An agent found you. Can it buy from you?
Most catalogs answer no, and nothing tells them. There's no error, no bounce and no lost-cart email — the product simply isn't in the set the agent can complete a purchase from. Here's the sequence that changes it, in two weeks, done by a person.
Delivered in 5 working daysRefunded if it turns out you don't need the two weeksI work on your site, never in your payment account
What agentic commerce optimization is — and why it isn't SEO
Search optimisation answers a question about attention. When a person types something, does your page appear, and is it persuasive enough to click? Every technique follows from a human reading a result.
Agentic commerce answers a question about capability. When software is sent to buy something on a person's behalf, can it establish what you sell, agree terms, and complete a payment without a human touching anything? Nothing about that is persuasion. The agent isn't reading your headline or being reassured by your testimonials — it's checking whether the operations it needs are present and answering correctly.
That's why a beautiful, well-ranked shop can be entirely absent from the agent channel, and why fixing it is engineering rather than copywriting. The two disciplines share a vocabulary and almost nothing else.
The protocol layer, in one minute
Three standards govern this, published by Google and OpenAI, and they're independent of each other. Each solves one problem in the order an agent meets it — so implementing one of the three doesn't leave you partly ready. It leaves a specific, silent hole.
A manifest at a well-known address declaring what you sell, what it costs, how you ship and which transports you speak. Without one, an agent arriving at your domain finds nothing it can act on. That isn't ranking low — it's being absent from the set.
On its own: Discoverable, but nothing can be bought.
Agents don't fill in forms; they negotiate. This governs that conversation — comparing offers, applying a buyer's preference for speed over cost, settling terms — as a stateful exchange rather than a page flow.
On its own: Able to sell, but nobody can find you.
The mandate that proves an agent may spend its principal's money, verifiable by the bank. It's what makes the other two trustworthy enough to complete.
On its own: Found and negotiated with, then stopped at the last step.
A shop with all three is reachable end to end. A shop with two is reachable until the step it's missing, and it will never be told which step that was.
If you are the person who has to implement this rather than the one paying for it, the step-by-step version is on its own page: how an agent buys from a site — the six things an agent does in order, what it needs to see at each one, and the published version of every specification above with the day we last read it.
Why now — other people's numbers, not ours
These are third-party projections, attributed and dated. We publish them separately from anything we measure, because a company that sells measurement should never let a forecast and a finding wear the same clothes.
dollars in global orders influenced by AI agents — about a fifth of all transactions in that window.
Salesforce, Cyber Week 2025
of B2B purchases expected to flow through AI agents.
Gartner projection, by 2028
of Google searches already end without a click. The channel most businesses were built for is contracting while this one opens.
SparkToro and Datos
Which agents you're optimising for
Not one, and that's the practical point. The assistants people already use — ChatGPT, Gemini, Perplexity, Copilot and the rest — plus the shopping and procurement agents being built on top of them, and the coding agents a developer points at your site to integrate with it.
They don't share a ranking algorithm you could court. They share protocols. That's genuinely good news for a smaller business: you can't outspend an incumbent on authority, and you can absolutely publish a correct manifest before they do.
The sequence
- 01Make the offer readable (Week 1)
Terms, privacy and a return policy at stable URLs, because a buying agent checks all three before it spends its principal's money. Offer markup on every product with a real price and an ISO-4217 currency — a bare dollar sign is two different prices to an agent quoting for Montréal and for Delaware. - 02Publish the discovery file (Week 1)
A profile at /.well-known/ucp so agent-commerce clients can establish what you sell and how to deal with you. Without it you aren't ranked low in the set of merchants an agent considers — you're absent from it. - 03Build the catalog feed (Week 2)
The three files the rails ingest, built from the markup you already publish. Nine of eleven required fields is the minimum accepted for physical goods; a missing field drops the whole ROW, so the product vanishes from the channel with no error you'd notice. Every blank comes back marked, never guessed. - 04Fix what the feed exposes (Week 2)
Dead product links and 404 images, prices without currencies, promotions with no end date. These are cheap to fix and each one silently removes products from the channel until it is. - 05Prove it, then hand it over (Week 2)
I re-run the audit and you see the level move. What has to be switched on inside your Stripe account is written up for you or your developer to do — I work on your site, never in your payment account.
What you get, and what you don't
Will this get me into ChatGPT's checkout?
+
No, and anyone promising that should be asked to show you the programme. Instant Checkout is a separate, US-only arrangement with its own approval. What this does is make you buyable on your own rails, which is what an agent falls back to and what you keep whatever any one platform decides.
Do you need access to my Stripe account?
+
No. I work on your site and hand you files and instructions. Your payment account stays yours, and I never hold a key that can read your money.
What do I need to give you?
+
Your domain and the questions your buyers actually ask. If you're not sure what those are, I draft them and you correct me — that's ten minutes, and it's the part that decides whether the rest is useful.
What if I'm already fine?
+
Then I tell you and refund the audit. It happens — some catalogs are further along than their owners think, and the two-week sequence would be work for nothing.
Key takeaways
- Being buyable by an agent is a different job from ranking. One is a protocol question, the other is a content question, and doing the second well does nothing for the first.
- The three standards are independent. Implementing one leaves a specific hole, not partial progress — and the hole is silent.
- There's no error message. An agent that can't complete a purchase doesn't bounce, complain or fill a lost-cart report. It picks somebody else and you never learn it happened.
- It's a fixed, finishable piece of work — closer to two weeks than two quarters.
Start with the number.
The audit — $499 — tells you your level from 0 to 4, the exact gates between you and the next one, and where your competitors sit. If the answer is that you don't need the two weeks, that's what it will say, and the audit is refunded.