Google-Extended does not opt your site out of AI Overviews

Google now documents several controls that contain the word “AI” somewhere in their explanation, but they do not govern the same system. That distinction has become operationally important as Search AI, Gemini products and model training draw on web content through related but separate pathways.

The most expensive implementation error is to treat Google-Extended as a universal opt-out. A disallow rule for that token can leave a page eligible for Google Search, AI Overviews and AI Mode. Conversely, excluding a property from generative features in Search Console does not turn off AI training. The controls are intentionally narrower than the phrase “AI use” suggests.

Search AI and Gemini reuse are different publishing decisions

Google’s crawler documentation describes Google-Extended as a standalone product token rather than a separate HTTP user agent. Google continues to fetch with existing crawler identities and uses the robots.txt token as a control. Teams searching their logs for a new “Google-Extended bot” can therefore reach the wrong conclusion: the absence of a dedicated request string does not prove the policy is inactive.

The Search-side rule is just as explicit. Google says the Google-Extended preference is not a Search inclusion or ranking signal. AI Overviews and AI Mode remain part of Google Search, where Googlebot access, index eligibility and Search-specific controls determine whether a page can participate.

One page can reach four different control surfaces

The practical control plane is easier to understand as a switchyard than as one master permission. A publisher first identifies the undesired use, then changes only the route that serves it. The infographic separates four decisions that are often collapsed into one robots.txt conversation.

Google-Extended governs reuse after Google has crawled the page

A publisher can disallow the token for an entire site or limit the rule to a path. The resulting preference governs whether content Google crawls may be used for the documented Gemini training and grounding purposes. It does not tell Googlebot to stop maintaining the Search index.

User-agent: Google-Extended
Disallow: /research/

This is a policy boundary, not a promise that previously processed systems change instantly. It also does not replace controls for other companies, agents or crawlers. A publisher that wants a multi-provider policy needs a named decision for each provider instead of assuming one Google token creates an industry-wide contract.

Search Console controls participation in generated Search answers

The Search generative AI control is rolling out to a subset of verified site owners. Its default state includes the property. Choosing exclude prevents the site’s links and content from appearing in the supported generative Search features and prevents that content from grounding their responses. Google lists AI Overviews, AI Mode and generative AI features in Discover.

The exclusion does not change ranking or inclusion in other parts of Search. It also does not affect AI training; Google directs publishers back to Google-Extended for that decision. Google says a live control change generally reaches the applicable Search AI surfaces within one or two days, although caching and propagation can make some content take longer.

This is the control that complements our earlier analysis of Search Console’s generative AI performance report. The earlier article explains how to interpret exposure after a property participates. This one explains which control changes participation before those impressions can exist.

Preview controls can narrow direct input without removing the URL

These directives still require Google to crawl the page and read them. Blocking the URL in robots.txt can prevent Googlebot from seeing a meta rule placed on the page. The same logic applies to data-nosnippet: valid markup matters, and Google advises against adding or removing the attribute from existing nodes through JavaScript because extraction may happen before or after rendering.

Granularity changes the trade-off. Page-level nosnippet can make a URL ineligible as a supporting link because Google requires a page to be eligible for a snippet in order to appear in AI Overviews or AI Mode. Text-level data-nosnippet is narrower: it can protect a licensed passage, answer block or sensitive excerpt while leaving the rest of the page available for preview.

The Control Cascade keeps property inheritance visible

Search Console adds another scope problem that robots.txt does not have: parent and child properties can inherit the generative AI setting. A top-level domain decision can quietly reach subdomains and URL-prefix properties unless a closer child property is manually configured to stop inheriting.

This hierarchy is why the Search Console change deserves configuration management. A team responsible for a subdirectory can see an inherited state without knowing who changed the parent. Another team can override a child and unintentionally create a pocket of participation inside an excluded domain. The effective setting is a path decision, not merely an account preference.

Deploy the narrowest control Google can verify

A safe rollout begins with the intended business outcome rather than the directive name. Removing a licensed paragraph from previews is a different request from excluding a property from Search AI, and both are different from preventing a page from appearing in Search at all.

Google’s current generative Search guidance also removes two tempting distractions. No special schema is required for AI Overviews or AI Mode, and Google Search does not use llms.txt as a visibility mechanism. A publisher can maintain such files for other services, but they neither improve nor reduce visibility in Google Search. The control plane still begins with crawlability, index eligibility, previews and the explicit product controls Google documents.

Operational verdict: do not ask whether a site is “opted out of AI.” Ask whether Googlebot may crawl it, whether Search may index and preview it, whether Search AI may link to and ground on it, and whether separate Gemini systems may use the crawled content. Four precise answers are cheaper than one ambiguous rule.

Questions about Google-Extended and Search AI controls

Official Google control sources