01 The research
Google credited almost no shared websites across two languages
Whether a search engine names your website as a source is far less settled than whether it gives a good answer.
Google’s answers barely changed in meaning while the list of websites getting credit for them changed almost entirely.
Five researchers drew 500 real search questions from a public research set, asked each one of Google and of Baidu in English and again in translated Chinese, and recorded which websites each answer showed as its sources. Google produced an AI answer for 97.6% of the Chinese questions and 97.0% of the English ones. Baidu produced one for 82.0% of the Chinese questions and 4.0% of the English, too few in English to analyze its sources at all.
Then the uncomfortable part. Google in English and Google in Chinese showed 5,341 different websites as sources between them, and only 179 of those appeared on both lists, an overlap of 3.4% that was the closest any two of the three settings came. The worst pair, Baidu in Chinese against Google in English, shared 25 websites out of 3,095.
The answers themselves were nowhere near that far apart. Compared by meaning rather than by wording, matched answers scored between 0.701 and 0.813 on a zero-to-one scale, which is close without being identical. So the engines were saying roughly the same thing while crediting almost entirely different websites for it.
One more figure is worth keeping. Baidu drew its Chinese answers from 443 websites and sent 27.1% of that credit to a single one it owns. Google spread it much wider, its most-used website taking 6.6% in Chinese and 5.5% in English, leaning on Wikipedia, Mayo Clinic and ScienceDirect.
Two limits belong right beside all of this. The questions came from a general research set, not from people looking for an IT and cybersecurity firm or a software company, and Chinese-language search is not your market. What does carry over is the mechanism: source credit is decided separately from the answer, and it does not hold still across platforms or even across two languages on one platform.
What to do about it
Ask an AI assistant your two most common customer questions, then read only the list of sources beside the answer instead of the answer itself. Write down which websites are on it. That list, not the wording of the answer, is what your business is competing to get onto.
Forever Cited sells the claim that being named as a source is what decides whether a business gets recommended, so a paper measuring exactly that is one we have every commercial reason to quote. The honest note is that it measured general questions in English and Chinese, and that nobody has yet run the same audit on the questions your customers actually type.
Source
arXiv, Yibo Li, Enci Guan, Yuedan Cai, Geng Liu and Francesco Pierri, “Auditing Source Exposure in Baidu and Google AI Search,” September 21, 2026 · arxiv.org
02 The operational change
Google will remove Business Profile posts with unverified numbers
A tracking number or a direct line in a Google post can now be taken down without anyone telling you first.
Google’s new rule covers phone numbers, email addresses and social handles, and the test is whether Google can tie them to your listing.
Google updated two of its Business Profile help documents this week. The new wording reads: “To avoid fraud or abuse, we may remove posts with unverified contact information. This includes phone numbers, email addresses, and social media handles.”
A second document is more specific. “Posts that include a phone number in the post description might get rejected if they cannot be verified as being connected to the business on which they are posted.” So the question is not whether your number is real. It is whether Google can match it to that listing.
This lands hardest on anyone running call tracking. A campaign-specific or call-tracking number is by definition not the number on your listing, which is precisely the mismatch the rule is written to catch. The same goes for a direct line for one attorney, one adviser or one surgeon inside a group practice.
It also catches a habit that looks harmless. A post telling people to email an address, or pointing them at a social account, can now be removed on the same grounds, and removal is not the same thing as a warning.
What to do about it
Open your Google Business Profile posts and look at every one carrying a phone number, an email address or a social handle. If the number is not the number on the listing itself, replace it with a link and a call-to-action button.
Removal takes the whole post rather than just the offending line, so a post you spent real time on is worth ten seconds of checking.
Source
Search Engine Roundtable, Barry Schwartz, “Google Posts Cracks Down On Unverified Contact Information,” September 25, 2026 · seroundtable.com
03 The research
An AI that read more studies found more false benefits
When a customer asks an AI whether something works, the answer gets more confident and less reliable the harder it looks.
Fred Sun and Shangqi Guo measured how often a searching AI wrongly concludes a treatment works, and across the range they tested the rate doubles.
Published research leans positive. A study that finds an effect tends to get written up and a study that finds nothing often does not, so anything reading the published literature is reading a skewed sample before it starts.
Fred Sun and Shangqi Guo measured what that does to an AI that searches for evidence and then draws a conclusion. On 140 questions built from Cochrane reviews, the systematic summaries clinicians treat as the top tier of medical evidence, they counted how often the system decided a treatment helped when the true effect was nothing at all. As they raised the number of searches it was allowed from 3 up to 20, that false-benefit rate climbed from 7.9% to 15.7%.
The direction is the whole finding. Searching harder made the conclusion worse rather than better, and the paper argues this is structural rather than a quirk of their setup: under a standard model of publication bias, the false-positive rate keeps climbing as the searching deepens.
