Amazon is blocking nearly every AI shopping agent. My read is that this is a negotiating position, and with the largest catalog in retail, Amazon can afford to hold it. It did something similar with Buy With Prime and Shopify, blocking access first and waiting for everyone to come to the table on its terms.

Goldman Sachs opens its 64-page report on agentic commerce with that fact, and it's the right place to start, because most of the industry is going the other way.

Shopify has never met an agent it didn't like. Walmart is selling inside ChatGPT, though nobody is reporting what those sales look like, while Amazon keeps its data behind a wall.

Jessica Lesesky sees it more charitably. Her point on the show was that the companies building these agents haven't earned much trust with payment data, and Meta has a long record of accounts getting hijacked on its platforms. Handing an agent your cards and your bank login is a big ask. I agree with her, but I think Amazon would make the same call if every agent were perfectly secure.

The section of Goldman's report labeled TBD covers who pays when an agent buys something it wasn't supposed to. In a store, when a fraudster uses a stolen card, the issuing bank eventually absorbs the loss, and nothing equivalent exists for agents yet. Jessica compared it to self-driving cars, where insurers are still arguing over whether the passenger or the manufacturer is responsible for a crash. Her bet is that the cost lands on the merchant or the consumer, and I think she's right. Amazon has fought regulators over whether it's responsible for dangerous products that third parties sell on its marketplace, so I don't expect the AI labs to volunteer for this one. I'd also rather the FTC settle it than whoever writes the terms of service.

Retail media was a $170B business last year, and every dollar of it assumes a person is looking at the page. An agent comparing products across a hundred sites never sees a sponsored listing.

That money will go somewhere, and Google, OpenAI, and Meta all want it. The labs keep saying their ads will be helpful to shoppers, which reminds me of the top-five mattress lists on blogs, most of them paid placements. The bigger issue for brands is that nobody can see what's influencing an agent's recommendation, and consumers can see even less than brands can.

Mondelez found that Oreos appeared in roughly 10% of AI cookie recommendations, a strange result for one of the best-known cookies anywhere. It hired AEO and GEO firms, and according to Digiday, Oreos now show up around 70% of the time.

Most brands haven't checked where they stand. Checking properly means testing how the agents themselves answer and doing technical work on your own site, since some brands are turning away bots they'd want to let in. Telling good bot traffic from bad is hard right now, and it gets harder as the volume climbs.

I don't know yet whether Amazon's wall holds or whether it's the opening offer in a negotiation.

Shopify is going after the money merchants spend on design help. Building a custom store used to mean writing code or paying someone to write it, and Canvas hands that work to Sidekick, which reviews screenshots of its own changes before the merchant sees them.

Ben Sehl, Shopify's Director of Product and a Kotn co-founder, says it took him two weeks to get a rough store working twelve years ago, and he knew how to code. He says a merchant can now build a better one in twenty minutes. I'd want to see conversion rates on those twenty-minute stores before giving that much weight, since the claim is Shopify's own. Sehl also admits Canvas isn't replacing the existing editor yet.

Agencies and theme developers have made a good living from Shopify being hard to customize. If merchants stop paying them, I don't know whether that budget moves into ads or ends up back with Shopify.