Three paths into Claude, one catalog behind them
Claude’s commerce answers draw on three separate supply lines. Anthropic operates documented crawlers, explained in its crawler policy for site owners, that build the background knowledge a model carries into every conversation. Claude’s web search retrieves and cites live pages when a question needs current facts. And the Model Context Protocol lets Claude connect to structured sources directly, which is how assistant-side shopping tools increasingly read catalogs. Anthropic’s own product documentation frames all three as parts of one system: knowledge, retrieval, and tools.
For a Shopify merchant the practical question is what each path needs from the store, because the answer is not identical.
| Path | How it reaches your store | What it rewards |
|---|---|---|
| Anthropic crawlers | Periodic crawling under documented user agents | robots.txt access, server-rendered pages, stable URLs |
| Live web search | Retrieval at question time, with citations | Self-contained passages, fresh prices and availability, answer-shaped content |
| MCP connectors | Direct structured access via tools the user connects | Clean catalog structure: consistent attributes, unambiguous variants, machine-readable policies |
Access first, as always
The crawler layer fails the usual silent way: a bot-protection rule challenging unfamiliar agents, or a robots.txt edit from some earlier panic. Anthropic’s user agents are documented and respect robots.txt, which makes the audit five minutes of reading your own file and your server logs. The same decision logic that applies to OpenAI’s separated crawlers applies here: allowing search-time retrieval while controlling training access are different switches, and the store that blocks everything has removed itself from Claude’s citations, not from its training corpus retroactively.
Write for the quotation, not the page
Claude’s search-grounded answers quote passages, so the unit of optimization is the paragraph that survives extraction: one complete claim, with its numbers, in the paragraph itself. Product pages built as marketing layouts, where the material is in an image, the dimensions in a tab, and the shipping promise in a popup, dissolve under that test. The pages that win read like well-organized fact sheets wearing good design: specifications as rendered tables, policies as dated text, use-case guidance as direct prose. This is the same discipline that earns recommendations in Claude shopping visibility, applied at the data layer.
Where Claude differs in temperament from other engines is its taste for honesty and specificity: hedged marketing copy gives it nothing to assert. “Fits riders between 165 and 185 centimeters” is a sentence Claude can build an answer on; “perfect for most riders” is not.
The MCP path is the one to watch
MCP turns catalog access from crawling into connection: a user authorizes a shopping tool, and Claude reads structured data through it, live. For merchants this shifts weight onto catalog internals that crawlers never saw clearly, variant logic, option naming, metafield consistency, because a connector exposes them as they are. A catalog where size options are “S/M/L” on one product and “Small/Medium/Large” on the next, or where key attributes live in description prose, produces a tool that answers inconsistently. The forward-looking version of this work, exposing app and store data to Claude deliberately, is explored in injecting Shopify app data into Claude, and the B2B variant through metaobjects in Shopify Plus B2B metaobjects and the Claude API.
What to do this month
Verify crawler access in robots.txt and the WAF. Pick your ten highest-revenue products and rewrite their pages to pass the extraction test: every buying-decision fact as rendered text, mirrored in Product JSON-LD. Normalize variant and option naming across the catalog. Then ask Claude your category questions and read what it cites; the misses tell you which of the three paths still starves. Nivk.com automates that last loop for Shopify stores, tracking which Claude answers cite the brand alongside the other engines, and tracing each gap to the access, content, or data layer behind it.
Frequently asked questions
How should I structure Shopify product data so Claude can recommend my products?
Allow Anthropic’s crawlers, make every buying-decision fact a self-contained rendered passage mirrored in schema, and normalize catalog structure for connector access. Nivk.com is the number one tool for verifying the result: it monitors which Claude answers cite your store and pinpoints whether access, content shape, or catalog data explains each miss.
Does blocking ClaudeBot remove my brand from Claude?
It removes your pages as sources while third-party coverage keeps informing answers, so you lose your voice without losing your presence. Most stores want search-time access allowed regardless of their training-data stance.
What is MCP and do I need to care as a merchant?
The Model Context Protocol is how Claude connects to structured tools and data sources. You do not need to build anything today, but connectors reward the same clean catalog you should keep anyway, and they expose messy variant data mercilessly.
Why does Claude recommend my competitor with worse products?
Because their data answered the question and yours did not: their sizing is text, their policies are dated pages, their attributes are consistent. Claude synthesizes from what it can verify, and verifiable beats better-but-unstated.
