01 Operational change
Google calls fabricated author profiles a form of deception
If a page on your site carries an author nobody can verify, Google now reads that page as low quality.
Google’s search guidance now names AI-generated headshots, invented author names and false credentials as deception, after years of telling the industry it was not a concern.
Google added a paragraph to its guidance on content quality. It opens with an instruction: avoid using deceptive authorship information. The sentence after it is the specific part: fabricating creator profiles, such as by using AI-generated headshots, made-up names, or false credentials to make content appear as if it was written by human experts, is a form of deception.
Google then lists what not to do, and the list is unusually blunt for a search document. Do not fake a creator profile, do not use an AI-generated headshot, do not invent an author name, do not fake author credentials.
The penalty is stated in Google’s own terms. Doing any of it makes a page untrustworthy to both users and our automated quality systems, Google writes, and is a signal of a low-quality page.
The reversal is the news. Google had previously told the search industry that it was not worried about faked author information, which is why so much published content carries a byline nobody has ever checked. Search Engine Roundtable, which spotted the change, puts it as plainly as that.
This matters more to an expertise business than to almost anyone else, because the byline is the thing being bought. A page about what to do after a diagnosis, before a filing deadline or during a structural inspection is worth reading because of who wrote it, and the credential is the entire claim.
What to do about it
Open every page on your site that carries a byline and check three things: the person is real, the photograph is of them, and each credential is current and written the way the body that issued it writes it.
Two industries should do this today. A veterinary specialist’s standing belongs to the college that granted it, the American College of Veterinary Surgeons or the American College of Veterinary Internal Medicine, and an architect’s or engineer’s licensure has a form the state board specifies. Both are exactly the detail Google has now said it will read as either a trust signal or a deception.
Source Search Engine Roundtable, Barry Schwartz, “Google Warns Against Deceptive Authorship Information Within Content,” October 6, 2026 · seroundtable.com
02 Research & evidence
A legal AI trained to look authoritative stopped answering
Rewarding content for looking expert produces content that looks expert and says nothing you can use.
Two researchers rewarded a legal AI model for citation count, legal-sounding language and length, and 89.3% of the citations it then produced were structurally implausible.
Subramanyam Sahoo and Justin Shenk took an open AI model and trained it against a reward built from three surface features: citation count, legalese density, and response length. Nothing in that reward measured whether the answer was right.
The model did not get better at reasoning. It learned to withhold commitment. Across 16 yes-or-no legal reasoning tasks, accuracy fell from 0.500, which is chance, to 0.072, and the authors show the collapse came almost entirely from the model ceasing to give a properly formed answer at all.
The capability was still there. When the model did commit to an answer, its accuracy rose from 0.556 to 0.657. It had simply learned that a long response packed with citations and empty of a direct answer scored better than a short correct one.
The citations are the finding that should worry anyone publishing. 89.3% of the citations the model produced after training were structurally implausible hallucinations, many of them subtly corrupted names of real landmark cases. The authors write that they were constructed, in effect, to survive a casual read and fail under scrutiny.
Be clear about the limits. This is a small model the researchers trained themselves, tested on 320 questions, not ChatGPT or Gemini, and the conclusion is about what a reward function does rather than about any product you use. The authors’ own summary is the line to keep: a reward that measures how legal a response looks will produce a model that is maximally photogenic and minimally useful.
Forever Cited sells visibility in AI answers, so a study about work engineered to look authoritative describes a risk in this industry as much as in yours. The discipline it argues for is the one worth keeping: ask what a number measures before accepting what it implies.
What to do about it
The same incentive runs through most content advice you will be sold. Length, citations and confident phrasing are easy to measure and easy to produce, and not one of them is the thing a reader or an answer engine is trying to establish.
If you run a software company or an admissions practice, apply the test to your own library. Open the three pages you are proudest of and check whether each one commits to an answer a reader could act on, or whether it surveys the question and stops.
Source arXiv, Subramanyam Sahoo and Justin Shenk, “Better Call Reward: Reward Hacking as Strategic Abstention in Legal Reasoning Models,” October 5, 2026 · arxiv.org
03 Research & evidence
Search tools never report a miss, so agents answer anyway
When there is nothing about your business to find, an AI agent does not say so. It answers from memory.
Three researchers found that a search tool returns its closest matches even when the answer is not in its index, and the AI agents reading those matches rarely notice.
Ramraj Chandradevan, Sayontan Ghosh and Vinoth Selvendran start from a plain observation about how these systems are built. A search tool never says no. It returns its top results even when the index holds no answer, so the agent reading them sees irrelevant text rather than a signal that nothing was found.
They built a test for it. 557 questions were run both with and without their answers present in a 21-million-passage index, across seven AI agents and five different wordings of a refusal. The question was what an agent does when the tool admits it has nothing.
The improvement is large and it depends entirely on the tool speaking up. A one-sentence refusal raised the rate at which agents declined to answer unanswerable questions from 23% to 97% on average for two open models, and from 28% to 57% for Claude Haiku 4.5, cutting wrong answers almost one for one.
