The AI training opt-out and the first independent price of a ChatGPT ad click landed in the same news cycle, and the market read them as two stories. They are one. On October 30, 2026, Anthropic, OpenAI, Google and Microsoft agreed to honor a single robots.txt line that lets a publisher refuse AI training. Days later, a working paper priced a ChatGPT ad click at about $0.26 against a known price of $1.72. The industry has standardized a switch for the archive and priced the slot beside the answer. Neither touches the publisher page the answer is actually built on.

  • The common __usage_ai-training token, backed by Anthropic, OpenAI, Google and Microsoft on October 30, 2026, covers model training only; grounding, the live use that builds AI answers, sits outside it.
  • Kaiser and Schulze's October 2026 working paper finds a paid ChatGPT visit earned $0.89 in US traffic against $2.45 for paid Google search, putting the parity price of a ChatGPT click at $0.26.
  • Organic ChatGPT referrals earned $3.27 per US session, the highest of the five channels measured, so the value sits inside the answer, and the answer is assembled from publisher content.
  • Brands are paying more than six times parity for a card outside the answer, and publishers can now refuse the one use that never paid them anyway. Both sides are pricing the wrong unit.

What the training and grounding split actually means

AI training is the use of web content to change a model's learned parameters; AI grounding is the use of that same content at answer time to build one specific response, usually with a citation pointing back. The October 30 common signal, drafted by Anthropic engineer Kevin Kelley for the IETF AI Preferences group, covers training. Only Microsoft published a separate grounding preference: page-level, Bing and Copilot only, live from February 1, 2027, with an explicit warning that opting out reduces visibility in AI answers. Microsoft also states that neither preference creates a licensing or payment arrangement.

The crawl-for-traffic contract versus the ground-for-free default

The open web ran on one implicit contract: allow Googlebot, get ranked, get the click, sell the pageview. robots.txt was the publisher's only lever, and the price of content was discovered in the ad auction that followed the click. In the agentic web the crawl still happens, but the output is an answer rather than a referral, and the auction sits beside that answer on a card most users ignore. The publisher now has an interoperable way to refuse the archive, and still the same allow-and-hope default on the live use that builds every answer. The link economy priced the click. The citation economy has priced the slot. Nobody has priced the source.

What Search Engine Journal and Nieman Lab reported, and what we verified

Search Engine Journal (November 2, 2026) reports the Kaiser and Schulze finding that US paid ChatGPT visits earned $0.89 against $2.45 for paid search, converted at 0.37% versus 1.17%, and that 73 of 87 dual-channel advertisers recorded zero sales from ChatGPT ads. We cross-checked the figures against the authors' own site, Search Engine Land and PPC Land; they match, with the caveat that this is a version-one working paper under review, its data window closed on August 22, 2026, and one advertiser carries 76% of US paid ChatGPT revenue, which widens the parity range to $0.05 to $0.57. Nieman Lab (November 2, 2026) reports the four-company commitment to the common robots.txt signal; the IETF draft confirms it is a training-preference token, and Microsoft's own October 30 post confirms grounding is a separate, proprietary control.

Why a switch on the archive and a price on the slot still leave the answer unpriced

The strongest case for the common signal, and for patience on ChatGPT ads

The standard is real progress. One line replaces a dozen crawler rules, Anthropic already honors it, and Meta endorsed it publicly while delaying its own broad-scraping terms. On the ads side, OpenAI reported on August 31 an e-commerce advertiser with three times return on ad spend over 28 days, DoubleVerify reported Weight Watchers' cost per acquisition 15.3% below benchmark, and Criteo says LLM-referred shoppers convert at close to twice search rates in several retail categories. Every new channel looks bad on last-click in its first year. Paid search did in 2002.

Why the cited page, not the sponsored card, is the unit that carries value

The paper's own numbers locate the value. Organic ChatGPT visits, the ones that originate inside the answer, earned $3.27 per US session, more than paid search. The card beside the answer earned $0.89. The authors conclude the card is cheap because it is separated from the answer, and that moving ads into the answer would buy value per click with user trust. That is a false binary. The answer is already built from publisher pages; a brand present on the cited page is inside the answer without corrupting it. Meanwhile the opt-out standard formalizes the publisher's right to refuse the use that pays nothing, training, while leaving the use that produces $3.27 sessions, grounding, on a default of free. A switch on the wrong unit is not leverage. Microsoft said it plainly: no inclusion, traffic or payment is guaranteed either way.

What this means for brands and for publishers

For CMOs, media buyers and agencies: stop paying six times parity for the card nobody buys from

  • Run the parity-price calculation on your own account: ChatGPT revenue per visit divided by paid search revenue per visit, multiplied by your search CPC. The study's median advertiser lands near $0.26.
  • Treat the sponsored card as a reach test, not a performance line; the median website in the paid panel recorded 0.00% conversion from it.
  • Move the AI visibility budget toward the sources: map which publisher pages ground the answers in your category and buy presence there. That is where the $3.27 sessions start.
  • Keep paid search funded. The same study shows it performs in line with direct traffic on every financial measure.

For publishers: take the training opt-out, then price the grounding

  • Add the __usage_ai-training line. It costs nothing, it is interoperable, and it closes a loophole Meta's terms were about to widen.
  • Do not mistake it for a grounding opt-out. Only Bing offers one, it is page-level, and Microsoft warns it cuts your visibility in AI answers.
  • Measure grounding: which engines cite which pages, for which prompts. That inventory is what brands will pay to sit beside.
  • Demand a model that pays per citation, not a preference file that, by its own wording, pays nothing.

Three signals that decide who pays for grounding by 2027

First, whether OpenAI and Google follow Microsoft and expose a grounding preference, which would make the citation something a publisher can withhold and therefore price. Second, whether OpenAI moves ads into the answer; the paper estimates it would need roughly 60 ads per commercial prompt at today's click value to reach its 2030 target, so the pressure to raise value per click is structural, not tactical. Third, the Wikimedia amicus brief reported by PPC Land on November 2, 2026, asking a court to rule that ChatGPT's web retrieval is not fair use. If the market does not price grounding, litigation will.

Conclusion

Hold on to this: the AI training opt-out gives publishers a switch on the archive, and the ChatGPT ads paper gives brands a price on the slot, yet the unit that carries the value, the cited page inside the answer, has neither a switch nor a price. Smalk AI builds the market for that unit: Generative Engine Advertising, native ads for AI agents placed on the publisher pages AI engines ground their answers on, so brands buy presence where the recommendation forms and publishers are paid per citation rather than per preference file. What to watch next: the first engine after Microsoft to expose a grounding preference. That is the moment the citation becomes withholdable, and withholdable means priceable.