01 The research
Gemini named the same brands without ever searching
If an assistant leaves your business off its list today, the same list is very likely still there three weeks from now.
Of 202 cases where every early answer agreed on a brand, 194 still agreed 20 days later and not one flipped.
Dmitrij Żatuchin published a test on October 14 of whether counting which brands an AI names is stable enough to act on. The method is the reason to take it seriously. The prediction and the rule for judging it were frozen, hashed and registered publicly before the first question was sent.
The same 162 requests went to Google’s Gemini in three waves across about ten hours on September 22 and 23, then again on October 13, 20.3 days later, with identical wording. Fifty-four questions in English, Estonian and Finnish, covering three categories of software, with twelve candidate brands in play and 648 answers in every wave.
The list held. Of 202 cases where every early answer had agreed on whether a brand was named, 194 were unchanged three weeks later, and not one reversed. A forecast built from the earlier answers beat every simpler guess in all nine question families, at odds against chance of about one in 250.
Then there is the line that should change how a business thinks about this work. Search was switched on for every request, and not one answer ran a search. Gemini named brands out of what it already held, not out of anything it looked up that morning.
Counting which businesses an assistant names is the measurement this whole industry sells, Forever Cited included, so a test of whether that count means anything cuts both ways. It holds, which also means a business missing from the list is missing for weeks rather than for an afternoon.
What to do about it
Write down the three questions a customer would ask an assistant before hiring someone like you, with your city in them. Ask Gemini and ChatGPT each one and save the names that come back. On this evidence a monthly check is enough, and the names on that list are the competitors you are actually being compared against.
Source arXiv, Dmitrij Żatuchin, “Twenty Days Later: A Pre-Registered Test of History-Based Forecasts of Brand Mentions in Gemini Answers,” 14 October 2026 · arxiv.org/abs/2610.19608 (a preprint, not peer reviewed; one model, one 20-day interval and twelve software brands, and the author states it estimates no rate of decay)
02 The consumer data
One in five ChatGPT health chats rated high risk
The more an assistant knew about the person asking, the less safe its advice became, and that person is your next appointment.
About one in four answers that looked safe in general turned out to be unsafe for the particular person who asked.
Shreyasi Roy Chowdhury and Kiran Garimella did something nobody had done before. Every earlier safety study of AI health advice used questions written for the purpose; these two scored real ones. They worked from 202,590 conversations donated with consent by 1,252 people in India, Nigeria, Brazil and Pakistan, and found 15,555 of them were about health.
Roughly one in five of those health conversations was rated high risk. The failures were ordinary rather than exotic: unsafe medication or dosage guidance, harmful self-diagnosis, missing safety warnings, encouraging people to lean on the model, and failing to tell anyone to see a professional.
Every answer was then scored a second time, with everything the assistant knew about that user applied. The person’s own history raised the high-risk share by 13 to 17 percentage points, and about one in four answers that scored safe on their own were unsafe for the specific person asking.
Read that as an owner rather than as a clinician. The answer your customer walks in holding was never a general answer. It was shaped by what the assistant knew about them, and the shaping is what made it worse. A pet owner reciting a dose to an emergency veterinary hospital, or a client telling a tax practice what an assistant said about their own return, is repeating something tuned to them by a system that owes them nothing.
The limits are real and the authors state them. The four countries studied are not the United States, the scoring was done by another model rather than by clinicians, and the work is a preprint. What survives all three is the direction: making the advice personal made it less safe, not more.
What to do about it
Ask the next three customers who mention an assistant what it actually told them, and write the answers down word for word. That sentence is what you are now competing with in the first ten minutes of an appointment. Correcting it is professional judgment and it belongs to the licensed person, not to a marketing channel, and nothing you publish should repeat an assistant’s claim in order to argue with it.
Source arXiv, Shreyasi Roy Chowdhury and Kiran Garimella, “How Safe Is ChatGPT Health Advice? Evidence from Real User Conversations in Four Global South Countries,” 13 October 2026 · arxiv.org/abs/2610.18465 (a preprint, not peer reviewed; safety was rated by a language model against a published rubric, not by clinicians)
03 The platform
Bing now names the questions that skip your site
Microsoft will show you, free, the exact questions where AI answers cite other businesses in your field and never cite you.
The report is a beta, the figures are published as ranges, and it only opens if your site already has enough data to fill it.
Microsoft announced a report on October 13 inside Bing Webmaster Tools, which is free to use, called Topics Portfolio. Microsoft’s word for the searches behind an AI answer is grounding queries, meaning the questions its own system runs in the background while it gathers sources to answer a person. The report groups the ones your site turns up in into themes.
Microsoft’s own examples of a theme are football, taxes and AI chatbots. Each theme reports how many citations it drew across all sites in the period, how many different sites were cited inside it, and how many separate questions it contains. An accounting practice can see on one screen whether the AI demand sits in the subjects it already writes about, or somewhere else entirely.