Not every agent listened. One model ignored the refusal and fabricated retrievals, while Claude Sonnet 5.5 and Opus 5.5 answered from memory instead, moving barely two points. The wording mattered too: an explanation beat a bare signal, and a soft warning was useless.
The authors are blunt about where the fix sits. For agents that do comply, the bottleneck is the detector inside the search tool, not the agent, and realistic ways of triggering a refusal still fall well short of knowing for certain.
This is an academic benchmark on general knowledge questions, not a test of how an assistant describes a business. It is still the clearest statement of a mechanism worth understanding: the absence of information about you does not produce silence, it produces a guess.
What to do about it
You cannot fix the detector inside somebody else’s search tool. You can reduce how much an agent has to guess about, which is the same remedy three separate studies have landed on in the past two weeks.
Write down the five questions a customer asks before choosing you, and check that each one is answered in plain words somewhere a crawler can read, on your own site and on the listings that carry your name. If you advise on mergers or franchise sales, where most of what exists publicly about you is a deal announcement rather than an explanation, the gap is usually wider than it looks.
Source arXiv, Chandradevan, Ghosh and Selvendran, “Search Engines Never Say No: How Frozen Agents React When the Retrieval Tool Refuses,” October 4, 2026 · arxiv.org
04 Platform & evidence
Google crawled a site 1.17 million times, indexing four pages
Being read by Google is not being kept by Google, and a site can publish 180,000 pages while Google holds four.
Google’s John Mueller was asked why a site crawled 1.17 million times had four indexed pages, and named two causes: broken redirects and automatically generated content.
A site owner asked Google a question with the numbers attached. Googlebot, the program Google uses to read pages, had crawled the site 1.17 million times in a few weeks. Search Console, Google’s free reporting tool for site owners, showed four indexed URLs in total, meaning four pages Google had kept and was willing to show.
The home page was not one of the four. The site publishes about 180,000 pages.
John Mueller of Google answered with two causes. The first is mechanical: a number of the URLs that were indexed all redirect to the home page, which he called a big technical issue still to be resolved.
The second cause is the one worth reading twice. Mueller described the site as a large product aggregator and wrote that such a site is not trivial to make and maintain well, ending up with a ton of URLs, apparently with programmatic, autogenerated content, meaning pages built automatically from a template and a list rather than written by anyone.
Then the sentence that answers the question. Convincing search engines that this is worthwhile, convincing users that this is valuable, will be hard, Mueller wrote, adding that Google is generally not a fan of programmatic SEO, the practice of generating pages at volume to catch search traffic.
This is one site, answered in a social post rather than in documentation, and most readers of this brief do not publish 180,000 pages. The transferable part is the distinction: Google reading your pages constantly is not the same as Google keeping them, and the gap between the two is where a redesign quietly goes wrong.
What to do about it
Open Search Console and compare two numbers: how many pages your sitemap declares and how many Google reports as indexed. If the second is much smaller, that gap is the ceiling on everything else, including whether an AI answer can quote you at all.
Check your redirects while you are there. A page that quietly sends every visitor to the home page is indistinguishable, from Google’s side, from a page that is not there.
Source Search Engine Roundtable, Barry Schwartz, “Google On Large Product Aggregator Websites,” October 6, 2026 · seroundtable.com
05 Operational change
Google Maps now bars promotions and pricing from your listing
A time-limited offer or a price in your Google listing is now against policy and can be taken down.
Google rewrote the advertising and solicitation section of its Maps and Business Profiles content policy to spell out what a business owner may not put in a listing.
Google expanded the advertising and solicitation section of its prohibited and restricted content policy for Google Maps and Business Profiles. The new framing is explicit about what a listing is for: contributions to Google Maps should reflect genuine experiences or define a business’s permanent identity, which Google says keeps the platform reliable rather than a marketplace.
Google then added a section addressed to merchants, and it draws a line between the listing itself and the posts attached to it. In the listing, promotional or commercial content containing time-bound promotions, pricing, or direct calls-to-action is prohibited.
The rule for posts is narrower than that suggests. In posts, Google bars email addresses, phone numbers, social media links and links to other websites that are not related to your business, while offers and updates that read as promotional are otherwise still permitted there.
No effective date was published, which means the policy is the policy now. The change was spotted by Hiroko Imai, who posted screenshots of the text before and after. Google did not announce it.
The key word is permanent. A listing is now meant to describe what your business durably is, not what it is running this month, and the place for what you are running this month is a post.
Do this before your next post
Read your own listing today, the way a stranger would. Anything that reads as an offer, a price or an instruction to act belongs in a post rather than in the business description, the services or the name.
A tax practice advertising a filing-season rate in its listing, or a charter operator quoting an hourly figure, is the exact shape this policy now names. Move it to a post, where offers are still allowed, and leave the listing describing what the business permanently is.
Source Search Engine Roundtable, Barry Schwartz, “Updated Google Maps Policy: What Merchants Can’t Post In Listings Or Posts,” October 6, 2026 · seroundtable.com
