Google gave every site an AI off switch. Don't use it.

What actually shipped on 31 August

Google had been running this as a limited release since June, starting with a subset of UK sites. That phase is over. In Google's own words, quoted by Search Engine Land, "as of August 31, 2026, we've rolled out this control to all websites worldwide". The generative AI performance reports went global on the same date.

The control sits in Search Console under Settings, then Search generative AI, and it has three states. Choosing Exclude means your "content is prevented from being visible to users in Search generative AI features, including being linked to within these features". It covers AI Overviews, AI Mode and AI Overviews in Discover as one bundle, with no way to separate them, and the change takes effect within 1 to 2 days of going live.

That last assurance is the one people distrust, and reasonably so, because it is Google marking its own homework. There is a stronger reason to believe it than Google's word: the UK regulator that forced this release also requires Google not to penalise the publishers who use it, as PPC Land reported. Retaliating would put Google in breach of a regulator's requirement rather than merely its own word.

This switch was built for publishers, not for you

Google did not wake up feeling generous. On 3 June 2026 the UK Competition and Markets Authority imposed a publisher conduct requirement obliging Google to "provide publishers with effective controls over the use of their search content in generative AI" and, separately, to give them "clear and detailed metrics on user engagement with their search content in search generative AI features". Google was given nine months to implement, with compliance reports every six months in the first year. Google's own announcement names the CMA as one reason the tools exist, describing itself as "engaging with regulators like the UK's Competition and Markets Authority".

The publishers who won the control are not impressed with it. Paul Bannister of Raptive told Digiday that a front-end toggle "is not real control; it is Google offering publishers a light switch while keeping the power plant running". Stuart Forrest, formerly of Bauer Media Group, put the gap precisely: "They've broken out search impressions about AI features, but they haven't given us search clicks for AI features". Phil Andraos of Reuters Digital said the decision is stuck on evidence rather than principle: "There's not enough data. If there was, we could make that decision very quickly."

Those quotes tell you who the switch was built for. A large publisher sells ads against pageviews, so a summarised article that nobody clicks is close to pure loss, and withdrawing is at least arguable. If you sell software, rooms, legal advice or garden furniture, the AI answer is the shelf your product sits on, and nothing in the summary replaces the thing you are selling.

Being cited beats not being cited, on the same queries

Start with what nobody knows. No one has published what happened to a site's traffic after it used this control, and Search Engine Journal says as much. Anyone telling you opting out costs a specific percentage is making it up. The argument below is inference, and the inference runs through the one comparison that has been measured properly: what happens to the same query when an AI Overview cites you against when it cites someone else.

Seer Interactive ran that comparison across 53 brands, 5.47 million queries and 2.43 billion organic impressions. Across 2025, where the brand was cited in the AI Overview, organic clickthrough ran at 2.07%. Where it was not, 0.94%. Seer states the difference plainly: being cited "delivers +120% more organic clicks per impression versus when you are not cited". Over 2025, the not-cited number kept falling, down 67% across the year. This is agency client data rather than a random sample of the web, so treat the level as indicative and the gap as the finding.

Excluding your site does not put you back in the pre-AI world. It moves you into the 0.94% column, permanently, on every query where an AI Overview appears. And that column is getting worse independently: Ahrefs now finds that the presence of an AI Overview correlates with a 58% lower average clickthrough rate for the top-ranking page, using December 2025 data. That is a different study from the 34.5% figure over 300,000 keywords we quoted back in July, which compared March 2024 with March 2025. Different windows, different samples, and the newer one is worse.

The honest counter-argument comes from Pew, who tracked 900 US adults through March 2025 and found people clicked a link inside the AI summary itself on just 1% of visits. So the citation link barely gets clicked, and on its own it is a poor prize. What it buys you is placement further down the same page. In the same Pew study, users clicked any traditional result on 8% of visits where a summary appeared, against 15% where none did, and Seer's gap shows where those surviving clicks land: on the brand the summary has already named. The fight is over the shrunken pool, and the citation is how you win a share of it.