One table is the reason to open it. Top Opportunity Queries lists, in Microsoft’s words, “the strongest queries across your whole portfolio where other sites appear and yours does not.” No platform has handed a business that list before, for nothing, from inside the system doing the citing.
Three honest limits. This is Microsoft’s view of Microsoft’s own AI answers, and no equivalent report exists for Google’s or OpenAI’s. It is a beta, the citation volumes appear as ranges rather than figures, and Microsoft labels the rest approximate. And it only shows up, Microsoft says, if there is sufficient data for your site, which quietly excludes the smallest ones.
What to do about it
Check whether your website is verified in Bing Webmaster Tools. If it is not, that is a ten minute job for whoever manages the site, and it is the only free window of its kind. If it is, open the AI Performance report, look for Topics Portfolio, and read the Top Opportunity Queries table before anything else on the screen.
Source Search Engine Roundtable, Barry Schwartz, “Topics Portfolio In Bing Webmaster Tools AI Reports,” 13 October 2026 · www.seroundtable.com/topics-portfolio-bing-webmaster-tools-42288.html (reporting and quoting the Bing Webmaster blog announcement by Krishna Madhavan, Saral Nigam and Vivek Jain of Microsoft)
04 The deadline
Google Ads will move your budget in February 2027
If you run Google Search ads with the Display Network switched on, your money moves in February unless someone changes two settings.
Google’s own instruction is to set the Display Network to low priority with a 50% cap in order to keep what you have today.
Google announced two optional controls in Google Ads on October 14. Channel prioritization lets you mark each place your ads can run as high or low priority, and campaigns fund the high priority ones first before any money reaches the rest. Budget allocation then caps a low priority channel at a percentage of the daily budget.
Underneath the new features is a date. “Starting in February 2027, for Search campaigns with the Google Display Network turned on, the current spend allocation setup will go away, and advertisers may see a shift in channel spending,” Google writes. The instruction that follows is unusually specific: set the Display Network to low priority with a 50% budget allocation to keep the current configuration.
Plenty of accounts have the Display Network switched on without anyone having chosen it, which is how a Search budget ends up buying image placements on other people’s websites. An insurance brokerage or a real estate team running a few thousand dollars a month through Search campaigns is exactly who will see spending move without touching a single setting.
Two caveats worth knowing before you act. Google describes the cap as best effort rather than a hard limit, and the controls roll out over the coming months, so they may not be in your account yet. And a Display placement is an image advertisement, which carries your own profession’s advertising rules exactly as your Search text does: your state insurance regulator, state real estate license law and the Fair Housing Act, your state bar, your state medical board.
What to do about it
Open Google Ads, go through each Search campaign, and note whether Search partners and the Google Display Network are switched on. Write the answer down with today’s date beside it. If the Display Network is on and you want nothing to change, the setting Google names is low priority with a 50% allocation. Put a reminder in January, because the controls arrive over months rather than today.
Source Google, “Guide your campaign spending with new priority and allocation features,” 14 October 2026 · blog.google · reported by Search Engine Roundtable (a company announcement; the February 2027 sentence is quoted from it in full)
05 The money
Google will now test whether your ads caused the sale
The only Google test that proves an advertisement caused a sale needs 1,000 conversions, which most practices will never have.
Holding your ads back from part of your audience is the only way to separate advertising that caused a sale from advertising that merely came before it.
Google rewrote a help page on October 8 and changed who can run its strongest measurement. Conversion Lift no longer needs a Google representative to set it up for Search and Performance Max campaigns. The test works by holding your ads back from part of your audience and comparing what the two groups did.
That comparison is the only thing on offer anywhere that separates cause from coincidence. Every other number in an advertising account tells you a sale happened after someone saw an advertisement, which is not the same claim. Withholding the ads is how you find out whether they did any work at all.
Then read the entry requirements. Google asks for at least 1,000 observed conversions and a minimum campaign budget of $5,000, and says conversions carrying supplementary data do not count toward the thousand. A campaign can sit in only one study at a time.
For a real estate brokerage or a software company those numbers are routine. For a surgical practice, an architecture studio or an admissions consultant, a thousand conversions is years of work, which means the one honest measurement Google offers is reserved for its largest spenders and everyone else keeps the attributed figures. That is worth knowing before you accept any account’s report of what your advertising achieved.
What to do about it
Look up how many conversions your account recorded in the last 90 days. If it is nowhere near a thousand, stop treating the conversion figures in your account as proof that the advertising caused anything, and judge the channel on booked work instead. If you are above the bar, Conversion Lift is now something you or your agency can set up without waiting for a Google representative.
Source Search Engine Roundtable, Barry Schwartz, “Google Ads Self-Services Conversion Lift Measurement,” 8 October 2026 · www.seroundtable.com/google-ads-conversion-lift-measurement-self-service-42250.html (quoting Google’s rewritten help documentation, spotted by Hana Kobzová)
