Is it possible to opt out of AI training without leaving the search engine? Webmasters should understand Google and Bing’s new rules

The most difficult question for website operators to answer is no longer simply “Should we allow crawlers to access our site?” but rather “After crawling, what can the content be used for?” An original tutorial might appear in search results, serve as a reference in answering questions, or be used to train future models. The value exchanged for these three uses is not the same, yet in the past, they were often lumped together under the vague phrase “allow AI to use it.”

On October 30, 2026, Google announced plans to support shared AI training preference signals, with a target date of 2027; the existing Google-Extended controls will continue to coexist. This announcement outlines the direction of future support, not an immediate, site-wide switch.Official Announcement from Google

On the same day, Bing announced more detailed implementation plans: preferences for training and response citations will be expressed separately. The new mechanism will take effect upon crawling starting February 1, 2027; users of the old “NoArchive” and “NoCache” settings should note the migration deadline of March 1, 2027. Eligibility for inclusion in traditional Bing search will not change as a result of these usage preferences, but visibility in AI-powered search results may be affected.Official Announcement from Bing

First, let's separate these three categories: discovery, citation, and training.

From the perspective of content management, these three types of usage can be understood as three types of relationships.

Search results are the gateway through which readers find you. Users first receive a link and then decide whether to visit the website. The website can typically capitalize on this visit through the page experience, further reading, and service conversions.

When you reply to a quote, your message appears in the other person's answer. Readers may find the answer directly on the Q&A interface, or they may continue browsing by following the link. While this provides brands with new opportunities to be discovered, it cannot be assumed that every mention results in a valid visit.

Model training, on the other hand, involves the development of the model's future capabilities. It does not necessarily generate a trackable link in every response. Therefore, a webmaster’s decision to be open to search engines while retaining data for training purposes is not contradictory: it is actually a matter of adopting different strategies for different purposes.

The key issue is not which use case is inherently better, but whether operators can choose one over the other. The more granular the tools are, the more likely website owners are to express reservations about a particular use case without sacrificing their overall visibility.

Shared signals reduce management costs; they do not automatically resolve all disputes.

The shared solution mentioned by the two companies in this instance uses policy tags in robots.txt—classified by purpose—to express training preferences. It aims to reuse existing file structures as much as possible, rather than requiring each website to build a new set of complex interfaces. However, the IETF website currently classifies the relevant documents as Internet-Drafts submitted by individuals and explicitly states that they are not finalized IETF standards. To refer to them as a “unified, mandatory new standard” would be an exaggeration.IETF Drafts and Status

This distinction is important. Common grammar addresses the question of “how do I express my meaning clearly”; the way participants actually adopt, interpret, and act upon it addresses the question of “how does the other person respond.” These are not issues on the same level.

You can think of it as a set of usage guidelines that are easier for machines to understand. With standardized guidelines, website maintainers won’t have to replicate a set of similar rules for each company; however, simply because these guidelines are documented, they do not serve as a technical barrier preventing any visitor from copying content. Nor should a preference statement be interpreted as a finalized commercial contract.

Therefore, it is worth paying attention to whether the scope of participation has been expanded, whether the platform’s explanations are sufficiently clear, and how conflicts will be handled—rather than simply focusing on how many major companies are mentioned in a single announcement.

For small websites, the most valuable feature isn't the “Close All” button

An AI navigation site may feature a combination of basic tool information, original in-depth reviews, paid courses, and user-submitted content. Applying exactly the same strategy to all content may not align with business objectives.

Basic introductions may place greater emphasis on being discovered; original research with significant investment may prioritize stable attribution and return traffic; and paid content requires protection through actual access controls, rather than relying on crawler preferences as a paywall. Different content types warrant separate discussion; a general statement like “resist training” or “embrace AI” cannot replace a nuanced assessment.

A reasonable decision-making process should start with content assets: Which content is used to build the brand? Which content directly constitutes the product? Which content contains third-party licensed material? And which content requires restricted access? Only once these questions are clarified can technical configurations be implemented with a clear purpose.

For websites that rely on search engines to attract readers, another risk is accidentally changing the “usage control” setting to “prohibit search crawling.” A setting intended to restrict subsequent training may end up preventing pages from being discovered, resulting in business impacts that are completely at odds with the original intent. What’s truly needed here isn’t more toggles, but fewer misunderstandings of their meaning.

Now is the time to get ready; this isn't the time to just copy and paste a configuration and call it a day.

At this stage, website administrators can undertake four low-risk tasks.

First, take stock of your existing rules. Compile a single list of all relevant settings—including robots.txt, page-level tags, server response headers, SEO plugins, and any controls that might be added by a CDN—to avoid checking only one file while overlooking conflicting settings that already exist at other levels.

Second, define your content classification objectives. Instead of selecting technical parameters first and then justifying them, start by clarifying which content should be prioritized for wider distribution, which should be reserved for training purposes, and which must remain private. You can create a concise matrix organized by section or content type to reduce the burden of maintaining each page individually in the future.

Third, prepare contingency and rollback plans. For any adjustment that affects search entry points, it is advisable to first retain the original settings, assign responsibility, and conduct testing on a small scale. In particular, do not delete all historical controls at once if you do not understand the relationship between the platform’s old and new rules.

Fourth, pair calendar milestones with verification actions. It is not enough to simply remind users that “a rule takes effect on a certain date”; you should also specify which platform document needs to be reviewed at that time, who is responsible for verifying the actual page output, and what evidence will be used to determine whether the setting has been identified. The announcement date, website deployment date, and platform access date may differ, so these must be distinguished in operational records.

These recommendations do not require websites to change their stance immediately. Even if they ultimately decide to continue allowing the use, knowing exactly what kind of use they are permitting is more reliable than letting default settings make the decision for them.

The next round of competition may center on measurable value exchange.

From a business analysis perspective, the evolution of training preference mechanisms has long-term significance: discussions between platforms and publishers have the opportunity to shift from “open versus closed” to “uses, boundaries, and returns.” However, more granular control also entails more granular maintenance work, and small teams in particular need to avoid letting configuration complexity end up eating into content production time.

Follow-up evaluations can focus on three types of metrics: whether the website is still being found by the target audience; whether the traffic generated by AI-driven scenarios is of actual quality; and whether the cost of maintaining the rules is affordable. Do not attribute results solely to a single fluctuation in traffic, nor should you equate a platform’s offer of choice directly with a promise of traffic.

For content creators, what’s most worth fighting for isn’t a seemingly all-purpose “reject” button, but the right to make choices that are understandable, maintainable, and verifiable. This round of announcements from Google and Bing has made the issue more concrete; however, there is still a long way to go in terms of implementation and fine-tuning before all participants can follow the same set of guidelines.