01 The research
Two researchers matched 2 million AI citations to 10,000 pages
The page tweaks sold as AI optimization did not explain who got cited, and the website’s own standing did.
The checklist held up until the researchers compared pages sitting on the same website, and then it did not.
Ben Moore and Liam Dunne published a study on September 28 matching about 2 million citations from four AI assistants, named in the paper as ChatGPT, Claude, Google AI and Gemini, against 10,000 crawled pages from nineteen business software companies, collected over six months.
They tested more than sixty features of those pages through nine separate statistical methods, and kept only the findings that survived all nine. The first is the one that will be quoted. FAQ sections, the hidden labels that describe a page to a search engine, and page-speed scores all looked helpful until the researchers accounted for which website each page sat on, at which point those effects reversed or fell to zero.
What survived is harder to buy. The website’s overall standing with AI assistants outweighed the strongest page-level feature by a factor of six. The single best predictor of a page being cited was how closely its wording matched the full range of questions people were actually asking in that field, including the questions that never produced a citation at all.
Two limits matter before anyone repeats this. Every page came from a business software company, so a veterinary hospital or an architecture studio sits outside what was measured, and the study is a preprint that has not been through peer review. Forever Cited sells AI visibility work, and a good deal of what this industry sells is page-level. A study suggesting that page-level work is not what moves citations is worth repeating even when it points at us.
If the finding holds outside software, the thing that gets a business named in an AI answer is what the whole site is known for, not what was added to one page last month.
What to do about it
Write down the ten questions a new customer actually asks you, in their words, and check whether your website answers any of them. The strongest single predictor in this study was the match between a page and the full set of questions being asked in its field. That is a decision about what you publish, not a setting you switch on.
Source
arXiv, Ben Moore and Liam Dunne, “What Drives Citations in Production Large Language Models? An Observational Multi-Method Study of Two Million AI Citations Across Ten Thousand Web Pages,” September 28, 2026 · arxiv.org
02 The platform change
Google opened AI Mode monitoring to every user globally
Any customer can now ask Google to watch your category and tell them when something changes, without ever searching again.
The feature was limited to paying subscribers until Sunday, when Google’s search product lead said it was going out to everyone.
Robby Stein, Google’s Vice President of Product for Search, announced on September 28 that the monitoring feature inside AI Mode is rolling out globally to everyone. Until now it was available only to Ultra and Pro subscribers. AI Mode is the tab in Google Search where an assistant answers in prose instead of returning a list of links.
Monitoring works the way it sounds. You tell AI Mode what to watch for, and in Stein’s description it checks websites, forums, social posts, Google’s real-time data sources and a shopping catalogue of more than 60 billion products, then alerts you in the Google app when something changes.
The examples Stein gave are consumer ones: new restaurants or pop-ups opening nearby, local holiday activities to do with children, and price drops. The version that matters to a business is the one nobody demonstrated. A person can now ask Google to tell them when a new practice opens in their area, or when someone nearby starts offering a service they have been waiting for.
That changes the shape of a search. A search is a moment and a monitor is a standing instruction, and the business named when the alert fires is one Google can already describe. Google announced this on X rather than in a blog post or Help Centre article, and that post is the only description of how it works, so nothing published says how Google decides which businesses qualify to appear in an alert.
What to do about it
Ask AI Mode to monitor your own category in your own city for a week. Whatever it names when the alert fires is the list Google can currently describe, and if your business is missing from it, that gap is the finding. It costs nothing and takes a minute.
Source
Search Engine Roundtable, Barry Schwartz, “Google Globally Rolls Out Monitoring Capabilities In AI Mode,” September 28, 2026 · seroundtable.com
03 The listing rule
Google now penalizes repeated unfounded Business Profile complaints
If you report a competitor’s wrong listing and cannot substantiate it twice, Google can stop reading your reports altogether.
The new policy names false reports and reports made in bad faith, and the penalty falls on whoever is complaining.
Google added a section titled “Misuse of support channels” to its Google Business Profile policy document, reported on September 29. A Business Profile is the listing that puts your business on Google Maps and in the panel beside your name in search results, and the support portal is how an owner reports a problem with one.
Google’s wording is that support paths must be used “responsibly” and that businesses should “avoid any misuse of the portal or flood them with invalid requests.” The two named examples of unfounded submissions are “false or misleading reports” and “reports made in bad faith.”
The consequence is one line and it is worth reading twice. “Repeated unfounded submissions can result in rejection of all further complaints without assessment.” That is the support channel closing to you, not a single report being turned down.
