01 The clinical evidence
GPT-4o matched physicians, then moved when the wording did
How a customer phrases a question changes the AI’s answer, so two people with the same problem get different advice.
Five AI models were put against practicing physicians on more than 6,000 triage cases, and the models’ recommendations moved on wording the physicians ignored.
The benchmark is large and the comparison is direct. Abinitha Gourabathina and four co-authors assembled more than 6,000 clinical scenarios, 7,000 physician annotations and 225,000 model responses, then rewrote each case in ways that left the medicine unchanged: swapping the patient’s gender, removing it, or making the patient’s own words more vivid or more hesitant.
At baseline GPT-4o was the most accurate model in the panel, at 91.7%, slightly ahead of the clinicians at 90.5%. Then the same cases were reworded. On the question of whether the patient needed to be seen, GPT-4o fell from 96.0% to 86.6% when the patient sounded uncertain.
The physicians did not move. Across all twelve combinations of task and rewrite the clinician figure shifted by at most 2.7 points, and by 0.9 on average, and that is the careful comparison, because it counts only the cases where the physicians themselves agreed the rewrite had not changed the medicine. The wording changed the model’s answer, not the question.
The spread across models was wide. Qwen2.5-32B lost 15.91 points under the more vivid phrasing and DeepSeek-R1-32B lost more than 12 points under every rewrite tested, while MedGemma-27B was the steadiest, at no more than 3.62 points, from the weakest starting accuracy in the panel.
Nothing here is about medicine specifically. It is about the gap between a careful question and a nervous one. An independent insurance broker fielding a claim question and a luxury home builder answering a worried email both hear the same situation described twenty different ways all day, and a professional reads through the phrasing in a way these models did not.
What to do about it
Ask your own question the way a worried customer would, not the way you would. Type the version with the hedging and the maybes, then type the calm version, and compare the two answers you get about your own field.
Two industries should try this first. An independent insurance broker and a luxury home builder usually get their first contact as an anxious description of a problem, and that is exactly the register where the models in this benchmark moved most.
Source arXiv, Abinitha Gourabathina, Haoran Zhang, Yuexing Hao, Walter Gerych and Marzyeh Ghassemi, “Sense and Sensitivity: Benchmarking LLM Clinical Triage Recommendations with Physician Experts,” September 29, 2026 · arxiv.org/abs/2609.38600
02 Research & evidence
Withhold the facts and identity explains the money advice
When a prospect gives an AI no details, the advice it hands back is built almost entirely from who they appear to be.
Researchers ran 96,600 money questions through an open AI model, holding the finances identical and changing only who the investor was said to be.
The design is the useful part. Saanvi Khetan and Sankar Balasubramanian built 100 financial profiles, 138 invented investor identities and seven different amounts of financial detail, then asked for a recommended stock allocation while grading the financial facts in the question from eight down to none. The finances never changed, only the stated identity.
With all eight facts supplied, two identities with identical finances got advice 4.78 percentage points apart. With no facts supplied, the gap was 10.34 points. Recounted so that a repeated question is only counted once, the no-facts gap came out 2.16 times the full-disclosure one, on a range the authors put at 1.69 to 2.79 times.
The share is the number to remember. Identity explained 5% of the variation in advice when the facts were complete, and 96% when they were absent. Leaving a single fact in the question still left the gap at 1.69 times its full-disclosure size.
With no facts at all, the model’s one-line explanation cited incomes, debts and savings it had never been told, and those invented finances turned adverse more often for larger households. Household size was the only attribute whose influence grew reliably as evidence was withdrawn.
One limit belongs in the headline of any summary of this paper. The model tested is a small open one, Llama-3.1-8B-Instruct, not the assistant most people actually use, and the authors claim nothing beyond it. What travels is the mechanism: a prospect who asks a money question without supplying details gets an answer assembled from assumptions about them, then arrives at a tax practice or an accounting office holding it.
What to do about it
Write down the three facts a stranger would have to give you before your answer to your most common question would change. Those three facts are the ones an AI assistant is inventing when a prospect does not supply them.
Then put those facts in your own published answers. A tax practice or an accounting office that states the conditions under which the answer changes gives the assistant something to use instead of a guess about the person asking.
Source arXiv, Saanvi Khetan and Sankar Balasubramanian, “Thin Evidence, Thick Priors: How Language Models Substitute Identity for Missing Financial Facts,” October 6, 2026 · arxiv.org/abs/2610.07798
03 Platform change
Google appears to have dropped review stars for healthcare
The star rating you earned may no longer appear under your healthcare pages in Google results, and nobody announced it.
One vendor reports review snippets disappearing from healthcare pages since late May, with a second drop in early August, and Google has confirmed nothing.
Review snippets are the star ratings that appear under a search result when a page carries review markup. Andrea Badder at Schema App reports that they have been vanishing from healthcare pages: “Across our healthcare customers, we’ve observed a drop in Review Snippet visibility since late May, with another drop occurring in early August for some organizations.”
