01 The rules Google wrote down
Google says unreviewed AI text shows little to no effort
If someone generated your pages with AI and nobody checked them, Google’s own guidance now counts that as little to no effort.
Google rewrote two quality documents on October 1 and 2, and named the four attributes its reviewers judge a page on.
On October 1 Google updated its guidance on using generative AI content, and wrote: “It is critical to manually factcheck and review all AI-generated content.” Its reason is stated plainly: “generative models don’t retrieve facts, but predict a likely sequence of words based on their training data,” so outputs may contain inaccuracies. Google rarely uses the word critical in a help page.
The next morning Google rewrote a second document, on helpful content, and added something it had never published there. It named the four attributes its human Search Quality reviewers are trained to judge: effort, originality, talent or skill, and accuracy.
The definition of effort is the one to read twice. Google wrote that “using generative AI to produce large amounts of text without manual oversight or curation represents little to no effort”, and added that “Attribution or giving credit to other sources doesn’t replace the need for original effort.”
Accuracy names the reader of this brief almost exactly. For topics that could significantly impact people’s lives or well-being, Google now says the content “must be highly accurate and consistent with established expert consensus.” That covers an accounting practice writing about a filing deadline, a veterinary hospital writing about a drug dose, an insurance brokerage writing about what a policy pays, and a tax adviser writing about anything at all.
One caveat, because it changes what this is. Google said it made the change to “get our documentation in sync with our presentations we use at our developer events”, so this is written-down practice rather than an announced ranking change. Barry Schwartz, who found both edits, notes they landed during the September spam update and wonders aloud whether the two are connected, which is his reading rather than anything Google has said.
What to do about it
Open the last three things published under your name. Ask one question about each: did a person who actually knows the subject read it line by line before it went up, and could they defend every number in it?
Where the honest answer is no, that is the page to fix first. The standard Google just wrote down for anything touching money, health or legal exposure is expert consensus, not a plausible-sounding paragraph.
Source Search Engine Roundtable, “Google Helpful Content Doc With New Main Content & EOT/SA Sections,” October 2, 2026 · seroundtable.com · the AI-content guidance, October 1
02 The deadline
Google will let some ads carry a different brand name
From October 16 you can advertise under the brand people know you by, if you can prove you are tied to the website it opens.
Google’s business name policy changes on October 16, and it shuts out resellers, booking intermediaries and affiliates by name.
Google has told advertisers it is loosening one of its strictest ad rules. Google wrote: “In October 2026, Google will update its Business name requirements policy to allow for differences between the business name and the destination domain in certain limited cases.” The email Google sent advertisers puts the date at October 16, 2026.
Until now the name shown on the ad had to match the site the ad opened. Three conditions replace that. The business name has to “accurately reflect the recognized name or brand of the advertiser”, there has to be “a verified direct relationship between the advertiser and the domain owner”, and the advertiser’s products or services have to be directly offered on the destination site.
This matters most where the name people know you by and the name on your website are genuinely different. An independent insurance brokerage advertising under a carrier brand, a real estate agent advertising under a brokerage brand, a practice that joined a group and kept its own name: all three have been fighting this rule for years.
The exclusion list is where to be careful, because Google names categories rather than companies. “Third-party resellers, independent booking intermediaries, affiliate distributors, and secondary sellers do not qualify,” and may not use the standalone brand name of what they sell. A charter broker selling seats on somebody else’s aircraft is an independent booking intermediary, and the brand on the tail is not theirs to advertise under.
What to do about it
If you advertise under a name that does not match your website, write down now which of the three conditions you can actually evidence, and who holds the domain. The verified relationship is the one people cannot produce on request.
If you sell somebody else’s brand rather than your own, read the exclusion list before October 16 rather than after. It is written by category, and the category is what Google will apply.
Source Search Engine Roundtable, “Google Ads To Allow Some Advertisers To Use Different Business Names,” October 2, 2026 · seroundtable.com
03 The research
Five researchers argue the first source cited wins the conversation
Whoever an AI names first is far more likely to be named again as the customer keeps asking, which is not how anyone measures this.
The paper is a mathematical argument rather than a measurement, and what it argues is that the way this industry counts AI visibility is the wrong way.
Junwei Yu, Jieyu Zhou, Mufeng Yang, Yepeng Ding and Hiroyuki Sato of the University of Tokyo published a paper on September 30 about how a business becomes the one an AI assistant names. Their claim is short: “a source cited early becomes substantially more likely to be cited again.” They call it conversational capture.
They describe two ways it happens, and both are familiar once stated. The machine keeps the earlier answer in view and leans on the same source for the next one, and the person, now holding that name, asks their follow-ups about it. Between them the first citation gets paid twice.
Here is the honest limit, and it is a large one. There is no sample: no businesses were tested and no real assistant was measured. The paper is a mathematical model whose figures come from what the authors call a model-derived illustration, so treat it as an argument about measurement rather than evidence about any engine.
What that argument says is still worth your morning. In their illustration the knock-on effect becomes larger than the original citation, and ranking businesses on a single question produces a different order from ranking them across a whole conversation. The two orders, they report, line up only weakly.
