01 The research
Seven AI models took their preferences from the buyer
Two customers with the same need can get different shortlists from one assistant, because it weighs what each of them said.
If the shortlist moves with the person asking, being the better business is not the whole of being the recommended one.
Two researchers built a choice experiment rather than a survey. Seven large language models each chose among five campaigns whose quality signals and founder names were randomized independently, under three different descriptions of the person asking and one with no description at all. The result is 40,248 recorded choices.
By default, the founder’s name barely mattered. Under a budget-conscious description of the buyer and under a delivery-focused one, the measured effects of gender and of a Black-signaled name were, in the authors’ words, “statistically equivalent to zero.”
The number moved when the person asking stated a preference. Told the backer wanted to support underrepresented creators, the models favored Black-signaled founders by 8.1 percentage points and female-signaled founders by 5.1, and the first of those was positive in all seven models. The effect was strongest among near-tied campaigns, which is where most real shortlists get decided.
The models almost never named gender or ethnicity in their reasons. The authors read this as an agency account of the behavior: “an aligned assistant takes its demographic preferences from its user rather than from its training data.” The preference is not buried in the model; it arrives with the question.
The limits are real and the authors state them. The campaigns were synthetic and the subject was crowdfunding, not professional services. For a wealth advisor the practical version is still worth sitting with: two prospects ask the same assistant for a fee-only fiduciary in the same city, one of them mentions wanting somebody who understands business owners, and the two lists that come back are not the same list.
What to do about it
Ask your question the way your prospects actually ask it, including the part where they describe themselves. Run it three or four ways and keep what comes back.
If your name appears on one phrasing and not another, that is a gap you can describe and write for. It is not a mystery about the model.
Source arXiv 2611.03993, Mirza Samad Ahmed Baig and Syeda Anshrah Gillani, “Whose startup does the AI back? An algorithm audit of reputation and demographic signals in LLM-assisted crowdfunding,” submitted 2 November 2026 · arxiv.org/abs/2611.03993 (seven models, 40,248 responses, five synthetic campaigns, three described backers plus one with no description; a preprint, not peer reviewed; the campaigns are crowdfunding projects rather than professional practices, and the abstract does not say which seven models were used)
02 The platform change
Google is asking AI Mode users to connect their apps
Google is wiring each searcher’s own account into their results, which makes one ranking for everybody the wrong thing to ask for.
A popup asking searchers to connect their Google apps is the mechanism underneath every other finding this week.
Searchers in AI Mode are being shown a popup headed “Supercharge Search with Personal Intelligence,” with a button to set it up. The body text reads “Make Search uniquely yours by connecting more of your Google apps, all while safeguarding your info,” and adds that the searcher is in control and can change it at any time.
Switching Personal Intelligence on also switches on connected apps inside AI Mode. Search Engine Roundtable’s Barry Schwartz, relaying a screenshot posted by Sachin Patel, called the prompt a little extreme while noting that Google has done similar things before.
Google has announced nothing here. The report names no apps, no data types, no regions and no dates, which makes this a popup somebody photographed rather than a product launch. Treat the mechanism as real and the scope as unknown.
Personalized results are what makes a single ranking meaningless. For an accounting practice the version that matters is this: the partner checking whether the practice comes up for a local tax query is signed into their own Google account, and so is every prospect, and no two of those sessions are the same test.
What to do about it
Stop treating one search from your own logged-in account as a measurement. Check from a signed-out window as well, and expect the two to disagree.
If you report AI visibility to anybody, say which of the two you used. A number without that detail cannot be repeated by the person reading it.
Source Search Engine Roundtable, Barry Schwartz, “Google Pushing AI Mode Users To Opt In To Personal Intelligence,” 6 November 2026 · seroundtable.com/google-personal-intelligence-popup-42468.html (built on a screenshot from Sachin Patel; Google has made no announcement, and the report states no apps, no data types, no regions and no dates; Search Engine Watch, where the screenshot was first shared, does not resolve from this session and was not read)
03 The measurement problem
Five AI judges ranked the same systems in different orders
A visibility score produced by asking one AI model to judge is not a stable number, and this industry sells those.
If swapping the judge swaps the ranking, a score built on one model is a measurement with no stated margin.
Two researchers asked a narrow question with a preregistered design. Five judge models from three providers re-judged 4,000 recorded head-to-head comparisons that already carried human votes, each one in both response orders, covering nine systems and 33 pairings.
Agreement with the human verdict was low, at a median Cohen’s kappa of 0.22, on a scale where 1.0 is perfect agreement and 0 is chance. More to the point, the rankings did not survive a change of judge. Rank correlation with the human ordering ran from 0.17 to 0.89, and the preregistered expectation that rankings would hold was not supported.
Checking a conclusion against a second cheap judge from another provider did help, by 0.10 in precision. The authors trace that gain to filtering out one badly aligned judge, and comparisons involving that judge produced a usable conclusion only 35% of the time. Some judges never settled on 26% to 70% of the pairings even at 100 verdicts each.
