When Agents Do the Shopping

I bought something online last month without thinking about how I found it. A recommendation appeared. It felt relevant. I clicked, I compared, I bought. Normal.

But the thing that found it for you might not have been a search engine. And it definitely wasn't a friend.

During Cyber Week 2025, one in five online orders was influenced by an AI agent. Software that compared options, evaluated claims, and narrowed down a shortlist before a human ever saw it. Sixty-seven billion dollars moved through that layer in a single week. Salesforce tracked the data across 1.5 billion shoppers. Global online sales hit $336.6 billion, up 7% year over year, and agents were involved in a fifth of it.

That number is not a forecast. It is a receipt.

The question isn't whether AI agents will reshape commerce. They already did. The question is whether your domain is legible to the thing that's now doing the shopping.

Your Domain Is Your Resume

When a hiring manager reads a resume, they don't browse it for vibes. They scan for credentials. They check consistency. They verify claims against references. If something doesn't add up, they move to the next candidate. No rejection email. They just stop reading.

AI shopping agents evaluate domains the same way.

They query structured data. They check whether the entity behind the domain is consistent across platforms. They evaluate content depth. They verify claims against external signals. They aren't "visiting your website." They're reading your resume.

And the volume of commerce flowing through this evaluation is not trivial. McKinsey projects that agentic commerce, transactions mediated by AI agents acting on behalf of humans, could redirect $3 trillion to $5 trillion in global retail spend by 2030. Nearly $1 trillion of that from the U.S. alone. The word "redirect" matters. This isn't new spending. It's existing commerce flowing through a new intermediary. Morgan Stanley's research puts the U.S. figure at $190 billion to $385 billion by 2030, translating to 10-20% of all online retail. They also found that 23% of Americans have already bought something via AI in the past month.

Projections vary widely, reflecting genuine uncertainty about the pace of adoption. But even the conservative figures represent a structural shift.

This isn't early-adopter behavior. It's a structural shift in how commerce moves.

The domain-as-resume is the shift from domains as human-facing destinations to domains as machine-readable credentials. When the primary evaluator is an agent, your domain's value is determined by its legibility, verifiability, and structural completeness. The same way a resume is judged by credentials, not aesthetics.

Here's the thing. Agents aren't a secondary layer of evaluation. They're becoming the primary one, the intermediary that decides whether the human ever sees you at all. When the machine is also the buyer's proxy, legibility isn't about visibility anymore. It's about transactions.

If the agent can't read your resume, you don't get the interview.

The Agent Doesn't Browse

Amazon Rufus does not open a browser and scroll through product pages. It queries product data, compares structured attributes, reads review sentiment, checks price history, evaluates inventory, and presents a recommendation. More than 250 million customers have used it. Monthly users are up 140% year over year, interactions up 210%. Customers who shop through Rufus are 60% more likely to complete a purchase. Amazon's CEO projected on the Q3 2025 earnings call that Rufus is on pace to generate $10 billion in additional annual sales.

Let that settle for a moment. A quarter of a billion people are already shopping through an agent on a single platform. And they're converting at 60% higher rates than people browsing on their own.

This is the behavior model: query, verify, compare, transact. No scrolling. No homepage. No "above the fold." The agent sees data, not design. And the behavior works. For the shopper and for the platform.

Rufus isn't alone. Perplexity Shopping, Google Shopping AI, and ChatGPT with retail integrations for Target, Instacart, and DoorDash all share the same fundamental pattern. None of them browse.

Now the infrastructure is catching up. Google announced the Universal Commerce Protocol at NRF 2026, the biggest retail trade event in the world. Sundar Pichai introduced it as an open standard for agentic commerce, designed to work across the entire shopping journey: discovery, buying, and post-purchase support. It was co-built with Shopify, Walmart, Target, Etsy, and Wayfair, and endorsed by over twenty additional retailers. UCP is compatible with existing protocols including Agent2Agent and Model Context Protocol (MCP).

