Agentic Web · WebMCP · GEO
Respuesta directa
An agentic website is one that AI agents can read (llms.txt, markdown versions), navigate (a well-formed accessibility tree) and operate (WebMCP tools: book, quote, contact). At Dribba we audit and implement all three layers — and we prove it at home: dribba.com is one of the first websites in Spain with WebMCP in production.
The shift
In May 2026, Google announced WebMCP at I/O and shipped it as an origin trial in Chrome 149, with Gemini built into the browser as the first agent able to use the tools a website exposes. OpenAI is pushing in the same direction with its browser, and ChatGPT already completes tasks on third-party websites. This isn't a prediction: a growing share of web traffic is already agents acting on behalf of people.
The pattern repeats with every channel shift: websites that adapted early to mobile or to SEO captured an advantage that lasted years. The difference this time is that the new "visitor" doesn't forgive ambiguity: if an agent can't read your catalogue, can't cite your prices or fails to fill in your form, it doesn't retry — it recommends another website. And its user will never know you existed.
The good news: the bar is incredibly low. Almost no Spanish website has llms.txt today, fewer still serve markdown versions, and websites with WebMCP in production can be counted on one hand. Being first in your sector doesn't require rebuilding anything — it requires adding three layers on top of what you already have.
Anatomy
Each layer works on its own and adds value on its own. Together they turn your website from "content agents try to interpret" into "a service agents know how to use".
01
llms.txt · markdown mirrors · Content-Signal · schema.org
An agent reading your HTML burns 90% of its tokens on markup before reaching a single fact. We publish an llms.txt index of your site, markdown versions of every page (10–20× fewer tokens, via content negotiation) and explicit usage signals for AI crawlers. The result: ChatGPT, Perplexity and Gemini read you fully, fast and without noise.
02
accessibility tree · HTML semantics · CLS · focusable states
Agents navigate your website through the accessibility tree, not the pixels. A mislabelled modal, an aria-hidden element with focusable children or a layout that shifts while loading make the agent click the wrong thing — or give up. We audit and fix the navigation layer with the same checks Lighthouse's Agentic Browsing audit uses.
03
WebMCP · document.modelContext · tools with JSON Schema
The layer almost nobody has: with WebMCP (the standard driven by Google and Microsoft, in origin trial since Chrome 149), your website exposes actions as tools the agent invokes directly — request a quote, book an appointment, check availability, start a purchase. No scraping, no interpretation errors: your business defines what the agent can do and under which conditions.
Live case
Before offering this service we implemented it fully on dribba.com. An agent with WebMCP can browse our services, request a ballpark quote, get the booking link and leave us a lead — without touching a single form. Try it yourself: request this very page as markdown with Accept: text/markdown.
4 WebMCP tools in production
list_services, get_project_estimate, book_meeting and submit_contact_request — an agent can request a quote or leave a lead without touching a form.
llms.txt + llms-full.txt
Complete site index for AI systems, advertised via Link headers on every response.
Markdown mirrors of every page
Any agent sending Accept: text/markdown gets the clean page — no nav, no footer — with a token estimate included.
Verified accessibility tree
Zero errors in Lighthouse's Agent Accessibility audit: dialogs with inert, no hidden focusable elements, CLS under control.
Content-Signal and schema.org
Explicit policy for AI crawlers (search=yes, ai-input=yes, ai-train=no) and an Organization + FAQ graph in JSON-LD across the site.
How we work
01
1 wkWe measure how agents see your website today: AI crawler access, readability (tokens per page), accessibility tree, citability in ChatGPT/Perplexity, and action opportunities (what an agent should be able to do on your site but can't).
02
1–2 wksWe implement llms.txt, markdown mirrors, schema.org and crawler signals. We fix the accessibility-tree blockers and the CLS issues that disorient agents.
03
2–4 wksWe define with you which actions your website exposes (bookings, quotes, leads, catalogue queries), and implement them as WebMCP tools with validation and limits, reusing your existing APIs.
04
OngoingWe monitor agent traffic, citations in AI answers and agent-originated conversions. The standard moves fast — we keep your implementation current with every Chrome release.
Who it's for
Agents already compare prices and complete purchases. If your catalogue isn't readable and your checkout isn't operable, the recommendation goes to the competitor whose site is.
Clinics, restaurants, workshops, consultancies. "Book me a table on Thursday" only works on websites where the agent can check availability and act. That's the difference between appearing in the answer or not.
B2B buying cycles increasingly start with a question to ChatGPT or Perplexity. Being citable (GEO) and demonstrable (agent-operable demos) puts you on the shortlist before the shortlist exists.
If AI systems cite your content with a link, you gain authority and referral traffic. If they can't read you properly, they cite you wrong — or cite someone else. You set the policy (ai-train=no included).
