Opting out of AI answers looks like the safe choice. For a brand, it rarely is

There has never been a better week for leaving AI. On October 30, Google, Microsoft and OpenAI said they will read one shared robots.txt line for AI training. Microsoft added a page-level switch for grounding, the step where an assistant reads live pages to answer a question. Google already lets every site opt out of AI Overviews and AI Mode in Search Console, a control PPC Land says arrived in July and reached all sites by August 31.

These controls are good news. Publishers asked for them for years, and clear switches beat guesswork about which crawler does what. But a clear switch also makes it easy to flip one without thinking through the cost. For a packaged goods brand, our view is that the cost is usually higher than it looks, and a small test published this week gives one more reason to think so.

The test that started the conversation

On October 29, Search Engine Land wrote up a case study that SEO consultant Cyrus Shepherd shared on LinkedIn. A site opted out of Google's AI search features. Three things followed:

• AI impressions fell to zero right after the opt-out.
• Based on the change in total clicks before and after, Shepherd estimated the AI features had a click-through rate of 0.9%.
• Traditional organic clicks also declined after the AI features were turned off.

The third finding is the surprise. If AI answers only took clicks from ordinary results, removing yourself from them should have left your ordinary clicks flat or higher. Shepherd's explanation is what he called a possible "Search SERP Halo Effect": appearing both in the AI answer and in the regular results builds trust, so the AI citation helps the regular listing earn the click even when nobody clicks the citation.

AI citations may drive more clicks than direct AI referrals show.

Why we would not lean on it too hard

We want to be careful here, because the test is thin. Shepherd described it as small and said results may vary. Search Engine Land notes that he did not disclose the site's traffic volume, the test dates, the size of the organic decline or the opt-out method. Without dates, a seasonal dip or a ranking update could explain some or all of the drop. Without volume, we cannot tell whether the change is outside normal week-to-week noise.

So this is not evidence that opting out costs organic clicks. It is a plausible mechanism plus one data point. Anyone who tells you the halo effect is proven is overselling it.

Still, the idea fits with something we have argued before: that AI referral traffic is the wrong number to judge AI search by. The direct click from an AI answer is small for most sites. If being cited also nudges other clicks, branded searches and later purchases, then measuring only the referral undercounts the value of being present. And if you opt out on the strength of the referral number alone, you may be giving up more than you measured.

The companies are telling you the cost

The more persuasive evidence comes from the platforms themselves. Microsoft's announcement is unusually blunt for a product launch. If a site chooses DisallowExtended or Disallow for grounding, Microsoft says, "you will see reduced brand visibility in AI experiences." A caution box further down the same post puts it more strongly: these settings "will significantly reduce your content's visibility in AI-generated experiences, without affecting eligibility for Bing search results." And it adds that "neither choice guarantees inclusion, traffic, or payment."

Google's post on the shared training token has a quieter version of the same message. A site that opts out through the shared token without naming Google-Extended will be treated as having disallowed Google-Extended, and Google says that "websites that specifically allow Google-Extended are eligible to appear in responses and receive clicks via grounding in Gemini Apps." Read the other way round, a blanket training opt-out at Google can also cost a brand its links in Gemini. We explain the mechanics in our guide to the shared opt-out.

Table: The AI opt-outs now on offer, and what each one gives up

Control	Who	What it stops	What it costs
Search Console AI features opt-out	Google	Showing the site in AI Overviews and AI Mode	AI impressions and citations in Search; possibly some organic clicks, per one small test
Shared token __usage_ai-training	Google, Microsoft, OpenAI; Anthropic implementing	Use of pages for model training, from 2027	At Google, also Gemini Apps grounding unless Google-Extended is explicitly allowed
AI Grounding DisallowExtended	Microsoft	Grounding outside Microsoft's own experiences	Visibility in third-party AI tools built on Microsoft's grounding
AI Grounding Disallow	Microsoft	All grounding, including Copilot	What Microsoft calls significantly reduced visibility in AI experiences

Sources: Google Search Central Blog and Microsoft Bing Blog, October 30, 2026; PPC Land, October 31, 2026; Search Engine Land, October 29, 2026.

Publishers and brands want different things

The case for opting out is strongest for publishers, and it is a real case. A newsroom's product is its reporting. When an assistant summarizes a story well enough that nobody visits, the publisher loses the visit, the ad view and the subscription prompt. Asking to be paid, or asking to be left out, is a reasonable response.

A brand is in the opposite position. Its product sits on a shelf or in a warehouse. The website, the ingredient list and the usage instructions exist so that the facts travel. A shopper who reads in an AI answer that your conditioner is silicone-free, and then buys it at Target without ever seeing your site, is a success, not a loss. The AI answer did the job your product page was built for.

That is why we read the click data differently for brands than for publishers. A 0.9% click-through rate on AI features would worry a publisher whose revenue depends on visits. For a brand, the click was never the main prize. The mention was.

What leaving actually leaves behind

There is a second problem with opting out, and it matters more for brands than the halo effect. When your own pages leave an AI answer, the answer does not go blank. The assistant still describes your product, using whatever else it can find: retailer listings, reseller pages, old reviews, Reddit threads and comparison articles.

Some of those sources are excellent. Others describe the formula you changed last year, the pack size you discontinued or a price from a promotion that ended. Your site is usually the one source that is guaranteed to be current. Taking it out of the pool does not take your brand out of the answer. It takes your version of the facts out.

This is also why we would be wary of a blanket robots.txt block added by a CDN or security vendor. A setting chosen to protect a blog from scraping can quietly remove a product catalog from the places shoppers now ask questions.

When opting out does make sense

We are not arguing that every brand should leave every door open. There are good reasons to use these controls narrowly:

• A paid or licensed content library, such as a recipe archive behind a membership or a sold research report, is a publisher asset inside a brand, and can be scoped out by path.
• Pages with regulated claims that must appear only with full context, such as some health or supplement content, may justify a page-level grounding choice after legal review.
• Old campaign microsites full of expired offers are better removed or redirected than left for assistants to quote.

What these have in common is scope. Each is a decision about specific pages, which is exactly what the new controls allow. The mistake is treating a site-wide opt-out as a neutral, cautious default. Microsoft's own warning says it is neither.

What we would measure before deciding

If a team is still tempted to leave, we would ask it to measure first. Count how often the brand is named and cited in AI answers for the questions shoppers ask, in Google, Bing and the main assistants. Note what the answers get wrong. Then compare branded search demand and organic clicks over the same weeks. Tools from companies such as XENA Intelligence, whose SearchPanel keeps assistant answers product by product, make this faster, but a fixed list of questions checked monthly is enough to start.

If the answers are wrong, the fix is usually better source pages and better retailer listings, not absence. If the answers are right and sending little traffic, that is the AI search model working as designed for a brand. In both cases, the brand is better off in the room.

The new switches are welcome. They make a hard question easier to act on. For most brands, the right action is to read the switches carefully, check that nothing has flipped them by accident, and leave them where they are.