Hold off on the new AI opt-out line until you know what it does at Google. Paste it into robots.txt and several AI companies are told one thing: don't train on these pages. It sounds like the safe move for any store. At Google, it can do more than that.
One line can say "don't train on my pages." At Google, it can also say "don't cite them."
What changed on Oct 30
OpenAI, Google and Microsoft each said they will honor a shared robots.txt token, __usage_ai-training, to mark pages as off limits for AI training. OpenAI calls it a "voluntary effort" that "makes it easier for publishers and website operators to communicate the same preference to multiple participating providers at once."
The token comes from an individual IETF draft. PPC Land reports that it is "not endorsed by the IETF," and that Anthropic's engineer, the draft's sole author, said Anthropic has already started building support. Per PPC Land, Meta, Amazon and Apple haven't committed.
Nobody honors it yet. Microsoft starts Feb 1, 2027. Google's crawler docs say early 2027. OpenAI says "coming months." Anthropic gave no date.
The Google catch
Google says that if your site uses the shared token and has no Google-Extended rule, it will read the token as a Google-Extended disallow. Google-Extended covers more than training. In Google's words: "Websites that specifically allow Google-Extended are eligible to appear in responses and receive clicks via grounding in Gemini Apps."
Grounding means your page gets pulled in and cited, with a link, when someone asks Gemini a question. So a store that pastes the line to protect its blog could also take its product pages out of Gemini answers. Google's post doesn't spell out that consequence. It's PPC Land's reading of Google's docs, along with the point that Google has no single setting for "cite me and keep me out of training."
Per PPC Land, AI Overviews and AI Mode in Search run on Googlebot and a Search Console control. This token doesn't touch them.
Microsoft handles it differently. It keeps training and grounding as separate choices, and grounding is set page by page. Microsoft also says: "Neither choice guarantees inclusion, traffic, or payment."
The decision, in six steps
These steps are our suggestions, built on what Google, Microsoft and OpenAI published.
1. Name the three uses
Training changes a future model. Grounding cites your page in an AI answer today. Crawling for search is ordinary Googlebot and Bingbot. The new token speaks to training only, as a preference, and it doesn't block crawling.
2. Sort your pages by type
Product and collection pages: most stores want these cited when a shopper asks an AI. Blog posts, guides, original photos and size or fit research: some owners want these kept out of training. Cart, checkout and account pages are a crawl question, and your robots.txt likely handles them already. Write one line per page type saying what you want for training and for citations.
3. Read what your robots.txt says today
Open yoursite.com/robots.txt. Per the draft text quoted by PPC Land, a blanket User-agent: * block doesn't count as a training preference. The token has to be named in its own group.
On Shopify, the file comes from a template. Shopify's help page says to go to Online Store, click Edit code on your theme, add a new template and pick robots, which creates robots.txt.liquid. Shopify calls this an unsupported customization that its Support can't help with, and warns that mistakes can cost you all your traffic. Have a developer or Shopify Partner make the edit.
4. If you opt out of training, add the Google line on purpose
Google says an explicit Google-Extended rule takes precedence over the shared token. So if you want out of training elsewhere and still want Gemini citations, add a Google-Extended group that allows your pages:
User-agent: Google-Extended
Allow: /
Know the tradeoff. Google-Extended covers both Gemini training and grounding, so allowing it for citations also allows it for training at Google. Pick the one that matters more to you for each page type.
5. For Bing and Microsoft products, set grounding per page
Microsoft's grounding control lives on the page, as a meta tag or an X-Robots-Tag header, with three values: Allow (the default), DisallowExtended and Disallow. Microsoft says DisallowExtended applies to its own products and leaves out its third-party grounding APIs. If your site uses NoArchive or NoCache today, Microsoft says those stop counting as AI preferences after March 1, 2027, and that it will disregard them as soon as it detects the new preferences on your site.
6. Put the dates on your calendar, and write it down
- Dec 15, 2026: Bing Webmaster Tools transition opens, for sites using NoArchive or NoCache (Microsoft).
- Feb 1, 2027: Microsoft starts honoring the token.
- Early 2027: Google's target.
- March 1, 2027: NoArchive and NoCache stop counting at Microsoft.
Nothing changes the day you paste the line, so there's time to get it right. Then write one paragraph: which pages AI may train on, which it may cite, who owns robots.txt, and when someone last checked it. Keep it in the same doc as your agent access policy, which covers the agents that visit and shop.
What this doesn't tell you
The token is a draft, and none of these companies honors it today. The Gemini grounding consequence comes from PPC Land's reading of Google's docs. We haven't seen the Anthropic message ourselves; it reaches us through PPC Land. And this note covers what AI companies do with your page content. Agents that browse and probe your store are a separate question, covered in AI agents are already probing your store.
What it does give you is time to decide, page type by page type, before anyone edits the file.
This week
List your page types. Mark each one for training and for citations. Read your current robots.txt. If you opt out of training, add a Google-Extended allow on purpose. Put the 2027 dates on the calendar. Write the one paragraph policy.
If you want a second pair of eyes on your robots.txt before anyone pastes the line, get in touch.
Linked sources
- OpenAI: Helping websites express AI training preferences (2026-10-30): https://openai.com/index/helping-websites-express-ai-training-preferences/
- Google Search Central (John Mueller): AI training control (2026-10-30): https://developers.google.com/search/blog/2026/10/ai-training-control
- Microsoft Bing (Krishna Madhavan): Your content, your choice: clearer preferences for training and grounding (2026-10-30): https://blogs.bing.com/search/2026/10/your-content-your-choice-clearer-preferences-for-training-and-grounding/
- PPC Land (Luis Rijo): Google, Microsoft, OpenAI and Anthropic back one AI training opt-out (2026-10-31, secondary analysis): https://ppc.land/google-microsoft-openai-and-anthropic-back-one-ai-training-opt-out/
- Shopify Help Center: Editing robots.txt.liquid: https://help.shopify.com/en/manual/promoting-marketing/seo/editing-robots-txt
- Related article: Write your store's agent access policy this week: https://22atlas.com/blog/write-your-stores-agent-access-policy-this-week/
- Related article: AI agents are already probing your store: https://22atlas.com/blog/assume-ai-agents-are-already-probing-your-store/
Get our Articles and posts by email.
Up to one email a day with new Articles and posts from 22atlas.com and sean-weldon.com, plus occasional news about 22Atlas services.
Also on the desk
Technique
Write your store's agent access policy this week
BCG surveyed big retailers and found most have already picked a stance on outside shopping agents. Pick yours on purpose, draw the data line, and put it on one page.
Technique
AI agents are already probing your store. Run this drill
NorForge watched cheap AI agents pick e-commerce targets, check that orders worked, and push for a foothold on the server. Patch StyleSmuggler, list every script on your checkout, and check what your own Always On Agents can touch.