AI publishers gain more choices—and advertisers need tighter control

A shared training preference is taking shape

On October 30, Google and OpenAI announced support plans for a common AI-training preference in robots.txt. The associated IETF Internet-Draft proposes a shared policy token, __usage_ai-training, so publishers can communicate preferences to participating providers without maintaining only company-by-company instructions.

This is work in progress, not a universally deployed standard. Google targets 2027; OpenAI says implementation is intended in the coming months and existing crawler controls will remain.

Google’s planned interpretation deserves particular attention. Without an explicit Google-Extended rule, a general training opt-out would be treated as disallowing Google-Extended. Google says explicitly allowing Google-Extended preserves eligibility for grounding in Gemini Apps—the use of web content to inform responses. Eligibility does not guarantee an appearance.

Illustrative application: a Realtor publishing original neighborhood research may want visibility in AI answers while limiting model training. That business needs provider-specific choices before someone copies a generic “block AI” configuration onto its site. ^1,2,3

Microsoft sets a transition timetable

Microsoft’s October 30 announcement adds independent training and grounding preferences, without changing eligibility for traditional Bing results. The new controls start taking effect as Microsoft crawls from February 1, 2027.

Publishers using NoArchive or NoCache need particular care: those older signals stop serving as training or grounding preferences after March 1, 2027. Microsoft describes both replacement settings and a temporary grounding transition in Bing Webmaster Tools.

For an ecommerce publisher, the practical task is to inventory existing restrictions and decide whether product information should inform generated answers. Allowing that use does not promise citations, visits or payment. Restricting it can reduce AI visibility even while ordinary search eligibility remains. This is a planned migration to review, not evidence that a client site has already changed. ⁴

ChatGPT Ads adds visual options and automatic variations

OpenAI introduced a multi-image ad format as an open beta on October 26. It accepts four to six square images; each image links to the same destination, with “Learn more” as the current call to action. A merchant could illustrate several uses of one collection, but should not assume each image has its own product link. ⁵

A separate change took effect October 30: creative enhancements became automatically enabled for new and existing self-serve and BPO campaigns, preserving existing opt-outs. This default-on change does not affect sales-managed accounts. Enhancements can change copy, translations, image sizing and destination URLs. Every variation is not previewable before delivery.

That makes accurate landing pages an operational requirement. Before testing, an eligible retailer should check campaign settings and whether automated wording or destination changes are acceptable. Paid placements remain separate from ChatGPT’s answers; campaign performance changes alone cannot establish an enhancement’s effect. ⁶

For Realtors, OpenAI’s existing prohibition on advertisements for individual housing sales or rentals still applies. This is prior policy, not a new restriction announced this week. ⁷

Generated pages need an intentional search policy

Google’s October 26 guidance update addresses platforms hosting user-created pages, generated outputs and interactive tools. It recommends keeping content intended only for direct sharing out of search and watching for unusual indexing or query activity.

For a report generator, the distinction is practical: a temporary customer output and a reviewed public resource serve different purposes. Decide which deserves discovery before hundreds of URLs accumulate. This guidance update does not announce a new ranking algorithm. ⁸