An AI visibility score is the number every agency in this field now sells, Forever Cited included, and most of those scores come from asking a single model to judge. This paper is the reason we put the same questions to five engines and report the spread instead of one figure.
The authors name their own ceiling: one human vote per comparison cannot separate a bias the judges share from ordinary disagreement between people. And this measured chatbot answers against each other, not businesses inside a recommendation.
What to do about it
If somebody hands you an AI visibility score, ask which model produced it and what the number does when a different model is asked the same thing.
A score that moves when the judge changes is still useful as a direction. It is not a position, and anyone presenting it as one is overselling it.
Source arXiv 2611.04090, Sangdo Han and Moonsoo Kim, “Swapping the Judge Changes the Ranking: A Preregistered Study of Cross-Provider Replication in LLM-Judged System Comparisons,” submitted 3 November 2026 · arxiv.org/abs/2611.04090 (five judge models from three providers, 4,000 Chatbot Arena battles, nine systems, 33 pairs, both response orders; preregistered, with a post-registration reduction of the judge set disclosed by the authors; a preprint, not peer reviewed; it compares chatbot systems rather than businesses)
04 The operational change
Google Ads added a telemedicine certification this month
If your practice advertises telemedicine on Google, an outside accreditation now sits between you and the ad you want to run.
A new certification route opens this month, and the accreditation it depends on has to come first.
Google Ads is adding a telemedicine certification option through the National Association of Boards of Pharmacy’s Healthcare Merchant Accreditation Program. It joins the existing routes rather than replacing them, and LegitScript, NABP, Digital Pharmacy Accreditation and .Pharmacy all remain available.
The change takes effect in November 2026 and covers Canada, the United Kingdom and the United States. Advertisers can apply once the policy is published.
The ordering is the part that costs time: an advertiser has to hold NABP accreditation before applying to Google. An accreditation from a standards body is not a form you fill in on the afternoon you decide to run the ads.
For a surgical practice or a concierge clinic advertising remote consultations, this is a gate in front of the channel rather than a setting inside it. The general version travels well past medicine. An advertising platform can require a credential from somebody else before it will take your money, and your own board’s advertising rules still govern the ad on top of whatever the platform asks for.
Google’s own policy pages do not resolve from this session, so the certification name, the three countries and the November timing here rest on Search Engine Roundtable’s reading of the policy update rather than on the document itself.
What to do about it
If your practice runs or plans telemedicine ads on Google in Canada, the United Kingdom or the United States, find out this week whether you already hold NABP accreditation. That step comes before the Google application, not after it.
If an agency runs the account, ask which certification route it is on. And have whoever signs off your advertising check the copy against your own board’s rules, which this change does not touch.
Source Search Engine Roundtable, Barry Schwartz, “Google Ads Gains A New Telemedicine Certification Option,” 5 November 2026 · seroundtable.com/google-ads-telemedicine-certification-42460.html (reporting a Google Ads policy update; Google’s support pages do not resolve from this session, so the accreditation requirement, the three countries and the November 2026 timing are as Search Engine Roundtable states them and were not read at the primary source)
05 The money
Researchers found automated visibility tuning lost to a handwritten prompt
One rewrite of your page does not lift you everywhere, because the measured gain depends on what the customer is deciding.
The money question is whether a single rewrite helps on every question, and the measured answer is no.
Four researchers tested something the AI visibility market mostly assumes, which is that a page rewritten to be more citable becomes more citable in general. They rebuilt the test around what the customer is actually trying to decide, across four subject areas and ten different states of need and decision progress.
The gains from existing optimization methods varied by what the user wanted. One description serves people with different goals, so its benefit is, in the authors’ framing, dependent on the request rather than fixed in the page.
The finding with the money in it is that automated prompt optimization did worse on average than the handwritten prompt it started from. Tracing the updates showed why: the automated version drifted away from explicit recommendation instructions and toward generic promotional tactics.
Their own method holds the instructions in a fixed core and learns only what gets appended, and they report a 35% improvement in average citation-rank gain over the strongest previous learning-based approach, leading on nine of the ten states of need. This is a method paper proposing its own method, which is a reason to read the comparison carefully rather than the claim.
For an architecture practice the useful version is that optimizing the site for AI search is not one job with one outcome. The page written for somebody choosing between restoration specialists is not the page for somebody who has not yet decided to hire an architect at all, and the measured gain follows that difference.
What to do about it
Before you pay for AI visibility work, ask which customer question the work is meant to win and what the answer to it looks like today. Work that cannot name the question cannot be checked afterwards.
If any part of it is an automated optimizer, ask to see what it actually wrote. The drift this paper measured was toward promotional language, and promotional language is the kind your own advertising rules govern.
Source arXiv 2611.04225, Runlin Lei, Shiqi Zhang, Xiaokui Xiao and Zhewei Wei, “IntentGEO: Intent-aware Generative Engine Optimization for Recommendation,” submitted 3 November 2026 · arxiv.org/abs/2611.04225 (four subject areas, ten states of need, plus a separately adapted test on Doubao, a commercially deployed closed-source assistant; a preprint, not peer reviewed; the authors propose the method they report as strongest, and the abstract gives no query or trial counts)
