Tuesday, September 22, 2026 Five things that moved Read time: 8 min

Today’s theme · What decides who gets named

ChatGPT named a business 2.8% of the time. With its website cited, 49%.

Two analyses posted to arXiv last week put numbers on what actually makes an AI assistant name a business, and the target is narrower and more fragmented than the advice being sold around it.

Most advice about getting found by AI assistants is sold without evidence behind it. Two analyses posted to arXiv on 19 September put numbers on the question, and the numbers are unusually specific about what moves.

The first measured 34,960 answers and found that whether an assistant names a business turns mostly on whether that business’s own pages were pulled into the answer. The second put the same questions to four assistants and found they almost never cite the same page.

Read together they describe a target that is both narrow and fragmented. Your own site is the thing that gets you named, and it has to earn that separately on every assistant your customers use.

The other three items are about changes around that target. Google is testing answer links that lead further into Google, it is switching off review collection without warning, and it is indexing thousands of spam pages made inside its own product.


01 The evidence

ChatGPT named businesses 2.8% until their websites were cited

If your own pages never get pulled into the answer, you are effectively absent from the recommendation.

An analysis of 34,960 AI answers found that the strongest signal of being named is something a business owns outright.

Benjamin Tannenbaum analyzed 34,960 answers from 75 monitored projects, covering 2,854 questions put repeatedly to ChatGPT and Google Gemini between June and September 2026. The question behind the work is a plain one: when does an assistant actually name a particular business?

When neither the business name nor its own website appeared in the material the assistant pulled up, it named that business 2.8% of the time on ChatGPT and 3.8% on Gemini. When the business’s own website was cited, those rates rose to 49.0% and 58.4%. When the assistant both cited the site and went looking for the brand by name, they reached 91.4% and 100%.

The pattern held inside individual projects across repeated runs, where adding that own-site citation moved the mention rate by 40.2 points on ChatGPT and 49.0 points on Gemini. History carried weight of its own. A business not named last time, with no citation this time, was named 1.6% of the time on the next run.

Two cautions belong with these numbers. The data comes from 75 projects on a single AI-visibility monitoring platform rather than an independent audit, it is a preprint that has not been peer reviewed, and the author writes that the model is predictive and observational, not a causal description of how the engines work inside.

What to do about it

Ask three or four real buying questions in ChatGPT and Gemini, and look at the sources listed under the answer. The question is not whether you are mentioned. It is whether any page you own is in that list.

Source arXiv, Benjamin Tannenbaum, “From Prompt to Recommendation: A Fitted Stage Model of Brand Visibility in AI Search,” September 19, 2026 · arxiv.org

02 The evidence, continued

Four AI assistants cited almost no pages in common

Being named by ChatGPT tells you almost nothing about whether Perplexity or Copilot will name you at all.

An audit of 589 citations across four assistants found the overlap between them was close to zero.

The same author ran a separate audit on 6 June 2026, putting 15 commercial questions to ChatGPT, Microsoft Copilot, Google and Perplexity and recording every page each one cited. That produced 589 citations covering 528 distinct pages across 356 websites.

The four assistants almost never agreed with each other. 84.9% of engine pairs shared no cited page at all, and among the top five citations there was no overlap in any of the 60 comparisons made.

96.4% of the pages cited appeared on only one of the four assistants. Any single assistant showed between 11.4% and 42.6% of the pages the four produced between them, so checking one and stopping tells you about a fraction of what your customers see.

A separate comparison of the same engine on 5 and 6 June found 67.0% of the cited pages changed from one day to the next. The author’s conclusion is narrow and worth keeping: a score calculated without asking a live engine can describe how good a page is, but it cannot tell you whether an engine will show it.

What to do about it

Check more than one assistant before you draw any conclusion about your visibility, and check on more than one day. A single look at a single engine is the weakest evidence available, and it is the evidence most visibility reports are built on.

Source arXiv, Benjamin Tannenbaum, “Scoring With the Engine: Retrieval Exposure, Cross-Engine Divergence, and the Limits of Engine-Agnostic GEO Scores,” September 19, 2026 · arxiv.org

03 The platform change

Google is testing AI Overview links that open AI Mode

A link inside Google’s answer may now lead to another Google answer instead of to your website.

Links that look like citations are being tested as routes deeper into Google rather than out to the web.

Google is testing links inside its AI Overviews that do not lead to a web page. They open AI Mode, Google’s fuller chat-style answer, with a follow-up question already loaded.

