01 The research
Four researchers ran 37,927 AI buyer questions
If an assistant cannot open and read your website, it answers about you from whatever else it can find.
The businesses it could read were recommended almost twice as often, and the gap held across all four assistant setups tested.
Ido Finder, Assaf Elovic, Gad Shalev and Liad Yosef published a study on September 28 covering 37,927 agent journeys, their term for one buyer question about one business, run across four independent assistant setups over 1,056 real businesses. The businesses were matched on how well known they were and on what the models already knew about them before any comparison was made.
The first finding is about where an answer comes from at all. Only 7 to 10 percent of a finished answer came from the model’s own training knowledge, whether or not the business’s website could be read. The rest was assembled live, out of pages the assistant opened during the conversation.
Then the split. Businesses whose own site the assistant could open and read had the answer built from their own pages 78 percent of the time, against 56 percent for the rest, and were clearly recommended 1.9 times more often. Answers built from the business’s own site were 41 percent more accurate.
The failure mode is not the one everyone worries about. It is omission rather than invention: answers assembled from elsewhere were 3.7 times more likely to contain none of the facts the buyer actually asked for. Someone asking about a surgical practice’s approach to a specific procedure, or asking whether an accounting practice handles multi-state returns, gets a fluent answer with the question quietly unanswered.
Three limits before anyone repeats this. The study is a preprint and has not been through peer review, the four setups it tested produced clear-recommendation rates that varied sevenfold from one to the next, and one of its four authors co-founded a company that sells web-search tools to AI agents, which is the direction the finding points. Forever Cited sells AI visibility work, and this paper argues that much of what this industry sells off-site matters less than whether a site can simply be opened and read.
What to do about it
Open your own website the way a machine does. Turn JavaScript off in your browser, load your three most important pages, and read what is left. Anything that appears only after a script runs is a fair proxy for what an assistant may never see. Start with the pages that answer a buyer’s actual questions: what you do, who you do it for, and what happens first.
Source
arXiv, Ido Finder, Assaf Elovic, Gad Shalev and Liad Yosef, “AX is the New AEO,” September 28, 2026, revised September 29 · arxiv.org
02 The listing change
Google Maps now requires a sign-in to read reviews
Your reviews are the most persuasive thing written about you, and reading past the first few now costs a customer a login.
Sorting reviews, opening all the photos and leaving a review each now prompt a Google account sign-in on Maps.
Barry Schwartz reported on Tuesday that Google Maps now asks for a Google account sign-in before you can see all the reviews on a business listing. The same prompt appears when you try to sort the reviews, search them, browse every photo or leave a review of your own.
Google’s own wording inside the prompt is “Read and search through every review and photo”. The first handful of reviews still show without an account. Everything past that now sits behind the login.
Schwartz says twice in his own write-up that he is guessing at the reason, and his guess is that this is aimed at bots, scrapers and AI platforms pulling review data out of Maps. Google has not said why, so the motive is speculation and the change is the fact.
The effect on you does not depend on the motive. For a surgical practice or an independent insurance broker, reviews are the single most persuasive thing a stranger reads about you, and a customer comparing three options now has to sign in to get past the top few. So does anything reading them on that customer’s behalf.
What to do about it
Look at your own Maps listing today from a browser where you are not signed in, and see what a stranger actually gets. Then make sure the reviews you would most want read are quoted, in plain text, somewhere on your own website, where nothing has to log in to reach them.
Source
Search Engine Roundtable, Barry Schwartz, “Google Maps Now Requires Sign In To Read All Reviews & More,” September 30, 2026 · seroundtable.com
03 The money
Google pays about 100 publishers 0.1% of ad revenue
Being the source an AI answer is built from now has a price, and for almost everyone that price is close to nothing.
One participant earns more than a million dollars a year from the pilot while small sites report under a thousand across several months.
This brief covered Google’s AI contribution pilot on September 14, when it had no public number attached to it. The Information has now reported one: about 100 publishers are in the program, and for several small and midsize sites the payments come to “less than one-tenth of 1% of their advertising revenue”.