That's not a whitepaper. It's a launched standard with institutional co-signers. The infrastructure that enables agent interaction, entity consistency, structured data, machine-readable identity, isn't a future requirement. It's being standardized now.

What the Purchasing Agent Evaluates

So the infrastructure exists. The protocols are live. But what, specifically, does a commerce agent look at when deciding whether to recommend your product?

The checklist is structural, not aesthetic. And it's worth being specific, because vague calls to "optimize for AI" aren't useful.

Schema markup (JSON-LD). Product, Offer, AggregateRating, Organization. These are the structured data formats that protocols like UCP are built around. They are how an agent parses what you sell, at what price, with what reputation.

Entity consistency. Does the name, description, and address match across Google Business Profile, Bing Places, social profiles, and the domain itself? If your schema says one thing and your business profile says another, the agent registers a discrepancy. It doesn't penalize you. It ignores you.

Content depth. Is there substantive, current content that answers real questions? Not placeholder text. Actual subject matter that an agent can use to build confidence in what the domain represents.

Review signals. Do external sources corroborate the domain's claims? Third-party validation is how agents verify assertions they can't independently confirm.

Freshness. When was this last updated? An agent evaluating a domain that hasn't changed in three years is evaluating a snapshot, not a living record.

Domain history. How long has the domain existed, and what is its signal history? Age is a proxy for credibility in the absence of other indicators.

But I want to be honest about the nuance here, because the Analyst in you will catch it anyway. The evidence on schema markup and AI visibility is genuinely mixed. BrightEdge research found pages with comprehensive schema markup are three times more likely to appear in AI-generated responses. Others suggest that large language models may not parse hidden schema markup the way traditional crawlers do. They process visible text, not metadata tags.

The honest assessment: schema is necessary infrastructure. It's the language that protocols like UCP speak. But the real signal is entity coherence, whether your domain tells the same story everywhere an agent looks. Schema, business profiles, social presence, content, reviews. If your schema says one thing and your Google Business Profile says another, the agent flags the discrepancy. It doesn't ask for clarification. It moves on.

What this amounts to is a trust architecture. Agents don't trust pretty websites. They trust verifiable entities, domains where the claims stack up across multiple independent checkpoints. The shift from "does this look legitimate?" to "can this be verified?" is the foundational change that agentic commerce introduces.

Google's UCP formalizes this. Schema, product data, and entity verification aren't optional enhancements in the UCP framework. They're protocol requirements. The standard assumes agents will check these things. It was built to make that checking fast, reliable, and universal.

So you know what agents check. You know the protocols exist. The next question is harder. What happens to the domains that don't pass?

The Brands That Agents Can't See

Here's what absence actually looks like.

If your domain has no structured data, the agent cannot parse your products. If your entity profile is inconsistent, the agent cannot trust your claims. If your content is thin or stale, the agent has nothing to verify against. If your domain is parked, if it's sitting dark with a registrar placeholder page, the agent sees a blank resume. There's nothing to evaluate. There's nothing to compare. You're not in the decision set.

The Amazon-Perplexity dispute makes this concrete. In November 2025, Amazon issued a cease-and-desist to Perplexity, then filed suit in California federal court. The claim: Perplexity's Comet shopping assistant was secretly logging into user accounts and masking machine actions as human clicks. Amazon's legal argument was fundamentally about agent identity: who is this, and who are they representing? An agent that doesn't identify itself is, in Amazon's view, indistinguishable from fraud. Perplexity's defense: the agent has "the same permissions" as the human user it represents.

The legal argument is about agent identity. But the mirror-image problem is the one that affects most domains. When an agent can't identify your domain, when your domain has no structured identity, no verifiable entity profile, no machine-readable content, the agent doesn't sue you. It doesn't flag you. It skips you. You're not blocked. You're invisible.

