On-site search used to be the cleanest expression of intent on a commerce site. A visitor typed what they wanted. If the catalog could answer, conversion followed. If it could not, the visit ended. That path still matters. People who know what they are looking for still use the box, and they still punish a miss.
A second high-intent path is now forming next to it.
Buyers use AI assistants to research products, compare specifications, shortlist options, and decide where to go. Some of that work never produces a click. Some of it does: a referral to a product page, a retailer, a part number. In a smaller set of cases, an agent may eventually transact. That last step is real in pilots and protocols. It is not yet the default way most commerce happens, and it should not be written as if it were.
The operating fact is simpler. Commerce businesses increasingly serve two retrieval surfaces: human search on the site, and machine or agent discovery off it or through it. Both need the same underlying product truth. Optimizing only the search bar is defending one door.
Intent Has Another Door. It Is Not the Only Door.
Adobe Digital Insights, measuring more than a trillion visits to U.S. retail sites, reported that traffic from generative AI sources rose 4,700% year over year in July 2025. In a companion survey of 5,000 U.S. consumers, 38% said they had used generative AI for online shopping. The most common tasks were research and recommendations, not checkout. In that same July data, AI-referred visits were still 23% less likely to convert than other traffic — down from a much wider gap earlier in the year — and those visitors spent more time and viewed more pages once they arrived.
Read that carefully. AI is already in the research path. It is not, on that evidence, the dominant purchase path. Shoppers often arrive more informed. They do not automatically buy. Conversion still depends on whether the page they land on, and the catalog behind it, can complete the job the assistant started.
IBM's overview of agentic commerce makes the same structural point: a 2026 Institute for Business Value study found 45% of consumers already use AI for part of the buying journey, and agents evaluate products against structured data such as attributes, availability, and constraints. Treat those figures as evidence of a split path, not as a forecast that storefronts are finished.
SEO is not dead. On-site search is not obsolete. A second retrieval surface has been added. The companies that treat that as a channel report, rather than as a product-information problem, will optimize the wrong layer.
What Both Surfaces Require
A human query and an agent request fail in the same ways.
They need a trusted product identity, not three codes that do not map. They need structured attributes, not a paragraph that only a person can parse. They need price the business is actually willing to honor for that buyer. They need availability that will still be true at order. They need relationships and compatibility when the asked-for item is not the item that should ship. They need policies — returns, lead time, substitution rules — that a machine can read if it is going to recommend with confidence.
Product content is now operating infrastructure is that record. Generic search breaks when the catalog is operationally complex is how that record is filtered and retrieved for a knowledgeable human. This article is the conversion implication of serving both a typed query and an assistant from the same truth.
Do not confuse the two. Facets, fitment, and nested filters are catalog modeling. Rank, referral, and agent-readable offers are interfaces onto that model. If the model is weak, improving either interface just publishes the weakness in a new place.
For industrial sellers, what agentic commerce means for manufacturers and distributors is the sector argument: machines will favor complete, structured, authorized catalogs. The conversion article for everyone else is the same mechanism without the fastener story. If an assistant cannot tell whether you have the spec, the pack, and the stock, it will shortlist someone who can.
Conversion Leaks at the Handoff
The new failure mode is a successful research step followed by a failed commercial step.
An assistant names a product. The referral lands on a page with a different description, a list price the account will not pay, or an in-stock badge the warehouse cannot support. The human searcher who typed the same spec hits a keyword match that is the wrong variant. In both cases intent was high. The record could not close it.
On-site, that still looks like search conversion: did the query find a buyable SKU. Off-site, it looks like missing demand you never saw, or like traffic that bounces because the page cannot confirm what the assistant claimed.
A B2B version of the same leak is an assistant or a procurement tool that cannot see contract price, pack, or allocation, so the buyer is sent to a list-price page they already know is wrong. High intent, unusable answer.
The response is not to abandon merchandising, paid search, or the search box. It is to make the product, price, and availability that humans retrieve the same facts an agent would retrieve. Machine-accessible catalog, inventory, and policy data are becoming part of conversion architecture, even while most orders still complete with a person in the loop.
Arizon Digital's Agentic Commerce Readiness Index frames discoverability and agent interfaces as capabilities, not as a prediction that autonomous checkout is already the main channel. That is the right altitude. Get the record right for today's searcher. The same record is what a later agent will use.
Measure Both Doors Without Abandoning the First
Leaders should keep watching on-site search: zero-result rate, conversion after query, whether the returned SKU is the one operations can ship. They should add a second set of questions. Can an external system describe this product unambiguously? Are attributes, price, and availability available as data, not only as HTML? When AI-referred visits arrive, do they find a page that agrees with the research that sent them?
If on-site search still fails knowledgeable buyers, fix that. It is still the highest-intent behavior you fully control. If on-site search works and assistants still cannot represent the catalog, the missing work is structured product truth and access — not another widget in the search bar.
Do not wait for autonomous checkout to become common before doing that work. The research path is already sending people to pages. Those pages either confirm the assistant or contradict it.
The highest-intent path is split. The product record should not be.