While OpenAI and Google race to become the place where shopping happens, Anthropic just released the tools for merchants to keep it happening on their own turf. The distinction in posture matters more than the launch itself.

Last week Anthropic dropped a set of blueprints for building shopping and merchant agents on its Claude models, timed neatly for retailers staring down their holiday planning calendars. On the surface this reads as the third major AI lab planting a flag in agentic commerce — the emerging category where software agents do the searching, comparing and cart-building that humans used to do with seventeen browser tabs and a growing sense of regret.

But filing this as "another entrant in a crowded race" misses what's actually interesting. Anthropic isn't running the same play as OpenAI and Google. It's running close to the opposite one, and for those of us selling things online, the difference matters considerably more than any single feature in the release.

The blueprint comes in two halves, delivered as reference code, patterns and guardrails designed to get a working commerce agent up and running quickly, with deployment across Amazon Bedrock, Microsoft Foundry and Google Cloud Vertex AI.

The first half is a shopping agent that lives inside a retailer's own app or website. A customer describes what they want in plain language — say, a tent, a sleeping bag and a stove for a weekend away with two kids — and the agent searches the catalogue, assembles the set, tailors suggestions, shows comparisons inside the conversation and builds the cart. It also handles service queries about orders, returns and refunds in the same thread. When the cart is ready, it hands off to the retailer's existing checkout. It does not complete the purchase itself. Payment stays with the merchant (a detail we'll come back to).

The second half is a merchant-facing agent that answers questions about what's selling, flags inventory problems before a promotion drains stock, and drafts pricing and campaign suggestions from the store's own sales history. A human approves any change before it goes live. So it's less "autonomous AI running your business" and more "extremely diligent intern who actually reads the data before suggesting things."

Anthropic says an early partner saw cart sizes rise roughly 30–35% and shoppers become about 60% more likely to complete a purchase. Those figures come from a single unnamed partner and haven't been independently verified, which makes them encouraging rather than canonical. Named collaborators include Shopify, Priceline and Wix on the retailer side, with Accenture, Mastercard and Visa positioned to carry the blueprints into their own client networks.

To see why this matters, we need to separate two things that agentic commerce quietly folds together: discovery and conversion. Discovery is the top of the journey — working out what you want and which product to buy. Conversion is the bottom — actually paying for it. For most of the web's history these happened in different places. A search engine or marketplace handled discovery and sent the traffic onward; the retailer's site handled conversion on its own terms.

We've tracked across several issues of this newsletter how AI is pulling those two stages apart and relocating them.

OpenAI and Google have both moved to own discovery. Shopping increasingly starts inside ChatGPT or inside Google's AI-generated answers, where the assistant itself becomes the shop window. Retailers don't compete for a click anymore — they compete for selection, which is a polite way of saying they compete for the model's approval. Both companies have built commerce protocols (standardised ways for an agent to read a catalogue and, in Google's case, push a purchase along) to formalise that role. The strategic effect is that the moment of discovery migrates onto a surface the platform controls, and the merchant becomes a supplier to it.

Anthropic has declined that lane entirely. There is no Claude shopping destination for consumers to visit. No assistant of Anthropic's own competing to be where the journey begins. Instead the company is selling the engine and letting the merchant run it on their own storefront.

In the language of technology strategy, this is the toolmaker's position — selling the picks and shovels rather than staking a claim. It's a deliberate refusal to become a demand aggregator (the kind of platform that gathers shoppers' attention and then rents access to it back to sellers, which is a business model that has historically never, ever resulted in escalating fees and diminished leverage for the seller — oh wait).

The appeal is immediate, because it answers the fear that the discovery shift created in the first place. When a general-purpose assistant makes the recommendation, the retailer may still get the sale, but the relationship happens somewhere else. The questions, the preferences, the comparison behaviour, the reason to come back — all of it accrues to the platform rather than the seller. An agent that runs on the retailer's own site promises the reverse: capable AI intelligence, with the customer, the data and the return visit all kept in-house.

