An AI shopping agent is a software system — usually built on a large language model — that acts on behalf of a buyer to discover, evaluate, select and (increasingly) purchase products. Agents differ from classical search tools in three ways: they reason about intent, they compare across sources, and they can take action.
In 2026, AI agents in commerce exist on a spectrum. Most today are advisory (they recommend, the human buys). A small but growing minority are transactional (they buy within scoped permission). All of them change what merchants must expose.
A taxonomy of commerce agents
| Agent type | What it does | Examples (2025–2026) |
|---|---|---|
| Answer-surface shopping | Recommends products inside an AI answer (chat, search overview) | ChatGPT Shopping, Perplexity Shop, Google AI Overviews, Amazon Rufus |
| Browser / computer-use agent | Drives a real browser on behalf of a user | OpenAI Operator, Claude computer use, Comet, Rabbit |
| Embedded platform agent | In-app assistants inside marketplaces or merchant apps | Amazon Rufus, Shopify Sidekick, Klarna AI Assistant |
| Dev-built vertical agent | Custom agents for niche tasks (procurement, gifting, restocking) | Enterprise buyers, indie agent apps |
| Agent-pay autonomous | Transacts with delegated credentials | Stripe ACP demos, Visa IC pilots, Mastercard Agent Pay pilots |
The canonical agent purchase flow
Stripped to essentials, most agent-mediated purchases follow the same eight-step shape. Name the steps, and the merchant mandate becomes obvious.
- Intent formation. The user states a goal ("I need running shoes for a half-marathon under €150, waterproof preferred").
- Decomposition. The agent decomposes intent into filterable attributes, constraints and preferences.
- Retrieval. The agent queries sources — search APIs, answer-surface indexes, marketplace APIs, retrieval of pre-ingested catalog data.
- Comparison. Offers are compared on price, match score, policy fit, delivery time, merchant trust.
- Disambiguation. The agent surfaces 1–3 candidates to the user, or narrows autonomously if scope permits.
- Commitment. The user selects, or the agent proceeds on delegated permission.
- Transaction. Payment is initiated — either by handing the user to a checkout, or via an agent-pay token.
- Post-sale. Tracking, delivery, returns — ideally observable by the agent for the user's future reference.
A more granular walk-through is in Anatomy of an agent purchase.
What agents actually consume from merchants
Across the taxonomy above, agents read three broad categories of merchant data:
- Catalog data. Product titles, identifiers, attributes, categories, imagery, price, availability. See product catalogs for AI.
- Policy data. Shipping regions and windows, returns, warranty, age/geographic restrictions, subscription terms.
- Trust data. Merchant identity, ratings, provenance, certifications, age of domain, payment acceptance.
The best agents today blend multiple sources: structured feeds where available (Google Merchant Center-style), structured pages (schema.org), open APIs, and, as a fallback, HTML scraping with vision models. Each fallback is more expensive, slower, and less reliable — a ranking penalty for the merchant in practice.
Where agents get stuck
The failure modes are worth enumerating because each is a merchant opportunity.
- Identifier ambiguity. Two listings for the same SKU with different titles. Agent cannot deduplicate.
- Price flicker. Price on the PDP differs from price at checkout. Agent loses trust in the merchant.
- Consent walls. GDPR consent blocking rendering. Agent treats the page as empty.
- Dynamic availability. In-stock on the PDP, out-of-stock at checkout. Agents penalize merchants whose feeds lag behind.
- Policy gaps. Returns page is prose only. Agent cannot verify "free returns within 30 days".
- Region mismatch. Page served in English but offer only valid in a specific country. Agent misinterprets.
Agent-pay rails
Agent-pay is the transactional plane of the protocol. Three overlapping initiatives as of 2026:
- Stripe Agentic Commerce — agent-scoped tokens, spending controls, audit trails.
- Visa Intelligent Commerce — agent-bound card credentials with delegated authority.
- Mastercard Agent Pay — agent identity + tokenized credentials, merchant-side verification.
- PayPal agent tooling — transactional endpoints for agent apps.
These rails converge on a shared need: allow an agent to present a payment credential that is scoped, revocable, auditable and identifiable as an agent. Merchants will increasingly see agent-paid orders as a distinct class. We track the comparison in standards, schemas and protocols.
Consequences for merchants
Operationally:
- Assume your PDP will be read by at least one agent per week within 12 months; plan structure accordingly.
- Assume orders will start arriving with agent-pay tokens; verify your PSP supports the scheme.
- Assume your returns policy will be parsed; represent it as data.
- Assume your competitors with stronger feeds will appear first in answer surfaces; invest in feed hygiene.
- Assume the answer surface is a new acquisition channel, not a threat; design for the handoff.
Frequently asked questions
Are AI agents buying from small merchants today?
Yes, selectively. Answer-surface shopping (Perplexity, ChatGPT) routes to direct merchants whose catalogs are retrievable. Autonomous agent-pay is in pilot with enterprise merchants as of early 2026.
Do I need to build a chatbot to be ready?
No. The infrastructure work — catalog, policies, feeds, identifiers — is separate from any conversational UI. Many merchants never need to build an in-house agent.
Can I block AI agents?
You can signal (robots.txt, user-agent strings, fingerprinting) but the practical cost of being unreachable is rising. The better default is selective exposure with clear commercial terms.
Where to go next
- Concrete scenarios in use cases.
- The standards and protocols powering the above.
- Operator-grade preparation in the readiness checklist.