From Recommended to Transacted — AEO Is the On-Ramp, Agentic Commerce Is the Destination

Getting recommended is the on-ramp. Being transactable is the destination. ChatGPT-referred ecommerce converts at roughly 15.9% against 1.76% for Google (Adobe, 2025) — the highest-intent traffic on the internet — and most brands are pouring it into a checkout an agent cannot complete. This is the final piece in the series, and it is about the last, most valuable inch: turning a recommendation into a purchase without the shopper ever leaving the chat.

A growth lead at a D2C brand asked us, near the end of a call, what he clearly thought was the closing question: "So if we win AI visibility, the sales just follow?" The honest answer is no — not by themselves. Winning visibility gets an agent to name you. Winning the sale requires the agent to be able to buy from you. Those are two different pieces of infrastructure, and this series has spent fifteen pieces on the first one. This last one is about the second.

This is the sixteenth and final piece in our Winning in AI Visibility with Amazon series, and it sits on the same measurement spine as the rest: a locked panel of 425 real buyer prompts (mixer grinders) re-run monthly against ChatGPT's shopping surface and stored in our geo_vis schema, plus the beauty-category expansion. Everything before this was about being seen. This one is about being bought.

The number that reframes the whole series

Start with the figure that should change how every retail team budgets. Adobe Analytics, tracking generative-AI referral traffic across US retail through 2025, found that shoppers arriving from ChatGPT converted at a dramatically higher rate than those arriving from traditional search.

That is roughly a 9x gap. It is not a rounding artifact and it is not our number — it is Adobe's, on public retail data. And the mechanism is intuitive once you have read the rest of this series: the assistant does the interpreting, the query fan-out, the comparison and the shortlisting before the shopper sees anything — as the teardown of ChatGPT's shopping surface showed, an entire pre-ranked shelf is assembled behind the scenes. The person who lands on your offer has already been pre-qualified by the model. They arrive later in the funnel than any click has ever delivered them.

Here is the trap. A 9x conversion premium is only collected if the purchase can happen at, or very near, the moment of recommendation. If the agent recommends you and then hands the shopper a link to a website it cannot navigate, you have re-inserted the entire leaky click funnel back into the one channel that had removed it. The premium evaporates in the handoff.

Recommended is the on-ramp — and the on-ramp leaks

Everything this series has measured lives on the visibility side of that handoff. In our locked panel, Amazon's presence — the share of queries where it appears at any rank — sat around 80%, roughly 340 of 425 prompts, and climbed from there. That is a genuinely strong on-ramp. But presence is where most of the industry's attention, and every rank-tracking tool, stops.

Recommendation narrows fast after that. Being present is not the same as being the cited evidence behind the pick, which we covered in the two-layer model: in the same graph, own-citation ran in the mid-teens as a share of cited prompts and thinned from there. And even a citation is not a sale. The funnel keeps narrowing — present, cited, chosen — and then, at the very end, it hits the wall this piece is about: can the agent actually complete the transaction?

For almost every brand today, the answer is no. The agent recommends, the shopper is bounced to a mobile web checkout, and the assistant — which just did all the hard work of discovery — cannot follow. That is discovery without transactability, and it is where the 15.9% quietly reverts toward the 1.76%.

The destination is transactability — and the plumbing already exists

The reason this is a "now" problem and not a "someday" essay is that the transactability layer has stopped being theoretical. The protocols are shipping and merchants are live.

• The protocols. Three standards are converging on the same job — letting an agent complete a purchase against a merchant it does not own: ACP (Agentic Commerce Protocol)">Agentic Commerce Protocol), UCP (Universal Commerce Protocol) and AP2 (Agent Payments Protocol). They differ in sponsor and detail, but the shape is the same: a machine-readable catalog, an agent-drivable cart, and a payment step the agent can execute with the shopper's authorization.
• The proof it works. Nykaa went live inside ChatGPT on June 16, 2026, via OpenAI's Agentic Commerce Protocol, powered by Stripe. A shopper can be recommended a product and check out without leaving the assistant. That is the full arc — recommended to transacted — running in a real category today.
• The movers. Quick commerce did not wait. Zepto shipped a public MCP server, and Swiggy runs live MCP servers exposing its catalog to ChatGPT and Claude. Their inventory is already agent-callable. Most brands' catalogs are not, which means an agent literally cannot query them, let alone buy from them.

