01 The market
AI answers cited news sites twice as narrowly as search
An AI answer draws from a shorter list than a page of search results, so being on the long list stops counting.
Carragher and Carley counted news outlets rather than professional practices, which is the limit to hold on to before anyone sells you the finding.
Peter Carragher and Kathleen M. Carley published a measurement on 29 October of who owns digital news and who gets cited. They built an ownership record for 9,761 digital news outlets drawn from Wikipedia and Wikidata, derived concentration figures for 77 countries, and compared those against what AI systems actually cite.
The finding that matters most is a comparison. Citation distributions from major AI systems are up to 2.08 times more concentrated than traffic, which means the answer hands a larger share of its mentions to a smaller number of owners than the open web hands out visits. Counting outlets instead of owners understates it, with owner-level figures a median 1.30 times higher.
Two further results are worth stating plainly. The authors calculate that if Google AI Overviews mediated all organic search, the rise in United States concentration would be several times that associated with highly regulated media mergers, and they are explicit that this is a hypothetical rather than something observed. Separately, the systems preferentially cited state-controlled outlets when the question was asked in that country’s own language.
Forever Cited exists because of the pattern this paper measures: ask an assistant who to hire, and the answer is two or three names with everyone else absent. The honest caveat is that this study counted news publishers and not accounting firms, so the figure does not transfer to your own market until somebody measures your own market.
What to do about it
Do not accept a concentration figure for your industry on the strength of a news-media study. If a vendor quotes you this paper, the question to ask is which businesses were counted.
If you want the number for your own market, it has to be measured there: the same questions put to the same assistants, with the names in your market listed in advance.
Source arXiv, Peter Carragher and Kathleen M. Carley, “Who Owns Digital News, and Who Gets Cited: Measuring Media Concentration Across 77 Countries and Seven AI Systems,” 29 October 2026 · arxiv.org/abs/2610.40164 (an ownership dataset of 9,761 digital news outlets with traffic-weighted concentration indices for 77 countries; the Google AI Overviews figure is a stated hypothetical, and the scope is news publishers rather than professional services)
02 The research
Physicians marked Google’s AI health answers at 38.2%
Being the source an AI answer names is not the same as being quoted correctly, and the size of that gap is now measured.
Google ranks health pages on credibility, and Google’s own AI answers cited less credible pages than Google’s ordinary results did.
Sam Agnoli, Brigid Reilly, Lily Barna and Emőke-Ágnes Horvát evaluated 3,611 benchmark health questions, published 29 October, comparing each AI Overview against that question’s top five ordinary results. An overview appeared for 96.4% of the questions, so this is not an edge case.
The first number undoes an assumption almost every business holds. Only 44.4% of the top five results were cited by the corresponding AI Overview, which instead overrepresented commercial health publishers such as WebMD and Healthline at the expense of academic and government sources such as the Cleveland Clinic and the National Health Service. Ranking in the top five stopped being the same thing as being on the shortlist.
Then the sourcing. After correcting for classifier error the authors estimate that 56.7% of cited statements were supported by the page they cited, and that a further 20.8% of all statements were factual claims carrying no citation at all. Pages cited by an overview scored lower than the ordinary results on four of six established credibility criteria.
Graded against rubrics written by physicians, the overviews earned 38.2% of the available points, with accuracy at 54.2% and completeness at 21.6%. Among the overviews whose rubric had anticipated a specific inaccurate or potentially harmful claim, 22.2% contained the harmful content. The authors note that accuracy was unrelated to how well statements were sourced, so a well-cited answer was not a more accurate one.
For a surgical practice or a clinic this is the argument for owning the page an answer is built from. For everyone else the transferable number is 44.4%: whatever an AI answer is assembled from, it is not simply the first page of results, and a vendor reporting your search position is not reporting your position in the answer.
What to do about it
Put a question your customers actually ask into Google and read the answer against the pages it names. You are checking two separate things: whether you are cited, and whether what the answer says about you is what your page says.
If the answer states something your own page does not support, the work is on your page. That is the only side of this you control.
Source arXiv, Sam Agnoli, Brigid Reilly, Lily Barna and Emőke-Ágnes Horvát, “Does the Citation Matter? Quality, Grounding, and Physician-Rubric Accuracy of Google’s AI Overviews for Medical Information,” 29 October 2026 · arxiv.org/abs/2610.39339 (3,611 benchmark health questions; the sourcing figures are corrected for classifier error by the authors; health information only, so the accuracy numbers describe medical answers and not answers about other fields)
03 The platforms
Google, Microsoft, OpenAI and Anthropic adopted one training signal
One instruction in one file will soon tell four AI companies at once whether they may train on your site.
The internet standards proposal behind this is not finished, and one of the people writing it said so in the thread announcing it.
Google announced on 30 October that it plans to voluntarily adopt a proposal under consideration at the Internet Engineering Task Force, the body that sets internet standards, alongside Bing, OpenAI and Anthropic. In Google’s words the control “will help websites express their AI training preferences across multiple AI providers in a single step”.
The mechanism is a shared token in robots.txt, the small public text file at the root of almost every website that tells automated visitors what they may do. Instead of writing a separate instruction for each AI company, a site owner writes one. Google’s crawler documentation now carries a note that it intends to support the shared signal, written as user-agent: __usage_ai-training, in early 2027.
The part to read twice is what else that one instruction moves. Google says that when a site uses the token to opt out of training and does not separately specify Google-Extended, it plans to treat that as if the publisher had disallowed Google-Extended. Google also says that “Websites that specifically allow Google-Extended are eligible to appear in responses and receive clicks via grounding in Gemini Apps”, so a single-step refusal of training can reach the control tied to appearing in Gemini’s answers.
