01 The evidence
Researchers found AI invents doctors when web search is off
If an assistant is not searching the web, your practice is not in the running, and a fabrication is.
The audit covered the 100 largest United States metro areas and checked every recommendation against the official registry for its field.
Two researchers, Hazem Ibrahim and Yasir Zaki, published an audit on September 16 that asked AI assistants to recommend providers across the 100 largest United States metro areas, then checked every name against the official registry for that field. Without web search, only 4% of one model's recommended doctors and 11% of the other's matched a clinician in the city that was asked about. With web search turned on, 64% to 71% matched.
The failure is not random. Advisory firms recommended without web search carried misconduct disclosures with the Securities and Exchange Commission at 3.6 times their share of the registry, even after adjusting for firm size. With search on, the rate fell below the registry's own base rate.
For a surgical practice or a registered investment adviser, that difference decides whether you are in the answer at all. When an assistant answers from memory, most of the names it gives are invented, so being genuinely excellent and genuinely licensed does not get you named. When it retrieves live pages, real providers appear, and what it can find about you starts to decide things.
One finding deserves its own line. In restaurants, where quality and visibility can both be measured, recommended businesses had three to five times the review count but a rating advantage of at most a tenth of a star. Volume of reviews predicted being recommended, and being rated better barely did.
The authors are blunt about the part nobody can see. An answer produced without retrieval carries no sign that its recommendations were never verified. Measuring how often AI assistants name a business is what this field sells, Forever Cited included, and this study lands on that work: a visibility figure that does not say whether web search was on is describing two different machines and reporting one number.
What to do about it
Open the assistant your customers actually use, ask it to recommend someone in your field in your city, and note whether it shows live web sources or answers from memory. Both modes exist inside the same product. The one that retrieves is the one where your own website can earn you a mention.
Source
arXiv, Hazem Ibrahim and Yasir Zaki, “Understanding AI Provider Recommendations in Local Service Markets,” September 16, 2026 · arxiv.org
02 The evidence
ChatGPT and Gemini share 5.4% of their sources
Being named by one assistant tells you nothing about the others, because they are reading almost entirely different pages.
The same shopping question put to two assistants returned almost no overlapping sources, and in most comparisons none at all.
A team led by Lucas Uberti-Bona Marin audited what AI assistants say when people ask what to buy. They gathered 2,528 real commercial-advice questions and graded 1,536 answers from ChatGPT, Google Gemini and Google's AI Overviews, in both the consumer apps and the developer versions.
Start with confidence. ChatGPT gave a first-person pick, its own favorite, in 79% of the answers that recommended a product, against 7% for Gemini and 2% for AI Overviews. The products it named often changed when the same question was asked again.
The number that should worry you is about sources. Asked the same question, the ChatGPT and Gemini apps showed only 5.4% of the same web domains on average, and in 76.7% of comparisons they shared no domain at all. The developer versions diverge again, sharing 12.0% and 14.8% of domains with their own consumer apps.
Read that twice. Two assistants answering one question are reading almost entirely different parts of the internet. A tax practice that appears well in one has learned nothing about the other.
One limit worth stating. This audit covered product recommendations, not professional services. The source behavior it measures sits underneath both, but nobody has published the same test for who to hire.
What to do about it
Pick the two assistants your customers are most likely to use and ask each the same question about your field. Compare which sources each one puts on screen. Those two lists are your actual visibility problem, and they are not the same list.
Source
arXiv, Lucas G. Uberti-Bona Marin and colleagues, “‘If I Had to Buy Just ONE: Galaxy S26 Ultra’: Auditing AI-Generated Product Recommendations,” September 16, 2026 · arxiv.org
03 The platform
Google is turning your Business Profile into an AI Overview
The description a searcher reads about your business is about to be written by Google, not by you.
The panel that currently shows your hours and phone number gains an AI summary and a chat box.
Google is converting the local knowledge panel, the box that appears when someone searches your business name and shows your hours, phone number and reviews, into an AI Overview. The panel now carries an “AI Overview” label and a “Show more” button that opens an AI Mode chat about the business. Barry Schwartz reported it on September 17 after Ben Fisher surfaced the change.
