Nearly every vendor in commerce and search now has an agentic story. Many of them are the same product as eighteen months ago, with a new slide added to the deck. That’s not necessarily a knock on the underlying technology — it’s a reflection of how quickly “agentic” has become the label everyone reaches for, whether or not the architecture underneath has actually changed.
We spent a recent masterclass, Designing Agentic Commerce: The Five Capabilities That Matter, unpacking exactly where that line sits. This post picks up two of the questions from that conversation that matter most for a large enterprise deciding what to trust.
Search Was Already Changing Shape
It helps to start with what’s actually driving this, because it’s a real shift, not just a trend vendors invented to sell something new. Shoppers no longer arrive expecting to translate what they need into the right two or three keywords. Increasingly, they describe a problem, a vague idea, or a specific comparison in their own words, and expect guidance back, not just a ranked list of matches.
That expectation didn’t form in a vacuum. Many shoppers now arrive at a site having already asked ChatGPT, Perplexity, or an AI Overview in Google Search the same question they used to type into a search box. The risk for a brand isn’t that these shoppers can’t find its products — search has always been reasonably good at that. The risk is that they get a better answer somewhere else first, and finish the purchase wherever that answer was.
That’s pushing search itself to change from a single-purpose lookup tool sitting in the header of a website to the entry point for a much wider range of intent: a precise SKU search, an open-ended question, a multi-step comparison between products, a request to narrow results down by something that isn’t even a catalog attribute — all arriving through the same box, and all needing an answer.
That’s the real substance behind “agentic commerce”: not a chat feature added to an existing site, but search evolving to carry a broader range of what shoppers actually want to ask.
The trouble is that this shift is easy to imitate on the surface and hard to build properly underneath. Which is why, for a large enterprise evaluating this space, the more useful question isn’t “is this agentic,” but something more concrete: if this system gives a shopper a wrong or off-brand answer, does the business have a way to catch it, explain it, and prevent it from happening again?
That question tends to separate genuine architecture from repositioning fairly quickly, and it comes down to two things: whether the system can actually be governed, and whether it’s a real evolution of search or a separate thing running alongside it.
Can You Actually Govern What It Says?
An agent that reasons over natural language and decides what to show a shopper is, by definition, exercising some autonomy — which is exactly the part that deserves scrutiny if it isn’t paired with real controls. Autonomy without governance just means unpredictability delivered more fluently, and for a brand with a compliance function or a regulated product line, that’s not a small risk.
There’s a useful way to think about what governance actually requires, borrowed from how enterprises already onboard people: when a new employee joins, they’re handed a handbook — pricing rules, what they can and can’t say, brand tone, escalation paths, what’s off-limits entirely. An agent needs the equivalent: not a system prompt with some friendly instructions bolted on, but an actual layer of business logic it operates inside, the way a well-trained employee operates inside company policy rather than improvising it fresh with every customer.
In a search and discovery context specifically, that layer looks like a handful of concrete things:
A layout system that separates what the AI is allowed to vary from what stays fixed — a comparison response might differ in which products it pulls in, but the structure of that comparison, the attributes it’s allowed to compare, and the ranking logic behind the results are set by the business, not improvised by the model in the moment.
Business rules that take precedence over the AI outright, so a campaign, a brand, or a regulated category can be locked to specific behavior no matter what the model would otherwise choose — merchandising rules win, not the AI’s best guess.
Content guardrails that block certain topics, claims, or language as hard constraints, including making sure the agent never exposes its own internal instructions to a shopper, which matters more than it sounds given how often that exact failure has made headlines. And just as important, clear rules for what the agent should decline to answer at all, which in a regulated industry can matter more than what it’s built to handle well.
If a vendor can’t point to where that governance layer actually lives in the architecture — not conceptually, but enforced — it’s worth asking what happens the first time the system gets something wrong in front of a real customer, not a demo audience.
Is This a Real Evolution of Search, or a Separate System Bolted On?
The most common failure mode here isn’t a bad model. It’s fragmentation — a conversational agent built as its own standalone surface, sitting beside the existing search and discovery experience rather than genuinely extending it.
It’s an understandable path, because it’s the fastest one to ship: stand up a chatbot, connect it to a language model, give it some retrieval access, launch it as a separate experience. The tradeoff is that a separate experience also means separate relevance logic, separate ranking signals, a separate permissions model, and no visibility into the search tuning, merchandising rules, and promotional logic the business has already spent years refining. A shopper who searches, gets a few results, then opens a chatbot to ask a follow-up question about those same products effectively starts over, because the two surfaces were never built to share what the shopper already established.
The more defensible architecture treats search and conversation as two outcomes of the same system rather than two systems that happen to share a homepage. A precise query — a known product, a model number — still resolves quickly, the way search always has, with the speed shoppers expect. An open-ended or comparative question resolves conversationally instead, but through the same underlying engine, carrying forward the same catalog, the same content, the same permissions, and the same business rules that applied a moment earlier.
The industry is converging on a protocol for exactly this — often called AG-UI, for Agent-to-UI — where the agent streams structured instructions to the frontend (render a product grid here, a comparison table there, a set of next-step actions after that) instead of generating a block of text, or worse, generating interface code live at runtime. A frontend library renders those instructions as real interface components, which is what allows the experience to shift from search results to a guided conversation and back without the shopper ever having to choose a mode or start over.
That distinction shows up in fairly ordinary moments that reveal a lot. A shopper narrows a product down by an attribute that isn’t formally in the catalog, asks for a comparison between two or three specific options, then asks what other customers thought of one of them — three different tasks, but one session, one catalog, one content index, and one set of guardrails carried through all three. If a vendor’s version of this requires the shopper to leave search and enter a separate chatbot to ask that same sequence of questions, that’s the fragmentation a real architecture is meant to eliminate — and worth asking about directly in any vendor conversation.
What This Actually Comes Down To
None of this is an argument against adopting AI in search and discovery, or a case for moving slowly — that shift is real, already underway, and not something a large enterprise can afford to sit out. It’s a case for applying the same scrutiny to “agentic” that these organizations already apply to any other system placed in front of customers: can it be governed the way a real employee would be, and is it a genuine evolution of search rather than a separate experience running next to it.
That’s a higher bar than most vendor conversations are built to clear on their own, which is exactly why it’s worth asking about directly. The organizations that get this right in the next few years likely won’t be the ones that moved fastest — they’ll be the ones that asked the harder questions before they signed.
If you want to go deeper on what separates a governed, unified architecture from a chatbot with a new coat of paint, we recently walked through this in more detail in our masterclass, Designing Agentic Commerce: The Five Capabilities That Matter, including a live look at how this plays out on an actual storefront.
