On March 25, 2026, Anthropic’s Model Context Protocol crossed 97 million installs in sixteen months — making it the fastest-adopted AI infrastructure standard since the REST API. That number signals a structural shift in how AI systems access the web. The question for website owners is no longer “how do I get cited in AI answers?” It’s “how do I get discovered by AI agents?” The two problems require different protocols, and most sites are only solving the first one.
Here are the seven protocols that determine whether AI agents can find, read, and act on your site — ranked by current adoption and near-term impact.
1. Model Context Protocol (MCP) — 97 Million Installs
MCP is the de facto standard for connecting AI agents to external systems. Introduced by Anthropic in November 2024 and adopted by OpenAI, Google DeepMind, and Microsoft within 18 months, it gives AI agents a structured way to access tools, APIs, and data sources using a consistent interface. Over 7,600 public MCP servers exist as of January 2026, covering use cases from GitHub access to database queries to real-time web data.
For website owners, MCP matters because agentic AI systems — tools that autonomously complete multi-step tasks — query MCP endpoints when they exist rather than scraping HTML. A business that exposes its product catalog, booking system, or FAQ database through an MCP server becomes programmatically accessible to every AI agent that supports the protocol. Sites without MCP exposure are accessible only through HTML parsing, which is slower, less reliable, and produces lower-quality results for the agent.
Implementation path: Build or deploy an MCP server that exposes your core business data. The directory at mcp.so lists pre-built servers for common platforms (Shopify, Notion, HubSpot) that require minimal configuration. Custom implementations use the official TypeScript or Python SDK.
2. llms.txt — 0.13% Adoption, Highest Early-Mover Advantage
llms.txt is a plain-text file placed at yoursite.com/llms.txt that gives AI agents explicit instructions about your site structure and priority content. It functions like robots.txt but for large language models: it tells AI systems which pages matter, what your site covers, and how to navigate your content hierarchy. Anthropic’s Claude, Perplexity, and a growing number of agentic systems actively check llms.txt during crawls.
Adoption is at 0.13% of sites, per a June 2026 analysis of 1.2 million domains. That number represents a rare window: implementing llms.txt today puts your site in a group that AI agents are structurally prioritized to encounter. The file takes 30 minutes to write correctly. A well-formed llms.txt includes your site’s primary topics, a priority list of 10–15 key pages, and a brief description of what your business does — written for an AI reader, not a human one.
Implementation path: Create /llms.txt at your domain root. List your most important pages with one-line descriptions of what each covers. Add an llms-full.txt for extended context if needed. Test with Anthropic’s official llms.txt validator.
3. OpenAPI / Swagger Specifications — 40%+ Adoption Among Developer-Facing Sites
If your site has any API surface — even a simple contact form endpoint or a product search API — an OpenAPI specification makes it accessible to AI agents programmatically. OpenAPI is a machine-readable description of your API’s endpoints, parameters, and response formats. AI agents with tool-use capabilities (ChatGPT with plugins, Claude with tools, Perplexity with integrations) can read OpenAPI specs and use your API directly, without HTML scraping or manual integration work.
For B2B SaaS companies, an accurate OpenAPI spec means AI agents can query your product’s features, pricing, and availability in real time — and surface that information in responses to user queries without the agent needing to visit your site at all. This is the highest-value agent accessibility path for transactional sites.
Implementation path: Generate your OpenAPI spec using tools like Swagger Editor or Postman. Publish it at /openapi.json or /api-docs. Add a reference to it in your llms.txt to make it discoverable by AI agents checking your site structure.
4. Schema.org Structured Data — 40% Adoption (Basic), 12% Full Implementation
Schema markup is the most broadly adopted AI-accessibility protocol, but most implementations are incomplete. A July 2026 analysis of Google AI Overview selection factors found that structured data correlated with AI Overview selection at 0.92 — the highest single factor in the study. Despite this, fewer than 12% of sites implement schema beyond basic Article or Product markup.
The highest-impact schema types for AI agent discoverability are: Organization (with sameAs identifiers linking to LinkedIn, Wikipedia, and Wikidata), FAQPage (question-answer pairs that AI agents extract for direct answers), HowTo (step-by-step processes that agents use in task-completion responses), and Product with complete offers markup (pricing, availability, SKUs). Each type gives AI agents structured extraction targets rather than requiring them to parse prose.
