Until August 31, a site couldn't leave AI Overviews without leaving Google Search, unless it gave up its snippets with nosnippet. Now there's a Search Console setting that removes a site from AI Overviews, AI Mode and the generative features in Discover and leaves its classic rankings alone.
That makes three separate switches on one URL: Search, the AI features inside Search, and Gemini. None of them affects the other two.
TL;DR
- The Search Console "Search generative AI" setting removes a site from AI Overviews, AI Mode and Discover AI without touching rankings. It joins
noindex(classic Search) andGoogle-Extended(Gemini) as the third separate control on one URL. - Blocking
Google-Extendeddoes nothing to AI Overviews or AI Mode, and the new setting does nothing to AI training. - The setting is property-level, not page-level. Excluding one directory means creating a URL-prefix property for it.
- The companion report shows AI impressions only. No clicks, no queries, no position. And those impressions already sit inside your Web totals.
- Don't leave by default. The AI answer still gets written; it just cites whoever stayed in.
What Google shipped
Two things landed together: a setting and a report.
Exclude means your URLs stop appearing as links in the AI features, and your content stops being used to ground their answers. Google says the setting isn't a ranking or inclusion signal anywhere else in Search.
The report sits as a tab under Performance and counts how often your URLs appeared in AI Overviews and AI Mode, by page, country, device and date.
On June 3 the UK's Competition and Markets Authority imposed a conduct requirement on Google: publishers must be able to keep content out of AI search features without being penalized for it. Google began testing the toggle with UK sites the same day, and hasn't said whether the toggle is meant to satisfy the order. The help doc is explicit about what it doesn't cover:
"This control doesn't affect AI training."
Google, Search Console Help: Search generative AI control, updated August 31, 2026Training was already its own decision through Google-Extended. What August 31 added is a separate decision for the AI features inside Search, which until then were bundled with classic results under noindex. Same URL, before and after:
The AI features inside Search got their own switch. The other two didn't move.
Three switches, and what each one doesn't do
These aren't nested. The strictest one doesn't cover the rest.
| Control | Where it lives | Classic Search | AI Overviews, AI Mode | Gemini grounding | Gemini training |
|---|---|---|---|---|---|
| noindex | Meta tag or X-Robots-Tag, per page | Removes | Removes | Removes (not in the index) | No effect |
| Search generative AI | Search Console, per property | No effect | Removes | No effect | No effect |
| Google-Extended | robots.txt token, per path | No effect | No effect | Blocks | Blocks |
The new setting removes the nosnippet tradeoff. What it costs you is precision: it works per property, so per domain, subdomain or URL prefix, not per page, at least until the CMA's March 2027 deadline for page-level controls.
A robots.txt Disallow for Googlebot isn't in the table because it doesn't select a product. It blocks the crawl entirely, so Google can't read a noindex on the page either. The URL can still get indexed from links, with no snippet and no content for anything else. If the goal is to rank but stay out of AI, robots.txt can't do it.
Where can this page still appear?
Flip the switches. The four surfaces update.
Impressions, and nothing to divide them by
The report shows impressions and nothing else: no clicks, so no CTR; no queries, so no idea which prompts you're answering; no position, so no idea whether you were the first inline link or the eighth. The CMA's own interpretive notes say Google should provide click-throughs and click-through rate for these features, separated from organic. Three months after the order, the global rollout still ships without them. And the impressions it does show are already inside your Web totals, so adding the two reports counts the same appearance twice.
The data has already broken once. From August 13, a logging error cut impressions in the generative AI report. John Mueller confirmed it was a logging issue, not a visibility change. We covered the last time Search Console impressions were wrong, for eleven months. Same lesson: Search Console numbers are inputs, not KPIs.
Should you leave?
Rankings aren't the cost of leaving. What changed is that one crawl now feeds several products and you get a permission for each; the doc says the covered feature list will grow, so expect more switches.
That makes a sitewide robots policy the wrong unit of decision. Your pricing page and your licensed dataset don't share a job or a risk profile, and property-level control is enough granularity to treat them differently.
Decide by the job of the page
| Job | Pages | Setting |
|---|---|---|
| Create demand | Product, pricing, comparisons, integrations, use cases | Include everything. These are the pages you want quoted when a buyer asks AI Mode which tool to pick. |
| Build authority | Docs, guides, original research, changelogs | Include Search AI. Grounding is where citations come from, and citations are the authority signal. |
| Protect value | Licensed data, paywalled reports, sensitive security detail | Restrict per product. Move it under its own URL prefix, exclude that property, and check Google-Extended separately. |
First, the obvious conflict: we're an SEO and GEO agency. Getting companies into these answers is what we sell, so we were never going to tell you to leave. Judge the argument on Google's wording, not ours. Leaving doesn't remove the AI answer. The doc says content from other sites will still be available in those features and "may appear similar to yours." The question gets answered either way. The only variable is whether the source is you or the competitor who stayed in.
What to do this week
- Open Settings, then Search generative AI, on the domain property and confirm it reads Include. Check the child properties too. Inheritance flows down, so one Exclude on a parent silently covers everything under it.
- Audit
Google-Extendedon its own. A block there keeps you out of Gemini, not AI Overviews; decide whether that's still what you want. Our robots.txt checker shows what's actually in the file. - Sort pages into three buckets: demand, authority, protect. Only the third is a candidate for exclusion, and only through a URL-prefix property.
- Pull the generative AI report and read it next to clicks, branded search and pipeline. Never add it to Web impressions.
- If you still want to leave, test one directory first. Create a URL-prefix property, set it to Exclude, and compare it with a matched directory for a few weeks. If AI impressions vanish and nothing downstream moves, they weren't worth much. If leads dip, switch it back.
Bottom line
Don't leave by default. For most B2B sites that means Include everywhere, Google-Extended open, and the new report read next to numbers that move revenue. Content whose value depends on controlled access gets its own property and its own decision.
Sources
- Google, Search Console Help: Search generative AI control. Updated August 31, 2026.
- Google, Search Console Help: Generative AI performance report (Search). Updated August 31, 2026.
- Google, The Keyword: New opportunities, control and insights for website owners. June 3, 2026, updated August 31, 2026.
- Google, Crawling infrastructure: Google-Extended. Updated July 14, 2026.
- Google Search Central: Block Search indexing with noindex.
- UK Competition and Markets Authority: Google Search publisher conduct requirement. June 3, 2026.
- Search Engine Journal, Matt G. Southern: Google Gives Sites AI Search Opt-Out, But Not The Data To Use It. June 6, 2026.
- Search Engine Roundtable, Barry Schwartz: Global: Google Search Console AI Performance Report and AI Control. September 1, 2026.
- Search Engine Roundtable: Google Search Console Performance Reports Drop In Impressions and Clicks, with John Mueller's confirmation of the August 13 logging error.