AI search & product discovery — eCommerce

Only 1 in 10 searchers
found what they were looking for.

A bath & lighting brand operating in 55+ countries — one of the world's largest branded-showroom chains in the category — had a site search that converted just 10–15% of searchers. Precise queries returned nothing relevant. Broad queries buried shoppers in noise. We trained a custom ranking model on their own catalog and query logs, and shipped it in three months.

Talk to an engineer Read the full story

Queryhand faucets for kitchen sink, 1 meter long, with long flexible tube & wall hookreal customer query

Keyword search

10–15% converted

Kitchen Sink, Stainless Steelno match
Wall Hook Set, Chromeno match
1m Copper Pipe Fittingno match

AI ranking model

80% more accurate

Pull-Out Kitchen Faucet, Flexible Hoserank 1
Wall-Mount Faucet, 1m Extensionrank 2
Kitchen Tap w/ Wall Bracket Hookrank 3

The story

Keyword search couldn't survive the catalog's own complexity

The client is one of the world's fastest-growing bath and lighting brands, operating across 55+ countries and running one of the industry's largest chains of branded showrooms. Their site search — the standard keyword-matching kind — was quietly costing them sales at scale: only 10–15% of people who searched actually converted.

The failure mode ran in both directions. Precise queries like "hand faucets for kitchen sink, 1 meter long, with long flexible tube & wall hook" returned unrelated products because no keyword index could parse that much intent. Broad queries like "kitchen faucet" did the opposite — dumping too many undifferentiated options on the shopper and leaving them to do the filtering themselves.

We didn't reach for a bolt-on search plugin. The fix required understanding intent, not just matching text — so we assembled a team of nopCommerce developers (their site's platform, hosted on Oracle Cloud) alongside AI/ML consultants, and spent the first week mapping the catalog, the content, and how people actually searched before writing any model code.

The trade-off we made explicit up front: a rules-based synonym system would have shipped faster but hit a ceiling fast on a catalog this varied. We trained a model on the client's own historical search logs and product data instead — slower to start, but it's what let the system keep improving after launch instead of degrading as the catalog grew.

Client: Bath & lighting brand, 55+ countries, luxury to value segments
Core Stack: nopCommerce, Oracle Cloud, MS SQL Server, Python NLTK
Models: TF-IDF, Word2Vec, Bag-of-Words → transformer ranking model
Deployment time: 3 months from kickoff to live search engine
Ongoing: Model retrains continuously on live user feedback

How it was built

Five stages, from raw catalog
to ranked results

Assembling and cleaning the data
Collated the product catalog, articles, and manuals alongside real search logs — actual queries and what shoppers clicked on. Standardized formats and resolved discrepancies before any modeling began.
Annotating and analyzing the text
Tagged tokens with parts of speech, semantics, and roles so the model could understand relationships between words. Ran exploratory analysis on query patterns to catch data quality issues before they became model bugs.
Vectorizing the catalog and queries
Converted text into numeric vectors using TF-IDF, Word2Vec, and Bag-of-Words, then used cosine similarity to find the closest-matching products for a given query — the mechanism that makes "hand faucets... 1 meter long" resolvable at all.
Training the ranking model
Fine-tuned a transformer-based ranking model on the client's own content and query logs, iterating on architecture and hyperparameters against recall, precision, and relevance metrics until results held up in practice, not just on paper.
Shipping the interface and the feedback loop
Built the auto-complete and search UI with spelling correction and intent detection, then wired live user feedback back into the model so it keeps improving after launch instead of freezing at deployment.

improvement in search results accuracy: 80%
higher search-to-conversion rate: 72%
higher product listing click-through rate: 57%
of conversions from cross-sell & upsell recommendations: 30%

How we scope search projects

Why this wasn't a plugin

We Won't
sell a third-party search widget as an "AI upgrade" when the real fix requires a model trained on the client's own catalog and query behavior. Off-the-shelf tools plateau fast on catalogs this varied.
We Won't
claim personalization is working without a real feedback loop behind it. If user clicks and conversions aren't retraining the model, "personalized" is just a marketing word.
We Will
show the recall and precision numbers we evaluated the model against — not just the headline conversion stat. A ranking model with no visible evaluation metric hasn't been tested rigorously enough to trust.