Recommended by AI, impossible to buy from.

There is a moment in an agent-mediated purchase that nobody is measuring, because it produces no data anywhere. The buyer asks for something. The assistant recommends your brand — genuinely, on merit, because the model knows you and thinks well of you. The buyer says "order it." And then the whole thing quietly falls over.

The agent reaches your site and cannot establish a price for this customer in this market. It cannot confirm a delivery date without a session. It cannot check whether your product satisfies the constraint the buyer actually stated. It cannot create an order without a browser. So it does the only thing left: it hands the intent back to the human with a link.

Sometimes the human follows the link. Often they do not, because by then the assistant has offered them a competitor who could complete the sentence. The demand was real, it was yours, and it converted somewhere else — and the only trace in your analytics is a slightly odd session that your bot filter may well have blocked before it ever became a session at all.

Eight layers of new plumbing, laid at once

Between 2025 and 2026, that infrastructure stopped being an architecture problem for individual companies and became a standards problem for the industry. Google published a commerce protocol. OpenAI, Stripe and Meta published a different one. Google donated its payment authorization protocol to the FIDO Alliance. Visa published an agent identity protocol. Cloudflare pushed cryptographic bot signing onto IETF-track standards. Coinbase revived an HTTP status code that had sat unused since 1997. The European Commission started building a registry of digital product passports.

None of it was coordinated. All of it answers one part of the same question: what does a business have to expose for an autonomous buyer to transact with it safely?

• 01Product knowledge — what exists, and what is verifiably true about it. Feeds, Schema.org, GS1, ETIM, the Digital Product Passport.
• 02Commerce — how it can be bought. UCP and ACP.
• 03Agent access — how an agent reaches your systems at all. MCP.
• 04Agent collaboration — how agents delegate to each other across company lines. A2A.
• 05Payment authorization — whether a human actually sanctioned this spend. AP2, Verifiable Intent.
• 06Machine payment — how software pays software. x402.
• 07Agent identity — telling a buying agent from a scraper. Trusted Agent Protocol, Web Bot Auth.
• 08Interface — what the buyer sees when someone else renders your product. A2UI, AG-UI.

The instinct on seeing thirteen acronyms is to wait for the standards contest to resolve. That instinct is expensive, because the layer doing most of the damage today is not contested at all.

The failure is silent, and it starts underneath the protocols

Search taught everyone a comforting model of failure: you rank lower, you see it in a dashboard, you do something about it. Agent-mediated commerce does not work like that.

When a buyer states a constraint — fits model X, waterproof, under 12 kg, delivered before Friday, REACH compliant — an agent can only apply that filter to products where the answer is machine-checkable. If your data cannot confirm it, you are not ranked lower. You are dropped from the comparison, silently, with no impression to explain it. Your competitor with the duller website and the complete attribute set survives to the shortlist.

This is why the highest-leverage work available to most brands right now is not protocol adoption. It is the unglamorous layer underneath: identifier discipline, classified properties with real units, claims that trace to a source, compliance data attached to the product record rather than filed in a drive, and price and availability fresh enough to be quoted rather than guessed.

The field guide: The Agentic Commerce Stack

We have published the whole methodology, free and without an email wall, in the same way we published The Agent-Operable Enterprise last month.

And the engagement: the Agentic Commerce Readiness Sprint

The guide gives away the method deliberately. The value was never in knowing the steps exist — it is in running them across five functions that do not currently talk to each other.

That is the honest diagnosis of why this work stalls in most organizations. Product data sits with merchandising. Commerce capability sits with digital. Identity sits with security. Authorization sits with payments and legal. Presentation sits with brand. Five owners, no shared meeting, no shared vocabulary, no shared metric — and a capability that fails at whichever of the five is weakest.

It joins the auxfirst canon as the fourth fixed-scope engagement, and it has a deliberate relationship to one of its siblings. The Agentic Shelf Audit asks whether AI models recommend your brand. This sprint starts where that one ends: whether an agent, having recommended you, can actually complete the purchase.

What we would do first, if it were our brand

Three things, none of which require choosing a protocol.

Look at what you are blocking. Ask your CDN or bot-management vendor for a breakdown of AI and agent traffic, including what was challenged or mitigated. Most teams have never looked. Most are surprised. Until verified agent traffic appears in a dashboard someone reviews, it does not exist organizationally, and no one can make the case for anything else.

Test your constraints, not your completeness. Write down the ten things buyers actually say in your category, then check whether each one is machine-checkable in your data — by category, weighted by revenue. Overall completeness scores hide the gap; the gaps are never evenly distributed.

Ask whether your commerce core is callable. Every protocol in this stack reduces to three demands: resolve, quote, commit. Return the authoritative record for an item. Return the real total for this basket, this address, this customer, today. Create the order. If those are callable services, adding a protocol is an adapter and a test suite. If they are trapped inside a session-bound checkout, adding a protocol is a project — and you will be doing it more than once.

The organizations that struggle over the next two years will not be the ones that picked the wrong standard. They will be the ones that waited for the contest to resolve before improving data they needed anyway.

---