Maximizing Product Discovery in E-Commerce Using LLM for E-Commerce Technologies

How does LLM catalog enrichment work?

LLM catalog enrichment is an extraction-and-normalisation pass in which a language model reads each product's existing text — titles, descriptions, supplier feeds, spec sheets — and writes back the structured data discovery depends on: consistent attributes, complete facet values, and use-case tags. It attacks the root cause of bad search results, which is thin and inconsistent product data, not the ranking algorithm.

• Attribute extraction — The model pulls material, fit, capacity, compatibility, and dimensions out of free-text descriptions into typed fields, so filters and facets stop silently excluding half the catalog.
• Taxonomy normalisation — "Crimson", "wine", and "burgundy" collapse into one filterable value; sizes, units, and category assignments align across suppliers who each shipped their own schema.
• Use-case tagging — SKUs get tagged with the situations shoppers search — "office chair for back pain", "gift under $50" — which is exactly the vocabulary embeddings match against.
• Query-aware copy — Descriptions are rewritten to answer what shoppers ask, in brand voice, grounded in the source record so no claim appears that the underlying data cannot support.

The guardrail matters as much as the generation. Every enrichment pipeline we ship gates output through a human review queue with per-category approval rules, and tracks review pass rate — the share of generated fields approved without edits — as its quality metric. It is the same grounded-drafting, human-approval pattern our LLM-based outreach systems use before any message reaches a customer.

Enrichment compounds. The same structured attributes improve search filters, recommendation quality, shopping-ad feeds, and the grounding corpus a conversational assistant answers from — one data investment lifting four surfaces, which is why it sits second in the rollout sequence, directly after the baseline.

Which discovery surface should you start with?

Start with site search on your highest-volume category: it carries the highest purchase intent, its baseline can be recorded in two weeks, and everything it forces you to build — the enriched catalog, the embeddings, the vector index, the reranker — is reused unchanged by recommendations and by any assistant added later. Recommendations and conversation are downstream of the same work, not alternatives to it.

The arithmetic picks the category, not the enthusiasm. A store handling 2 million site searches a year at a 15% zero-result rate sends 300,000 sessions to a dead end. Recovering half of those at a 3% search-to-order conversion and a £70 average order value is £315,000 of orders that already had intent behind them. Run the same three lines per category and start with the biggest number — then re-run it after the pilot with measured figures instead of assumed ones.

What is conversational product discovery?

Conversational product discovery is a storefront or messaging assistant that narrows a shopper from vague intent to a shortlist through dialogue — asking clarifying questions, applying constraints, and answering product questions from catalog data. It replaces the query-refine-scroll loop with a guided exchange, which is why assistants measurably compress product discovery time.

The behavioural shift is not speculative: shoppers are moving query volume from search boxes to assistants, and analysts priced the move in before it happened.

Production-grade means grounded. The assistant retrieves from the enriched catalog and answers only from what it retrieved, with confidence thresholds, a hand-off path to human support, and full conversation logging. Architecturally it is retrieval-augmented generation over the same vector index the search stack already built — a conversational and generative AI system in the standard pattern, not a raw model behind a text box.

Discovery also extends past the storefront. The same grounded-generation pattern that answers an on-site question can open a conversation: outreach systems draft recommendation and re-engagement messages from CRM and catalog data, gate them behind review queues, and measure replies per segment. That is how a discovery layer becomes a revenue channel on both the inbound and outbound sides of the funnel.

Where LLM product discovery fails

LLM discovery fails in five predictable places: thin catalog data, exact-match regressions, relevance judged by opinion, merchandising rules lost in the rerank, and latency and cost per query. None of them show up in a vendor demo on a clean sample catalog; all five show up in week one of production traffic, and all five are cheaper to find in a scoped pilot than after a replatform.

• Thin product data — Embeddings encode whatever text exists. Given a three-word title, the model clusters products by brand rather than by purpose, and semantic search returns confident nonsense. Enrichment comes before indexing — never after, when the index has already trained the merchandising team to distrust it.
• Exact-match regressions — Pure vector search blurs near-identical SKUs, so a shopper typing a model number gets the product family instead of the item. This is the single most common reason a semantic launch is rolled back, and keeping the lexical path is the whole fix.
• Relevance judged by opinion — Without a rated judgement set — 300 to 500 query-product pairs scored by people who know the catalog — tuning becomes an argument between merchandisers and engineers. Build the set before touching the reranker, and re-rate it quarterly as the catalog turns over.
• Merchandising lost in the rerank — Margin, stock, campaign, and compliance rules have to survive the merge. A reranker optimised on relevance alone will cheerfully promote an out-of-stock item at the top of the page. Apply business rules after the model, as explicit constraints, not as prompt suggestions.
• Latency and cost per query — An embedding call plus a cross-encoder on every keystroke adds both milliseconds and cost. Embed the catalog once and cache it, embed the query once per submitted search rather than per character, and rerank only the top hundred candidates — a discovery layer that costs more than the margin it recovers is not a discovery layer.

There is a sixth failure mode that is organisational rather than technical: nobody owns relevance after launch. Catalogs turn over, seasons change the meaning of "warm", and new suppliers arrive with new schemas. Give one named person the weekly review of top queries, zero-result queries, and search conversion, and the system keeps earning; leave it unowned and it decays into the engine it replaced.