Agentic Commerce 101: How AI Shopping Agents Decide What to Buy

What is agentic commerce?

Agentic commerce is what happens when an AI assistant doesn't just suggest a product — it actually buys it.

A shopper types "find me a hypoallergenic baby formula under $35, nickel-free, available for same-day delivery" into ChatGPT or Perplexity. The AI doesn't return a list of links. It reads product feeds, checks inventory, compares compliance data, and places the order. The human approves or the agent acts autonomously, depending on how much trust they've delegated.

That's the shift. Buying decisions are increasingly made by software, not by humans browsing pages.

The cost of being invisible to AI agents

When a shopper asks an AI assistant for a product recommendation, the agent pulls from machine-readable sources — structured data feeds, verified catalogs, protocol-compliant product listings. If your product data is incomplete, the agent doesn't downgrade you. It skips you entirely and moves to a competitor whose feed it can read.

Traditional SEO got you found by humans. Agentic commerce requires software to find you, read your data, and trust what it reads. That's a different bar.

A few things that get merchants skipped:

• Missing or invalid GTINs (Global Trade Item Numbers, the barcodes that uniquely identify products)
• Stale or inaccurate inventory signals (an agent won't recommend something it can't confirm is in stock)
• Thin product descriptions without structured attributes (material, certifications, age-grading, regulatory compliance)
• No machine-readable feed in a format the agent's protocol can parse

This isn't a future problem. ChatGPT has native shopping integrations. Perplexity runs a merchant program. Google's AI Overviews surface product data differently from organic search. These channels are live, and the merchants who show up are the ones with clean, structured, verifiable data.

How an AI agent actually buys a product

The process isn't magic. Agents follow a consistent pattern: discover, evaluate, then transact. Understanding each step tells you exactly where poor data costs you a sale.

1. Discovery: the agent finds your product

   The agent starts with a purchase intent — from the user, or from an automated workflow. It queries product catalogs via structured feeds, protocol APIs, or search indexes. Products that appear here have GTINs, categories, and feed URLs registered with the channel the agent is using (Google Merchant Center, a UCP-enabled store, an MCP server, or an ACP-compatible endpoint).

   If your product isn't in a feed the agent queries, it doesn't exist for that transaction.

2. Evaluation: the agent decides whether to trust the product

   This is where most merchants fail. The agent checks whether the product data is credible and complete: Is the inventory count real-time or stale? Do the safety claims have backing data (certification numbers, lab reports, regulatory flags)? Is the material composition structured, or buried in a paragraph? Are there compliance signals for the buyer's jurisdiction?

   Regulated products — supplements, baby items, medical devices, jewelry with metal content — face extra scrutiny. An agent helping a parent find a "nickel-free baby teether" needs a nickel content field, not a sentence that says "safe for babies." Unstructured claims don't parse.

3. Price and availability check

   Before committing, the agent checks current price (not the listed price from a stale feed) and real-time availability. Merchants with live inventory feeds and accurate pricing win here. Merchants with daily batch-update feeds sometimes lose orders because the agent sees "in stock" on a product that sold out three hours ago.

4. Transaction: the agent completes the purchase

   If the evaluation passes, the agent places the order through whatever protocol it's using — directly via ACP, through a UCP-enabled store (the Google–Shopify standard), or via a checkout API. The merchant's job at this step is to have a machine-accessible checkout path. Some protocols handle payment directly; others hand off to a human to confirm.

The protocol landscape: ACP, UCP, MCP, and beyond

Merchants currently face four main channels where AI agents transact. The architecture and data requirements differ across all of them.

What "regulated verticals" means for agents

AI agents evaluating products in regulated categories do more than check a price. For an agent processing a request like "find an FDA-registered zinc supplement with no proprietary blends, third-party tested, and in the $20-$40 range," your product data needs to answer each of those requirements with a structured field — not prose.

Arbling covers eight regulated verticals, and each one has its own trust requirements:

• Jewelry — metal composition, stone sourcing, hallmarks, conflict-free attestations, nickel content for EU compliance
• Supplements — FDA registration numbers, third-party testing certificates, ingredient lists in structured format, no unauthorized health claims
• Baby products — CPSC compliance, age-grading, material safety data, choking hazard flags
• Beauty — ingredient INCI names, EU/US banned substance flags, cruelty-free certifications
• Electronics — FCC certification, RoHS compliance, energy consumption data
• Medical devices — FDA 510(k) clearance numbers, intended use, contraindications in structured fields
• Luxury watches — movement type, case material, water resistance rating, authentication documentation
• Furniture — GREENGUARD certification, VOC emissions data, load ratings, country of manufacture

If your product is in one of these categories and your feed is missing these fields, an agent evaluating a safety-conscious buyer's request will pass. Not because your product is unsafe — because the agent can't confirm it isn't.

Is your catalog agent-ready? A quick self-check

Run through these before worrying about which protocol to integrate:

• Every product has a valid GTIN (UPC, EAN, or ISBN where applicable)
• Inventory counts update at least hourly (real-time is better)
• Price and sale price are always accurate in your feed
• Product type and category follow a recognized taxonomy (Google's taxonomy is a safe default)
• Regulated products have structured compliance fields, not paragraph-form claims
• Material composition is a structured attribute, not buried in a description
• Your feed validates against the spec for at least one channel (Google Merchant Center's diagnostic tool is free)
• You have no active policy violations on Google Merchant Center (these block AI surfaces too)

If you're hitting three or more gaps, your catalog is likely invisible or deprioritized across AI shopping channels already.

Sources

• OpenAI: Buy it in ChatGPT — Instant Checkout and the Agentic Commerce Protocol — OpenAI's ACP announcement
• Google: an open standard for agentic commerce (UCP) — Google and Shopify's Universal Commerce Protocol
• Model Context Protocol specification — Anthropic's open spec for MCP, including how tools and data resources are exposed to AI models
• Google Merchant Center product data specification — Official Google feed spec, including required and recommended fields
• Google AI Overviews and Shopping — How Google's AI surfaces use Merchant Center data
• Perplexity for Merchants — Perplexity's merchant shopping program overview

Ready to see where your catalog stands? See how Arbling helps you get found by AI.