This lands hardest where a wrong listing is a live problem and reporting it is the only remedy. Veterinary hospitals and medical practices competing against duplicate or out-of-date entries, real estate brokerages whose agents each generate listings, and insurance brokers in markets full of lead-generation pages all report listings as routine work. The policy names no number, no appeal path and no definition of “repeated.”
There is also no stated effective date. Barry Schwartz reported the change on Tuesday and quoted the live policy text, which means the published document is the rule as of now.
What to do about it
Before your next report, gather the evidence first: a screenshot of what is wrong, the correct information, and a sentence on how you know. A report you cannot substantiate now carries a cost it did not carry last week, and the cost is your ability to report anything later.
Source
Search Engine Roundtable, Barry Schwartz, “Google Business Profiles Policies Adds Misuse Of Support Channels,” September 29, 2026 · seroundtable.com
04 Who gets quoted
Researchers audited 30,000 AI answers built from Reddit comments
When an assistant summarizes what people say about your field, it keeps the confident top comment and drops the lived experience.
Formal, directive language survived into the summary, and first-hand accounts of what happened to someone did not.
Agam Goyal, Wang Claire and Eshwar Chandrasekharan audited Reddit Answers, the feature that replies to a question by summarizing Reddit discussions. They ran 10,000 questions drawn from 20 advice and support communities, three times each, producing 30,000 answers on top of 14.68 million comments. The paper was submitted on September 13.
Three findings set up the fourth. The differences between three runs of the same question came mostly from which comments were retrieved rather than from how the answer was written. Answers routinely combined evidence from different communities. And selection strongly favoured comments that were already visible and already at the top level of a thread.
The fourth finding is the one that reaches your business. Formal and directive language was more likely to be surfaced, while experiential voice was less likely to survive selection and was weakened further during the writing of the answer, with first-person singular language declining sharply. The person saying “here is what happened when I tried this” loses to the person saying “you should do this.”
If people discuss your field in public, this is the filter between what they said and what a stranger reads. A single confident top comment can outweigh a dozen accounts of real experience, and none of those accounts is deleted, they are simply not selected. The authors’ own conclusion is that community-grounded generative search “is not neutral summarization.”
It bites hardest where people ask strangers before they ask a professional: admissions consultants, whose market researches on forums first; specialty veterinary hospitals, where owners post symptoms before they call anyone; and luxury home builders, where one bad thread about a contractor can define a region.
What to do about it
Search your own city and category on Reddit and read the top-level comments rather than the whole thread. Those are the ones an assistant is most likely to use. If the top comment is wrong about your business or your field, correcting it in public is worth more than another page on your own site.
Source
arXiv, Agam Goyal, Wang Claire and Eshwar Chandrasekharan, “The Wisdom of the Loudest: A Large-Scale Audit of Generative Search on Reddit,” September 13, 2026 · arxiv.org
05 The eligibility change
Google lowered its Search profile threshold to 10,000 followers
A verified page about you inside Google’s own results is now open to anyone with ten thousand followers on one platform.
Search profiles are the verified panel Google shows for a named person or brand, and the bar for claiming one has moved.
Ibrahim Badr of Google announced three changes to Search profiles on September 16. A Search profile is a verified space inside Google Search that gathers a person or brand’s social accounts, recent articles and links in one place, and the owner controls what it says.
The third change is the one that opens a door. Google lowered the United States eligibility threshold to “10,000 followers across YouTube, Instagram, X, or TikTok,” in its own words. The post does not say whether that is a combined total or a count on any one of the four, which is worth knowing before anyone counts their own.
The other two are smaller. A media company can now claim and manage profiles for all of its sub-brands from a single login, and article headlines inside a profile are longer with better thumbnails.
One caution about relying on it. Google removed the Follow button from Search profiles on desktop around September 25 and it reappeared four days later, with no announcement either way. Damien Andell, who spotted both, reported on September 29 that it was back. The profile is worth claiming. It is not worth building a plan on.
Eligibility is limited to United States publishers and creators, and Google’s post defines neither word. Nothing in it says a professional practice qualifies, and nothing says it does not.
What to do about it
If anyone at your business has ten thousand followers on any one of those four platforms, try to claim a Search profile in their name this week. It is a page you control that Google can show beside your name, and the only cost of trying is finding out you are not eligible yet.
Source
Google, Ibrahim Badr, “3 new ways we’re improving Search profiles for publishers,” September 16, 2026 · blog.google · Search Engine Roundtable, September 28, 2026 · seroundtable.com