The behavior was spotted by Gagan Ghotra, who posted a video of it working. Google has not announced the test, and a test is not a rollout, so it may never ship in this form.

The direction is consistent with the rest of this year. Google keeps the reader inside its own answer for longer, and this particular test removes the last step where a reader crosses over to somebody’s website.

For a business the practical effect is that a citation is worth less than it looks. Being named in an AI Overview was already worth less than a visit, and a link that returns the reader to another Google answer removes even the chance of one.

What to do about it

Stop treating a mention in an AI answer as traffic. Count it as reach, the way you would count a billboard, and make sure the pages that do get visits are the ones built to turn a visitor into a call.

Source Search Engine Roundtable, Barry Schwartz, “Google AI Overview Links Pushing You To AI Mode Over Web Pages,” September 22, 2026 · seroundtable.com

04 The operational change

Spam spikes now freeze new reviews on Google listings

Your listing can stop accepting new reviews for days, without warning, because of reviews you never asked for.

Google has begun emailing owners to say that new ratings and reviews on their listing are temporarily switched off.

Google has started emailing business owners when its systems detect a spike in spam reviews on a Google Business Profile, the free listing that shows up in Maps and beside a brand search. The email says the fake reviews were removed before they affected the rating.

It also carries something owners will feel more directly. New ratings, reviews and other contributions are temporarily switched off, and Google says it lifts the pause once the risk passes, typically within a few days.

No action is required from the owner, and legitimate reviews removed by mistake can be appealed. The pause is automatic, and that is the part worth planning around rather than reacting to.

It is already catching real businesses. Amy Toman described a client who had just opened a new location and kept a steady flow of reviews coming in, and said the filter needs refining. A wealth advisory practice or an insurance brokerage that asks every satisfied client for a review can look, to an automated filter, much like one buying them.

What to do about it

Spread review requests out rather than sending a batch, especially around a new location or a busy month. If your listing goes quiet, check your email before assuming customers stopped writing.

Source Search Engine Roundtable, Barry Schwartz, “Google Business Profiles Email: Detected A Spike In Spam Reviews,” September 22, 2026 · seroundtable.com

05 The noise

Google indexed over 12,000 spam Gemini Notebook pages

The pages competing with yours to shape an AI answer now include thousands nobody wrote for a reader.

Public pages made inside Google’s own research tool are being indexed in the thousands and carrying spam.

Google is indexing over 12,000 public Gemini Notebook pages carrying spam. Gemini Notebook is Google’s own research tool, previously called NotebookLM, and anything made in it can be published to the open web.

The indexed content includes coupon codes, peptides, pornography and sex toys. Charles Floate found it first, and Gagan Ghotra described it as black hat search marketers using public Gemini Notebooks as a new parasite tactic, meaning spam that rides on a trusted website’s reputation.

Glenn Gabe, who amplified the finding, expects the pages to disappear before long. Google has not removed them yet. The pattern is a familiar one, where a new product ships and the spam arrives ahead of the safeguards.

This matters to an ordinary business because the material an assistant draws from is that same index. Every spam page is one more item competing for the space where an answer about your industry gets assembled, and it carries a google.com address for as long as it sits there.

What to do about it

Nothing here needs a response today. It is worth knowing the next time someone shows you a competitor outranking you on a page that reads like it was written by nobody, because sometimes it was.

Source Search Engine Roundtable, Barry Schwartz, “Gemini Notebook Pages Spamming Google Search,” September 22, 2026 · seroundtable.com

If you do one thing this week

Find out whether your own website appears in the answers your customers are already getting. Ask three or four real buying questions in ChatGPT, Gemini and Perplexity, and write down whether any page you own is among the sources.

If none of them is, that is the gap the first study measured, and it is the only item on today’s list you control directly. If you appear on one assistant and not the others, that is the second study, and it is the normal state rather than a fault to fix.

Forever Cited sells the case that being the cited source is what gets a business named, which makes today’s lead finding the one we would most like to be true. Both studies share a single author, neither has been peer reviewed, and the author states plainly that the model is observational rather than causal.

Read the full archive →

Forever Cited · AI Visibility, Engineered by Industry · One brand per market.
Sources are linked in full above. We link to primary documents and original research wherever they exist.

The daily brief

Get this briefing every morning — free.

Five sourced items on AI search and visibility, written for business owners. No pitch, no spam. Confirm by email (double opt-in); unsubscribe anytime.

← All AI Industry Updates  ·  forevercited.ai