RafalAI brief · 2 September 2026

Search is one job. A browser agent is another.

Google Search Central and Chrome are both publishing AI guidance. They are not describing the same lever.

Written from official Google and Chrome pages RafalAI already watches. Not a client case study. Not an agency pitch.

This file exists so a person and a browser agent can open the same document. ChatGPT can explain it and put it on the First Look screen. A person still decides what happens next.

What we actually watch

RafalAI runs a recurring scan of public developer sources: Google Search Central, Chrome’s agent and WebMCP docs, and related measurement posts. The output of that work is internal — dated notes, source catalogs, and weekly synthesis. This brief is the public write-up of what those sources currently say.

It is not a client win. It is not a ranking report. It is the difference between two jobs that keep getting sold as one product.

What Google Search said

On 15 May 2026, Google Search Central published a guide for generative AI features in Search — AI Overviews and AI Mode. The through-line is old Search advice in a new surface: useful, non-commodity pages; crawlable HTML; people-first content. The new features still retrieve from the Search index.

On 15 June 2026, Google added the clarification people kept asking for. llms.txt and similar AI-only files do not help or hurt Google Search. They do not change visibility, rankings, AI Overviews, or AI Mode. Google said it is fine to keep those files for other systems. Search itself does not use them.

The same week, Search Console started rolling out dedicated Search Generative AI performance reports: impressions in AI Overviews, AI Mode, and generative Discover, with pages, countries, devices, and dates. That is first-party measurement. It is not a third-party “GEO score.”

Google’s AI features page is equally plain. To show up as a supporting link in AI Overviews or AI Mode, a page has to be indexed and eligible for a snippet. There is no extra AI-only markup requirement.

Through late August 2026, Search Central’s changelog kept moving — favicons, preferred sources, site reputation, social and video platform properties in Search Console. None of that reversed the June line on llms.txt.

What Chrome said about WebMCP

Chrome is documenting a different problem. A browser agent that guesses from the DOM is slow and wrong. WebMCP is a proposed way for a page to declare tools: a name, a description, a schema, and code that runs on the page a person can see.

Chrome’s WebMCP overview, last updated 7 August 2026, calls this a reliability layer. Tools can be declared in HTML on a form, or registered in JavaScript with document.modelContext.registerTool. The origin trial is listed from Chrome 149. Local testing uses a Chrome flag.

Chrome’s own examples are support tickets, travel booking, and checkout. Those are retailer-shaped demos. The same API can declare a smaller job: explain this page, open a document, run a look, fill a note. The page still decides what the tool is allowed to do. A send that changes the world can stay on a human button.

WebMCP is not a Google Search ranking signal. Chrome does not claim that. Search Central does not claim that. If someone sells you “add WebMCP to rank in AI Overviews,” they are mixing the two jobs.

Why the two jobs get mashed together

Both stories use the word AI. Both mention agents. Both mention files and markup. That is enough for a vendor deck.

The work is not the same.

Search needs a crawlable page, useful public content, and measurement you can defend.
A browser agent needs labels, a stable page, and — if you offer them — tools with honest descriptions and a human gate on side effects.
A context file can help a system that chooses to read it. It does not create ranking rights, and it does not create permission to act.

Google already told people to ignore AEO/GEO hacks aimed at Search: chunking pages for machines, inventing AI-only files, chasing inauthentic mentions. Chrome is telling people to declare tools when an agent should stop guessing. Those can both be true at once.

What a first look can prove

A first look on a URL can read the published HTML, title, meta, headings, form names, whether tools were declared, /llms.txt, and robots.txt. That is a real fetch. It is not a full audit, a rank tracker, or a rewrite of the CMS.

If llms.txt is present, that is a fact about the site. It is not a Search win. If WebMCP tools are present, that is a fact about how an agent can act on that page. It is not a ranking win. If robots.txt blocks a training crawler and allows a search crawler, that is a policy choice, not a GEO tactic.

The useful artifact is the list of those facts, in front of the person who asked. Then a note, if they want one. Then a person still sends it.

How we know this

The Search claims come from Google Search Central: the 15 May 2026 AI optimization guide, the 15 June 2026 llms.txt clarification on the documentation updates page, the 3 June 2026 Search Console generative-AI reports post, and the AI features documentation. The August 2026 Search changelog was checked on 2 September 2026; it did not walk back the June Search line.

The WebMCP claims come from Chrome for Developers: the WebMCP overview (updated 7 August 2026), the February 2026 early-preview post, and the secure-tools and imperative API pages. WebMCP is a proposed standard in an origin trial. This brief does not treat it as a finished web platform feature.

The sentence “RafalAI watches these sources” is about RafalAI’s own public-lab scanning. It is not a claim that Google or Chrome endorse this page. Dates and quotes should be checked against the live URLs below; documentation pages move.