The product was in stock. The shopper couldn't find it.

What they asked for

"Build conversational discovery agents grounded in catalog and inventory to guide purchases."

They didn't have a recommendation problem. They had a findability problem. Keyword search matched words, not intent, so a shopper who knew the need but not the product name hit a dead end and left.

Intent, not keywords. The assistant understands the need, not just the words a shopper types.

The situation

A US online retailer had a deep catalog and healthy traffic, yet too many shoppers arrived, searched and left without buying. Conversion and average basket size lagged the targets, and the working assumption was that the recommendation engine needed tuning.

When we looked at the search logs, a different story emerged: shoppers were typing what they wanted in plain language, the search was returning nothing or nonsense, and they were leaving, walking past products that were sitting in stock the entire time. It was not a recommendation problem. It was a findability problem.

What we found

Our diagnostic surfaced three compounding issues.

Search matched keywords, not intent. A query like "warm waterproof jacket for autumn hikes" returned nothing because those words did not appear in the product titles, so any shopper who could not name the exact item simply bounced.

Recommendations weren't grounded. Generic "you may also like" panels routinely surfaced out-of-stock or off-need products, so shoppers learned to distrust and ignore them.

The undecided got no help. The shoppers most likely to abandon (the ones weighing two or three options) had nobody to guide them, so browse-to-buy stayed low and baskets stayed small.

The scale of the problem is well documented. 94% of consumers have abandoned a shopping session because of irrelevant search results, and search abandonment costs US retailers on the order of $300B a year. The maddening part is that the product is usually right there in the catalog; the search simply does not understand the shopper. (Source: Google Cloud and The Harris Poll, 2021.)

This is exactly the gap a grounded, conversational agent closes, and it shows up in the field: consumer-retail brand SharkNinja runs agentic shopping guidance that helps customers find new products, compare their options and, in its own reporting, put more items per cart. So we treated this as a findability problem and set out to understand intent, not match strings.

What we did

Understand intent. A conversational GenAI discovery agent interprets the underlying need (the use, the constraints, the occasion) and maps it to catalog attributes, so a plain-language request returns the right products rather than an empty results page.

Ground every answer. The agent recommends only products that are actually in stock and genuinely fit the need, grounded in live catalog and inventory, so its suggestions earn trust rather than eroding it, and a shopper is never sent toward something they cannot buy.

Guide the whole need. It asks clarifying questions, compares options in plain language and, where it genuinely helps, completes the set (the jacket and the waterproofing spray), which lifts basket size naturally rather than through a pushy upsell.

How we rolled it out

We didn't boil the ocean. We launched the agent on the highest-traffic categories first, grounded it in that catalog, proved that conversion and basket rose while dead-end searches fell, then widened coverage, keeping merchandisers in the build, so the agent reflected how they actually range and position the products.

The shift

Indexed to before = 100, conversion rose to roughly 118 and basket size to roughly 115 once shoppers could find what they came for and were guided to the whole need.

Outcomes

• ~18% higher conversion for shoppers who use the discovery assistant.
• ~15% larger average basket: the assistant completes the need, not just the item.
• ~40% fewer dead-end searches: intent understood, so "no results" stops losing ready buyers.

The win wasn't a smarter recommendation engine. It was a shopper who found what they came for, and one more thing they needed.

An illustrative engagement. The scenario and figures are representative, drawn from outcomes across comparable agentic AI product-discovery deployments, not the audited results of a single named client.