01 The research
Seven models changed their advice when the wording changed
Two customers with the same problem can be pointed at different businesses because one of them phrased it as a risk.
Across 240 situations in healthcare, finance, careers and the law, the wording alone moved which recommendation came back.
A paper posted on 21 October built 240 high-stakes situations drawn from healthcare, finance, careers and education, and legal and consumer rights, then wrote every one of them twice. One version stated the odds as a chance of gaining something. The other stated the identical odds as a chance of losing it.
Across seven large language models and 2,880 prompts, the category of recommendation changed in 32.4% of the comparisons where every version returned a usable answer. Most of those changes were not straight reversals. The models moved on how certain they sounded and how firmly they pushed, which the authors checked by having people read the answers and rate them.
Two further tests make this harder to dismiss. The effect survived rephrasing the question conversationally, so it is not a quirk of stiff wording. And it was larger than the ordinary variation you get from asking one model the same thing twice, which is the comparison almost nobody runs.
For a tax practice or a wealth advisor this is close to home, because the two framings are how clients actually talk. “What are my chances of keeping the deduction?” and “what are my chances of losing it?” describe one set of odds. On this evidence the second phrasing can produce a more cautious answer, and a different recommendation.
Every AI visibility number sold in this field, Forever Cited’s own included, is produced by putting questions to a model and counting the answers. This paper is a direct warning about that method, and the honest consequence is that a score is worth only as much as the published list of questions behind it.
What to do about it
When you check how an assistant describes your business, ask the same thing two ways, once as the upside and once as the risk. If the two answers differ, that difference is what your customers are meeting, and only one of the two is in any report you have been shown.
Source arXiv, “Same Probabilities, Different Advice: Gain-Loss Framing Effects in High-Stakes LLM Decision Support,” 21 October 2026, 14 pages · arxiv.org/abs/2610.29294
02 The research
Models recommended artificial intelligence in 59% of answers after tuning
Machines carry preferences learned long before they ever heard of your business, and a customer’s wording decides how strongly those show.
Removing the 7% of training examples that mentioned artificial intelligence cut the rate by 14 points, below the pretrained model.
Asked what to study, which career to take or what product to build, models lean toward answers involving artificial intelligence. A paper posted on 21 October asked where that lean comes from, by comparing models before and after their training was finished and by changing how the question was put.
Models that had only been pretrained named artificial intelligence in 27% of their recommendations. The instruction-tuned versions released to the public did so in 59%. Most of that preference is added by the last stage of training rather than carried by the raw model.
Then the authors retrained models themselves, and general instruction tuning left the rate unchanged. Removing the 7% of training examples that were about artificial intelligence cut it by 14 points, below where the pretrained model started. Data that rewarded being helpful raised it by 9 points, and data ranked by the preferences of varied simulated users lowered it by 9.
The last part is the one that matters for anyone being measured. With the model held completely still, asking for the “best” option raised the rate by 8 points and asking what to do “right now” raised it by 18. Telling the model it was an AI advisor added 4 points.
Read that across to your own category with care. The paper measured recommendations of artificial intelligence, not of brokerages or clinics, so the size of the pull elsewhere is unmeasured. What it does establish is that a recommendation can move by several points on nothing but the words a buyer chose, which is a finding about questions rather than about any one industry.
What to do about it
You cannot change how a model was trained. You can write the page that answers the question a buyer actually types, in their words rather than yours, so that when the wording moves the answer there is something of yours for it to land on.
Source arXiv, “Where Do LLMs Learn to Recommend AI?” 21 October 2026, 18 pages · arxiv.org/abs/2610.29785
03 The research
Models recognized revised guidance far more than they followed it
An assistant can know your details were updated and still answer a customer with the version you replaced.
Two clinicians checking 300 answers by hand found the gap wider than the models’ own scoring did, at 40.3 points.
Medical guidelines change. A paper posted on 22 October assembled 1,709 individual revisions to clinical recommendations across 12 kinds of change, then tested whether models use the new version or merely notice that it exists.
The paper’s main model recognized that a revision had happened in 90.8% of responses and was clinically correct in 80.2%. It fully incorporated the revised guidance in 58.5%. That is a 32.3-point gap between knowing and doing, and it held across 14 models.
Two clinicians then read 300 of the answers and scored them. In that checked sample the gap was wider, at 40.3 points. The failures also cluster: changes to how strongly something is recommended, and changes to who qualifies, were the ones most often answered with the outdated version.
