Stop being scraped — start being queried
APIs and feeds beat scraping
When an agent has to scrape an HTML page to extract a price or check stock, two things go wrong. The agent gets the answer slowly, and it gets it wrong some non-trivial fraction of the time. A clean API or product feed inverts both problems. A small ecommerce site that exposes a product feed at /products.json or supports the existing Schema.org Product offers vocabulary correctly is dramatically more usable to an agent than one that does not. Larger sites publish OpenAPI specifications describing inventory, pricing, availability, and booking endpoints. Either way, agents prefer the structured route, and merchants who provide it get prioritised in agentic answers.
Model Context Protocol (MCP) is the common standard
Anthropic's Model Context Protocol, released as an open standard in late 2024, has become the common standard for AI agents to talk to tools and data sources. (For a plain-language introduction see What is MCP? and the best MCP servers of 2026.) An MCP server exposes a set of tools and resources — search, read, query, write — over a simple JSON-RPC interface. Major AI assistants now include MCP clients out of the box. For a website, the practical implication is that exposing an MCP server gives any compatible assistant — ChatGPT, Claude, Copilot, Gemini, and a growing list of agent platforms — first-class access to the site's data without needing per-vendor integrations. Microsoft's NLWeb project, discussed below, makes this even more concrete.
A2A, AP2, and ACP are the commerce protocols
Three further protocols matter for any business that sells. Google's Agent-to-Agent (A2A) Protocol standardises how two agents talk to each other directly, useful when both consumer and merchant operate their own. Google's Agent Payments Protocol (AP2), launched September 2025 and backed by Mastercard, PayPal, American Express, Adobe, and Alibaba, lets agents make verifiable purchases through cryptographically signed mandates that link intent, cart, and payment. OpenAI's Agentic Commerce Protocol (ACP), co-developed with Stripe, allows users to complete purchases inside ChatGPT directly. None of these is yet universal, but adopting one or more — usually through an existing payments or commerce provider — gets a merchant onto the rails the major AI surfaces are building on.
NLWeb makes a website conversational
Microsoft's NLWeb project, led by Schema.org and RSS creator R.V. Guha, lets a site owner stand up a natural-language query interface on top of existing Schema.org markup, RSS feeds, or product databases. Out of the box every NLWeb instance exposes itself as an MCP server, so any compatible assistant can query the site directly. Early adopters include Shopify, Allrecipes, and Tripadvisor. The practical attraction for non-developers is that NLWeb does not require rewriting a site; it builds on signals already covered in earlier lessons (schema markup, feeds, sitemaps). For most small business sites in 2026, the right strategy is to wait until a hosting provider or platform integrates NLWeb or a comparable bridge — but to make sure the schema, feed, and sitemap groundwork is in place so the integration is trivial when it arrives.
UCP and WebMCP: the newest additions to the stack
Two further protocols are worth tracking. The Universal Commerce Protocol (UCP), referenced in Google's May 2026 generative-AI guidance, is an emerging standard meant to let search agents handle a wider set of commerce actions (comparison, configuration, ordering) without the brand integrating separately with every consumer-agent vendor. WebMCP, in Chrome early-preview as of 2026, is a proposed web standard that lets a site declare an MCP server in the browser itself, so any compatible agent on the user's device can call its tools without a separate hosting step. Neither is universal yet. Both are deliberate attempts to lower the integration cost of being agent-callable. For most sites the right posture is the same as for the other protocols in this lesson: keep the underlying data clean, watch which protocols the site's platform adopts, and avoid bespoke implementations until a clear winner emerges.
Could an AI agent ask your website 'is this product in stock today, in size medium, blue, for delivery to Manchester?' and get a clean machine answer? If not, what stands between the agent and that answer?