How to Prepare Your Site for AI Agents in Agentic Search: Actions, Data, Checkout

Quick answer

To prepare your site for AI agents, Canlah AI recommends in this guide exposing three layers a model cannot infer from marketing copy: the facts it reads, the actions it can call and the checkout it can complete. Six agent-facing standards shape those layers in 2026, and three of them, UCP, ACP and AP2, apply only to sites that sell online. Most sites fail at layer two, because a page that a human can click is not an action an agent can call.

Canlah AI measured the gap on 50 cross-border DTC storefronts in August 2026 and published the raw data. Of the 47 reachable storefronts, 25 served a UCP manifest, all 25 of those manifests were issued by Shopify rather than the brand, 0 exposed an agent-payments endpoint, 1 exposed an MCP endpoint and 9 carried Product schema on the sampled product page. Agent readiness in that sample is almost entirely a platform default, not a brand decision.

Treat any single readiness check as directional, because one crawl cannot prove that an agent will choose your site or that a fix caused a transaction. Fix the blocking layer first, re-run the same check and only then decide what to build.

Disclosure: Canlah AI publishes this guide and runs the agent-readiness audit described below. Protocol details for Google, OpenAI, Stripe, Anthropic, Microsoft and Cloudflare come from each organisation’s own public documentation, reviewed on September 23, 2026, and were not tested through paid accounts or merchant approval. The storefront figures are Canlah AI’s own open dataset and have not been independently audited.

What does an AI agent need from your site?

An AI agent needs six things from your site before it can finish a task on your domain: access, facts, actions, checkout, identity and evidence. The agent arrives as a caller rather than a reader, fetches what it can reach in one or two requests and abandons any step that requires a human to interpret a layout.

• Access: whether robots.txt allows named agents and no firewall, rate limit or edge bot rule blocks them.
• Facts: whether price, availability, location, hours, policy and entity identity sit in served HTML or a structured file, not only in an image or script-rendered widget.
• Actions: whether search, filtering, availability lookup and booking are reachable as a callable endpoint rather than only as a click path.
• Checkout: whether an agent can create a cart, price shipping and tax and then complete payment under a protocol your payment provider supports.
• Identity: whether you can tell a signed, verifiable agent from a scraper claiming the same user agent string.
• Evidence: whether your logs separate agent sessions from human sessions, so the work can be measured at all.

Canlah AI treats access, facts and checkout posture as pass-or-fail gates, because a failed gate makes every downstream content decision unmeasurable.

What an agent-ready page must expose

An agent-ready page answers seven questions without a human in the loop, and Canlah AI checks all seven on every template it audits.

1. Which entity publishes the page, as Organization or LocalBusiness structured data with a stable identifier.
2. What is offered, at item level, with an identifier never reused for a different item.
3. What it costs on the date the page is served, in major currency units, with the currency named.
4. Whether it is available now, using an explicit status rather than an absence of one.
5. What the buyer must accept: shipping, returns, delivery windows and any regulated disclosure.
6. Which endpoint performs the next step: a search API, an MCP server or a checkout session.
7. When each fact was last confirmed, so a stale answer traces back to a stale page.

What 47 cross-border storefronts exposed in August 2026

In Canlah AI’s August 2026 sample, only 25 of 47 reachable storefronts served an agent-discoverable commerce manifest, and an agent-payments endpoint was not found on any of the 47. Canlah AI collected the sample across 50 Chinese-origin cross-border DTC brands with United States-facing storefronts, re-verified the issuer attribution on August 21, 2026 and published all 50 rows and 22 fields under CC BY 4.0.

Agent-facing endpoints and structured data on 47 reachable storefronts, collected August 2026.

Signal	Storefronts	Share	What the number means for an agent
UCP manifest at /.well-known/ucp	25 of 47	53.2%	An agent can discover commerce capabilities, but the profile describes the platform’s implementation
UCP manifests issued by Shopify	25 of 25	100%	No brand in the sample issued its own manifest; the capability arrived with the platform
Agent-payments endpoint	0 of 47	0%	Not found at collection, so no storefront published a way to settle an agent-initiated payment
MCP endpoint	1 of 47	2.1%	Not found at collection on 46, where an agent must fall back to reading pages
llms.txt present	26 of 47	55.3%	A plain-text site summary exists, which helps reading and does nothing for transacting
Product schema on the sampled product page	9 of 47	19.1%	Price and availability are machine-readable on fewer than one page in five
Review schema on the sampled product page	4 of 47	8.5%	Rating evidence is not extractable where an agent looks for it
robots.txt naming any AI crawler	6 of 47	12.8%	Access policy is implicit for the other 41, and 2 of the 6 block at least one agent

Source: Canlah AI open dataset, Agent-Readiness of 50 Cross-Border DTC Brands, version 1.0.0, collected August 2026. The sample is purposive and should not be projected onto DTC brands generally. “Not found at collection” means the endpoint did not respond at its published path when the storefront was fetched in August 2026. It is not proof that the brand has no agent strategy.

