Google, Microsoft and OpenAI Built One AI Opt-Out Switch. Most Food Brands Should Leave It Alone.

Google, Microsoft and OpenAI agreed to honor a shared robots.txt line that tells AI companies not to train on your website. A small food brand should mostly leave it off. In an October 2026 Euromonitor test of ChatGPT and Gemini, brand-owned websites supplied 3 percent or less of the citations. Blocking your site hides the part AI barely reads.

Updated November 2, 2026.

What did Google, Microsoft and OpenAI agree to?

They agreed to read the same opt-out line in a site's robots.txt file, the plain text file that tells crawlers what they may and may not touch. Search Engine Journal reported on November 1 that the line is a user-agent group called __usage_ai-training. Put a Disallow under it and you've told every participating company, in one place, that you don't want your pages used to train its models.

Anthropic, the company behind Claude, submitted the draft to the Internet Engineering Task Force, the group that writes internet standards. Paul Thurrott wrote on October 30 that all four companies have signed on. Thurrott also called the change "largely symbolic," because the companies had already collected the web at scale long before the switch existed.

None of it works yet. Google is targeting early 2027. Microsoft starts honoring the new preferences on February 1, 2027. OpenAI says "in the next few months." Search Engine Journal reported that none of the three companies supports the line today.

What does the line actually switch off?

Training, and only training. That's the part most of the coverage skips, and it's the part a food brand needs to get right.

Training means a company uses your pages to change what its model knows. Search is a different thing. So is grounding, the step where an assistant goes and reads a live page to answer somebody's question right now. Google says the shared line will be treated like a block on Google-Extended, which covers Gemini training and grounding in the Gemini apps and does not touch Google Search. AI Overviews and AI Mode sit on a separate control in Search Console. Microsoft set grounding up as its own preference, and it defaults to Allow.

One more detail that will catch somebody. A wildcard group, the User-agent: * line at the top of most robots.txt files, can never match the new token. If you already disallow everything under the wildcard, that doesn't count as a training opt-out. You'd have to write the new line on purpose.

Why haven't I blocked the bots on my own site?

I pulled gallucci.net/robots.txt this morning. I've owned the domain since 1996, and the sitemap lists 552 blog URLs as of today. GPTBot, ClaudeBot, Google-Extended, Applebot-Extended, Meta-ExternalAgent and about two dozen other AI crawlers are named at the top of the file. Every one of them gets the same rules as everybody else. Stay out of the account pages, the site search and a few system folders, and read whatever else you like.

That's a choice, and I made it on purpose. My site runs on Squarespace, which has a checkbox in its crawler settings to block known AI crawlers. Mine is unchecked. I write those posts so that people who hire for social media strategy and content creation can find out what I think. If an assistant learns that from my archive and repeats it to a founder in Austin at 11 at night, the post did its job. I don't sell the articles. They're the sample.

Most food and beverage brands are in the same spot. Your product pages, your recipe page and your founder story aren't the product. The jar is the product. Every page on your site exists so somebody buys the jar, and an assistant that knows your jar exists is working for you for free.

Who should flip it?

The people whose words are the product. A recipe developer who earns a living on ad views from recipe pages has a real reason to say no. So does a food writer, a cookbook author, a photographer licensing images, or a publisher with a paywall. If the page itself is what you sell, the training opt-out is worth writing as soon as the companies start reading it.

A brand that sells chips, salsa, coffee or canned water doesn't fit that list. The risk to a brand like that is staying invisible, and nobody copying the jar off a web page is the problem.

Where do assistants learn about a food brand, then?

Mostly somewhere else. Euromonitor ran 1,000 scalp care prompts through ChatGPT and Gemini in the UK and US, and Rabia Yasmeen's October 7 write-up found that 58 percent of Gemini's citations came from retailer product pages. Brand-owned sites were 3 percent or less on both assistants. That study is scalp care, not snacks, but I'd bet the retail-page pattern looks the same in a grocery aisle.

So the page an assistant reads about your salsa is probably your listing on a retailer's site, written by someone in a hurry from a sell sheet. That's the page to fix first. If the retailer listing says "great taste, made with quality ingredients," the assistant has nothing to repeat except the category. If it says the chiles come from a farm in Hatch and the recipe has five ingredients, the assistant has your story. I made the same case for auditing what the engines already say about you, and for what Google shows when someone searches your brand name.

What to do this week

1. Open yourdomain.com/robots.txt in a browser. Look for GPTBot, ClaudeBot, Google-Extended and CCBot. Done when you know whether each one is allowed or disallowed, and who on your team made that call.
2. Decide what your site sells. If it sells the jar, leave training open. If it sells the words or the photos, plan to write the __usage_ai-training line once Google and Microsoft start reading it in 2027. Done when the decision is written down in one sentence.
3. If you do opt out, opt out of training only. Leave Googlebot, Bingbot, OAI-SearchBot and Claude-SearchBot alone, because those are how you show up when somebody asks a question today. Done when the search crawlers still have access after your edit.
4. Rewrite your top three retailer listings so they carry the ingredients, the place and the person. Done when a stranger could tell your product from the house brand by reading the listing alone.
5. Ask ChatGPT, Gemini and Claude the question your customer asks, word for word, in English and in Spanish. The answers change with the language. Done when you have the three answers saved with today's date, so you can check them again in 90 days.

The big brands will have a lawyer write their robots.txt. An independent brand gets one decision out of this, and it's a cheap one. Let the assistants read about you, then make sure what they read sounds like you wrote it. If you're paying a food and beverage marketing agency, ask who owns your robots.txt and your retailer copy, because those two files now tell AI what your brand is.

I fix the retailer listing first with a new food brand. If yours reads like the house brand wrote it, DM me.