There are two catches, and we should hold both in view.

The first is a quieter dependency. Keeping the customer relationship is real, but the reasoning still runs on someone else's model. A brand swaps the risk of renting the audience for the risk of renting the intelligence. That's a better trade for most merchants (the customer relationship is the harder asset to rebuild), but it isn't independence. It's a different flavour of reliance, packaged in more comfortable language.

The second catch is structural, and it's the one worth sitting with. A storefront agent defends the conversion end of the funnel. It does very little for discovery. A superb agent waiting on your website is of limited use if the shopper never arrives, because they asked ChatGPT or Google's assistant instead and bought from whichever product those surfaces chose to name. The two plays are not competing for the same ground. OpenAI and Google are contesting where shopping begins; Anthropic is helping merchants win the shopper who has already turned up. A brand can adopt the Anthropic blueprint in full and still lose the customer upstream, before its storefront agent ever gets to speak.

It's the retail equivalent of hiring the world's best floor staff for a shop that fewer people are walking past.

For sellers whose primary surface is Amazon, the tension is starker still, because the central promise of the blueprint — run the agent on your own turf — is not available to them. A third-party seller doesn't own the storefront; Amazon does. The conversational agent shoppers meet there is Rufus, which answers to Amazon's ranking and merchandising logic rather than the brand's intentions.

The option to "keep the customer on your own ground" assumes you have ground of your own. The sellers with the least control over their surface are the ones this particular answer can't reach, which leaves their exposure sitting squarely with whatever Amazon's own agent decides to show (and how little of that decision Amazon chooses to explain, which — based on precedent — will be approximately nothing).

One feature of the release is worth reading as more than a product detail. Anthropic has built in deliberate limits. The agent recommends but does not transact on its own. The merchant agent requires human approval before changes take effect. Guardrails tie prices and products to real catalogue data and discourage manipulative upselling.

These constraints are being positioned as prudence, and they are. But they're also an admission. Keeping a human at the point of payment and at the point of pricing is the responsible design precisely because a long-running agent left to optimise a target without supervision can't yet be relied on to behave. The limits are simultaneously a feature and a confession, which is a combination that more AI companies could stand to adopt (looking at you, every startup whose demo works flawlessly and whose production deployment does… something adjacent to that).

The contest that matters isn't which lab has the most capable shopping agent this quarter. It's whether the branded storefront survives as a place shoppers still choose to go.

Anthropic is betting that it does, and selling merchants the tools to make their own site worth arriving at. OpenAI and Google are betting that the storefront becomes a back-end that fulfils orders decided elsewhere, with the interesting part of the journey happening on surfaces they control. Both positions can't be entirely right.

For those of us operating in this space, the practical read is straightforward: a storefront agent and a discovery strategy are answers to different halves of the same problem, and adopting one doesn't excuse neglecting the other. You need to be findable where the AI assistants are looking and you need to convert brilliantly once the shopper arrives.

The open question is whether a brilliant agent on your own storefront turns out to be a genuine moat — or a beautifully appointed room in a house that fewer and fewer shoppers still walk into.

P.S. The fact that Anthropic deployed across Amazon Bedrock, Google Cloud and Microsoft Foundry simultaneously is the corporate equivalent of telling all three suitors you're "keeping your options open." Respect.

P.P.S. If you're an Amazon seller reading this and thinking "great, another tool I can't use because I don't own my own storefront" — you're not wrong, and we should probably talk about that. The D2C conversation just got a lot more urgent.

🤖 CodeRabbit - AI code review that leaves context-aware feedback on pull requests.

🖥️ Gamma - Generate polished presentations, docs and sites from a prompt.

🌐 Framer - Design and launch fast websites with AI, no code required.

🌪️ Zapier - Connect thousands of apps and add AI steps to automate workflows.

😍It was great, all I needed

🙂Good, not great

🙄It sucked