One typo, no results
Somebody types "recieve" or "safty" and gets nothing at all. On mobile, where most people are typing with a thumb on a moving bus, this is a large share of queries.
Fix: fuzzy matching with a sensible edit distance.
Sites that load fast and convert
Custom builds and integrations
Bounded, auditable, human-approved
Being found, staying fast, staying safe
Someone typing into a search box has already decided they want something and is telling you exactly what. On most sites they get an empty results page, or twelve items sorted by the date they were added, and they leave.
Better search is not about a clever model. It is about the boring failures: no tolerance for a typo, no idea that "torch" and "flashlight" are the same thing, no way to filter by the attribute that actually matters, and nobody at your company has ever looked at what people are searching for. That last one is usually the most valuable thing we hand over.
Drag left and right over the panel ↓Diagnosis
None of these are exotic. They are the default behaviour of the search that shipped with your platform, and every one of them is fixable.
Somebody types "recieve" or "safty" and gets nothing at all. On mobile, where most people are typing with a thumb on a moving bus, this is a large share of queries.
Fix: fuzzy matching with a sensible edit distance.
Your catalogue says "torch". Half your customers say "flashlight". The database has no idea these are related, so the search returns nothing and the customer assumes you do not stock it.
Fix: synonyms, plus meaning-based matching.
Results ordered by date added, or alphabetically, or by an internal ID. Relevance never enters into it, so the best match sits ninth and nobody scrolls.
Fix: a real relevance score, then business rules on top.
"Something to keep coffee hot on site" contains none of your product words. Keyword search cannot help, and the customer had told you precisely what they wanted.
Fix: semantic matching alongside keywords.
Four hundred results and no way to narrow by size, rating, availability or the one specification that decides the purchase. Faceting is table stakes and frequently absent.
Fix: facets built from real attribute data.
Your search box is the only place customers tell you, in their own words, what they want and cannot find. On most sites that data is collected by nobody and read by no one.
Fix: a weekly report of top queries and zero-result queries.
Interactive
A small trade-supply catalogue, searched three ways at once. Type anything, or use one of the sample queries — each is chosen to break one of the three.
The lesson is not that meaning-based search wins. It is that each approach fails in a different direction: keyword search is exact and brittle, semantic search is flexible and vague about specifics like part numbers and certifications. Production systems run both and combine the scores.
Matches the letters people typed. Fast, exact, and completely defeated by a synonym or a typo.
Matches what the query means. Handles description and intent, and gets vague about codes, sizes and certifications.
Both scores combined, then re-ranked. Slightly more to build and run, and it is what production systems actually do.
The stack
Most of this problem was solved by search engineering long before language models existed. The models add meaning-matching on top of infrastructure that already handles ranking, faceting and speed at scale.
Search engines
Vector & analytical stores
Embeddings & language
Ranking & machine learning
Commerce platforms
Content & front-end
Interactive
A customer is looking at a cordless impact driver. Here is what each recommendation strategy would put underneath it — and how each one goes wrong.
Most sites pick one strategy, wire it up once, and never look at it again. The useful version chooses per slot: complements on the product page, substitutes in search results, popularity on a cold home page. Which strategy sits where is a merchandising decision, and it should be measured rather than assumed.
currently viewing 18V cordless impact driverbasket emptyvisitor first visit
Applications
Ordered roughly by how quickly the difference shows up in the numbers.
Visitors who search convert at a markedly higher rate than those who browse, which makes the search box one of the highest-leverage surfaces on the site.
Usually the first thing we fix
Thousands of parts where customers know the application but not the part number. "Fits a 2019 model" is a query keyword search cannot answer and yours should.
Parts, industrial, wholesale
Publishers, associations and service businesses whose archive is large, valuable and effectively unsearchable through the CMS default.
Where the archive is the asset
Promoting stocked items over backorders, margin-aware ordering, seasonal pinning — applied as deliberate rules rather than by fighting the ranking.
Control without breaking search
Complements on the product page, alternatives in a dead-end search, popularity where nothing else is known — each chosen for the moment, not applied everywhere.
Basket size, not just clicks
The queries with no results are a product roadmap, a content plan and a merchandising brief, written by your customers, in their own words, for free.
