01 The deadline
Google gives you four days to reject a listing edit
A stranger can change your hours, address or phone number, and silence on your part counts as agreement.
Google has put a four-day clock on customer-suggested edits, and a business that misses the notification does not get a say.
Anyone can suggest a change to your Google Business Profile, the panel that shows your name, address, phone number and hours when a customer searches for you. Google has now documented what happens next. You are notified by email or an in-product notification, and you then have four days to accept or reject the suggestion.
If nobody answers inside that window, Google says it may automatically publish the update if it is supported by other publicly available information, such as your business website. It also states plainly that it may apply suggested user edits to your business information without prior review.
The risk here is not the malicious edit, which is rare. It is the stale edit that happens to agree with a stale source. An old address still sitting on a directory page, or on a page of your own site nobody has touched in two years, is exactly the corroboration that lets a suggestion through, which is how a correction you already made gets quietly undone.
It lands hardest where the address and the hours are the booking. An emergency veterinary hospital, a surgical practice, a title office with a short counter day: a wrong hour on the listing is a customer who drives to a locked door. For a charter operator or an accounting firm whose clients call one main line, a changed phone number is a quiet week before anyone works out why.
What to do about it
Confirm today that Google notifications reach a person who reads them daily, rather than a shared mailbox nobody opens. Then check the name, address, phone number and hours on your profile against what your own website publishes, because your own site is one of the sources Google uses to decide whether a suggested edit is credible.
Source
Search Engine Roundtable, Barry Schwartz, “Google Business Profiles Gives You Four Days To Reject User Edits,” September 23, 2026 · seroundtable.com
02 The evidence
Researchers found pages padded with unsupported detail ranked higher
A competitor who invents specific-sounding detail can outrank your honest page, and the tools meant to catch that failed their own test.
An evidence-paired benchmark of 50 shopping queries found that rich but unsupported claims gained rank, and that no automatic checker was reliable enough to stop them.
Yihan Xia, Huiling Fan, Kangrong Zhong and Taotao Wang posted a study to arXiv on September 21 testing something no business owner would think to ask. When a generative search system ranks products and services, can it tell a specific claim that is true from a specific claim that is merely written confidently? They built a benchmark of 50 shopping queries and 1,950 cases, pairing each candidate page with the evidence that would or would not support it.
On one of the ranking models tested, pages padded with rich but unsupported detail gained significant rank over clean pages, while pages whose detail was properly supported, and neutral pages, did not. The authors put that gain at 0.065 to 0.092 on a normalized rank scale. The effect was model-dependent: marginal on a second model and absent on a third.
The defense side is where it gets uncomfortable. Given perfect evidence labels, a scorer cut the rate at which unsupported padded pages reached the top three from 0.65 to 0.43. Measured against a 370-claim human gold standard, every automatic judge tested failed the reliability gate the authors had registered in advance.
The authors state the limit themselves, and the posture is worth more than the finding. The results do not validate an automatic defense, and one arm of the human comparison remains pending adjudication. It is a preprint, it has not been peer reviewed, and the honest summary is that the problem is now measured and the fix is not.
What to do about it
This is not an invitation to pad your own pages, but a reason to make every specific claim on your site checkable: the year, the number, the named credential, the source. An architecture practice naming a project and its completion year is making a claim someone can verify; one calling itself award-winning is not. That difference is what an evidence-aware ranker is built to notice, and it is what a customer reads too.
Source
arXiv, Yihan Xia, Huiling Fan, Kangrong Zhong and Taotao Wang, “GroundedGEO: Auditing the Evidence Gap in Generative Search Rankings,” September 21, 2026 · arxiv.org
03 The evidence, continued
Two researchers found AI agents excuse their platform’s paid listings
When an AI recommends a business, the rules that make paid placement obvious to people may never reach the customer.
A preprint argues that sponsorship disclosure now lands on the machine doing the choosing rather than on the person being advised.
Davood Wadi and Yu Ma posted a study to arXiv on September 16 asking a question advertising law has never had to answer. When an AI assistant picks a business for a customer, whose interests is it serving? The assistants now recommending hotels, contractors and services are built by platforms that also sell advertising to those same businesses.
The experiment turns on a single change. In one condition the system prompt names a traveler as the assistant’s principal; in the other it names the booking platform, and nothing else moves. Naming the platform significantly reduced the penalty the assistant applied to sponsored listings, and weakened the skepticism the disclosure label produced in its stated reasoning.
