The last piece ended on a question: when an agent does the shopping, whose agent is it? Four camps were each defining agentic commerce as the slice of the journey they already controlled, and the size of the companies doing the defining (OpenAI, Google, Stripe) was doing a lot of the work in deciding which definition sounded natural.

Step away from the vendors and look at it from the two vantage points that actually settle this: the merchant’s and the shopper’s. From there, the paradigm looks different from the one being sold.

Start with the part everyone got wrong

Payment was supposed to be the hard problem. For shoppers, it mostly isn’t. Anyone who finds checkout annoying already solved it years ago with browser autofill or a password manager; anyone who hasn’t solved it that way is not about to hand their card to an autonomous agent. The friction was real, but it was never the friction that stopped a purchase.

The market has now said the same thing out loud. OpenAI retired Instant Checkout in March 2026, roughly six months after launch, and repositioned ChatGPT around product discovery, with checkout handed back to merchants’ own environments. Walmart’s own numbers pointed the same way: checkout inside ChatGPT converted materially worse than a click-through to walmart.com, even as ChatGPT delivered a much higher rate of new customers than search. The least we can say is that end users are not ready to check out inside their AI assistant yet. Your agent can live wherever you want, but checkout between agents is still far from mainstream. [1] [2] [3] [4]

Furthermore, delegating to agents still carries risks for both merchants and users: is the price accurate? Does the promotion exist, and is it correctly applied if it does? Any hallucination here would break the trust of both sides.

Discovering through AI is becoming the new standard, while checkout stays on the merchant’s site.

Put the funnel back on the table

The old e-commerce funnel is still the most honest map available: awareness, interest, consideration, conversion, retention. Run the agentic era through it and the picture sharpens fast.

Awareness is genuinely being rewritten. What used to run on SEO and paid social now leaks toward the growing share of product questions people put to an assistant, and toward the young discipline of getting recommended by one. And the monetisation arrived faster than most brands planned for: ChatGPT Ads launched in the US in February 2026 and expanded to 31 European markets on 24 August, bringing the total to around 40. [5]

Worth keeping the nuance, though, because the caution is visible in the product itself. Ads sit in labelled slots, they’re claimed not to influence the answers, and advertisers don’t get the conversations. In Europe, personalised targeting requires explicit consent. For sure, an important part of ad budgets will move from Meta and Google to OpenAI, but it’s still very hard to predict whether advertisers will get more efficiency through these new channels, with higher conversion from AI agent ads than they had on the incumbents. [6] [7]

In the meantime, the problem of understanding and retaining their own customers remains unsolved for brands that don’t get more data from their own customers.

In our good old e-commerce funnel, that leaves us in the middle: interest and consideration, a part less publicly fought over, and the one part brands can still own. That’s not a consolation prize. It’s where the decision is made.

Not every purchase is worth defending

Of course, plenty of purchases will be intermediated, as they have been for 20 years with Amazon, and the world will be fine for it. Replenishment, commodities, the printer ink you reorder without thinking, the seller whose business is essentially marketplace distribution. For all of those, an agent buying on your behalf is a genuine improvement, and fighting it is wasted energy.

But there’s a difference between what gets bought and what gets shopped. Shopping is the category where desire is the subject: discovery, taste, identity, the fragrance, the jacket, the bag you’ll keep for ten years. People actively seek out brands they find desirable, and the act of choosing is part of what they’re paying for. It’s also often a higher average basket. That’s where brand equity lives, and it’s exactly where a generic agent flattening you into “here are three options” destroys value.

So the strategic posture is two-sided: accept intermediation on the buy tail, and defend the shop experience ferociously. The costly mistake isn’t being intermediated. It’s being a shop brand that lets itself be handled like a commodity.

Being described is not being represented

Ask any e-commerce team how ChatGPT talks about their brand and you’ll hear the same three complaints, almost word for word. The tone is wrong. The brand image doesn’t come through. And the products it pushes are not the ones the brand would put forward first.