The finding that matters is not the 53.2%. It is that 25 of 25 manifests were platform-issued, which means most of these brands are agent-discoverable by accident and would lose that discoverability with a replatform they did not think of as a marketing decision.

Six agent-facing standards, and which ones apply to you

Six standards define the agent surface in 2026, and three of them, UCP, ACP and AP2, apply only to sites that sell online. Canlah AI reviewed each maintainer’s public documentation on September 23, 2026 rather than testing the protocols end to end.

Agent-facing standards reviewed September 23, 2026.

Standard	Maintainer	Discovery path	What it lets an agent do	Who needs it	What to verify
UCP, Universal Commerce Protocol	Google, with Shopify, Etsy, Wayfair, Target and Walmart, endorsed by over 20 partners	/.well-known/ucp	Read your capability profile, then run discovery, cart, checkout and order management	Sites that sell online	Which protocol version your profile declares, and whether your platform or your brand issues it
ACP, Agentic Commerce Protocol	OpenAI and Stripe, Apache 2.0, status beta	Product feed plus merchant-hosted checkout endpoints	Complete a purchase inside ChatGPT while you remain merchant of record	Sites that want to sell inside ChatGPT	That your payment service provider implements the Delegated Payment Spec
AP2, Agent Payments Protocol	Google, announced September 16, 2025, donated to the FIDO Alliance at version 0.2	Extension to A2A and UCP	Carry cryptographic proof that the user authorised this specific purchase	Sites that accept agent-initiated payment	Who is accountable when a mandate is disputed, in writing, before launch
MCP, Model Context Protocol	Anthropic, adopted as a UCP transport option	Your own advertised server endpoint	Call your search, availability or booking functions as tools	Any site with a task an agent should complete	Authentication, rate limits and what the server is allowed to change
NLWeb	Microsoft, open source	Site-hosted natural-language endpoint	Query your own content conversationally and expose it as an MCP tool	Content-heavy sites	Which index it reads, and how often that index is rebuilt
Web Bot Auth	IETF drafts, implemented by Cloudflare	/.well-known/http-message-signatures-directory	Prove cryptographically that a request came from the agent it claims to be	Any site that allows, challenges or blocks named agents	Whether your edge provider verifies signatures or only matches user agent strings

Source: each maintainer’s public documentation, listed in Method and sources and reviewed on September 23, 2026; no protocol was tested end to end. “Who needs it” is Canlah AI’s scope judgment, not a capability claim, and a standard may be implemented by your platform without appearing on your own domain.

The five-step agent-readiness checklist

The agent-readiness checklist has five steps, run in order, because each step assumes the one before it passed: fetch, facts, actions, checkout posture and traffic separation.

Step 1: Confirm the agent can fetch the page at all

Fetch your key templates with each named agent user agent and record the status code, because a 403 at the edge overrides every other step. Check robots.txt, then check the firewall rule separately, since a permissive robots.txt and a blocking bot rule are a common and silent combination. Repeat the fetch without JavaScript and confirm the facts you want quoted are in the first response.

Step 2: Publish the facts an agent cannot infer

Put entity, price, availability, location and policy facts into structured data and into a plain-language page. Nine fields are required for a usable product feed under OpenAI’s product feed specification for ACP: item_id, title, description, url, brand, seller_name, image_url, availability and price. Use unknown explicitly where stock status is unavailable, because an omitted or unrecognised availability value rejects the row.

Step 3: Expose actions, not only pages

Decide which three tasks an agent should be able to complete on your domain, then expose each as a callable endpoint rather than a click path. For most sites those tasks are search, availability lookup and booking or enquiry submission. An MCP server or an NLWeb endpoint turns each into a tool call; without one, an agent must guess your URL parameters and will frequently guess wrong.