The authors tried a fix. Making the revision more prominent reduced outdated content and improved correctness, and still did not significantly improve full incorporation. The problem is not that the newer information is missing.
This is directly about medical practices, and the pattern will be familiar to anyone whose details have changed. A model can hold your old description and your new one at once and answer with the old one, having registered that it was superseded. The paper tested clinical guidance rather than business information, so the comparison is an inference rather than a finding.
What to do about it
Search your own most important detail the way a customer would, and check which version comes back. If an assistant returns something you replaced, publishing the new version is only half the job, because the old one also has to stop being available to it.
Source arXiv, “Large language models show systematic failures in incorporating evolving medical knowledge,” 22 October 2026 · arxiv.org/abs/2610.30631
04 The platform change
Microsoft now shows which rivals get cited instead
You can now see, free, the exact buyer questions where competitors are quoted and your business is not.
The opportunity list names the questions where rivals are cited, with citation volumes shown as ranges rather than counts.
Microsoft announced a beta feature called Topics Portfolio inside the AI Performance report in Bing Webmaster Tools on 13 October. It groups the questions that produced a citation of your site into themes, so an owner sees the shape of their presence instead of a list of single queries.
Each theme carries three figures. Topic Size is the estimated total citations for that theme across all sites, Sites Cited is how many sites were quoted in it, and Queries is how many distinct questions it contains. Between them they say how much demand a theme represents and how crowded it already is.
The part worth opening the tool for is Top Opportunity Queries, which lists the strongest questions across your whole portfolio where other sites are cited and yours is not. For an accounting firm or a veterinary hospital that is the closest thing yet to a free list of what to write next.
Read the numbers with the caution Microsoft itself asks for. The three figures are approximate, citation volumes in the opportunity list are shown as ranges rather than counts, and Microsoft says to treat them as directional signals for planning, not as a precise ledger. The tab appears only where there is enough data for a site, and there is no click data in it at all.
What to do about it
Open Bing Webmaster Tools, find the AI Performance report, and read Top Opportunity Queries. Write down the three questions on that list that a customer would really ask before hiring you. Those are the three pages worth writing before anything else.
Source Microsoft’s Bing Webmaster blog, “Introducing Topics Portfolio (beta) in AI Performance,” 13 October 2026, reported by Search Engine Roundtable (Barry Schwartz), 13 October 2026 · seroundtable.com/topics-portfolio-bing-webmaster-tools-42288.html
05 The deadline
Google will verify financial advertisers in early 2027
If you advertise anything touching money, an outside registry may soon decide whether your ads run at all.
Advertisers in scope will have to prove an outside registry authorizes them before certain financial ads can run.
In a post about scam advertising on 20 October, Google said it is bringing its Financial Services Verification program to the United States and Canada in early 2027. Advertisers in scope must demonstrate they are authorized by relevant regulators or recognized registries before certain financial ads can run.
In the United States, Google is working with the American Bankers Association and will use its registry to verify authorized banks before ads run. For Canada, Google says only that it is engaging closely with the Canadian Bankers Association as it prepares, and names no registry.
Be precise about who this reaches, because the announcement is not. The named mechanism verifies banks, while Google’s own wording about regulators and recognized registries is broader than banks alone, and the post does not say whether wealth advisors, insurance brokers or accounting firms advertising financial services are in scope. Anyone in those categories should read the guidance Google says it will publish before the rollout rather than assume either way.
The reason to care now rather than in 2027 is that verification programs of this kind stop ads rather than slow them. A campaign that has run for years can be refused at the moment the requirement switches on, and the remedy is paperwork with an outside body working to its own timetable.
Whatever Google ends up requiring, your own advertising rules apply to this channel exactly as they do to every other one. For a registered investment adviser that is the Securities and Exchange Commission’s marketing rule, for an insurance broker it is state licensing, and for a certified public accountant offering financial services it is the AICPA code. None of them is satisfied by clearing a check run by Google.
What to do about it
If you advertise anything that touches money, find out now whether a registry lists you as authorized, and decide who at your business would handle it. Early 2027 is a long way off for a marketing decision and a short way off for a registration.
Source Google, Keerat Sharma, “Expanding our defenses against financial scam ads,” 20 October 2026 · blog.google/products/ads-commerce/financial-services-verification-united-states-canada/
