Visibility #41: Publishers Get a Clearer Choice About AI Training


The Visibility Report​
Issue #41 | Week of October 26 - November 1, 2026

Publishers Get a Clearer Choice About AI Training

Publishers have been stuck with a blunt choice: let AI companies use their content or risk disappearing from AI answers. New controls from Google, OpenAI, and Microsoft make that choice a little less blunt.

Microsoft draws the clearest line between model training and grounding, which is the use of current web content in generated answers. A publisher can say no to training without automatically saying no to citations and referral traffic. That still does not create a right to payment or guarantee a place in an answer.

Training and search visibility are becoming separate decisions

Google said it plans to support an IETF AIPREF policy token in robots.txt. If a publisher opts out through the shared token and has not set Google-Extended separately, Google plans to read the signal as a Google-Extended disallow. Existing provider-specific controls remain available.

OpenAI also announced support for the shared approach. Microsoft went further by documenting separate AI Training and AI Grounding preferences. Microsoft says the new preferences begin taking effect as it crawls on February 1, 2027. Sites currently using NoArchive or NoCache must publish the replacement settings before March 1, 2027 if they want those restrictions maintained.

Do not block every bot and call it a strategy. List each crawler and control, make separate decisions about training and grounding, and test whether your CDN or CMS overwrites them. Microsoft says that allowing grounding only makes content eligible to inform answers. It does not guarantee inclusion, traffic, or payment. (Nothing can be that easy.)

People do not prompt the way they search Google

Cate Dombrowski at iPullRank matched 16,310 users across Google and AI-search desktop panels from November 2024. In that sample, questions represented 27% of AI prompts and 8% of Google searches. AI prompts were almost twice as long and contained more personal detail. AI search accounted for 2.7% of the users' combined query activity.

The study covers Perplexity, Copilot, and You.com, not every assistant, and it analyzes one month of US desktop activity. So the safe conclusion is narrow: adding a few words to your keyword list will not capture the qualifiers people volunteer when they ask an assistant for help.

Build prompt sets around the audience, the task, the constraints, and the decision stage. Vary the wording. A short, fixed list can produce a very tidy chart that does not resemble how customers actually ask.

​Source: iPullRank's user-level comparison​

One citation screenshot is not a report

Ahrefs tracked about 37,500 citations across five AI tools for 20 days. At least half disappeared by the next day. Google AI Overviews lost 50%, AI Mode 58%, and ChatGPT 59%; Gemini and Copilot were higher, partly because later answers often cited no source. Roughly 75% of disappeared citations returned within the 20-day window.

A screenshot captures a moment, not a pattern. Measure visibility across repeated runs and a rolling time window. Separate branded fact prompts from category discovery and recommendation prompts. The latter are the harder test of whether your content reaches new buyers.

​Source: Ahrefs citation-stability study​

What the report should show

Weekly reporting should show how often a brand appears across repeated checks, which kind of prompt produced the answer, and whether the brand was cited, mentioned, or recommended. Combining all three into one score hides more than it explains.

From Practitioners

  • ​Marie Haynes: Her review of Alphabet's earnings call connects Google's product language to a practical change in monitoring: AI Overviews, AI Mode, and traditional results are being managed as one Search product. That is her interpretation of the call, not a separate Google announcement.
  • ​Ethan Crump at Foundation: His Profound workflow starts with buyer questions, separates citations from recommendations, and turns visibility gaps into page and distribution work. It is an agency playbook built around one vendor's data, not a neutral comparison of tools.
  • ​Slobodan Manic: Shopify's developer preview can attach commissions to promoted products returned to shopping agents. Merchants and agent developers need to separate recommendation quality from paid placement and disclose the commercial relationship.

From the Tool Blogs

More From This Week

​Source: Search Engine Journal reporting on Reddit's statement​

What to watch

  • Whether the shared training preference reaches a final standard and how each provider maps it to existing controls.
  • Whether Google gives publishers a separate control for grounding, citations, or AI-answer visibility.
  • Whether AI Mode's ad expansion closes the monetization gap Google described without displacing useful source links.
  • Whether Business Profile agents show merchants the questions, answers, handoffs, and resulting leads.

What to do next

  • Inventory every AI crawler rule in robots.txt, CDN settings, and page-level directives.
  • Write separate policies for model training, live grounding, search indexing, and agent access.
  • Build prompt samples around real customer constraints, then rerun them over several weeks.
  • Report citations, mentions, recommendations, visits, and conversions as separate measures.
  • Check whether local profiles and shopping feeds contain the facts an agent needs to answer and hand off accurately.

One thing to do this week: audit your robots.txt and page-level AI directives, then write down separate decisions for training and grounding before the new shared controls reach production.

​Subscribe to The Visibility Report for a source-checked briefing every Tuesday.

The Visibility Report​
Will Scott
Search Influence

The Visibility Report

Join 500+ digital marketers getting weekly AI visibility tools, tactics, and search updates. Free • 5-min read • No spam.

Read more from The Visibility Report
A glossy white robot stands at a permission gate separating ordinary indexing from AI agents taking action.

The Visibility ReportIssue #43 | Week of November 9 - November 15, 2026 Google Search Is Querying More Than Your Query Google's newest Search model can split one complicated question into several related searches, then build an answer from what it finds. That makes the pages behind the answer harder to predict, but it does not make the work mysterious. Subscribe to The Visibility Report for a source-checked briefing every Tuesday. Google expands query fan-out with Gemini 4 Hydrogen Google...

A glossy white robot places a cyan verification marker into a brass sidewalk locator outside a warmly lit storefront while three arrow cards point toward the business.

The Visibility ReportIssue #42 | Week of November 2 - November 8, 2026 ChatGPT Webmaster Tools May Be Coming OpenAI appears to be testing something search marketers have wanted for years: a first-party view of how businesses appear in ChatGPT. Think of it as ChatGPT Webmaster Tools, with one important caveat. This is an unconfirmed, unfinished feature seen by a third party, not an OpenAI product announcement. Subscribe to The Visibility Report for a source-checked briefing every Tuesday....

A human-scale white robot stands beside a much taller lighthouse as its beam guides boats toward a harbor.

AI Traffic Stays Small as AI Answers Reach Buyers The Visibility Report #40 | Week of October 20-26, 2026 AI assistants still send very little identifiable referral traffic compared with Google. Marie Haynes found that ChatGPT sessions equaled 0.62% of Google organic volume across 100 sites and more than 145 million sessions. The same week, OpenAI explained how ChatGPT ads use a separate language model to read multi-turn intent, retrieve candidates, and run an auction. Those findings put the...