Agentic checkout: what brands should do now.

Agentic checkout generates more conference slides than purchases right now, which makes it easy to dismiss — and dismissing it is the wrong read. The buyer's journey has already grown two machine-mediated steps, and a third is in pilot. The practical question for an ecommerce lead is not "is the future agentic" but "which stage is live, and what work pays off at every stage regardless of which protocol wins." Here is the staged read as of mid-2026.

Stage one, fully live: AI shopping research

ChatGPT, Perplexity, Google AI Overviews, Gemini, Claude, and Copilot answer "best X for Y" questions with named products and cited sources every day. This stage is not speculative; it is measurable per SKU, and the mechanics are covered across this blog — how engines pick products, Perplexity's citation patterns. If your products lose this stage, the later stages never see them: an agent does not draft a cart around a product the answer never surfaced.

Stage two, live for participants: draft carts

ChatGPT Shopping and Perplexity's Buy with Pro pre-fill a cart on the buyer's behalf and hand back for confirmation and payment. The buyer still completes checkout; the agent does the selection and assembly. For the merchant, the qualifying bar is machine-readability: accurate price and availability in Product JSON-LD, parseable variant structure, and discoverable shipping and returns policies. A PDP that converts humans can still be unreadable to the agent doing the drafting — that gap is what the agent-readability score measures, separately from citation.

Stage three, in pilot: agent-completed checkout

Three protocols are racing to close the loop so the agent can complete the purchase itself.

Statuses per the published specs and partner announcements — OpenAI and Stripe document ACP publicly at agenticcommerce.dev. Shopify's Agentic Storefronts work gives its merchants a platform-level on-ramp. None of this is the dominant consumer path yet, and anyone telling you otherwise is selling urgency. The deep treatments live in the ACP guide and the UCP guide.

What to do now: the no-regret list

The strategic property of this moment is that the preparation is protocol-agnostic. Whichever standard wins, agents will read the same surfaces:

1. Complete Product JSON-LD per SKU — price, availability, GTIN, variants, return policy. The field-by-field post has the full block.
2. Admit the AI crawlers and action agents in robots.txt — OAI-SearchBot, ChatGPT-User, PerplexityBot, ClaudeBot. The robots.txt post has the starter policy.
3. Make policies discoverable — returns, shipping, warranty on linkable, crawlable pages, not buried in an accordion rendered client-side. Agents check policies before recommending a purchase.
4. Serve an llms.txt — cheap hygiene that some crawlers respect as of mid-2026.
5. Measure both scores per SKU — citation and agent-readability fail independently and get fixed by different teams. The agentic commerce page shows how the two-score model maps to the three stages.

Every item raises citation odds in stage one, qualifies you for draft carts in stage two, and becomes a hard prerequisite in stage three. There is no scenario where the work is wasted — which is rare enough in this category to be worth saying plainly.

What not to do

Do not re-platform for a pilot. Do not sign a multi-year "agentic commerce suite" contract to solve a problem that is currently a robots.txt edit and a schema sweep. And do not report agent-completed checkout revenue to your board as if it were a current channel; as of mid-2026 it is not, and credibility spent there is hard to win back. The honest framing for leadership: research and draft carts are live and measurable, full agent checkout is a 2026–2027 rollout to be ready for, not to bet the quarter on.

How to know when stage three goes real

Watch three signals rather than press releases: a major Shopify cohort enabling ACP or UCP checkout by default, an engine reporting agent-completed transactions as a standard merchant metric, and refund or dispute flows getting protocol-level treatment (the unglamorous part that decides whether agents can transact at scale — AP2's territory). When two of the three land, move agent checkout from "monitor" to "implement." Until then, track your catalog, fix the readability gaps, and let the free Agentic Readiness Grader tell you where a given PDP stands today. The product AI visibility guide ties all three stages into one operating model.

Key takeaways

• Agentic commerce is arriving in stages: research (live), draft carts (live for participants), agent checkout (pilot).
• ACP, UCP, and AP2 are the protocols to know; all are in pilot as of mid-2026.
• The no-regret work is protocol-agnostic: schema, crawler admittance, discoverable policies, llms.txt.
• Losing stage one removes you from every later stage — citation work comes first.
• Prepare without re-platforming; report the stages honestly to leadership.