Step 4: Decide the checkout posture before a platform decides it for you

Choose one of three positions, in writing: run agentic checkout, refer the agent to a human checkout or decline agent transactions. Each is defensible, and the failure mode is having no position, which in the August 2026 sample meant a manifest advertising capabilities the brand had never reviewed. If you choose agentic checkout, OpenAI specifies the launch tests: session creation with and without an address, shipping option updates, delegated payment tokenisation, order completion returning HTTP 201, order webhooks with a valid HMAC signature, recoverable error handling and idempotency behaviour under a repeated key.

Step 5: Separate agent traffic from human traffic

Split your logs by user agent and by signature verification before you report anything, because agent sessions have no scroll depth, no session duration and frequently no referrer. Keep user agent, timestamp, path, status code, client IP and content type. Where your edge provider supports Web Bot Auth, record whether the request carried a valid signature, since an unsigned request claiming a major agent name is the single clearest fraud signal available at the edge.

What to record at each layer

Canlah AI records six layers in every agent-readiness audit, and each record supports exactly one decision.

Agent-readiness record, six layers per audited template.

Layer	What to record	Decision it supports
Access	Status code per named user agent, per template, with the date	Whether to change a robots rule or an edge rule, and which one
Facts	Which of the nine required product fields are present, and the last-confirmed date	Whether the gap is a content task or a feed-pipeline task
Actions	Which tasks are callable, and by which endpoint	Whether to build an MCP server or accept read-only agent behaviour
Checkout	Declared posture, protocol version, manifest issuer	Whether the brand or the platform controls the commerce surface
Identity	Share of agent requests carrying a valid signature	Whether to allow, challenge or rate-limit unverified agents
Evidence	Agent sessions and outcomes, reported separately from human sessions	Whether any of the above changed anything measurable

Source: Canlah AI agent-readiness audit template, September 23, 2026; product fields follow OpenAI’s feed specification.

How to interpret five common results

Five results cover most agent-readiness checks, and each one points to a different first fix.

No manifest, no feed, no structured data: the site is readable and not transactable. This is the correct end state for a services business and an expensive gap for a storefront. Start at Step 2 and do not build an endpoint before the facts are correct.

Manifest present, issued by your platform: discovery works and belongs to someone else. Record the protocol version and the issuer, add the replatform risk to the technical roadmap and confirm which capabilities the profile advertises on your behalf.

Structured data present but stale: an agent will quote the stale price with the same confidence as a correct one. Fix the pipeline that generates the field rather than the page, then re-check a sample of items against the live cart.

Agent fetches the page and abandons the task: the facts are reachable and the action is not. Look at which step of the task requires a form post, a session cookie or a rendered widget, then expose that step as an endpoint.

Agent traffic present but unverifiable: requests claim a known agent name and carry no signature. Treat the traffic as unattributed until the edge can verify it, and do not report it as AI referral volume in the meantime.

What an agent-readiness check cannot prove

An agent-readiness check cannot prove five things, and a report that implies any of them is overclaiming.

• It cannot prove that an agent will select your site over a competitor that exposes the same endpoints.
• It cannot prove that a readiness fix caused a transaction, because agent behaviour, model versions and merchant eligibility all change in the same window.
• It cannot confirm that a platform-issued manifest reflects capabilities you have tested yourself.
• It cannot establish which party is accountable for a disputed agent-initiated payment, which is a contractual question rather than a technical one.
• It cannot replace legal review of the terms an agentic checkout integration asks you to accept.

When to move from a checklist to monitored agent traffic

Canlah AI recommends moving from a quarterly checklist to monitored agent traffic when any one of three conditions holds: agent traffic is a measurable share of sessions, the site transacts or several teams publish numbers that must stay consistent across a feed, a page and a manifest. Monitoring is also justified before a replatform, when platform-issued capabilities change silently.

Canlah AI is a Singapore-based SEO + GEO agency that measures and improves brand visibility inside ChatGPT, Gemini, Google AI Overviews and Google AI Mode, reporting AI visibility as re-verifiable ranges with timestamped evidence. On agent-readiness work its own scope stops at the discovery and facts layers: Canlah AI audits the manifest, the feed fields, the structured data and the access policy, and it does not implement checkout endpoints or payment tokenisation, which belong to your commerce platform and your payment service provider. Its own site publishes llms.txt, llms-full.txt and a robots.txt naming 21 user-agent groups, and serves no UCP profile at all, because it is an agency and not a storefront.