None of that is a bug. OpenAI’s own framing is that merchants share product feeds so their products are fully represented in conversations. But a feed is a description, not a representation. It carries attributes; it doesn’t carry judgement. Nobody outside your company knows which products to bundle, which upsell is worth making, which item you’d rather not push this month because the margin is thin or the stock is wrong. The data to make these decisions should still be the property of the brands: their margins, their strategy, and all the proprietary data left to them by their own customers on their website or even in their own retail stores. [8]

OpenAI’s own website states that building your own App (not strictly an agent, but the generalisation holds) gives you more control over the experience than simply uploading a feed. [2]

As a brand that wants to be shopped, you surely don’t want to be disintermediated in your relationship with your customer. What that means is: you want it on your store, with your tone of voice, and direct access to what is shared between them and any AI at that moment.

“But couldn’t Google just do this?”

It’s the first objection worth taking seriously, and the honest answer is that Google already has. Business Agent went live on 12 January 2026 with retailers including Lowe’s, Michael’s, Poshmark and Reebok: a conversational experience on Search that Google describes as a virtual sales associate answering product questions in the brand’s voice. Retailers activate and customize it from Merchant Center: colours, welcome message, suggested prompts, a tone-of-voice setting. And Google has said retailers will be able to train it on their own data, get customer insights, and eventually enable direct purchase inside the conversation. [9] [10] [11]

So the objection isn’t “can they build it.” They can. The question is what you actually get.

You get a tone picked from a menu, not a voice. You get grounding limited to what fits a product feed and a public website, not your merchandising logic, your margins, or what your best sales associate knows. You get an agent running on Google’s surface, in Google’s session, with the customer insight flowing back through Google. And today, you get it if you’re a US retailer.

It’s a rental. A good one, worth taking where it’s available. But you don’t own the logic, and you don’t own the relationship it creates.

What a brand agent actually has to encode

Owning it means encoding three things that only exist inside the company:

Identity. Built from the brand’s own content and its own rules about what may and may not be said: configuration that has no equivalent inside a general assistant.

Responsibility. What the brand is allowed to say, and what it must never say. Every brand carries commitments it cannot contradict and claims it cannot legally make. A cosmetics brand can’t promise a medical outcome, a retailer can’t invent a promotion or quote a price that isn’t live. Stocks may or may not be shared with end users. This is the guardrail layer. A brand-specific agent can be constrained, whereas a generic one can only be prompted nicely.

Expertise. Selling, which is two things at once. First, turning a stated intent into the right recommendation: a well-established machine learning problem (collaborative filtering, propensity modelling) that has to be reinvented now that the input is a conversation rather than a click. “It has to fit Ryanair cabin rules,” or a photo of someone’s skin, are inputs no ranking model was built for, and far richer signals than a session log ever produced. Second, knowing when to push and when not to: which products to bundle, which upsell is worth making, which item to leave alone this month because the margin is thin or the stock is wrong. That judgement comes from the brand’s existing sales discourse, its site data, and above all from past conversations: what landed, what didn’t, when pushing worked and when it cost trust. You train human associates by hand, one at a time. An agent deployed across a site improves from every conversation at once.

That last point carries the condition the whole thesis rests on. The compounding only happens if the conversations stay yours. An agent is a listening post before it’s a salesperson: it’s the first system that hears customers describe what they want in their own words, at scale. Hand that surface to a platform and you get the recommendation without the learning, and next year’s agent is no better than this year’s.

Some merchants can skip all of this, and should. But a brand with a strong identity has no real alternative to encoding its expertise into an agent it controls. It isn’t a break from what came before, either; it’s the logical next version of the search, upsell and cross-sell systems already running on the site, rebuilt around conversational data.

Which raises the obvious next question: what does that agent actually do first? The answer isn’t selling. It’s finding, and your search bar isn’t doing it yet. More on that in the next post.

Sources