AI agents are now doing the shopping
Online retail has been built around human shoppers who browse, compare, and complete purchases on websites. Increasingly, those shoppers are sending AI agents to do the work for them.
HUMAN's 2026 benchmark found that AI-driven traffic to retail sites nearly tripled across the year, with agentic AI traffic specifically growing 7,851%. Shopify says AI-driven orders rose fifteenfold in 2025, and its 2025 Global Holiday Report found that 64% of shoppers expected to use AI in their purchasing decisions, rising to 84% among shoppers aged 18 to 24.
The question for sellers is whether a machine can read their storefront when an agent is deciding what to recommend or buy.
What agentic commerce means
In agentic commerce, AI agents shop on behalf of customers. They compare product options across stores, check inventory and shipping policies, and complete purchases when authorized.
This changes what sellers should optimize for. Homepage design, category filters, and checkout flows tuned for human attention spans matter have no effect on the agent channel. Agents read structured data, policy text, inventory feeds, and third-party signals. When that information is missing, vague, or contradictory, the agent moves to another seller.
How agents connect to stores
Several standards define how AI agents discover products and complete transactions on behalf of buyers. The newest is Chrome's WebMCP early preview, which lets a website expose structured tools to a browser-based agent. The agent calls those tools directly (search the catalog, add to cart, check stock, retrieve order status), using the session and credentials of the user already signed in. For sellers, this means less scraping, more reliable transactions, and a seller-controlled interface that defines what an agent can do on a site.
Chrome describes WebMCP as a complement to the broader Model Context Protocol (MCP), not a replacement. WebMCP is early and other standards are emerging in parallel.
Four overlapping standards matter today. Plan to remain legible to all of them rather than betting on one.
| Protocol | What it does | Where it runs |
|---|---|---|
| MCP (Model Context Protocol) | Open standard for connecting AI models to external tools and data. A service exposes tools, an AI client consumes them. | Server-side |
| WebMCP | Chrome's browser-side companion. The model uses tools the website has registered with the browser, on a session the user is already signed into. | Browser-side |
| UCP (Universal Commerce Protocol) | Google's framework for normalizing commerce interactions across merchants. Gives agents a common vocabulary for products, carts, payments, and order status. | Cross-merchant |
| ACP (Agentic Commerce Protocol) | OpenAI and Stripe collaboration focused on agent-driven checkout. | Checkout layer |
Structured data decides whether agents can find your products
If a critical product attribute (material, dimensions, intended use, compatibility, return window) is buried in marketing copy or rendered only via JavaScript, an agent may not read it.
What sellers need in place:
- Schema.org product markup with attributes that match the vocabulary major AI surfaces use
- Machine-readable policies for shipping, returns, warranty, and compatibility
- Inventory and price feeds that update at the cadence agents expect
- Documented SKU relationships for variants, bundles, accessories, and replacement parts so an agent can traverse them
- Review and reputation signals in formats the major AI surfaces ingest
Agents build the shortlist before any human sees it. Clean, complete, accurate data is what gets a product on it.
AI-facing brand presence
The homepage is no longer the only entry point. Buyers and their agents are starting from ChatGPT product recommendations, Microsoft Copilot shopping suggestions, Google AI Mode results, in-browser agent flows, and partner apps that route through Shopping Graph.
Each online store needs:
- Verified merchant status
- A current catalog feed
- Reviews and reputation signals weighted by that surface
- Machine-readable policies an agent can use to answer buyer questions without redirecting to the seller's website
What changes for retail, B2B, and SaaS
B2C retail - Discovery and impulse purchasing are moving into AI chat tools. Merchandising loses leverage when an agent compares five sellers' versions of the same product against a buyer's stated criteria. Brand premium has to be justified by attributes the agent can verify.
B2B - Gartner found that 61% of B2B buyers prefer a rep-free buying experience, and that preference accelerates with agents in the picture. Procurement teams already use agents to gather quotes, check lead times, and assemble RFP responses. Sellers who require a sales conversation before disclosing pricing or specs are being filtered out of agent-driven shortlists.
SaaS - Evaluation is increasingly handled by agents that read documentation, pricing pages, and security information. Vendors that publish complete, structured information about pricing tiers, integrations, and compliance perform well. Vendors that gate that information behind sales calls get filtered out before a human sees them.
Five risks to plan for
Fraud detection breaks. Rules engines that flag unusual session behavior need to learn what legitimate agent traffic looks like. Bad actors will use agents to hide their activity inside that traffic.
Margins shrink for products that look the same to an agent. When several sellers offer what an agent sees as the same product, the cheapest one usually wins. That pushes prices down across the category. To keep pricing power, sellers need real differences a machine can confirm: proprietary materials, exclusive distribution, certifications, or better return policies.
Platform terms are still being written. Shopify notes that some platforms share only the data required to complete the order. Sellers may not see the customer's email, browsing history, or attribution path.
Direct customer data goes missing. Sellers who rely on first-party data for CRM and lifecycle marketing should plan for a significant portion of orders arriving without the customer information they're used to having. In agent-mediated commerce, the customer relationship often sits with the agent, not the seller.
A few platforms will dominate. If three or four AI platforms account for most agent-initiated discovery, the same power dynamics that shaped the search and social eras will repeat with new gatekeepers.
A 6-month action plan
The plan below is split into four phases of about three months each. The phases overlap in practice, so treat the timing as a guide rather than a strict sequence.
Weeks 1 to 6: Data foundation
- Audit structured product data against Schema.org markup and the attributes major AI shopping platforms use
- Convert shipping, returns, warranty, and compatibility information into machine-readable form
- Document SKU relationships so an agent can compare variants, bundles, accessories, and replacement parts
Weeks 6 to 12: Protocol readiness
- Evaluate WebMCP's early preview with a small set of test flows
- Submit or refresh feeds for Shopping Graph and equivalent platforms
- Run test queries against ChatGPT, Copilot, and Gemini shopping experiences to see how products appear today
Weeks 12 to 18: Trust and verification
- Pursue verified merchant status with the AI platforms that offer it
- Monitor mentions of your brand in AI tools, not just reviews on your own site
- Update fraud detection to handle agent-mediated traffic
- Decide how to handle customer service tickets that come from agents, including who counts as the customer of record when an agent is the buyer
Weeks 18 to 26: Measurement and optimization
- Track how your products appear in ChatGPT, Copilot, and Gemini by running test queries on a regular schedule and logging when, how, and alongside which competitors your products get recommended
- Update your marketing attribution model to account for AI-driven paths to purchase, even if the data is imperfect
- Test variants of structured data and measure how they affect product visibility on AI platforms
- Review protocol and policy changes every quarter. These standards change too often to review only once a year
Specific descriptions, machine-readable policies, current feeds, and accurate inventory are what protect a seller's visibility as more traffic becomes agent-mediated.
What this means for sellers
Agentic commerce are changing how purchasing decisions get made. The path from intent to purchase is shortening, with agents mediating or completing more of the steps. A seller's influence on the human-facing side is shrinking, and influence on the machine-facing side depends on the quality of structured data, the maturity of protocol integrations, and visibility on AI surfaces.
This is data-quality work, policy work, and integration work. Sellers who do it well will stay visible as buying continues to migrate into agent-mediated channels. Sellers who do not will see their products drop out of consideration before any human reviews them.
Talk to Graybox about agentic commerce readiness
Graybox helps mid-market and enterprise sellers prepare their storefronts for agent-mediated buying. Our practice areas align with the work this shift requires:
If you want an assessment of how your products appear to agents today and what to fix first, reach out for a consultation.