AI Search and Discovery in Ecommerce: What's Actually Working

A client of ours replaced their legacy Solr-based site search with an AI-powered semantic search engine in January. Within six weeks, their search-to-add-to-cart rate went from 3.2% to 5.8%. That’s not a typo — nearly doubled. And it wasn’t because the AI was magically understanding intent. It was because the old search was genuinely that bad.

This is the state of AI search in ecommerce right now. The gains are real, but they’re not coming from some fundamental breakthrough in artificial intelligence. They’re coming from catching up on a decade of neglected search infrastructure. Most B2B ecommerce sites I look at have search that hasn’t been meaningfully improved since 2018. Anything would be an upgrade.

So let me walk through what’s actually working, what’s still overhyped, and where I’d focus if you’re thinking about this for your own store.

What’s Actually Shipping

The biggest win right now is semantic understanding — the ability to match a query like “brass fitting for outdoor water line” to a product whose title is “3/4-inch NPT compression coupling, brass.” Old search would choke on that. It couldn’t bridge the gap between how a customer describes what they need and how a catalog names it.

This works because of embedding models. They convert both the query and the product data into high-dimensional vectors and measure the distance between them. The tech is mature — OpenAI’s text-embedding-3-small, Cohere’s embed v3, and Google’s multimodal embeddings all perform well. The hard part isn’t the model. It’s the product data plumbing that feeds it.

Algolia’s 2025 State of Search report found that retailers using AI-powered search saw a 1.7x improvement in search revenue compared to those on keyword-only systems. That number tracks with what I’m seeing in the field. But it’s worth noting that the same report found only 23% of surveyed retailers had actually deployed semantic search. Most are still testing.

Vector search also handles synonyms, misspellings, and regional language differences without you having to maintain massive synonym dictionaries. If your merchandising team has spent years building and tuning synonym lists, this is the part that feels like relief.

Where It Falls Short

Personalization in search results is still mostly aspiration. Everyone wants to show “relevant results based on browsing history,” but the signal-to-noise ratio is rough for most B2B catalogs. A distributor selling industrial fasteners doesn’t have the same behavioral data density as a consumer fashion brand. You might get 200 SKUs viewed per session on a consumer site. On a B2B parts catalog, you might get 3. There’s only so much personalization you can build on three data points.

Visual search — uploading a photo to find a matching product — gets a lot of attention at conferences. I’ve yet to see it perform well in a B2B context. Google Lens works great for finding a pair of shoes you saw on the street. It doesn’t work well for identifying a specific hydraulic valve from a grainy photo taken under a truck. The models aren’t the bottleneck here. The training data is. Nobody has a million labeled images of industrial components.

Faceted filtering combined with AI search is where I see the most practical value. Let the semantic model handle the fuzzy, conversational queries. Let traditional facets handle the structured narrowing — price range, material, thread size, voltage. They’re complementary, not competing.

The Data Problem Nobody Talks About

Here’s the part that catches people off guard: AI search is only as good as your product data. I know, I know — you’ve heard that before. But I mean it literally. If your product descriptions are two-word fragments copied from an ERP system, the embedding model has almost nothing to work with.

One distributor we worked with had 60,000 SKUs with an average product description length of 11 words. Eleven. The AI search they’d paid six figures for performed worse than their old keyword system because there simply wasn’t enough semantic content in the catalog to embed meaningfully.

The fix wasn’t a better model. It was enriching 8,000 top-selling products with actual descriptions — attributes, use cases, compatible applications. That took a content team six weeks. After that, search performance improved 40%.

Where Humans Still Matter

Search merchandising — deciding what shows up first for high-value queries — still needs a human in the loop. AI can rank by relevance. It can’t rank by margin strategy. It doesn’t know that you’re trying to push private label products in Q2 or that a particular manufacturer has supply issues next month.

The best setups I’ve seen use AI for the baseline ranking and give merchandisers an override layer on top. The AI handles 80% of queries well. The merchandiser steps in for the 20% that drive 80% of revenue. That’s the pattern that actually works.

AI-generated search synonyms and related product suggestions are good starting points, but they need review. I’ve seen models suggest a safety glove as “related” to a table saw. Technically correlated in purchase data. Practically a terrible recommendation that makes you look careless.

Where I’d Start

If your site search hasn’t been touched in two or more years, start there. You’ll get outsized returns just from moving to modern infrastructure. Don’t try to boil the ocean — pick your top 5,000 products by revenue, enrich their data, and deploy semantic search for those first. Measure search-to-cart and search-to-revenue before and after. The numbers will tell you whether to expand.

What’s your site search situation right now? Are you running something you touched this year, or something your team has been meaning to get to? I’m curious where most people fall on this.