Their own fix stops searching once the evidence stops improving. It cut the false-benefit rate from 15.7% to 6.4% while using 67% fewer searches, and lifted accuracy on the no-real-effect cases by 21 points, from 40.0% to 61.4%. Overall accuracy went from 61.4% to 69.3%.
The caveat is scope and it is a real one. This was biomedical evidence tested against Cochrane reviews, not questions about choosing a builder or an adviser. What carries over is that more searching is not automatically better searching, which is worth remembering the next time something is sold to you on the depth of its research.
What to do about it
If your field runs on published evidence, this is the one to watch. Physicians, surgical practices and specialty veterinary hospitals already answer questions from people who asked an AI first, and the finding here is that those answers lean toward saying something works.
The practical move is unglamorous. Wherever your own pages say a treatment, a method or a product works, state what the evidence is and where it stops, because a page that states its own limits is a page an AI can quote accurately.
Source
arXiv, Fred Sun and Shangqi Guo, “When More Evidence Hurts: Publication-Bias Drift and Principled Stopping for Biomedical Causal Search,” September 21, 2026 · arxiv.org
04 The counter-signal
Google tested moving its hotel AI answer off the top
Where the AI answer sits on the page is still being decided, so do not rebuild your site around this month’s layout.
Google’s hotel test is one screenshot on one category, and it is still the only change all month pointing back toward the links.
Every item in this brief for weeks has pointed the same way. The written answer takes the top of the page and the links move down. On hotel searches, Google is testing the reverse.
Lluc Penycate posted screenshots showing the AI Overview moved out of the top slot and into a panel on the side, underneath the hotel information box. Search Engine Roundtable reported it on September 24 and noted it resembles similar tests running in local panels.
Treat this as weak evidence, because that is what it is. It is a test rather than a rollout, on one category, observed by one person, and Google has said nothing about it. No number attaches to it at all.
It still earns a line, for one reason. Hotels are a local category with a rich information box, which is the same shape of result a clinic, a veterinary hospital or a charter operator appears in. If Google decides the answer belongs beside the listing rather than above it, those categories are where it would show up next.
What to do about it
There is nothing to do here, and that is the honest answer. Put a note in your calendar for a month and check whether the AI answer on your own category has moved, because a layout change is the one thing in AI search you can see with your own eyes instead of measuring.
It is in the brief as one screenshot with no confirmation, labeled that way, because a test written up as a shift is how a reader ends up rebuilding a page for something that never launched.
Source
Search Engine Roundtable, Barry Schwartz, “Google Tests AI Overviews On Hotel On Side Panel,” September 24, 2026 · seroundtable.com
05 The money
Google delayed the Dynamic Search Ads shutdown to February 2027
Automatically created assets and campaign-level broad match convert in your Google Ads account this month, and most advertisers do not know they are on.
Google’s April announcement said September, that sentence is still on the page, and a June update at the top of it says February 2027.
In April, Google said it would automatically convert Dynamic Search Ads to AI Max in September. Dynamic Search Ads let Google pick which of your pages to advertise against a search, working from your website rather than from a keyword list you wrote. AI Max is the replacement.
Then Google moved the date. An update stamped June 11 sits at the top of the same announcement: “we are extending the timeline for Dynamic Search Ads sunset and auto-upgrade, which will begin in February 2027.” The reason given is advertisers asking for more time.
Two things are still converting this month, and this is the part to check. “Campaigns using Automatically Created Assets (ACA) and campaign-level broad match setting will continue to be auto-upgraded starting in September 2026,” Google says. Automatically created assets means Google writing your headlines and descriptions off your website, and campaign-level broad match means Google showing your ads against searches well beyond the words you picked.
The page is genuinely confusing, which is why this gets a paragraph rather than a line. The April text underneath the June note still reads that starting in September, Dynamic Search Ads “will automatically be upgraded” and that Google expects “all upgrades for eligible campaigns to conclude by the end of September.” Nobody edited that sentence when the date moved.
Google's claim for AI Max is also narrower than it first looks. The 7% average is more conversions or conversion value at similar cost from using the full AI Max feature set rather than its search term matching alone, which compares two AI Max setups with each other and not AI Max with what you have now. On September 23 Google published guidance for running that comparison in your own account, with one limitation: a campaign made entirely of dynamic ad groups cannot use the ready-made experiment template.
What to do about it
Open Google Ads and check two settings rather than one. Look at whether automatically created assets is switched on, and whether any campaign has broad match set at the campaign level, because those are what convert this month.
Your Dynamic Search Ads campaigns now have until February 2027, so use the time. Google published the experiment guidance on September 23, and a 7% average drawn from every advertiser on earth says nothing about your account. That holds as much for an admissions consultant buying twenty clicks a day as for a franchise adviser buying two thousand.
Source
Google Ads & Commerce Blog, “We’re upgrading Dynamic Search Ads to AI Max,” April 15, 2026, updated June 11, 2026 · blog.google · also Search Engine Roundtable, Barry Schwartz, “Test Google Ads AI Max Against Dynamic Search Ads,” September 23, 2026 · seroundtable.com
