The agents arrived faster than most shop roadmaps: in October 2024 Anthropic released Computer Use as a public beta, in January 2025 OpenAI followed with Operator, which was folded into ChatGPT agent in July 2025, and in November 2025 Google shipped Gemini Agent with Gemini 3, initially for Ultra subscribers in the US. These agents navigate websites, fill in forms and prepare purchases on behalf of users.
At the same time, open standards emerged that structure agent-to-service communication without the agent having to parse the DOM: Anthropic's Model Context Protocol (November 2024), Coinbase's x402 (May 2025), Cisco's AGNTCY (at the Linux Foundation since July 2025), the Agentic Commerce Protocol from OpenAI and Stripe (September 2025) and Google's Universal Commerce Protocol (January 2026).
For e-commerce operators this is a double change. In the short term: a growing share of traffic comes from browser-automation agents that ignore images, block cookie banners and trigger anti-bot protection. In the medium term: the first brands are building agent-first endpoints that deliver structured product and order data directly — without a browser. Anyone failing to plan for both layers will lose ground systematically in the next generation of conversion paths.
What does an AI agent see on a product page?
On a product page, an AI agent mainly sees structured facts: product name, price, variant, stock status, delivery time and shipping cost. It captures hero images, videos, emotional copy and trust sections, but gives them little weight. A browser-based agent such as ChatGPT agent or Gemini Agent opens the URL, waits for the rendered DOM, reads the screenshot and the accessibility tree, and identifies price, variant selector, availability and the buy button. It has to get past cookie banners, pop-ups and anti-bot checks just like a human, and sometimes the session already fails there. The more cleanly the same facts are also available as Schema.org markup (Product, Offer, ShippingDetails, MerchantReturnPolicy) or through a product API, the less the outcome depends on layout. For shop operators this means every purchase-relevant detail has to be machine-readable, unambiguous and reachable without interaction. Otherwise the agent compares on incomplete data and, when in doubt, picks a competitor whose information is complete.
With MCP, ACP or UCP endpoints the same process becomes asynchronous and browserless: the agent authenticates at the domain's MCP server, structurally queries the product offering and receives JSON with the same fields — but canonical, without the layout lottery. An order runs via an ACP or UCP checkout flow or an x402 payment loop directly between agent and service API.
Browser agents launched by Anthropic, OpenAI and Google
open protocols: MCP · ACP · UCP · x402 · AGNTCY
over DOM — agents prioritize structured data
The protocols structuring agentic commerce
| Standard | Sponsor | Function | E-commerce relevance |
|---|---|---|---|
| MCP (Model Context Protocol) | Anthropic, open | Structured data exchange agent ↔ backend | Product catalogue, inventory API, search |
| ACP (Agentic Commerce Protocol) | OpenAI and Stripe, open source | Standardized checkout for agents | Order and payment, token-based authorization |
| UCP (Universal Commerce Protocol) | Google with Shopify, Etsy, Wayfair, Target, Walmart | Open standard across the shopping journey, compatible with A2A, AP2 and MCP | Checkout in AI Mode and Gemini (early access) |
| x402 | Coinbase, open | HTTP pay-per-action with status code 402 | Microtransactions, API pay-per-use |
| AGNTCY | Cisco (Outshift), Linux Foundation since 2025 | Agent-to-agent communication, discovery | Multi-agent marketplaces, procurement |
For most e-commerce providers the priority order is: product data and MCP first (complete feeds, agent-accessible product and inventory data), checkout protocols second (ACP or UCP, depending on the target platform), x402 third (for API services with microtransactions). AGNTCY matters for B2B procurement and multi-agent marketplaces, but is not yet a must for most shops.
OpenAI shows how fast the checkout layer is moving: it scaled back Instant Checkout, launched in September 2025, in March 2026. Merchants now complete purchases in their own checkouts, while ChatGPT focuses on product discovery and comparison based on merchant feeds (CNBC, March 24, 2026). Google is building a checkout directly into AI Mode and Gemini with UCP, currently through an early access program in Merchant Center. Clean product data and feeds are therefore the stable investment. Plan checkout integrations per platform and keep them reversible.
Is your stack agent-capable for 2026/27?
We check 16 signals: schema completeness, MCP endpoint, robots.txt strategy, cookie-banner behaviour, crawler rendering, anti-bot triggers, product API, agent checkout flow. Includes a 90-day roadmap.
Schema setup: the agent commerce stack
Six Schema.org types form the foundation. Every gap forces the agent to infer the missing information from the layout or to skip it:
1. Product with GTIN/SKU/MPN. GTIN is not optional — agents cross-check products across marketplaces, price comparison and review aggregators. Without GTIN a product is not entity-linked for the agent, so any comparison step with competitors systematically goes their way.
2. Offer with Price, PriceCurrency, Availability, ShippingDetails. ShippingDetails is critical in 2026: agents weight delivery time and shipping cost more heavily in comparisons than marketing copy. Hide shipping info inside the DOM and you drop out of the agent comparison set.
3. AggregateOffer and ProductGroup for variants. A product with five sizes, three colours and two material options belongs in a ProductGroup with clearly declared variant schemas. Otherwise agents make purchase decisions at a variant level they do not fully know.
4. MerchantReturnPolicy. Returns are a trust signal for agents — with concrete numbers (returnDays, returnMethod, returnFees). Generic trust badges without structured data are ignored.