A law firm that told its web developer a year ago to stop AI using its content may already have set that control, and nobody would have recorded it as a visibility decision at the time. Google says existing company-specific controls keep working alongside the new shared one, which is what makes the combination worth reading rather than assuming.
Nate Hake, who says he spent the past year participating in the working group, wrote in the same thread that “it’s not actually done at the ietf level” and that “some significant areas of dispute remain”. Google’s own wording is that it aims to support the control in 2027 and will provide updates, which is a plan rather than a shipped feature.
What to do about it
Open your own robots.txt now, before any of this arrives. Type your domain followed by /robots.txt into a browser; the file is public and reading it takes a moment.
Look for any line mentioning Google-Extended. If one is there and you did not ask for it, that is the line to raise with whoever maintains your site, because from early 2027 it will also be read as your answer to the shared signal.
Source Search Engine Roundtable, Barry Schwartz, “Google Adopts New AI Training Control,” 30 October 2026 · seroundtable.com/google-ai-training-control-42430.html (reporting Google’s statements, the note added to Google’s crawler documentation, and the working-group remarks quoted above; the proposal is a draft and Google dates its own support to 2027)
04 The deadline
Microsoft separated AI training from AI quoting, with dates
You can now let AI quote your business without letting it train on you, but only if you say which you mean.
1 March 2027 is the deadline, and it falls on a setting most site owners have never looked at.
Microsoft announced new Bing AI content controls on 30 October. Krishna Madhavan, Principal Product Manager, Microsoft Bing, wrote that they offer “newer, simpler, and clearer preferences that allow you to make choices about training, grounding, or both while remaining eligible for Bing search results”.
There are three settings, and they do different jobs. AI Training allows or prevents your content being used to train AI models, while AI Grounding allows or prevents it being used to help generate AI answers.
The third is narrower. Limited Grounding allows your content to support Microsoft’s own AI products but not its third-party grounding services, which is the setting for an owner who wants to be quoted by Copilot and nowhere else.
Those were one decision and are now two, which is the sentence to take away: “This means you can block AI training but still appear in AI answers”. The reverse carries a cost Microsoft states openly. Blocking AI answers from using your content “can reduce your visibility in AI experiences”, while traditional Bing rankings are unaffected.
Four dates came with it. On 15 December 2026 temporary grounding controls become available in Bing Webmaster Tools for eligible sites using the older NoArchive or NoCache directives, and on 1 February 2027 Microsoft begins honoring the new preferences as it crawls pages.
Then the two that end it. On 1 March 2027 publishers using the old directives must have configured the new preferences if they want to keep their restrictions, and on 30 September 2027 the temporary grounding preferences expire.
The March date is the one that creates work, and it creates it for people who do not know they are included. NoArchive and NoCache are old instructions a developer may have added years ago to stop search engines keeping a copy of a page, and a real estate brokerage whose listing pages carry one has a restriction that lapses rather than persists if nobody restates it.
What to do about it
Ask whoever maintains your site one question before December: does any page carry NoArchive or NoCache. If the answer is yes, 1 March 2027 belongs in your calendar; if it is no, you can stop here.
Bing Webmaster Tools also reports citations, cited URLs and grounding queries. That is the one place in this story where you can see what is already happening rather than set what happens next.
Source Search Engine Roundtable, Barry Schwartz, “Microsoft Bing Separates AI Training From AI Search Grounding With New Controls,” 30 October 2026 · seroundtable.com/microsoft-bing-ai-controls-42429.html (reporting Microsoft’s announcement and quoting its published timeline; the grounding controls it describes as temporary expire on 30 September 2027)
05 The money
OpenAI began rewriting advertisements on 30 October, in 89 countries
The advertisement OpenAI reformats on your behalf now runs in 89 countries, and no version of it needs your approval.
This brief gave readers the date on 21 October, and the useful question now is what to check rather than what to stop.
On 21 October this brief reported that OpenAI would enable creative enhancements for new and existing self-serve campaigns starting Friday, 30 October, and that declining had to be done before the date rather than after. Friday was 30 October. OpenAI’s weekly advertiser email that day listed creative enhancements as being auto-enabled, so the change went in on schedule.
The same email carried the scale. ChatGPT Ads went live in 26 more countries and are now available in 89 countries worldwide, with over 100,000 advertisers. Barry Schwartz, who published the email, noted that some of those figures were already old, which is a fair reading of an announcement restating its own milestones.
Two format changes matter more than the country count. OpenAI is rolling out a wider standard ad card that uses more of the conversation width on desktop, and says the format “uses existing campaigns and creative, so no advertiser action is needed”. Multi-image ads are now in open beta to all advertisers.
No advertiser action needed is the phrase to sit with, because for a regulated advertiser the action is the point. An attorney advertising in the United States answers for every version of an advertisement under the rule against false or misleading communications, and a registered investment adviser answers for it under the Securities and Exchange Commission marketing rule. An auto-enabled rewrite produces versions nobody approved, and your own advertising rules apply to the new format exactly as they applied to the one you wrote.
None of that is an argument for leaving the platform. The setting has quietly become the approval step, so knowing which one you are on now belongs in the record rather than in the technical notes.
What to do about it
Open OpenAI’s Ads Manager and read the creative enhancements preference at account level, then again at campaign level, because campaign settings take priority over account settings.
If your field requires a record of what you approved, capture the setting and the current creative today. The point of a record is that it exists before anyone asks for it.
Source Search Engine Roundtable, Barry Schwartz, “OpenAI’s ChatGPT Gets Wider Ads, Multi-Image Ads Expands, Expansion & More,” 30 October 2026 · seroundtable.com/chatgpt-ads-update-42427.html (reporting OpenAI’s weekly advertiser email; the author notes that several of the figures restate earlier announcements)