Google has been testing versions of this for a while, so treat it as a rollout in progress rather than a finished product. What is new is that the first description a searcher reads about your business becomes something Google assembles, rather than the fields you filled in.
For a real estate brokerage or an independent insurance broker, a search for your own name is the last step before someone calls. Someone typing your name has already decided to consider you, and an AI summary now sits between that person and your phone number.
Schwartz, who has covered this beat for two decades, put the discomfort plainly: he is not sure he wants AI describing his business when he could describe it himself. You cannot edit an AI summary the way you can edit a business description. What you can edit is the material it reads.
Two cautions. No figures have been published, and this is an observed test rather than an announced launch, so how widely it is running is unknown.
What to do about it
Search your own business name today and look at what sits in that panel. If an AI summary has appeared, read it for anything wrong or out of date, then correct the underlying facts on your profile, your website and your listings, because those are what it is drawing on.
Source
Search Engine Roundtable, Barry Schwartz, “Google Local Knowledge Panel As An AI Overview,” September 17, 2026 · seroundtable.com
04 The thing to do
A scam account uploaded fake phone numbers to 150 businesses
A photo anyone can upload can send your next customer's phone call to somebody else entirely.
Two people spotted it on a roofing company's listing, and Google removed the account about a day later.
Naomi Stevens and Sam Sarsten found an AI-generated photo uploaded to a roofing company's Google Business Profile. The image carried the business's own name and a phone number that did not belong to it.
It was not a one-off. The account behind it had made more than 150 photo contributions to unrelated businesses across the United States, with the same wrong number appearing on images for other businesses. Stevens described it as systematic spam meant to redirect calls.
Google removed the contributor account about a day after the report went up, which is the encouraging half. The discouraging half is that anyone can add a photo to your listing, and the defense that worked here was two people noticing.
For a medical practice or an insurance brokerage, the phone call is the conversion. A customer who dials the number in the photo instead of the number on your listing never reaches you, and never knows they did not. Nothing in your own dashboard flags a call you never received.
Keep the scale honest. This is one documented operation rather than a measured trend, and the fix took about a day once a person reported it. The exposure it demonstrates is open to every listing, including yours.
What to do about it
Open your Google Business Profile, go to the photos, and look at every image you did not upload yourself. Report anything carrying a phone number that is not yours. It takes five minutes, and it is the only check that catches this before a customer does.
Source
Search Engine Roundtable, Barry Schwartz, “Google Maps Phone Number in Photo Scam,” September 10, 2026 · seroundtable.com
05 The money
Judge Brinkema rejected a Google breakup, ordered six years' oversight
The ad market you buy from stays in one company's hands, now with a court-appointed monitor for six years.
The judge said a breakup was neither realistic nor needed and appointed a monitor instead.
Judge Brinkema issued her remedies on September 17 in the United States government's advertising-technology case against Google. She rejected a forced breakup as “neither realistic nor needed”, saying a divestiture would reflect a lack of trust that Google will comply with a court order.
What she ordered instead is behavioral. Google must share data, stop discriminatory bidding that favors its own products, and make its ad exchange work with rival publisher ad servers. A third-party Monitor and a Technical Committee will oversee compliance for six years, and the court can extend that if Google does not comply.
The Department of Justice had asked for more than twelve years of supervision. The plumbing that decides which ads appear where stays in one company's hands, with a referee attached.
If you buy search or display advertising, and a tax practice loading up before filing season or an IT and cybersecurity firm running lead generation both do, nothing changes in your account this month. The interoperability requirement is the part to watch, because it decides whether real alternatives to Google's exchange exist in a few years.
What to do about it
Nothing to change today. Put a note in your calendar for your next advertising review to ask whoever buys your ads whether any non-Google exchange has become worth testing. That is the question this ruling makes answerable in a year or two.
Source
Search Engine Roundtable, Barry Schwartz, “Google Ad Tech Monopoly Remedies: Behavioral Changes With Oversight,” September 17, 2026 · seroundtable.com