Which is uncomfortable here, so it is worth saying out loud. Every company selling AI visibility measurement, Forever Cited included, asks each question once and counts who gets named. If this paper is right, that method quietly flatters whoever happens to show up in a one-shot answer, and nobody in this industry has published a measurement across a real conversation yet, ourselves included.
What to do about it
Next time you check whether an AI assistant names your business, do not stop at the first answer. Ask the follow-up a real customer would ask, then the one after that, and watch whether the first name it gave keeps coming back.
If a competitor appears first and stays, that is the thing to work on, and it is the same work either way: being the source the answer was built from.
Source Junwei Yu, Jieyu Zhou, Mufeng Yang, Yepeng Ding, Hiroyuki Sato, “Conversational Capture: A Trajectory-Level Framework for Evaluating Generative Engine Optimization in Multi-turn Human-Agent Interaction,” arXiv preprint, submitted September 30, 2026. No empirical sample · arxiv.org
04 The ruling
Judge Amit Mehta dismissed Penske Media’s AI Overviews case
No court has now found that Google owes anybody traffic in return for the content it reads and reuses.
US District Judge Amit Mehta wrote that an expectation is not an agreement, and threw the claim out on that sentence.
About a year ago Penske Media, which owns Rolling Stone, Billboard, Variety and The Hollywood Reporter, sued Google over AI Overviews. The argument was that Google takes their work in exchange for visibility and then answers readers directly. On October 1 the case was dismissed.
US District Judge Amit P. Mehta found there was no deal to breach. The publishers, he wrote, “failed to plead any actual agreement whereby Defendants promised to ‘sell’ Plaintiffs any specific amount of traffic, or any traffic whatsoever, in exchange for ‘buying’ their content.” He put the whole ruling in five words: “an expectation is not an agreement.”
He was not unsympathetic. The court wrote that it “does not treat Plaintiffs’ alleged harms lightly” and acknowledged “the knock-on consequences to journalists, educators, and other online creators whose content Google takes and repurposes without compensation.” Jason Kint of Digital Content Next, which represents US publishers, called the result a dismissal of the last path available.
One thing to keep separate, because the two were reported in the same week. A different ruling on September 30 let Gannett, the Daily Mail and a class of digital publishers proceed against Google. That case is about advertising technology and has nothing to do with AI Overviews.
You are not a publisher, and this still settles the question sitting under every other item today. Nothing obliges an answer engine to send you anyone. The traffic a search engine used to pass along was a habit, not a contract, and a court has now said so in terms.
What to do about it
Stop planning around the hope that this gets corrected from above. It will not be, at least not through this route, and the companies waiting it out are the ones losing ground while they wait.
Spend the attention on the part nobody has to grant you: being accurate, specific and readable enough that an answer built without your permission still gets you right.
Source Search Engine Roundtable, “Google AI Overview Lawsuit Dismissed Over No Agreement With Publishers,” October 1, 2026 · seroundtable.com
05 The consumer data
Microsoft researchers read 40,000 conversations that started with a photo
Your customers are already asking AI about things they photograph, and nothing you publish is written for that question.
The study is about testing AI models rather than about businesses, and its finding is that photo questions range wider than typed ones.
Jinyi Ye, Scott Counts, Gaurav Verma, Kate Lytvynets and Weiwei Yang, four of them at Microsoft, published a study on September 30 of what people do when they hand an AI a picture. They analyzed “over 40,000 de-identified image-upload conversations from Microsoft Copilot” and sorted them into ten kinds of task.
Two findings are worth your attention. Most picture questions are not one question: the majority involve several different things at once, such as reading what is in the image and then reasoning about it. And picture questions cover a wider range of ground than typed ones, which the authors put as a broader and more diverse task space than text-only interactions.
Be clear about what this paper is for. It is about whether the tests used to grade AI models match what people actually do, and the answer it gives is no: across 253 existing benchmarks, the tests cluster on perception and fixed answers while real workflows go elsewhere. It makes no claim at all about businesses or recommendations, and the findings were checked against a separate ChatGPT dataset.
The reason it belongs here is the surface it proves exists. Someone photographs a crack above a doorway, a damaged roof, or a letter they do not understand, and asks an assistant what they are looking at. Each of those is a question answered before any human is called, and architecture and engineering practices, home builders and independent insurance brokerages sit at the end of all three.
Nobody is measuring that. Every AI visibility number in circulation, this brief’s included, is built from typed questions, because that is what can be sampled at scale. A picture question is harder to reproduce, so it goes uncounted, and uncounted is not the same as unimportant.
What to do about it
Take a photograph of the thing a customer would photograph before they call you. Hand it to an AI assistant with the question they would ask, and read what comes back.
If the answer is wrong about what it is looking at, or sends them to the wrong kind of business, that is the gap. It will not show up in anything you are being shown about search.
Source Jinyi Ye, Scott Counts, Gaurav Verma, Kate Lytvynets, Weiwei Yang, “From Images to Tasks: Characterizing Multimodal LLM Interactions in the Wild,” arXiv preprint, submitted September 30, 2026. Over 40,000 de-identified Microsoft Copilot image-upload conversations · arxiv.org