5. Review & AggregateRating. Validated review aggregates. Agents systematically pull reviews into comparisons when they are structurally available. Without schema markup, external platforms (Trustpilot, Google) are preferred — where the competitor may be better positioned.
6. Service / Action for bookings and configuration. Not only products: bookable services need ReserveAction or OrderAction schemas so agents can reach the booking endpoint directly.
The cookie-banner trap and the agent user-agent
Classical GDPR cookie banners are a first-order conversion brake for agents. Three scenarios appear regularly: (a) the agent clicks "reject all" and sees a reduced version of the page without tracking-induced personalization; (b) the agent fails at the banner and aborts the session; (c) the agent accidentally clicks "accept all" and triggers a GDPR risk because the human never consented.
One possible approach: detect agent user-agents (such as ChatGPT-User or Claude-User) and serve a technically complete variant that runs without consent-requiring cookies and tracking scripts. Whether the consent banner may then be dropped is a legal question: under the ePrivacy rules, in Germany Section 25 TDDDG, only strictly necessary access to the device is exempt from consent. Clarify this with your data protection advisers before implementing it. Commercially, a low-friction agent path is a significant conversion lever.
robots.txt strategy: differentiate, do not blanket-block
The question "block GPTBot or allow?" is too coarse in 2026. Three classes with different strategies:
Training crawlers (GPTBot, ClaudeBot, CCBot and the Google-Extended control token): check your licensing position. If training data is free to take, allow. If you target licensing through platforms like ScalePost or TollBit, block and point to the licensing marketplaces.
Search and retrieval crawlers (OAI-SearchBot, ChatGPT-User, Claude-SearchBot, Claude-User, PerplexityBot): allow. They bring direct citation traffic and are indispensable for visibility in ChatGPT, Claude and Perplexity (see the crawler documentation from OpenAI and Anthropic). Google-Extended does not belong in this group: according to Google, the token only controls Gemini training and grounding and has no effect on Google Search. AI Overviews and AI Mode use the regular Googlebot.
Operational shopping agents (ChatGPT agent, Claude Computer Use, Gemini Agent): allow and prioritize. These agents mediate purchase intent — blocking them means direct revenue loss. Ideally expose a dedicated agent API that detects agent user-agents and offers a faster path.
Agent commerce KPIs — what to track
| KPI | Definition | Meaning |
|---|---|---|
| Agent traffic share | % sessions with agent user-agent | Volume of agent mediation |
| Agent conversion rate | Conversion rate of agent vs. human sessions | Quality of agent optimization |
| Schema coverage score | % products with full mandatory schema | Structural agent readiness |
| MCP hit rate | % requests using MCP instead of DOM | API-first maturity |
| Cookie banner drop-off | Sessions aborting at the banner | Cookie-induced revenue loss |
| Agent cart abandonment | Cart drops in agent sessions | Friction in the agent checkout |
The 90-day roadmap to agent readiness
Days 1-30: schema audit and completion. Complete product schema with GTIN/SKU/MPN, Offer with shipping/returns, AggregateRating, ProductGroup. Validation via the Schema.org validator and Google Rich Results Test. More at Schema Implementation.
Days 31-60: agent user-agent detection and cookie banner bypass. Server-side detection of known agent user-agents and delivery of a reduced variant without cookie banner or anti-bot friction. Legal review with the data protection team: only strictly necessary access is exempt from consent.
Days 61-90: MCP endpoint pilot. Initial MCP server implementation with read-only endpoints for product, inventory and search. Structured JSON schema for agent consumption. Optional: checkout pilot via ACP or UCP for selected categories.
Conclusion: agent readiness is not optional in 2026
Agent commerce is small today, and there are no reliable market figures for it yet. The investment in agent readiness still pays off, even without agent traffic: the same schema improvements strengthen LLM citation, AI Mode visibility and Knowledge Graph trust. Lean consent and anti-bot paths help human users too. An MCP API is also headless-commerce infrastructure. Build agent readiness in 2026 and you are not building for a single scenario — you are building for the conversion paths of the next five years.
Sources
- Anthropic: Introducing computer use, a new Claude 3.5 Sonnet, and Claude 3.5 Haiku (October 22, 2024)
- OpenAI: Introducing Operator (January 23, 2025)
- OpenAI: Introducing ChatGPT agent (July 2025)
- Google: Gemini 3 brings upgraded smarts and new capabilities to the Gemini app (November 18, 2025)
- Anthropic: Introducing the Model Context Protocol (November 25, 2024)
- x402 Foundation: x402, an open standard for internet-native payments
- Linux Foundation: Linux Foundation Welcomes the AGNTCY Project (July 29, 2025)
- OpenAI: Buy it in ChatGPT: Instant Checkout and the Agentic Commerce Protocol (September 29, 2025)
- Google: New tech and tools for retailers to succeed in an agentic shopping era (January 11, 2026)
- Google Merchant Center Help: About the Universal Commerce Protocol (UCP) and UCP-powered checkout
- CNBC: OpenAI revamps shopping experience in ChatGPT after Instant Checkout (March 24, 2026)
- OpenAI: Overview of OpenAI Crawlers
- Anthropic: Does Anthropic crawl data from the web, and how can site owners block the crawler?
- Google Search Central: Google common crawlers (Google-Extended)
- German Federal Ministry of Justice: Section 25 TDDDG