Implementation path: Start with Article schema on editorial content, Organization on your homepage, and FAQPage on any page with question-and-answer content. Use Google’s Rich Results Test to validate. Add sameAs identifiers to your Organization markup pointing to LinkedIn and any industry directories where your business is listed.
5. Sitemaps with Priority and Changefreq Metadata — 65% Adoption
XML sitemaps are the most widely implemented discovery protocol, but most site owners don’t know that AI crawlers actively use them. Perplexity’s crawler and Google’s AI crawler both consume sitemaps to prioritize which pages to index for AI retrieval. A sitemap with accurate lastmod timestamps signals content freshness — a meaningful signal for platforms like Perplexity that weight recency heavily in citation selection.
The implementation gap is in maintenance: 65% of sites have a sitemap, but fewer than 30% update it automatically when content changes. A static sitemap that hasn’t been refreshed in six months tells AI crawlers your site is stale. For sites publishing new content regularly, an auto-generated sitemap (standard in most CMS platforms) is table stakes.
Implementation path: Verify your sitemap is auto-generated and includes lastmod dates. Submit it in Google Search Console and Bing Webmaster Tools. Reference it in robots.txt so AI crawlers can find it without additional configuration.
6. robots.txt with AI Crawler Directives — 30% Updated for AI Agents
The original robots.txt standard was designed for search engine crawlers, not AI agents — and most sites haven’t updated their robots.txt to account for the new crawler landscape. Major AI systems run distinct user agents: GPTBot (OpenAI), ClaudeBot (Anthropic), PerplexityBot (Perplexity), GoogleOther (Google AI), and Bingbot (Microsoft AI). Sites that block all bots (using Disallow: / for all user agents) are blocking AI indexing without knowing it.
Conversely, sites that explicitly allow these agents signal accessibility. The update takes five minutes and involves adding specific Allow directives for each AI crawler’s user agent. Check your current robots.txt — if it has a catch-all Disallow directive, you may be invisible to AI systems.
Implementation path: Add explicit Allow directives for GPTBot, ClaudeBot, PerplexityBot, and GoogleOther. If you want AI systems to access your content but restrict certain directories (admin panels, staging paths), scope your Disallow rules to those paths only.
7. Agent.json — Emerging Standard, Under 1% Adoption
Agent.json is an emerging protocol — similar in concept to llms.txt but more structured — that defines how AI agents should interact with a website. It specifies available actions (book, search, purchase, contact), required authentication, rate limits, and preferred interaction patterns. No single standard has fully emerged yet: variants include agent.json (proposed by several AI labs), ai-plugin.json (OpenAI’s original format), and a growing number of platform-specific formats.
Adoption is under 1%, but the trajectory is significant. OpenAI’s plugin directory, which used ai-plugin.json, processed over 4 billion agent queries in Q1 2026 before shifting to the newer GPT Actions format. Sites that implement some form of agent interaction manifest — even a minimal one — are position themselves for the next wave of agentic traffic before it arrives.
Implementation path: At minimum, create a JSON file at /.well-known/agent.json describing your site’s available actions and contact information. Watch the emerging standards from Anthropic’s agentic web working group for the format that consolidates.
Implementation Priority: Where to Start
If you’re starting from zero, the order is: llms.txt first (30 minutes, immediate effect on AI agent crawls), schema markup second (highest measurable impact on current AI citation rates), robots.txt update third (five minutes, removes invisible blockers). MCP and OpenAPI implementations require more development time but unlock the highest-value agent interactions — autonomous AI systems taking action on your site rather than just reading it.
The protocols with the widest adoption aren’t always the ones with the highest impact for your specific site. A B2B SaaS company gets more from an OpenAPI spec than from sitemap optimization. A local service business gets more from Organization schema with sameAs identifiers than from MCP. The right stack depends on the type of AI interaction your business needs to win.
To see which of these protocols your site currently implements — and which gaps are costing you AI citations and agent discoverability — run the free audit at ai-visibility.llmagnet.com. It checks schema, llms.txt, robots.txt AI directives, sitemap freshness, and structured data completeness in 30 seconds.