The vendor’s own troubleshooting ruled out its own work. Testing different ways of connecting the rating property, alternative property paths and multi-typing healthcare entities as products did not restore the snippets, and Barry Schwartz, who reported the change, writes that he does not see them appearing either.
What the customers did next is the part worth copying. Many declined to change markup that was semantically accurate just to chase a star rating, choosing instead to keep describing their physicians, locations, services and specialties correctly for search engines and AI systems.
This is one vendor’s observation of its own customer base, with no sample size published and no confirmation from Google. Forever Cited sells visibility in AI and search results, so the pull to repeat a finding like this as settled fact is ours as much as anyone’s, and naming what has not been verified is the only version worth reading.
Google has switched off rich result types before, and has done it for particular industries. If the stars are gone for medical pages, the markup underneath them still matters, because the same structured description is what an AI assistant reads when it is asked what a practice treats and where.
What to do about it
Open Google Search Console, go to the performance report and filter by search appearance for review snippets. If the line falls off a cliff in late May or early August, you are seeing what this vendor’s customers saw, and there is nothing on your side to repair.
Do not pull the markup out. A specialty veterinary hospital and a surgical practice both depend on that same structured description to tell an AI assistant what they treat and where, and that outlives any single star rating.
Source Search Engine Roundtable, Barry Schwartz, “Google Drops Review Snippets For Healthcare Web Pages,” October 7, 2026 · seroundtable.com
04 Operational change
Google Ads will let your name differ from your domain
From October 16 you may be able to advertise under the name customers know, even if your website address says something else.
Google is loosening the rule that an ad’s business name must match its destination domain, in three defined cases, and resellers and booking intermediaries are excluded.
Google will update its business name requirements policy in October 2026 to allow a difference between the business name shown on an ad and the website address the ad points at. Google’s own wording is that it will do so “in certain limited cases.”
There are three conditions and all of them have to hold. The business name must accurately reflect the recognized name or brand of the advertiser; there must be a verified direct relationship between the advertiser and the domain owner; and the advertiser’s products or services must be directly offered on the destination domain.
The exclusion is as specific as the permission. Third-party resellers, independent booking intermediaries, affiliate distributors and secondary sellers do not qualify, and Google prohibits them from using the standalone brand name of what they sell as their own business name.
The exact date is not in Google’s published policy text, which says only “In October 2026.” The October 16 date comes from an advertiser email that Arpan Banerjee forwarded to Search Engine Roundtable. Treat it as the date to check against rather than a published deadline.
This matters to any business advertising under a name its website does not carry. A franchise advisory practice trading under a brand while the site still sits on a founder’s old address, or a real estate brokerage advertising a team name on a parent company’s site, has been choosing between the recognizable name and the policy. From the middle of the month, that choice may not be required.
What to do about it
If your ads run under a name that does not match your website address, read the three conditions above against your own setup before the middle of October, and make sure you can evidence a direct relationship with whoever owns the domain.
If you sell someone else’s product or book on someone else’s behalf, the opposite applies. Resellers, affiliates and booking intermediaries are named as excluded, so a brand name you do not own is still off limits as your business name.
Source Search Engine Roundtable, Barry Schwartz, “Google Ads To Allow Some Advertisers To Use Different Business Names,” October 2, 2026 · seroundtable.com
05 Money
Google pays about 100 publishers a tenth of one percent
What your content earns when an AI answer is built from it is, for most sites, close to nothing.
The Information reports that Google’s AI contribution pilot pays participating publishers less than one-tenth of 1% of their advertising revenue, and the publishers cannot see how it is calculated.
Google runs a pilot that pays publishers for content used in its AI features, including AI Overviews, AI Mode and Gemini. The Information reports it covers about 100 publishers, and that for several small and midsize sites “the payments amount to less than one-tenth of 1% of their advertising revenue.”
The spread is wide and does not follow size in any obvious way. One publisher that joined a few months ago has earned about $50,000 to $60,000, while one that joined early is earning more than $1 million a year. Smaller sites have earned less than $1,000 over several months.
Payments are said to depend on how much each source contributes to an AI answer. Some participants told The Information they have no idea how Google calculates the payouts, and the amounts move from month to month with no reason given.
No business in the industries this brief is written for is in that pilot, and that is the point. It is the only price Google has published for content used in an AI answer, and for most participants that price rounds to zero.
So the case for publishing was never a licensing check. It is being the source the answer is built from, which is worth something only when the answer leads somewhere. An architecture and engineering practice or an admissions consultant writing a genuinely useful explainer is buying a citation and the call that may follow it, not a royalty.
What to do about it
Stop valuing a page by what a platform might pay for it, because the published answer to that is a tenth of one percent. Value it by whether somebody who read the AI answer it fed can find you and contact you.
A practical test: take your most useful explainer, ask an assistant the question it answers, and see whether your business is named or merely absorbed. Those are different outcomes and only one of them is worth writing for.
Source Search Engine Roundtable, Barry Schwartz, reporting The Information, “Google Paying ~100 Publishers 0.1% Of Ad Revenue In AI Contribution Pilot,” September 30, 2026 · seroundtable.com
