Generic ecommerce search is built for a certain kind of catalog. The shopper types a name. Filters offer color, size, brand, and price. Ranking is mostly textual relevance. That is adequate when products are few, attributes are obvious, and the buyer is browsing.
It breaks as soon as the catalog is how a manufacturer, distributor, or dealer actually sells.
A knowledgeable buyer does not look for "bolt, silver, medium." They look for a diameter, thread, material, grade, finish, standard, and pack. An automotive buyer works in fitment: make, model, year, engine, sometimes trim. A plant buyer may have a customer part number, an OEM number, and a distributor number for the same item. Substitutions matter. Units matter. The item they are allowed to buy may be a subset of what the company stocks.
The search problem is therefore not, first, "how do we make the search box better?" It is whether the digital product model can express how those buyers identify and purchase the product.
If it cannot, no amount of UI polish will make the catalog findable. People will call someone who knows. Or they will leave.
Search Fails at the Model, Not the Box
A filter can only expose a field that exists and is trusted.
If specifications live in a PDF, a title string, or a sales engineer's memory, faceted search has nothing honest to offer. If thread and diameter are free text, two products that are interchangeable will not collapse into one result set. If UOM is ignored, a buyer searching in feet will miss an item stocked in meters. If customer part numbers are not mapped, the query that would have converted looks like a miss.
This is the same record problem described in product content as operating infrastructure. Search is simply the first place the weakness becomes visible. A page can hide a missing attribute in a paragraph. A filter cannot.
Taxonomy is the other half. Consumer sites often force every category into the same facet pattern. Complex catalogs need more than one retrieval shape. Independent attributes — material, grade, size — belong in facets that combine. Hierarchical relationships — vehicle, then platform, then year — belong in nested filters that remove invalid paths as the buyer chooses. Mixing those patterns, or offering every attribute on every category, produces empty results and false confidence.
Empty results are not a small UX defect on a technical catalog. They are a signal that the navigation path and the product data do not agree. Dynamic facets that only offer values still in the remaining set are not decoration. They are how you stop a valid inquiry from dying at zero.
How Knowledgeable Buyers Actually Retrieve
Watch the queries and the phone calls.
They search manufacturer part numbers, distributor numbers, and their own item codes, often in the same session. They search standards as if they were brands. They search a competitor's number hoping for a cross. They use synonyms the internal catalog never adopted: generator for alternator in one line of business, a trade name for a generic spec in another.
They also search incomplete. They know the application and three of five attributes. Retrieval has to tolerate that without dumping the entire category. Relationships have to be explicit: accessories, replacements, superseded SKUs, kits, next size. If those relationships are only "customers also bought," a technical buyer will not trust the result.
Business ranking sits on top of match, not instead of it. In a result set of valid items, the company may need to prefer what is in stock, what the account is contracted to buy, what the margin policy allows, or what is not a dead substitution. That is commercial policy encoded as retrieval, not merchandising theatrics.
None of this requires calling the work "AI search." It requires a product model, a taxonomy, synonym and standard maps, and rules for what may appear. Performance at catalog scale matters because a 50,000-SKU industrial assortment will punish naive filtering. Speed is a consequence of the model and the index, not a reason to skip the model.
Customer-specific assortment belongs in the same model. A dealer or contracted account that can see the full enterprise catalog will search successfully and still order the wrong thing — or will be blocked later by operations. Retrieval that ignores who is allowed to buy is only half a search. The filter has to know the account as well as the spec.
If Humans Cannot Disambiguate, Agents Will Not Either
What agentic commerce means for manufacturers and distributors is that machines will increasingly ask the catalog questions a knowledgeable buyer already asks. They will not rescue a record that cannot answer.
If attributes are missing, the agent inherits the ambiguity. If fitment is a paragraph, the agent cannot apply it. If substitutions are tribal knowledge, the agent will either omit them or invent them. Poor human search and poor agent discovery fail in the same place: the product cannot be identified from data.
The highest-intent path is now split between human search and agent discovery. That article is about the two interfaces. This one is the catalog those interfaces both read. Do not skip the model and hope ranking intelligence will guess the spec.
Arizon Digital's Agentic Commerce Readiness Index includes product data and discoverability for this reason. A readiness score does not create facets. Completeness of identity, attributes, relationships, and assortment does.
The Test Is a Knowledgeable Buyer, Not a Demo Query
A search program on a complex catalog has worked when a buyer who knows the item can reach it without calling, and when a buyer who knows the application can narrow without hitting zero.
Ask a few operational questions. Can the site filter the way the inside sales team actually qualifies a request? Can a customer part number resolve? Can two equivalent specs meet in one result? Does a valid nested path ever empty? When stock is zero, is the substitute a modeled relationship or a hope?
If the answers depend on a person interpreting the catalog, generic search did not fail because it lacked features. It failed because the digital product never learned how the business sells.
Inside sales already performs this retrieval every day, with better attributes than the site has. The work is to make that qualification logic — spec, fitment, number, unit, substitute, entitlement — explicit in data so the digital channel can do the same job without a phone call.
The leadership question is not which search plugin to buy. It is whether the product model is complete enough that retrieval — by a filter, a query, or eventually an agent — can identify the same item a seasoned employee would.