The spread between participants is enormous. One publisher that joined early is earning more than $1 million a year. Another that joined a few months ago has made about $50,000 to $60,000 so far. Smaller sites have earned under $1,000 across several months.
Payment is meant to track how much a source contributes to an answer in AI Overviews, AI Mode and Gemini. Participants told The Information they have no idea how Google calculates it, and that the amounts move month to month with no explanation offered.
You are almost certainly not eligible, and that is the part worth keeping. Your content is being read and used at a rate almost nobody is paid for, and even the publishers who are paid cannot audit the figure. For an accounting practice or a wealth advisor writing genuinely useful explainers, the return on that work is being named in the answer, not a payment.
What to do about it
Keep publishing the explainers, and measure the right thing. The question is whether your name comes back in the answer, not whether anyone pays you for the sentence. Ask an assistant the three questions your best customers ask you, and write down whose names appear.
Source
Search Engine Roundtable, Barry Schwartz, “Google Paying ~100 Publishers 0.1% Of Ad Revenue In AI Contribution Pilot,” September 30, 2026, reporting The Information · seroundtable.com
04 The evidence
Three researchers tested 622,080 AI judgments of domain names
What an AI concludes about whether a website is legitimate can turn on the words in its address rather than the evidence.
Adding a claim like “official” to a made-up address moved one model’s alert rate by 45 to 66 points, in both directions.
Daiki Chiba, Hiroki Nakano and Takashi Koide published a study on September 24 analyzing 622,080 judgments across 64 brands and five AI models. They built fake domain names that impersonated real brands, inserted short claim words such as “official” or “not-phishing” into the address itself, and compared them against control names matched for length and punctuation.
The claims worked, and not consistently. Words denying risk inside the address cut one model’s alerts by 45.3 percentage points, even when that model had been handed the real brand and its genuine domain to compare against. Without that comparison, endorsement words in the same position pushed alerts up by 65.6 points in the same model.
The effect runs in both directions and differs by model, so there is no trick here to copy. The authors’ own conclusion is the useful part: an AI should look for independent evidence before treating a domain name as safe or authorized, because the name is asserting its own legitimacy and the model is partly believing it.
This lands hardest where impersonation is already a problem. A tax professional or an insurance broker whose clients receive messages from lookalike addresses now has a second audience reading those addresses, which is the assistant the client asks about them. The study is a six-page preprint and has not been peer reviewed.
What to do about it
Write down the exact address you use, then check it matches on your Google listing, your directory profiles and your email signature. The defense against a lookalike is that your real address is corroborated everywhere a machine can check, which is the same work that gets you named in an answer.
Source
arXiv, Daiki Chiba, Hiroki Nakano and Takashi Koide, “ClaimMirage: When Self-Claims in Domain Names Change LLM Threat Judgments,” September 24, 2026 · arxiv.org
05 The citation test
Google is testing AI Overview citations at the bottom
Where your name appears as a source in an AI answer is moving again, and this time it needs a click.
Glenn Gabe found a desktop layout with the right-hand source bar gone and the full citation list behind a button.
Google is testing a desktop layout that stacks the citation cards at the bottom of an AI Overview instead of in the bar on the right. Glenn Gabe spotted it and posted screenshots showing that the right side citation bar is gone, with a “show all” control expanding the full list underneath the answer.
Google has tested versions of this at least twice before, in September 2025 and again in March 2026, so this is a recurring experiment rather than a launch. Nothing has been announced and most people will not see it.
Barry Schwartz’s caveat is the part that matters. Until someone clicks “show more”, there are no citation cards at all, only the small inline link icons inside the text. Gabe’s read was that the links look more enticing down there, which is a different claim from being more visible.
A test is not a change, and nothing here is worth acting on. It is worth knowing because it is the third variation in a year on one question, which is how prominently an AI answer credits the businesses it was built from.
What to do about it
Nothing to do this week. If you keep screenshots of where your business appears as a source in AI answers, note that the layout itself is unstable, and do not read one week of missing citation cards as a drop in how often you are cited.
Source
Search Engine Roundtable, Barry Schwartz, “Google AI Overviews Testing Citations Cards At Bottom (Not Right Side) Again,” September 29, 2026 · seroundtable.com