And this invisibility is categorically different from the old kind. In the search-dominated web, being invisible meant you didn't appear on page one. You still existed. Someone could still type your URL directly. In the agent web, decision-set exclusion means something harder: you are excluded from the evaluation before any human is involved. The agent never considers you. The human never knows you were an option. You're not losing a position. You're losing the right to be evaluated.

If AI can't read a parked domain, agents certainly won't shop there. And the timing matters. A domain that is invisible to agents during the period when agent commerce is standardizing its protocols is missing the window when the rules are being written. Protocol conformance, entity profiles, and content records accumulate over time. The domains that start late start behind.

This is also true for domains with the opposite problem. Domains that are technically active but structured to sell rather than to serve. Agent legibility isn't just about being machine-readable. It's about being machine-trustworthy. A domain loaded with thin content and aggressive monetization may be technically parseable but structurally suspicious, which is exactly why monetizing too early poisons domains. Agents evaluate the cumulative record, not the current moment. The question isn't whether your domain ranks. It's whether it has agent legibility: can an AI agent read, verify, and act on what your domain represents?

Preparing for Agent Commerce

This isn't about adopting a new tool. It's not a checklist for "AI-proofing your website." It's about recognizing that the audience for your domain is changing, and the new audience reads differently than the old one.

The domains that will be legible in the agent era share structural characteristics. Not aesthetic ones.

Entity coherence. The same identity, consistently represented, across every platform an agent might check. This is the most important one, and the most commonly neglected. A domain might have excellent content and no schema at all but still be legible if its entity profile is consistent everywhere agents look. The inverse is also true: perfect schema with contradictory business profiles produces a confused record.

Structured data. JSON-LD that describes what the domain is, what it offers, and how to verify it. Not because schema is magic, but because protocols like UCP are written in schema. It's the language the infrastructure speaks.

Content depth. Substantive content that answers real questions. An agent building confidence in a domain needs something to evaluate. Thin pages, dated content, and placeholder text aren't neutral. They're evidence of absence.

Freshness signals. Evidence that the domain is active, maintained, and current. An agent evaluating an entity wants to know whether the record it's reading reflects the current state or a historical snapshot.

Protocol awareness. UCP, MCP, and Agent2Agent protocols are being standardized now. The structural requirements they impose, structured product data, entity verification, agent-readable endpoints, aren't speculative. They are being written into retail infrastructure by companies with the scale to make them standard.

None of this is entirely new. The structural argument has been building: domains need to be readable by both humans and machines, entity profiles accumulate over time whether you're building them deliberately or not, the evaluator for most decisions is no longer human, and the domains that accidentally became aggregators did so because they were already structurally legible. Agent commerce is the commercial consequence of those structural observations. When agents do the shopping, the domains that agents can read are the ones that get the business. The ones that can't be read don't get penalized. They simply don't exist in the evaluation layer.

The agent era is the beginning, not the end, of a decade in which domain identity determines domain value in ways we are only starting to understand. The protocols will evolve. The evaluation criteria will sharpen. But the foundational requirement won't change: to be considered, you must first be legible.

The SiteWarming perspective: When agents do the shopping, your domain is your resume. Warming builds the structured identity, entity coherence, and protocol readiness that agents actively check before they recommend you. Not design. Not branding. The machine legibility that determines whether you exist in the evaluation at all. What is SiteWarming?

The internet just hired a purchasing department. It doesn't browse your website. It reads your credentials, checks your references, and decides whether you're worth recommending. All before any human knows you exist.

Sometimes I think the domains that will do best in this shift aren't the ones that saw it coming. They're the ones that were already legible. Already doing the quiet, unglamorous work of being real. Consistent across platforms. Honest about what they offer. Present in the ways that matter, even when nobody seemed to be checking.

It turns out, something was always checking. We just didn't know its name yet.

And the protocols being written right now, UCP, Agent2Agent, the frameworks Visa and Mastercard are building, those aren't the end of the shift. They are the foundation for a decade of change in what domain identity means and who gets to participate.

Image Credits

Photo by Resource Database on Unsplash