User-Agent: __usage_ai-training Disallow: /The example from the IETF draft “Machine-Readable Training Preference Signal”, published October 30, 2026.
Those two lines are what OpenAI said on Friday, October 30, it will honor. One entry in your robots.txt, the small text file every website carries, and every AI company that signs on is asked to keep your pages out of its training. Don’t expect it to do anything yet. OpenAI’s own post says it aims to implement the signal “in the coming months”, and the draft it points to is marked as expiring in May.
Most of the write-ups stop at “one line for every provider”. They skip two things.
First, the draft is blunt that a wildcard doesn’t count. Its words: “wildcard groups (i.e., User-Agent: *) MUST NOT be interpreted as matching the policy token.” If your file only has the star, you’ve said nothing.
Second, nobody looked at what a normal site’s file actually says. I read the robots.txt on two of the sites I run this morning, scoy.ai and livenewschat.eu, and both carry only that wildcard. Here’s the whole of the first one:
User-agent: *
Allow: /
Sitemap: https://scoy.ai/sitemap.xml
Sitemap: https://scoy.ai/guides/wp-sitemap.xmlA stock WordPress install writes the same star, so yours probably reads the same. Does it? Open yoursite.com/robots.txt in a browser and look. It takes a minute.
What I’d do with that minute depends on what you want:
- If you’d rather keep your pages out of training, add the two lines at the top of this post. The draft lets you re-allow one folder underneath them, a press kit for example, with an Allow line.
- Keep any GPTBot or other crawler lines you already have. OpenAI says the new signal “will not replace the controls we already provide through robots.txt.”
- Don’t expect proof. The same draft says the token “is not a directive for crawling behavior” and that “private agreements MAY override” it. It’s a preference, not a lock.
What I can’t tell you is who else honors it, or when. OpenAI’s post names only the IETF working group it sits in. The lists naming Google and Microsoft come from secondary write-ups, so one source on that so far, and it isn’t the companies.
Separate story, and this one has a date you can plan around. Google’s Gemini API pricing page, read this morning, says Gemini 4 Argon costs double from December 1. That’s Google’s top model, the one a lot of “AI-powered” tools quietly run on. Every row on the page carries the same two dates:
| Standard tier, per million tokens | Through November 30 | From December 1 |
|---|---|---|
| Input | $2.00 | $4.00 |
| Output, thinking included | $10.00 | $20.00 |
| Cached input | $0.10 | $0.20 |
Tokens are the word-pieces the model counts, and Google bills every one, including the model’s own thinking. So a job that sends in a million tokens and gets a hundred thousand back costs three dollars today and six dollars in December. If a tool you pay for runs on Argon, has the vendor said anything about December? I’d ask. The page carries no promo wording, just the two dates, so I can’t tell whether Google extends it again.
One more date, one source. OpenAI’s help page, updated Saturday, November 1, says that from December the Plus, Pro and Business plans share one usage allowance between ChatGPT and Codex, its coding agent. Instant chat won’t count against it. The page gives no numbers, so if you lean on Codex, your own coding work will come out of the same pool as your chats, and nobody can tell you yet how big that pool is.
If you want me to look at your robots.txt with you, the free call is thirty minutes. Send the file, or just the domain name.