AI Is Rewriting Search. Here's How Your Website Writes Back.

On May 19, Google's CEO Sundar Pichai announced that AI Mode has crossed 1 billion monthly active users. AI Overviews – the AI-generated summaries that now appear above organic results in Google Search – have reached 2.5 billion. Those aren't projections. They're the current scale of the systems now sitting between your prospective clients and the web.

We implemented WebMCP on fieldway.org this week. Here's why.

What's happening to search traffic right now

The traffic impact of AI search isn't theoretical. It's documented.

Seer Interactive analyzed 25 million impressions and found that 93% of AI Mode sessions end without an outbound click, with AI Overviews showing an 83% zero-click rate in the same dataset. SISTRIX tracked position-1 CTR on queries where AI features appear and documented a drop from 27% to as low as 11% by March 2026. Kevin Indig tracked 450 million B2B SaaS impressions and measured a 56.6% click decline since the AI Overview rollout intensified.

That's the downside. Here's the other half.

Brands cited inside AI Overviews earn 35% more organic clicks and 91% more paid clicks compared to non-cited competitors on the same queries. (Digital Applied, March 2026.) The traffic isn't disappearing. It's concentrating around the sources AI systems choose to cite.

The question isn't whether AI search has already changed how websites attract clients. It has. The question is whether your website is in the cited group or the uncited group.

Why traditional SEO doesn't answer that question anymore

Organic ranking used to be the primary gate for search visibility. That relationship is eroding.

Ahrefs analyzed 16.975 million URLs cited by AI Overviews. In mid-2025, 76% of citations came from pages ranking in Google's top 10. By February 2026, that figure had dropped to 38%. The remaining 62% of AI citations came from pages ranking below position 10 – including 31% from beyond position 100.

So what actually predicts AI citation?

Profound analyzed 250 million AI responses. Traffic explains almost nothing about citation behavior (r² = 0.05). Backlinks explain even less (r² = 0.038). Entity richness – how clearly and explicitly your content identifies its subject matter, sources, and context – produced a 267% citation lift. Content closely matched to the structure of the query being asked: a 7.3x citation multiplier at high cosine similarity.

The translation is direct: AI systems cite websites that are legible to them. Content that makes its subject matter, authority signals, and structure explicit gets cited. Content that assumes a reader who navigates and infers doesn't.

Most professional services websites are built for human readers who navigate and infer. The AI systems now distributing referrals don't work that way.

What WebMCP is

WebMCP is a browser protocol that lets websites expose their capabilities and content directly to AI agents, in a form those agents can query without inference or guessing.

Announced by the Chrome team on February 10, 2026, and co-developed by engineers at Google and Microsoft, WebMCP introduces a new browser API – navigator.modelContext – that websites can use to publish a structured declaration of their available actions, content, and parameters. Chrome reads that declaration and makes it available to AI agents browsing the site.

Lawrence Hitches, who wrote the most detailed practitioner explainer on the protocol, draws the analogy that captures it: WebMCP is to AI agents what XML sitemaps were to search engines – a structured, machine-readable declaration of what a site does and how to use it. (The Web Standard for AI Agents, May 5, 2026)

Websites also publish a .well-known/webmcp manifest – analogous to robots.txt or sitemap.xml – so AI agents know what capabilities and content a site exposes before they navigate it. Two implementation paths exist. The Declarative API adds HTML attributes to existing forms and Chrome builds the schema automatically. The Imperative API handles more complex interactions via JavaScript's navigator.modelContext.registerTool(). Neither requires a back-end overhaul. Both run client-side within the browser.

The distinction from traditional SEO is worth naming directly. A search crawler reads your HTML, indexes your text, and infers what your page is about. WebMCP gives a site a way to answer that question without inference: here's what this site covers, here's what services are available, here's how to access the content you're looking for. Instead of an AI agent guessing from your copy, the site describes itself in terms the agent can work with directly.

WebMCP is currently in Chrome Canary (version 146+) under the Early Preview Program, transitioning from W3C community incubation toward formal draft. A stable, widely-implemented standard is a late 2026 to mid-2027 timeline at the earliest.

What we built – and why we built it now

I want to be precise about what I am and am not here: I'm not an AI protocol engineer. I'm a consultant who runs an advisory practice and wanted to understand whether this approach was actually practical to implement on a site like this one – small, practitioner-led, without a dedicated engineering team.

The answer is that it's buildable. Not trivial, but not out of reach.

I released claude-skill-webmcp on GitHub under MIT license. It's a Claude Code skill that walks through the audit and implementation workflow: scanning a codebase for WebMCP candidates, scoring them, and building the declarative or imperative implementation along with the .well-known/webmcp manifest. The Next.js App Router patterns are included because that's what fieldway.org runs. The repository is at https://github.com/mstublefield/claude-skill-webmcp.

I released it MIT because this space is still forming. I'd rather practitioners explore and learn from it than wait for a polished product to hand them the answer. The point of open-sourcing early work is to be useful while the work is happening, not after the conclusions are already obvious.

The immediate catalyst was a client question. We were in a consulting engagement when the Google I/O announcements landed. The client's concern was direct: if AI systems are becoming the primary interface through which prospective clients look for expertise, how does their website remain legible in that environment? That's the same question we were already working through for fieldway.org. WebMCP is the practitioner answer we're testing now.

What this means for your advisory website right now

WebMCP is worth implementing if you're a practitioner building your own site and want to stay at the leading edge. For most consulting and advisory websites, the more immediate lever is the content legibility problem that WebMCP formalizes – and that you can start solving without waiting for a stable standard.

Profound's analysis of 250 million AI responses points at what actually moves the needle for AI citation now:

• Entity richness – Does your content explicitly identify the specific subjects it addresses, with named organizations, people, concepts, and sources? AI systems need declared signals, not inferred ones.
• Content specificity – Can a specific question be answered in 40 to 60 words from your content? AI models extract passages, not pages. Answers buried in general context don't get cited.
• Structured formats – Do your pages use clear definitions, step-by-step structures, or comparison formats matched to the queries they're targeting?
• Freshness signals – Does your content carry visible publication dates? Multiple AI citation studies have documented a preference for dated content over undated.
• Source citations – Name your sources explicitly. AI systems evaluate evidential basis alongside content — naming the research, data, or authority behind your claims is a positive signal in citation behavior. Profound's analysis confirms that content clarity (including citation structure) is one of the strongest predictors of AI citation lift.

These aren't WebMCP-specific. They're the content fundamentals that WebMCP formalizes into a machine-readable declaration. Getting them right makes your site more legible to AI systems today, and positions it to take fuller advantage of WebMCP as the protocol matures.

I want to be careful about what I'm claiming here. Nobody knows yet exactly how AI search will settle – how citation patterns will evolve across different site types, whether WebMCP becomes a standard expectation or gets absorbed into some other approach, what traffic implications look like at scale. These are genuinely open questions.

What isn't open: the direction. AI systems are increasingly in the middle of the conversation between people and the web. A website optimized only to talk to a 2015 search crawler is optimized for a conversation partner that's becoming less central to how people find information.

We built this because a client had a real concern and we wanted to understand it from the inside rather than read about it from the outside. We're posting about it now, while the news is fresh, because being early to understand something has more value than being early to write about it after everyone else already has.

AI is rewriting search. WebMCP is how your website writes back.

If you want to talk through what this means for your practice's inbound strategy, email matthew@fieldway.org.

---

Related: Consultant AI Guides All Cover the Same Five Things. Here's the Sixth. | The $50K Alternative to McKinsey: Managed Intelligence for Advisors