A second study took the label apart. The gap between the two conditions widened when the paid placement was attributed to the platform, and although the stricter word Sponsored produced fewer paid choices than Promoted, it did not close that gap. The authors’ conclusion is the line to keep: disclosure rules written for human consumers cannot by themselves protect a customer when an AI is doing the choosing.
Two limits belong with this. It is a preprint that has not been peer reviewed, and the abstract reports the direction of each effect without publishing sample sizes or effect sizes, so how large the shift is remains unknown. What the study establishes is that the effect is there and holds across models and reasoning depths, not how big it is.
What to do about it
Nothing here is yours to fix, which is worth knowing on its own. The practical read is that a paid placement inside an AI answer will not announce itself to your customer the way a sponsored search result does. So when an insurance broker or a home builder is shown what share of AI answers mention them, the useful question is whether any of those mentions were bought.
Source
arXiv, Davood Wadi and Yu Ma, “Whom Do AI Agents Work For? Role Assignment Induces Sponsorship Bias in LLM Recommenders,” September 16, 2026 · arxiv.org
04 The regulator
UK regulator wants Google to show where answers came from
If this holds, the source of an AI answer becomes something a regulator checks rather than something Google chooses.
The Competition and Markets Authority decides on October 9 whether Google must attribute the publishers behind its search results.
The Competition and Markets Authority, the United Kingdom’s competition regulator, has proposed measures requiring Google to give Android and Chrome users in the UK a real choice of search service. Users would be offered that choice on first use, then prompted every year to pick the default they want.
One clause matters more to a business trying to be found than the choice screen does. The proposal would require search providers to fairly attribute publisher content so that users can easily access source content and know where their search results have come from.
That is a regulator writing attribution into the rules, and the CMA will make a final decision on this proposal by October 9. The measures also cover AI assistants that meet stated technical and security standards, which pulls the assistants inside the same regime rather than leaving them beside it. Sarah Cardell, the CMA’s chief executive, framed it as recognition that the way people search for information online is changing.
Two limits, stated plainly. This is a proposal in consultation, not a binding order, and it covers the United Kingdom, so a practice in California gains nothing from it directly. What changes is the direction: attribution is now something a regulator can require rather than a courtesy a platform extends.
What to do about it
Nothing to do this week, and one date to watch. If attribution becomes a requirement in a market the size of the UK, the machinery that supports it rarely stays in one country. A business that is already the clearest source on its own subject is the one that gains first from any rule forcing an answer to name where it came from.
Source
Search Engine Roundtable, Barry Schwartz, “CMA: Google Should Make It Easier To Choose Other Search Engine On Android & Chrome,” September 23, 2026 · seroundtable.com
05 The number
Peec AI’s chart puts AI Overview external links above 25%
The most encouraging number in AI search this week arrived with no sample size, no method, and an investor disclosure attached.
One chart on X carries a claim that would be genuinely good news, and none of the working behind it has been published.
Malte Landwehr of Peec AI posted a chart on X showing AI Overviews going from close to 0% containing external links to over 25% containing external links. Barry Schwartz reported it and noted that it runs against his own earlier reporting on how reluctant Google has been to link out of an AI Overview at all.
If it holds up, it is the most useful thing in this issue. An AI Overview that links out is an AI Overview that can still send a customer to your website, which is the whole difference between being described and being visited.
What has not been published is everything that would let anyone check it. No sample size, no date range, no list of queries, and no definition of how a link was counted as external. Schwartz also discloses that he holds a small investment in Peec AI and deliberately covers the company less often because of it, which is a disclosure worth more than most in this field.
Forever Cited sells the case that being a cited source is what gets a business found, so a chart showing AI answers linking out far more often is a number we have every commercial reason to repeat. That is exactly why it is worth saying that one chart with no method behind it is not yet evidence, and waiting for someone to show the working.
What to do about it
Do not change anything on the strength of this number. Do check it in the cheapest way available: run five real buying questions for your own market in Google, and write down whether the AI answer at the top links anywhere other than back into Google. Ten minutes on your own market beats a chart about everybody’s.
Source
Search Engine Roundtable, Barry Schwartz, “Google AI Overviews With Way More External Links; Report,” September 23, 2026, reporting data posted on X by Malte Landwehr of Peec AI · seroundtable.com
