Google AI Overviews opt-out: no ranking penalty, but other sites fill the slot

The claim that sites can now opt out of Google AI Overviews and AI Mode without losing rankings is Overstated. The switch is documented, and Google states the choice is not a ranking signal elsewhere in Search Google · 2026A. Opting out removes the site from the answer but leaves the answer in place, built from other sites Google · 2026A. A randomized experiment found that hiding AI Overviews entirely raised click-through by 8.8 points, a gain no single site can buy by leaving Wang et al. · University of Pennsylvania; Northeastern University · 2026A. The sections below set out what the switch does, what it leaves out and how the arithmetic differs for brands and publishers.

The Google AI Overviews opt-out is real and documented as ranking-neutral

Google's Search Console control has three states: include, exclude or inherit from a parent property, with include as the default Google · 2026A. It covers AI Overviews, AI Mode and generative features in Discover, and an excluded site gets no traffic or impressions from them Loew · Google · 2026A.

The no-penalty half of the claim is documented twice. Google states the control is not used as a ranking signal elsewhere in Search Google · 2026A. The Competition and Markets Authority (CMA) expects Google not to downrank opted-out content or strip its snippets Competition and Markets Authority · 2026A. Both are statements rather than measurements, and no before-and-after study of an opted-out site has been published.

The rollout is narrower than the claim implies. Google began testing the control with some UK site owners in June 2026 Loew · Google · 2026A. Its help page still described the rollout as reaching a subset of website owners on 24 August Google · 2026A. Practitioners have reported the control on some non-UK accounts since July, and Google has not announced that the global rollout is complete.

Opting out of AI Overviews removes the site, not the AI answer

Google's documentation states the consequence plainly. If a site excludes itself, content from other sites remains available in those features and may look similar to its own Google · 2026A. The excluded site's crawled content also stops being an input to the answer, so it loses any say in how the answer describes it Google · 2026A.

The traffic harm sits with the feature, so one site's exit does not undo it. Cohen and Davies, in a peer-reviewed competition-law article, call the opt-out an ineffective remedy because traffic diverted to Google stays diverted Cohen · University of Oxford; University College London · 2026A. They report that publishers not cited in an AI Overview see organic click-through almost three times lower. Their figures are cited from Pew, Seer Interactive and SSRN papers rather than measured, and Davies discloses Google employment from 2016 to 2022 Cohen · University of Oxford; University College London · 2026A.

The randomized evidence points the same way. In a preregistered field experiment, hiding AI Overviews raised external click-through by 8.8 percentage points Wang et al. · University of Pennsylvania; Northeastern University · 2026A. That arm removed the whole feature for a user, and after a Google HTML change it reached only 51.1% of Overviews. A site-level opt-out removes one source and leaves the feature live.

The AI Overviews opt-out ships without the click data to decide it

Search Console's generative AI report shows impressions for URLs used as sources, with no clicks, click-through rate or query data Google · 2026A. A site can count how often it appears as a source. It cannot see what those appearances return in visits.

The cost of staying in is measured far better. An SSRN field experiment found AI Overviews cut outbound organic clicks by 39.8% where they appeared Agarwal · Indian School of Business; Carnegie Mellon University · 2026A. The asymmetry is the problem: a causal estimate of the cost on one side, a raw impression count on the other.

AI Mode carries a larger estimate with a sharper caveat. Assigning users to AI Mode cut external click-through by 18.8 percentage points in the 2026 preregistered experiment Wang et al. · University of Pennsylvania; Northeastern University · 2026A. That arm imposed AI Mode as a seven-day default. The users' own baseline AI Mode use was 0.6% of searches, so the result does not describe voluntary use.

Even a before-and-after test of opting out would be noisy. After Gemini 3 became the default AI Overviews model in January 2026, 42.4% of previously cited domains stopped appearing, almost all in the long tail Deda · SE Ranking · 2026B. A drop in impressions after an opt-out cannot easily be separated from that kind of model churn.

The CMA expects click data and page-level opt-out, with no confirmed date

The control exists because a regulator required it. The CMA imposed its publisher conduct requirement on 3 June 2026, the day Google announced the control Competition and Markets Authority · 2026A. Its interpretative notes expect impressions, clicks and click-through rate per publisher, separated from organic search, and grounding opt-out at directory and page level Competition and Markets Authority · 2026A. The notes attach no dates to either expectation, so when Google must ship them is not confirmed.

The shipped control is finer than site-wide but coarser than page-level. Settings inherit from domain to subdomain to URL-prefix property, and a child property can be set on its own Google · 2026A. A directory verified as its own URL-prefix property can therefore be excluded today. That follows from the documentation; no public test has demonstrated it.

The AI Overviews opt-out does not cover AI training

Google states the Search generative AI control has no effect on AI training Google · 2026A. Training sits with Google-Extended, a robots.txt token that governs whether crawled content may train future Gemini models and ground answers in Gemini Apps Google · 2024A. Google states that Google-Extended does not affect inclusion or ranking in Google Search Google · 2024A.

The two levers have different scopes, and the claim treats them as one. Blocking Google-Extended leaves a site in AI Overviews, and excluding a site from AI Overviews leaves Gemini training untouched. The CMA expects Google to widen Google-Extended to cover training of its search AI features Competition and Markets Authority · 2026A. The same notes do not require Google to remove content already ingested, so any withdrawal is forward-looking Competition and Markets Authority · 2026A.

Brands and publishers run different arithmetic on the same opt-out

For a brand, the opt-out now reaches its own name. Ahrefs found AI Overviews on 82.91% of US desktop branded searches on 29 September 2026, across 232.8 million results Law · Ahrefs · 2026B. A second tracker dated the jump to late September, from 26.12% of branded keywords on 1 September to 82.06% on 29 September Grieselhuber · DemandSphere (Ginzamarkets, Inc.) · 2026C. That tracker does not disclose its keyword count; treat its levels as provisional and its timing as the useful part.

An opted-out brand leaves the answer about itself to other sources. On commercial-intent questions, vendor, product and long-tail pages drew 90.6% of 127,198 citations in one disclosed vendor study Khallad · SurfacedBy · 2026B. Shopify reports that AI-referred shoppers reaching a product page converted about 80% better than organic in Q2 2026 Risley · Shopify · 2026B. That is first-party data from a seller of AI-commerce products without sample sizes, so it shows direction, not proof.

Publishers earn from the click, so the opt-out looks better to them. Executives at USA Today, People Inc., Reuters and Time have called blocking Google entirely an option, though none had acted when reported The Wall Street Journal · 2026B. Even for a publisher, the opt-out moves a page into the uncited group, which Cohen and Davies associate with click-through almost three times lower Cohen · University of Oxford; University College London · 2026A.