A small brand may not need monitoring. A documented quarterly re-run of the five-step checklist is enough until the number of items, endpoints or markets makes the process unreliable. Whichever route you take, verify every protocol detail in this guide against the maintainer’s current specification and your own fetch logs before you build, because versions and required fields change between releases.

Frequently asked questions

Do I need a UCP profile if I do not sell online?

No. UCP describes commerce capabilities, so a site with nothing to transact gains nothing from a UCP profile. The two layers worth completing for a non-commerce site are access, meaning an explicit robots.txt policy for named agents, and facts, meaning structured data plus a plain-language facts page.

How do I prepare my site for AI agents if my technical SEO is already sound?

Keep the technical SEO work, because indexability, response codes, canonical tags and rendering serve agents and search crawlers alike. Then add the three checks a technical SEO crawl does not report: which actions are callable, whether a commerce profile exists and who issued it and whether agent requests can be verified cryptographically at the edge.

Is llms.txt enough to make a site ready for AI agents?

No. In Canlah AI’s August 2026 sample of 47 reachable cross-border storefronts, 26 published llms.txt and only 1 exposed an MCP endpoint, so the file is widely adopted because it is cheap. llms.txt helps an agent read a site summary. It does not let an agent call a function, check live stock or complete a purchase.

Will AI agents buy from my site without an agentic checkout integration?

Some AI agents will attempt a browser-driven purchase on the human checkout, and those attempts often fail on address forms, session handling and payment fields. A declared checkout posture is more useful than an accidental one. If you decide not to integrate agentic checkout, say so in your terms and keep the human checkout usable by a browsing agent.

Should I block AI agents instead of preparing for them?

Blocking AI agents is a legitimate posture for paywalled or licensed content, and it is a decision to take per agent rather than per category. Retrieval agents decide whether your pages can be cited in current answers; training crawlers do not. Blocking retrieval agents removes you from those answers.

What does an agent-readiness audit from Canlah AI actually cover?

A Canlah AI agent-readiness audit covers the access, facts, actions and evidence layers: per-agent fetch tests, structured-data and feed-field coverage, manifest presence and issuer, a callable-endpoint inventory and a log split separating agent from human sessions. It does not cover payment integration, merchant approval or legal review. Confirm the scope in writing before the audit starts.

Method and sources

Protocol behaviour in this guide comes from six maintainers’ public documentation, checked on September 23, 2026. No paid account, merchant approval or vendor demonstration was used, and no protocol was tested end to end. The storefront figures come from Canlah AI’s own open dataset, collected in August 2026, which is first-party, purposive and not independently audited. Protocol versions, required fields and launch tests change, so confirm them against the specification before building.

• Google Developers Blog, Under the Hood: Universal Commerce Protocol. UCP scope, governance and the partner list. developers.googleblog.com
• Universal Commerce Protocol specification, version 2026-04-08. The /.well-known/ucp profile, version negotiation and payment architecture. ucp.dev
• Agentic Commerce Protocol. Maintainership, Apache 2.0 licence and beta status. agenticcommerce.dev
• OpenAI, Products feed specification. The nine required basic product fields and the availability value rules. developers.openai.com
• OpenAI, Agentic commerce in production. The launch certification tests quoted in Step 4. developers.openai.com
• Agent Payments Protocol. Announcement date, version 0.2, the FIDO Alliance donation and the A2A and UCP integrations. ap2-protocol.org
• Cloudflare, Web Bot Auth. The key directory path and the IETF drafts it implements. developers.cloudflare.com
• Microsoft, NLWeb reference implementation. Scope of the natural-language endpoint and its MCP exposure. github.com/nlweb-ai/NLWeb
• Canlah AI, Agent-Readiness of 50 Cross-Border DTC Brands, version 1.0.0, CC BY 4.0. All storefront counts in this article. canlah.ai open dataset
• Canlah AI robots.txt and llms.txt, read September 23, 2026. The 21 named user-agent groups and the published site summary. canlah.ai/llms.txt

See what AI answers already say about your brand. Run the free AI visibility audit →

Related articles

• How to Do Technical AEO on Your Site: Crawl Access, Rendering and Server Logs
• What Is llms.txt? A Plain Guide
• How to Rank in ChatGPT Shopping for Ecommerce Products: Feeds, Reviews, Merchant Data
• How to Do GEO for Shopify Stores in AI Search: Themes, Apps, Product Data