Why Keyword Search Is Costing You Sales

A customer types "blue cocktail dress for wedding" into your search bar. Your keyword search engine sees four separate tokens: blue, cocktail, dress, wedding. It returns every blue dress, every cocktail dress, every product with "wedding" somewhere in its metadata. Half the results are irrelevant. A quarter are out of stock in the customer's size. The customer bounces.

This is not a hypothetical. At a major UK fashion e-commerce retailer, we audited the search experience and found that keyword search was silently haemorrhaging revenue. The zero-result rate was over 14%. One in seven searches returned nothing at all. And the searches that did return results were often so poorly ranked that customers abandoned them anyway. The search bar — the highest-intent action a customer can take on an e-commerce site — was actively losing sales.

The Problem with Keyword Search

Keyword search is fundamentally a string-matching exercise. It tokenises the query, stems the words, and matches them against product titles, descriptions, and metadata fields. If the words match, the product appears. If they don't, it doesn't. There is no understanding of what the customer actually wants.

This creates three compounding problems. First, the vocabulary mismatch. Customers don't use the same language as merchandisers. A customer searches for "going out top" — the product is tagged as "evening blouse." A customer searches for "work trousers" — the product is categorised as "tailored pants." Every mismatch is a missed sale.

Second, keyword search cannot handle multi-faceted queries. "Comfortable black shoes for standing all day" expresses a clear intent that any human would understand. Keyword search treats it as a bag of words and returns black shoes sorted by whatever static ranking rules were configured months ago. The "comfortable" and "standing all day" signals are effectively discarded.

Third, keyword search has no memory. It doesn't know that this particular customer always buys size 12, prefers mid-range price points, and has never purchased a premium brand. Every search starts from zero, ignoring everything the business knows about the customer.

How Neural Search Actually Works

Neural search replaces string matching with semantic understanding. Instead of comparing words, it compares meanings. Here is how it works at a high level.

Every product in the catalogue is passed through an embedding model — a neural network that converts the product's text, attributes, and image data into a high-dimensional vector. Think of it as a coordinate in a space with hundreds of dimensions, where similar products sit close together regardless of how they're described in words.

When a customer submits a search query, that query is also converted into a vector in the same space. The search engine then finds the products whose vectors are closest to the query vector. "Blue cocktail dress for wedding" doesn't need to match any specific keywords. It needs to be semantically close to products that match that intent — and it is, because the embedding model has learned from millions of examples what "cocktail dress" and "wedding" mean in context.

The result is search that understands synonyms, intent, context, and even negation. It handles misspellings gracefully because the embedding of "blck dress" is nearly identical to "black dress." It handles natural language because "something to wear to a garden party" maps to a region of the vector space that contains appropriate products, even though no product description contains the phrase "garden party."

Curious what your margin opportunity looks like?

Implementation: What We Actually Built

We didn't rip out the existing search infrastructure overnight. We ran neural search in parallel with the existing keyword engine for four weeks, logging results from both systems side by side without showing the neural results to customers. This gave us a clean comparison dataset.

The embedding model was fine-tuned on the retailer's own catalogue and historical search-click-purchase data. Off-the-shelf embedding models work reasonably well, but fine-tuning on domain-specific data — fashion terminology, brand semantics, the retailer's specific product taxonomy — improved relevance scoring by roughly 23% on our internal evaluation set.

Product vectors were stored in a vector index that supports approximate nearest-neighbour search at scale. With 50,000+ active SKUs, the index returns results in under 40 milliseconds. We re-index nightly to capture new products, price changes, and stock movements. Fast-moving signals like stock availability are handled via a real-time filtering layer on top of the vector results.

The ranking layer blends semantic relevance with business signals — informed by the same data warehouse that powers our in-house CDP — including stock depth, margin contribution, sell-through velocity, and return rate. A product might be semantically perfect for a query but have a 40% return rate in that size — the ranking layer suppresses it. This is where neural search becomes a personalisation tool, not just a relevance tool. The same query produces different rankings for different customer segments based on their demonstrated preferences.

The Results

We ran the A/B test for six weeks. 50% of traffic got the old keyword search, 50% got neural search. The results were not subtle.

Zero-result rate dropped from 14.3% to 2.1%. That alone is transformative — 12% of searches that previously returned nothing now surface relevant products. Every one of those is a customer who would have bounced and now has a reason to stay.

Search-to-purchase conversion lifted by 31%. Customers who used the search bar were meaningfully more likely to buy something. Average order value from search sessions increased by 8%, driven by better product discovery surfacing items the customer wouldn't have found through category browsing.

But the number I care about most is the merchandising hours saved. The old keyword search required manual intervention: boost rules, bury rules, synonym dictionaries, redirect maps. A merchandiser spent roughly 15 hours per week maintaining search configuration. With neural search, that dropped to under 2 hours per week of monitoring and exception handling. The algorithm handles the long tail better than any human can across tens of thousands of queries.

Why This Matters for Margin

Better search isn't just a conversion play. It's a margin play. When customers find what they're looking for on the first search, they buy at full price. When they can't find it, they either leave or start bargain hunting in the sale section. Every failed search is a nudge toward the kind of discount-dependent behaviour that erodes margin across the business.

Neural search also enables margin-aware ranking. The old keyword system sorted by static merchandising rules. The new system factors margin contribution into the ranking algorithm. Between two equally relevant products, the higher-margin option ranks first. This isn't about hiding cheap products — it's about breaking ties intelligently when relevance scores are similar.

This was one component of a broader transformation programme that delivered £6.4M in tracked P&L impact. Search alone didn't drive all of that — but it was one of the highest-ROI initiatives we ran, because the implementation cost was modest relative to the conversion uplift it delivered.

If your search bar is still running on keywords, it is actively working against you. Your highest-intent customers — the ones who know what they want and are ready to buy — are being served irrelevant results and zero-result pages. That is not a technology problem. It is a margin problem.