Frequently the biggest surprise
Honesty
Search and recommendation projects are unusually easy to declare successful without evidence. These are the traps.
Recommendations get credit for sales that would have happened anyway. Without a holdout group the reported lift is mostly attribution, and it is always flattering.
Ranking by what already sells makes what already sells sell more. The long tail becomes invisible and the catalogue effectively shrinks.
New products and new visitors have no behavioural data. A recommender that only learns from history will ignore your new range indefinitely unless it is designed not to.
Missing attributes, inconsistent categories and copied-in manufacturer text limit results far more than the algorithm does. Search exposes data quality mercilessly.
Part numbers, CSA certifications, exact dimensions. Meaning-based matching blurs precisely the details a technical buyer is filtering on, which is why hybrid exists.
Tracking behaviour to personalise carries privacy obligations, and the return is often smaller than the effort for catalogues under a certain size.
Search results must feel instant. A semantic layer that adds noticeable delay will be used less, and that cost rarely appears in the business case.
Recommendations that reveal too much inference about a person damage trust more than they lift revenue. Restraint is a design decision.
A meaningful share of enquiries are solved by synonyms, typo tolerance and better facets on the existing engine — no AI involved, far less money.
Scope
The query log audit comes first, because it usually rewrites the priorities.
What people actually search for, what returns nothing, and where they abandon. If no logging exists, adding it is the first task and often the most valuable.
Attribute completeness, category consistency and description quality across the catalogue — the ceiling on what any search can do.
Hosted or self-managed, decided on catalogue size, query volume, budget and who maintains it. Sometimes the existing engine is fine and only needs configuring.
Keeping the index current as products, stock and prices change, without a nightly rebuild that leaves eight hours of stale results.
Field weights, synonyms, typo tolerance and hybrid blending, tuned against a set of real queries with known good answers.
Built from genuine attributes, showing counts, and never leading a customer to a combination with zero results.
Deciding what appears where, which strategy drives each slot, and what happens when there is not enough data to fill it.
Search conversion, zero-result rate, click position and recommendation lift against a holdout — set up before launch so there is a baseline.
A way for your team to pin, boost and bury without calling a developer, plus a record of who changed what.
Questions
Often just better search. Typo tolerance, synonyms, sensible field weighting and real facets fix the majority of complaints, cost far less, and add no latency.
Semantic matching earns its place when customers describe what they want rather than naming it — which is common in trade, technical and gift-buying catalogues, and rarer in others.
Search matters from a few hundred items upward. Recommendations need enough behavioural data to learn from, which usually means meaningful traffic rather than a large catalogue.
Below that, hand-curated related products genuinely outperform a model, and we will say so.
Yes. Both expose the catalogue and both support replacing or augmenting the default search. The work is the indexing pipeline and the front-end, not the platform.
For very large catalogues an external search service is usually the right answer regardless of platform.
By measuring against a baseline captured before launch, and for recommendations by holding back a percentage of visitors who see the old behaviour.
Without a holdout, any recommender looks successful, because it takes credit for purchases that were already going to happen.
It can be. Session-based recommendations using only the current visit need no profile and avoid most of the difficulty. Cross-visit personalisation involves storing behaviour, with the consent obligations that follow.
The return on the second is smaller than most vendors imply, so we tend to start with the first.
Handled by attribute similarity rather than behaviour — a new item resembles existing items, so it can be recommended from day one and start gathering its own data.
Designing for this explicitly is the difference between a new range selling and sitting invisible for a quarter.
Yes, and they should. Pinning, boosting, burying and seasonal rules belong in an interface your merchandisers use, kept separate from the relevance scoring underneath.
Mixing the two is how search quality quietly degrades over a year of well-intentioned tweaks.
Fixing an existing engine — synonyms, typos, facets, logging — is often a matter of weeks. A hybrid semantic layer with a proper indexing pipeline is a longer project.
We would generally do the first, measure, and then decide together whether the second is justified.
Call back
Leave a number and a good time. We will call you back to talk about what you are trying to build — no charge for the conversation.
Or call us directly +1 (647) 385-5532