The incumbents are repositioning around exactly this. Amazon retired Rufus, its answer-box assistant, toward an "Alexa for Shopping" agent (announced around May 13, 2026) — a move from answering shopping questions to acting on them. When the largest marketplace rebuilds its assistant from recommender into agent, it is telling you where the surface is going.

What this means — for marketplaces and for D2C brands

If you are a marketplace or large retailer: your placement advantage is real but rented. You already win the buy-link layer on most cards; the Amazon shelf reversal showed how quickly that surface re-sorts. The defensible position is being the checkout the agent reaches for by default — the merchant of record the protocols route to first. That is won by being early and clean on ACP/UCP/AP2, not by out-optimizing a PDP.

If you run a D2C brand: this is the more urgent case, and also the bigger opportunity. You control your own catalog and your own checkout, which means you can become transactable without waiting for a marketplace to carry you — and you can do it while keeping your brand as the merchant of record and keeping the attribution. The brand that is agent-buyable in a category where competitors are only agent-findable collects the entire 9x premium alone. The gap Nykaa and the quick-commerce players opened is a gap, not a moat, and it is early enough to close.

This is where Unified Checkout fits. Winning citations and rank puts you on the shortlist; Unified Checkout is the transactability layer that lets the agent finish the job — a catalog an agent can query, a cart it can drive, and a checkout it can complete under the agentic-commerce protocols, with you as merchant of record and clean attribution back to you. Visibility products get you recommended. This is the one that gets you transacted.

What not to do

Do not treat an AI assistant as another link-out channel. The instinct is to bolt "traffic from ChatGPT" onto the existing web-analytics dashboard and optimize the landing page. That re-imports the click funnel you were just handed a way to skip. The assistant is a place to complete the sale, not to start another one.

Do not wait for a single winning standard. Teams are stalling on "which protocol wins — ACP, UCP or AP2?" The catalog exposure and checkout hygiene underneath all three are the same work, and it is the same work you would do anyway. Do the durable part now; the protocol wrapper is the cheap part to swap later.

Do not let channel-price and inventory drift make you un-transactable in practice. As covered in the price piece, the agent compares every offer before rendering one. An agent-reachable checkout showing an inconsistent price or a stale stock status doesn't just convert worse — it can get you routed around entirely.

Two qualifications keep this honest

First, the 15.9% figure is Adobe's US retail measurement, and conversion in your category and geography will differ — the shape (assistant traffic converting far above search) is the robust finding, not the exact multiple. Treat 9x as the direction, not a forecast for your P&L.

Second, transactability is necessary, not sufficient. Being buyable by an agent does nothing if the agent never recommends you — which is why the previous fifteen pieces exist. The checkout layer collects value the visibility layer creates. Build the destination, but do not stop feeding the on-ramp.

The series arc, in one paragraph

If you have followed the whole series, the shape should be clear now. It moved from visibility — being present on the shelf at all — to eligibility — crossing the gates that decide whether you can be cited — to levers — the price, review and evidence signals that move position once you qualify — and finally, here, to transactability — being able to complete the sale the visibility earned. Each stage is worthless without the ones before it and incomplete without the one after. Visibility with no eligibility is noise; eligibility with no transactability is a recommendation that dead-ends at a checkout the agent cannot use.

The measurement habit

The discipline is the same one this series has preached throughout, extended one column to the right. Keep the locked prompt set, re-run on a schedule, track presence and citation as separate scoreboards — and now add a third: transactability. For each recommendation surface you care about, can an agent actually complete the purchase, and is the attribution landing back with you? That column is empty for almost everyone today. The brands that fill it first will be collecting a 9x conversion premium while their competitors are still counting link-outs. If you want help wiring it up, book a demo.

We started this series with a question about why ChatGPT recommends someone else. We end it with a harder one: when it finally recommends you, can it buy from you? Getting recommended was always the on-ramp. Transactability is the destination — and, for the first time, the road is open.

---

That closes Winning in AI Visibility with Amazon. The next chapter is not another piece of analysis — it is the build: turning everything measured here into a catalog an agent can query and a checkout an agent can complete.