Scale settles how much this matters. Sundar Pichai said at Google I/O in May that AI Overviews passed 2.5 billion monthly active users, and that AI Mode had "already surpassed 1 billion monthly active users" within a year of launch. The visits that do arrive behave well: Adobe measured AI referrals to US retail sites in July 2026 converting 60% better than non-AI traffic, growing 62% year on year. That covers AI referrals broadly rather than AI Overviews, and US retail only, but it matches what we found on the value of an AI visit earlier this year.

Three misreadings, and one unconfirmed side effect

Most people reaching for this control want something it does not deliver. It is not a training opt-out. Google's help page is explicit that to limit training of the models behind these features, you use Google-Extended, which is a separate mechanism in robots.txt. It is also not a withdrawal of your content from Google: even when excluded, your pages "may still be used to power Google Search as a whole", helping Google understand language and queries. And it does not reach the standalone Gemini app, which sits outside the scope of the toggle entirely.

The side effect is the one that will catch news sites. John Shehata of NewzDash told Search Engine Journal that the Search Console opt-out "will likely remove publishers from Top Stories inside AI Overviews". Google has not confirmed it and nobody has tested it, but the shape of the risk is familiar. Sites that blocked every AI crawler to keep their content out of training runs took themselves out of AI search results in the same move, and only found out later. A control this broad has consequences its documentation does not list.

What the report shows, and the five places it stops

The generative AI performance report shows how many times links to your site were shown to a user in a generative AI feature on Google Search, split by pages, countries, devices and dates. Discover gets its own separate report, with pages, countries and dates but no device breakdown. We covered the first version of this back in July, in Google finally let you measure AI search. What is new is that every property now has it, and that we have had long enough to see where it stops.

It stops in five places, and you should know all of them before you build anything on it. There are no clicks and no clickthrough rate, so you can see attention and not outcome. There are no queries at all. Impressions are aggregated per property, so if two of your pages appear in one AI answer, they "count as a single impression", which means the metric understates how much of an answer you own. Search Labs experiments are excluded, so a growing slice of AI activity never appears. And the data starts on 18 May 2026 with no backfill, so nobody has a year-on-year comparison and nobody can reconstruct what their AI impressions looked like before Google started counting.

Add one practical limit on top: the report is only in the Search Console interface. It has no API, so it cannot feed BigQuery, Looker Studio or any scheduled monitoring you already run. Every month somebody has to open a browser and export it by hand. If your property looks empty, that may be normal too. John Mueller confirmed on Reddit that these reports, like Discover and News, are "only shown if there's sufficient data to display".

It is still the first Google-owned count of your presence in AI answers, and it is free. It is also unfinished. Trade press reporting from PPC Land and Search Engine Journal puts click data at December 2026 and page-level controls at March 2027. Neither date appears on the CMA's published requirement, but March 2027 is the nine-month mark from imposition, so the timing fits what the regulator did write down.

What you actually do

Four things, in this order, and the whole job takes about twenty minutes.

1. Confirm the switch is on Include. Search Console, Settings, Search generative AI. Include is the default, so for most sites this is a five-second check. Do it anyway, and if you manage a domain property with children, check that none of them is inheriting an Exclude somebody set in June while the UK release was live.

2. Open the report and read it by page. The Generative AI performance report opens from the Performance section of Search Console. Sort by pages first. The pages Google is putting into AI answers are rarely the pages you would have guessed, and they are the ones to keep current, structure properly and expand. Then check countries, because AI impressions distribute differently from classic impressions and often reveal a market you were not targeting.

3. Export it every month, starting now. There is no API and no backfill. The history you will want in a year exists only if someone saves it this month. A dated CSV in a shared folder is enough.

4. Supply the missing half yourself. Google gives you impressions and withholds clicks, and the reported CMA timeline puts the earliest change at December. Your own analytics already sees the visit that arrives, so the referrer is the one place the outcome is visible today. Baseline's AI Traffic attributes visits to ChatGPT, Perplexity, Gemini and the rest, which turns Google's impression count into something you can weigh against revenue instead of admiring on its own.

Then leave the switch alone. If one archive or directory of yours is being summarised in a way that costs you money, write it down and wait for the page-level controls due in March 2027. Today's control cannot do that job, and turning off the whole property to protect one corner of it is a bad trade.

See what your AI visitors did next

Questions people ask about this