Guide
An agentic website is a website prepared so that AI agents can consume and operate it with the same reliability a browser renders it with for a person. The term groups three concrete technical capabilities: readability for AI systems (llms.txt files, markdown versions of pages, schema.org structured data), navigability through the accessibility tree (the semantic representation agents use instead of pixels), and operability via WebMCP, the standard that lets a page expose invocable actions — book, quote, buy — with an explicit contract of parameters and validation.
The "why now" has a date: at Google I/O in May 2026 WebMCP was announced and Chrome 149 shipped it as an origin trial, with Gemini built into the browser as the first agent able to discover and use the tools a website registers. In parallel, ChatGPT, Perplexity and Claude already answer a significant share of commercial searches citing a handful of sources. Being in that handful — being citable — and being operable when the agent wants to act are the two new competitive positions. Whoever occupies them first in their sector inherits the momentum early SEO and early mobile adopters once inherited.
Done right, the implementation is additive: it requires no redesign and no migration. The reading layer is built as parallel infrastructure (an llms.txt index, an HTML→markdown converter served via content negotiation, Link headers so crawlers discover the clean versions). The navigation layer is an accessibility audit with an agentic focus: correct ARIA roles, dialogs that don't trap focus, stable layout during load. And the action layer reuses the APIs that already exist behind your website's forms, wrapping them in WebMCP tools with JSON Schema, rate limiting and human confirmation for sensitive actions. A mid-size website completes all three layers in four to seven weeks.
The prioritisation criterion is the usual one: where the money is. An e-commerce starts with a readable catalogue and an operable checkout; a booking business with queryable availability and executable bookings; a B2B SaaS with citability in the AI answers that open its sales cycle. And any website with valuable content should set its AI-usage policy (Content-Signal, robots.txt for AI crawlers) before others set it by default. The agentic web isn't a bet on the future: it's the 2026 version of "having a well-built website".
An agentic website is a website prepared so that AI agents (Gemini in Chrome, ChatGPT, Perplexity, Claude) can read it efficiently, navigate it without errors and execute actions on it — book, quote, buy, contact. It's built in three layers: readability (llms.txt, markdown versions, schema.org), navigability (a well-formed accessibility tree) and operability (WebMCP tools the agent can invoke).
WebMCP is the adaptation of the Model Context Protocol to the web: a standard proposed by Google and Microsoft (W3C Web Model Context) that lets a page register tools in JavaScript — with a name, description and JSON Schema — that browser agents can invoke directly. It's available as an origin trial since Chrome 149 (May 2026), with Gemini in Chrome as the first consuming agent. Instead of the agent visually interpreting your interface, your website tells it what it can do and how.
No: it extends it. Traditional SEO remains the foundation — Google is still the main discovery source. The agentic layer (sometimes called GEO, Generative Engine Optimization, or AEO) optimises for the growing channel: AI answers with citations, assistants that compare options and agents that execute tasks. Websites that rank well in classic SEO start with an advantage; the agentic layer decides who converts it into visibility inside AI answers.
Almost never. The three layers are added on top of your current website: llms.txt and markdown mirrors are new infrastructure that doesn't touch your design; accessibility-tree fixes are surgical (and improve your real accessibility and SEO along the way); and WebMCP tools reuse the APIs you already have behind your forms. What you do need is server-rendered content — if your site is 100% client-side rendered, that's the first fix.
It depends on the size of the site and how many actions you want to expose. The agentic audit is delivered in one week and includes a prioritised plan with a fixed estimate. A typical full implementation — reading, navigation and the first WebMCP tools — runs between 4 and 7 weeks. Request the audit and you'll have concrete numbers before committing to anything.
That's precisely WebMCP's advantage over scraping: you define the contract. Every tool declares its parameters, validates its inputs and passes through the same limits as your forms (rate limiting, origin validation, anti-spam). Sensitive actions — submitting a form, confirming a purchase — are flagged so the agent asks the user for explicit confirmation before executing them. An agent without WebMCP will try to do the same by interpreting your UI, but with no contract and no limits: more risk, not less.
It shows in your server logs: user agents like GPTBot, ClaudeBot, PerplexityBot, Google-Extended or Bytespider already crawl most commercial websites. In the agentic audit we analyse your AI crawler traffic over recent months, what they can see today (and what they can't) and whether your robots.txt is blocking them without you knowing — a surprisingly common mistake.
Because we don't just talk about it: we do it. dribba.com is one of the first websites in Spain with WebMCP in production — an agent can browse our services, request a ballpark quote and leave us a lead without touching a form. We've spent years building AI agents for clients (RAG, function calling, process automation) and now we apply that same engineering to the other side: websites agents can operate. Two sides of the same problem, same senior team.
The other side
We've spent years building AI agents for companies: RAG on private data, function calling, process automation with ERPs and CRMs. It's the same problem seen from the other side — and the same team.
Start with the agentic audit: within a week you'll know how AI agents see you today, what you're missing and what it costs to fix. With a prioritised plan and a fixed estimate.