A new spec lands. It has a W3C Community Group Draft. It is co-authored by Google and Microsoft. It is running in Chrome 146 beta. Within days the industry has declared it a fundamental shift in search — the next schema markup, the thing you need to implement right now before everyone else does.
The pattern is familiar. The question worth asking before reaching for the implementation guide is what the thing actually does — and whether what it does is genuinely new.
What the declarative API actually is
The declarative API annotates HTML forms with machine-readable descriptions so an AI agent can understand what the form does and how to interact with it. You add attributes — a tool name, a description, explanations of each input — and the agent reads them rather than guessing from labels and IDs.
This is structurally what browser autofill has been doing since the early 2000s. Autofill infers field purpose from labels, name attributes and element IDs. WebMCP makes that inference explicit rather than guessed. The article that introduced most people to WebMCP draws the comparison to schema markup — and it is fair. You are annotating what already exists so a new class of reader can understand it.
Lighthouse 13.3.0 shipped the Agentic Browsing category into the default config — no longer experimental, no flag required. The category checks WebMCP implementation, accessibility tree quality, layout stability and llms.txt presence. Within two weeks PageSpeed Insights inherited it. Since Screaming Frog pulls Lighthouse data via the PSI API, the agentic browsing audit is now surfacing in standard crawl reports without any additional setup.
The spec and the auditing tool were shipped by the same team. The "independent validation" Lighthouse appears to provide is Google validating Google's own standard. Worth noting.
Lighthouse 13.3.0 — May 7 2026Where it does get genuinely interesting
The imperative API is different. Rather than annotating existing forms it allows you to register tools directly in JavaScript — defining what your site can do, what information is needed and what the action produces. Tools can change based on state. The agent moves through a structured flow without having to guess what each step means.
The toolautosubmit attribute allows an agent to complete and
submit a form without an explicit user confirmation step. For low-stakes
actions that is a genuine convenience. For anything involving commitment —
money, time, preference — the question of what "on behalf of" actually means
is worth examining.
There is a less discussed implication. The same structured annotation that
makes legitimate agent interaction more reliable makes malicious agent
interaction more reliable too. Credential stuffing becomes more efficient.
Automated payment fraud through checkout flows with toolautosubmit
enabled becomes more structured. The spec notes it should only be used for
low-risk reversible actions — but that is guidance not enforcement. The
coverage has been almost entirely enthusiastic without this being mentioned once.
The holiday problem
The article promoting WebMCP argues trust adoption will follow the same curve as online shopping: scepticism, reluctant adoption, dependency. The comparison does not hold. Entering a card number is a single confirmable step. Booking a holiday involves dates, budgets, travel companions, flexibility and trade-offs — decisions that are genuinely yours to make with real consequences if made incorrectly. The trust gradient between "autofill my postcode" and "book four flights without asking" is not linear.
The llms.txt question
While WebMCP was being declared a game changer, the same Lighthouse update added llms.txt validation. Google's public position — confirmed by Gary Illyes and John Mueller — is that Google does not support llms.txt and has no plans to. Mueller compared it to the keywords meta tag. Lighthouse is a Google tool. It now checks whether your llms.txt follows recommendations at scale, automatically, in every standard crawl report.
Dixon Jones published a piece titled "A Cynical Read of Google's AI Optimisation Guide" — a careful, specific dissection of Google's recently published guidance on optimising for AI features. The TLDR was sharp: "a well-written, polite and strategically coherent document — for Google. It tells publishers to keep doing the work, trust the system, disregard competing standards and not ask hard questions about the economic shift underway."
On llms.txt specifically, Jones noted that Google's dismissal is suspiciously absolute. llms.txt is a community-led standard developed partly to give publishers clearer agency over how AI systems consume their content. A standard Google did not create and does not control is precisely the kind of thing Google would prefer not to legitimise. The phrasing "this doesn't mean the file is treated in a special way" tells you what Google does with it. It says nothing about whether publishers are right to want it.
The article disappeared from LinkedIn shortly after publication. Whether that was the author's choice or something else is not known. The argument stands regardless.
Dixon Jones — A Cynical Read of Google's AI Optimisation Guide, May 20 2026
While the WebMCP spec was being shipped and audited Google announced
Business Agent for leads — conversational AI built directly into Search Ads.
The connection the coverage missed: WebMCP is the mechanism that makes the
final step possible. The agent qualifies the user inside the ad, reads your
website content via WebMCP's structured annotations and at the natural end
of the conversation: "Would you like me to go ahead and apply on your behalf?"
The toolautosubmit attribute completes the form. Nobody clicks
anything. There is no click to count.
CPC as a pricing model assumes a click occurred. If the conversion happens inside the conversational interface it does not. The replacement models — cost per conversation, cost per qualified lead, cost per outcome — are all defined scored and reported by the platform receiving the payment. CPC was transparent in a way none of these are. One click, one cost, auditable in your own analytics. What replaces it is a number Google generates using a model Google owns.
The impression share inventory that brands spent years building — pushed toward 100% by Google's automated recommendations and the agency ecosystem built around its partner programme — now monetises the same interface. AI Overviews sit at the top. Conversational ads sit below them. The organic result that impression share was supposedly protecting is further down the page than it has ever been. The architecture is complete. The brands that paid to build the inventory are now being asked to invest in the next layer of the same stack.
See also: The Inventory Was Always the Point and Who Else To Ask Now Jeeves Has GoneWhat this has to do with SEO
WebMCP helps an AI agent interact with your site once it has already arrived.
It does nothing about discovery. An AI agent still needs to find your site
before it can interact with it. The content that establishes relevance,
the authority signals that establish trust, the technical fundamentals that
allow crawlers to understand what a site is about — none of that changes
because your contact form has a tooldescription attribute.
WebMCP is a conversion layer. Discovery still happens upstream. That part has not changed.
The architecture becomes visible
Taken individually each development looks incremental. Taken together they describe something more coherent: a complete end-to-end infrastructure for agentic web interaction built and controlled by Google.
Discovery — llms.txt. Interaction — WebMCP. Conversion — conversational ads
with toolautosubmit. Measurement — predictive attribution.
Each piece shipped separately, each announced as its own story, each
validated by Google's own tooling. The complete absence of MCP — the
Anthropic-originated protocol that has rapidly become the de facto standard
for agentic tool use — from Google's AI optimisation guide is equally telling.
As Jones put it: a guide that positions itself as definitive whilst ignoring
the dominant interoperability layer outside Google's ecosystem is not an
oversight. It is a position.
The original argument was: WebMCP is better autofill with a W3C draft. The fundamentals haven't changed. Discovery still happens upstream.
The revised argument is: all of that remains true and the speed at which the surrounding picture filled in is itself the point. Within weeks of the original post, Lighthouse moved the category to default, PSI inherited it, Screaming Frog started surfacing the data, a credible industry voice published a piece calling the whole strategy out — and that piece disappeared. The waters muddied faster than the ink dried.
The question the original post ended on — whether WebMCP is worth implementing while the barrier is low — has a clearer answer now. Yes, because the audit is already running in your standard crawl reports whether you asked for it or not. Not because the fundamentals changed. Because the instrument changed, and the instrument is now showing it to everyone.
May 2026Implement WebMCP if your site has forms that AI agents are likely to interact with. Understand the imperative API if you are building multi-step flows. Keep your own independent measurement alongside whatever Google's dashboards report. And read Google's guidance with one eye on who benefits from each recommendation.
The latest acronym still doesn't mean sustainable growth or success. But the architecture being built around it is worth understanding — because it is being built regardless of whether you are paying attention.