Google finished the global rollout of two Search Console features on August 31, and they arrived as a pair for a reason. One is the Search generative AI control, a property level switch that pulls your site out of AI Overviews, AI Mode and the generative features in Discover. The other is the AI performance report, which shows how you are doing on those surfaces. Together they look like a complete decision kit. They are not.
The control is a real lever. The report, as shipped, gives you impressions, pages, countries and dates, plus a device breakdown on the Search side. It does not give you clicks. Barry Schwartz confirmed the shape of both when the rollout completed, reporting it at Search Engine Land on August 31.
What the Search generative AI control actually does
Start with the scope, because the name invites two wrong readings.
It is not a training opt-out. Whether Google can use your content to train models is governed separately by Google-Extended, and flipping this switch does not change that. It is also not a crawl block. Googlebot still comes, your pages still get indexed, and your regular Search rankings are unaffected by the choice itself.
| WHAT THE CONTROL DOES | WHAT IT DOES NOT DO | WHERE IT APPLIES |
|---|---|---|
| Removes your site from grounding and appearing in Google's generative AI Search features | Does not stop Googlebot crawling or normal indexing | Property level, set in Search Console settings |
| Removes the impressions and traffic those surfaces would have sent you | Does not act as a ranking signal in regular web Search, per Google | AI Overviews, AI Mode, generative features in Discover |
| Takes effect for the whole property once set | Does not govern model training, which stays with Google-Extended | One switch, no per-directory or per-template control |
| Applies only to Google's surfaces | Does not affect ChatGPT, Perplexity, Copilot or any other engine | Google only, every other engine needs its own handling |
That last row is the one teams skip. A publisher who flips this because they are angry about AI answers has removed themselves from one engine's answer layer and left every other engine exactly as it was. If the goal is to stop being summarized, this does about a third of the job, and it does it on the engine most likely to send you a click afterwards.
The property level scope matters too. There is no version of this where you keep your product documentation visible to AI Mode and pull your subscriber archive out. It is all or nothing per property, which for most large sites means the decision has to be made by whoever owns the least valuable content and the most valuable content at the same time.
The report that cannot price the decision
Here is the structural problem. The new report tells you how often you appeared. It does not tell you what appearing was worth.
Impressions without clicks is not a partial view of the same thing. It is a different metric answering a different question. An impression in an AI Overview where your line is the third of five cited sources, below the fold, is not the same asset as an impression in AI Mode where a user clicks through to your pricing page. Both count once. Only one of them pays.
“A control without a click column asks you to sell an asset before anyone will tell you the price.”
This is the same measurement gap we walked through in the error bars around AI search reporting, and it is why the earlier generative AI logging behavior in Search Console mattered more than a reporting quirk usually does. When the only instrument you have counts one side of the ledger, every argument in the room becomes a vibe argument, and the loudest person wins.
None of that makes the report useless. Impressions by page is genuinely new information, and for most teams it is the first honest answer to which of your pages the answer layer is actually reading. Use it. Just do not mistake it for a business case.
The practical workaround, until clicks arrive, is to treat the impression data as a coverage map rather than a value estimate. Coverage questions it answers well: which templates appear, which sections never do, whether your appearance is broad and shallow or narrow and deep, and whether it moved after a publishing push. Value questions it cannot answer: what a given appearance was worth, whether it displaced a click you used to get, or whether the user who saw you went on to buy from anyone at all. Keep those two lists separate in every deck you build off this data, because a stakeholder reading an impressions chart will assume it is a traffic chart unless you say otherwise on the slide.
Who should opt out of the Search generative AI control
Very few sites, and the ones that should are not deciding it on traffic grounds.
| SITUATION | SHOULD YOU OPT OUT? | WHY |
|---|---|---|
| Subscription publisher whose paid archive is being summarized | Maybe, and it is a business decision, not an SEO one | The summary can substitute for the subscription. That is a licensing and product question your revenue team owns |
| Regulated content where a paraphrase creates compliance exposure | Possibly, with legal in the room | You cannot control how a generative surface restates a disclosure. Some legal teams will not accept that risk |
| B2B SaaS competing for comparison and best-tool answers | No | You are trying to be in that answer. Opting out hands the citation slot to a competitor or an aggregator |
| Ecommerce with product and category pages | No | Shopping-intent answers increasingly route through these surfaces. Removing yourself removes a shelf |
| Site frustrated by falling click-through rates | No | Opting out does not restore the old click. It removes the impression and leaves the behavior change in place |
| Anyone acting before reading their own impressions data | No | You have had the report for days. Read it first |
The frustration row deserves its own sentence, because it is the most common reason a team reaches for this and the worst one. Zero-click behavior is not caused by your presence in the answer. It is caused by the answer existing. Withdrawing your content does not put the click back, it just means somebody else's content grounds the answer that the user reads instead of visiting you. You have removed yourself from a surface that was going to appear either way.
The publisher case is different and it is real. If your product is the article and an AI summary is a functional substitute for reading the article, then appearing in that surface is closer to giving away inventory than to marketing. That is a legitimate position and we would not argue a media business out of it. It is just not an SEO decision, and it should not be made by the person who owns Search Console. For a B2B software company the calculus almost never lands there, because the summary is not the product, it is the thing that gets the product named.
The three questions to answer before you touch the toggle
If someone in your organization is pushing for this, these three questions settle it faster than a debate about the future of search.
There is a fourth question that only matters at enterprise scale: whose property is it. Large organizations often run separate Search Console properties for documentation, marketing and support, sometimes owned by three different teams. A property level control means one of those teams can withdraw a whole content class from the answer layer without anyone else finding out, and the first signal will be a slow impression decline in a report nobody checks daily. Write down who has settings access before somebody exercises it.
What to do this week
For almost every commercial site, the correct action on the control is to leave it alone and document that you looked at it. The correct action on the report is more involved and more useful.
Export the AI impressions by page now and keep it. This is the first week the data exists for everyone, so it is the earliest baseline you will ever have. Six months from now the question will be whether your answer layer presence grew, and the only way to answer it is to have kept September.
Then look at the page list against your own map of what matters. Most teams find one of two patterns. Either the answer layer is reading their documentation and support content far more than their marketing pages, which is a content investment signal, or it is reading a handful of old comparison posts, which is a refresh signal. Both are actionable this quarter without touching a single setting, and both are the kind of finding that shows up in our generative engine optimization work long before anyone talks about opting out of anything.
Set an internal rule while it is calm: nobody flips the control without the three questions answered in writing and a named owner for reversal. Rules like that feel like bureaucracy right up until a bad quarter makes a switch look like a strategy.
And keep the wider frame in view. Google shipped a visibility control for its own surfaces. It did nothing about anyone else's, which means the work of earning citations across engines is unchanged, and the argument for building citation strategy that does not depend on licensing deals is exactly as strong as it was last week. One vendor giving you an off switch is not the same as you having control of the answer layer.
Read the report. Export the baseline. Leave the switch alone unless your revenue model says otherwise, and if it does, make that call with the people who own the revenue model rather than the people who own the dashboard.
See where you are cited today
A free snapshot audit of your rankings and AI citations before we ever talk.
Josh leads work at the intersection of SEO and generative engines at Something Inc., helping B2B brands get ranked and cited across every major AI engine.