not want a PDP rebuild — they wanted to know whether conversational commerce could be validated quickly and cheaply, with enough realism to produce behavioural signal from real shoppers.

Twistag built a working prototype of a conversational product discovery layer embedded inside the existing product detail page, with no changes to the underlying e-commerce platform, deployed and user-tested in under six weeks. Three architecture decisions carried the work.

First, the conversational interface sits on top of the existing PDP rather than replacing it — the familiar layout stays, the conversation is an optional layer, and shoppers who prefer browse-and-filter are unaffected. Second, the component is headless and deployable on any product detail page without touching the platform, so the prototype could be tested in isolation and a production version could roll out page by page or per A/B segment. Third, the retrieval and state layers were split: Pinecone stores every product as a vector embedding so natural-language queries find the closest semantic matches regardless of keyword overlap, at sub-500ms query latency, while LangGraph manages conversation state so the assistant builds on stated preferences and coordinates multi-step retrieval when a shopper asks a comparison question.

The least glamorous and arguably most important piece was the fashion knowledge layer, which translates shopper language into product attribute language ("casual but not boring" maps to specific style characteristics), applies occasion logic ("gym to dinner" narrows to a different set than "weekend travel") and calibrates tone to the brand. The build ran six weeks: week one discovery, weeks two to five build with weekly demos, week six user testing with real shoppers on real product pages.

Stack: Pinecone, LangGraph, Next.js, Node.js. Services: AI agents and conversational systems, vector search and semantic retrieval, product engineering, product design.

Twistag delivered a functional prototype in under six weeks with three engineers and one designer. User testing with real shoppers confirmed the hypothesis: shoppers engaged more deeply, explored more products, and reported higher confidence in their selections. The brand advanced the concept to its production roadmap. Because the component is headless, a production rollout can proceed page by page or per A/B segment without a PDP rebuild. These are process and qualitative outcomes, stated as such — this case study reports no conversion, revenue or retention figure, because none has been published by the brand.

The engineering value isn't in the model. It's in the knowledge layer that translates between user intent and the brand's product reality.