Google's AI Search Rules: The Opt-Out and What You Can Stop Doing

What Google actually published

Three documents, and most coverage has picked up one of them.

1. AI features and your website, which lists every control you have and what each one costs you.
2. The generative AI performance report in Search Console, plus a separate one for Discover.
3. The AI optimization guide, which contains a section titled "Mythbusting generative AI search: what you don't need to do".

The third one is the interesting one, and almost nobody in the optimization industry is quoting it. We will, in full, including the parts that are inconvenient for us.

The opt-out control, and why the old warning is out of date

Search Console now carries a Search generative AI control under Settings. Switching it off keeps your content out of AI Overviews and AI Mode.

The sentence that matters is Google's own: the control "isn't used as a ranking or inclusion signal affecting other parts of Search."

That is a real change, and it invalidates advice that was correct until recently, including advice we have given. Before this shipped, the only ways to stay out of AI features were noindex, nosnippet or blocking Googlebot, and every one of them damaged your ordinary Search appearance as a side effect. "Do not opt out, you will lose your rankings" was true because opting out meant opting out of everything.

It is no longer true. So if you have a licensing position, a paywall, or a legal reason to stay out of AI answers, you now have a scoped instrument for it. Changes take a day or two to apply, and child properties inherit the setting from the parent by default.

We still would not use it, and the reason has nothing to do with rankings. The AI answer gets generated whether or not you are one of the sources it names. Opting out does not remove the answer, it removes you from the credits. You keep the same lost clicks and give up the mention.

Note also what it does not do: it does not stop training. That is Google-Extended, a separate directive covering grounding and training in Google systems outside Search.

The report you can look at, and what it will not tell you

The generative AI performance report covers AI Overviews and AI Mode. Read the limits before you plan a dashboard around it.

• Impressions only. No clicks. No position. You can see that you were shown and not whether it did anything.
• Search Labs excluded, because those experiments are still moving.
• Partial rollout. Google is releasing it "to a subset of website owners", and you also need enough impressions in AI features to qualify. Many sites simply do not have it.
• Search and Discover are separate reports. The Discover one adds top and bottom pages by impressions and where they came from geographically.
• The usual caps apply, including the 1,000 row limit, and recent data can still change.

Everywhere else, AI feature traffic stays folded into your ordinary numbers: sites appearing in AI features "are included in the overall search traffic in Search Console", reported under the Web search type. That is why you cannot split AI clicks from normal clicks in the main Performance report, and why nobody can honestly tell you what percentage of your traffic came from an AI Overview.

What to do with it this month. If you have the report, write down your current impressions. You cannot backfill this number later, and a baseline recorded now is worth more than a better report you get next year.

What Google says you do not need to do

Quoted directly from the mythbusting section, because paraphrase is how this kind of guidance gets bent.

"You don't need to create new machine readable files, AI text files, markup, or Markdown to appear in Google Search."

"Google Search itself doesn't use them."

"Doing so will neither harm nor help your site's visibility or rankings in Google Search."

"There's no requirement to break your content into tiny pieces for AI to better understand it."

"There's no ideal page length, and in the end, make pages for your audience, not just for generative AI search."

"You don't need to write in a specific way just for generative AI search."

"Structured data isn't required for generative AI search, and there's no special schema.org markup you need to add."

"Seeking inauthentic 'mentions' isn't as helpful as it might seem."

Read together, that is Google saying most of the tactical layer sold as generative engine optimization does nothing inside Google. The underlying reason is stated too: the AI features are "rooted in our core Search ranking and quality systems". They are not a separate index with separate rules.

We sell some of what is on that list, so let us be exact rather than defensive. Structured data still earns rich results and still helps a search engine tell which company you are, which matters a lot when your name is shared with other businesses. It is just not an AI-specific lever. And llms.txt, an optional plain text file at your domain root that summarizes your site for language models, does nothing in Google. Google says so and Google is right about its own product.

The distinction almost everyone skips

Google is describing Google.

Its guidance covers AI Overviews, AI Mode and Discover, all of which are grounded in Google's own index. ChatGPT, Perplexity and Claude are a different retrieval chain. They use their own crawlers, their own indexes and, in some cases, a third-party search backend. Nothing in Google's documentation describes them and nothing in it constrains them.

So the honest version of each claim looks like this:

The one row that is identical in both columns is the boring one: a crawler has to be able to fetch the page. You can check yours with our free AI crawler access checker, which reads robots.txt precedence rules across the search and answer bots.

So what actually works

Strip out everything Google just ruled out and the remainder is not small. It is just less mechanical than the tactics it replaces.

Be worth citing. Google's own recommendation is unique, first-hand, genuinely useful content, and it explicitly warns against material that "could easily be produced by a generative AI model". An assistant summarizing commodity content has no reason to name a source.

Write passages that stand alone. Not because a word count is rewarded, and Google says plainly there is not one. Because a paragraph that only makes sense after the two above it cannot be lifted into an answer, so the model reaches for a source where somebody wrote a complete thought in one block. This is a writing standard, not a schema.

Keep crawlers in. The one hard gate, and the one most sites fail without knowing. Ordinary Googlebot rules for Google, and separate named user-agents for the assistants.

Be a legible entity. Consistent name, address and profiles, so a model can tell which company you are. This is where structured data still earns its place.

Measure mentions, not only rank. Whether you are named is now a separate question from where you rank, and only one of them is on your dashboard by default. Our AI visibility checker scores the signals that decide it.

The other controls, and what each one costs

For completeness, because the trade-offs are not obvious and the documentation lists them plainly.

The pattern is worth internalizing: everything except the top and bottom rows is a blunt instrument that pays for AI exclusion with normal Search visibility. Only the generative AI control separates the two, which is exactly why its arrival changed the advice.

What we changed on our own site because of this

Publishing this obliged us to check our own pages against Google's documentation, and one FAQ needed fixing. In our post on whether ChatGPT can read your website, the answer about blocking AI crawlers listed Google-Extended alongside GPTBot and PerplexityBot as the crawlers that earn you citations. That is wrong: Google-Extended governs training and grounding in Google systems outside Search, and it does not decide whether you appear in AI Overviews. The control that does is the one described above, and it did not exist when we wrote that answer.

Both points are now corrected in place, with the date. The recommendation did not change: allow the search and answer crawlers, and treat pure training crawlers as a separate licensing decision. What changed is that one of the bots was filed in the wrong group, and that a scoped instrument now exists for the people who genuinely need to opt out.

Every quote and claim in this post was checked against Google's own documentation on 12 August 2026. Google updates these pages, so if you are reading this much later, check the source before acting on it.