The searches that returned nothing — and 44% better product tagging
Mining on-site and inbound search intent to find demand the catalogue could not answer, lifting product tagging accuracy 44% and sales 10%.
US retail chain
+44%product tagging accuracy
There is a report most retailers do not run, and it is one of the most valuable they could: which searches returned no products?
Every one of those is a customer who told you exactly what they wanted, in their own words, and received nothing. Either the product is missing from the catalogue, or it is there and tagged in language no customer uses.
Challenges
The retailer wanted to serve better recommendations and understand demand it was not meeting. Two gaps stood in the way.
Search terms were not being mined. The queries existed in logs, unused as a source of demand signal.
Catalogue language versus customer language. Products are tagged by merchandisers using merchandising vocabulary. Customers search using their own. The mismatch is invisible until you compare the two directly.
Data engineering
Two sources of search intent. Terms searched on the company's own website, and terms typed into external search engines that brought a user to the site. The second source is the one usually missed, and it captures intent formed before the customer arrived.
Segmented geographically. Terms were split by US state. Demand is not uniform across a country, and the same word can mean different products in different places.
Matched against the product catalogue. Search terms were matched to catalogue tags — the join that reveals both what is missing and what is mistagged.
Zero-result analysis. The specific focus: which keywords were searched most frequently while returning no product from the catalogue. That list is a demand signal and a merchandising to-do list in one.
Data science
Data science techniques were applied to the search terms to improve the recommendations served to customers — both while they were on the site and after they left. Executive reporting was built on the same foundation so the findings reached the people who decide what to stock.
Results
| Measure | Outcome |
|---|---|
| Product tagging | +44% improvement |
| Sales, from better recommendations | +10% |
Both figures are as recorded in the project's documentation.
The tagging number is the one worth dwelling on. A 44% improvement in tagging is not a data science achievement; it is the consequence of finally comparing what customers say to what the catalogue says. The modelling improved recommendations, but a large share of the value came from a join nobody had run.
What we would take from this
Zero-result searches are the highest-value report you are not running. Customers state their intent in their own words and get nothing. Every entry is either a stocking decision or a tagging fix.
Capture inbound search intent, not just on-site. What someone typed into a search engine before arriving tells you what they came for, which is often more specific than what they type once they are on your site.
Geography changes vocabulary. Segmenting by region surfaced differences that a national aggregate averages away entirely.
We do not name clients. Engagements are described by sector and scale because confidentiality obligations outlast the work, and consent we cannot produce is consent we do not have.