How a Claude shopping agent picks products

Anthropic's Claude for Commerce blueprint ships a shopping agent and no ranking layer. Claude in Chrome shops any site for a consumer and asks before it buys. Here is what Anthropic documents, what it leaves out, and who has filled the gap.

A Claude shopping agent is one of two things. Under Claude for Commerce, merchant-run code searches the merchant's catalog, and Anthropic documents no ranking method. In Claude in Chrome, Claude browses a store for the shopper and asks before buying.

Last updated: 12 October 2026.

Two things people mean by "Claude shopping agent"

The phrase covers two products that share a model and little else.

The shopper's agent: Claude in Chrome. "Claude in Chrome reads the page you're signed in to, then clicks, types, and fills forms while you decide what happens next", and it "stops before taking sensitive actions" (claude.com/chrome, read 10 October 2026). The pilot post set the rule: approval before "publishing, purchasing, or sharing personal data" (Anthropic, 25 August 2025). It shops any site, and what it sees is your rendered page.

The merchant's agent: Claude for Commerce. On 2 September 2026 Anthropic published "Building commerce agents with Claude" and opened anthropics/commerce-agents under Apache 2.0, reference code that "is not maintained and does not accept contributions". "The shopping agent lives inside your app or website" (Anthropic, 2 September 2026). A merchant runs it, Claude.ai users never meet it, and it sees whatever the merchant's backend returns.

Neither is a Claude storefront or a listing programme. The release is covered in Claude for Commerce explained and the brand view in Claude for brands. The merchant agent has the documented answer, so it comes first.

What Anthropic documents about product selection

Everything here is from the repository or the engineering deep dive by Ali Shazal and Matthew Koen, read 10 October 2026.

Search is yours. StorefrontBackend has "11 required methods, plus get_account_context, get_disclosure, and checkout_handoff"; search is one of them. The deep dive names the systems an agent's tools should call, "search, ranking, cart, inventory, fulfillment", and says "the tool boundary is where their logic ends" (Anthropic, 2 September 2026). Your search order is the agent's starting order. "A model-supplied result count is clamped to max_search_results" (docs/safety.md).

Rows, not pages. "A compact row is 70 to 120 characters, so a family holds up to about sixty variants"; detail is capped at 12,000 characters per family (docs/backends.md). Builders are told to "Return only fields the model needs" and warned against "image URLs on every search row" (deep dive). ReFiBuy's Vedant Gupta traced one prompt through the reference agent and landed in the same place: what the agent gets "is a short row of fields from the catalog", "the product page itself does not really come into it", and "For a lot of it, the information simply was not in the record" (via Scot Wingo, LinkedIn, 11 September 2026).

Options filter; price is a "from" price. "Search matches option values as well as attributes, so a filter like size = king works." "Storefront price is the lowest in-stock variant's (a 'from' price); a family is in stock while any variant is" (docs/backends.md).

Ids are server-issued. "Cart writes accept only product ids a catalog or order tool returned this session, or lines already in the cart", and "A term or a figure is stated only from a tool result in this conversation" (docs/safety.md).

Checkout and the web are out of reach. "The backend returns the URL and the host renders it; the model never sees it", and "Web search is registered only when enable_web_search is set" (README). The demo data is fictional: four verticals, retail, travel, telecom and entertainment (ticketing), where "the only company is ACME".

The wiring, and where MCP does and does not enter, is in Claude for Commerce for merchants.

What Anthropic does not document

Read on 10 October 2026, none of the claude.com post, the README, docs/safety.md, docs/backends.md or the deep dive contains a ranking algorithm, relevance score, boost, sponsored slot, paid placement or ad format; a crawler, index, product feed, GTIN or schema requirement; a cross-merchant comparison; or any route for a brand to "get into" Claude's shopping results, because Claude has none of its own.

On India, the primary sources hold one sentence: Akhil Bansal, Senior Engineering Manager at Zomato, says the blueprint's practices "are the ones we recognized from building Zomato's own agent" (claude.com, 2 September 2026); no Indian retailer, UPI rail or India availability statement appears.

Vendors and bloggers have filled the silence. These are their claims, not Anthropic's.

Claude in Chrome: the shopper's agent reads your page

Flip to the consumer side and the inputs invert. Claude in Chrome has no catalog contract; it reads the rendered page, so the price it compares is the one on screen and the variant it picks is the one in your selector. Anthropic's pilot numbers: prompt injection success fell from 23.6% to 11.2% with mitigations, and browser-specific attacks from 35.7% to 0% (Anthropic, 25 August 2025). A directory listing refreshed 30 September 2026 files Claude in Chrome under "AI shopping agent" (ecommerceguide.com), which is how the two meanings get mixed. The merchant agent cannot see your product page; the consumer agent sees nothing else.

What this means for your catalog

For a Claude for Commerce deployment the agent acts on fields. Prose and images do not reach it by default.

Three moves follow. Put size, shade and pack count in options, because an option is filterable and a paragraph is not. Keep variant price and stock live; the "from" price is computed from them. Make the shopper's constraints answerable from the record, since "the information simply was not in the record" was Gupta's finding (11 September 2026). Method: optimising your product catalog for AI agents. For Claude in Chrome the work is the page: a visible price that is the real price and a checkout an agent can finish, covered in enable agentic checkout.

How this compares with surfaces that publish their ranking

OpenAI has told merchants what it weighs; Anthropic has handed the weighing to the merchant, so a Claude for Commerce agent is the one surface where "how does it rank" is a question for your own engineers. We ran the same exercise for Meta in how Meta Muse picks products; the wider comparison is AI shopping agents: how they buy. MCP and the referral numbers are in Claude commerce: how Claude shops and how brands get in, not repeated here.

Where Tru Commerce fits

MCP App Builder deploys a brand's catalog as a live MCP server returning the fields in the table above: title, attributes as options, variant-level price and stock. Our server connects from Claude as a custom connector on web and desktop, and from Claude Code with claude mcp add, documented at /docs/mcp. We have run this on Gray Porter and a group of US DTC brands: we made the product pages and checkout readable and finishable by an agent, then placed a real Muse order on the store. Agent orders land in GA4 as Direct; for MyMuse, GA4 showed ₹11.5K/mo of AI revenue against ₹81.2K/mo actual, 7× hidden. To see what a Claude agent gets back from your catalog today, book a demo.