01 The research
ChatGPT cited the pages its own search retrieved
The page checklist your marketing budget pays for explains almost none of why one of your pages gets cited and another does not.
The finding that costs money is which half of the standard advice turned out to do nothing inside your own site.
Mahe Chen and Nick Koudas published a paper on October 10 asking what makes a generative search engine cite one page rather than another. Generative engine optimization, or GEO, is the trade that sells answers to that question. They collected 158,099 answers to commercial questions in two markets, ninety days each, carrying 468,037 citations, then replayed the questions through the engines.
The central result is about where the decision gets made. Comparing pages of the same website, which pages ChatGPT’s own search retrieved explained 32% of the difference in whether a page got cited, against 2.8% for the properties of the pages themselves. Every page ChatGPT cited was one its own search had already pulled up, and those searches are rewritten on every answer.
That reorders the standard advice. “FAQ blocks, structured data, summary boxes and page speed appear effective only across websites,” the authors write, so those things may separate a strong site from a weak one but not one of your own pages from another. The one content property that worked inside a site was plain: pages whose title and paragraphs match the questions people actually ask were cited more.
The second finding is about your name rather than your pages. ChatGPT writes brands it already knows into its own searches, prefers websites it knows, and names brands with no supporting page at all. In controlled tests, swapping the brand name reordered its rankings of products with identical facts, and the authors report the pattern held in a separate market of small vendors, closer to an architecture studio or an insurance brokerage than to a consumer giant.
The paper describes “a growing industry of generative engine optimization” advising publishers how to earn these citations, and Forever Cited is part of that industry. The honest reading is that most of the checklist sold under that name does little inside a single site, and that being retrieved and being remembered are not page edits at all.
What to do about it
Take the one page you most want an assistant to cite and read its title and first two paragraphs against the exact question a customer would type. If they do not match in plain words, that is the edit with evidence behind it. Leave the structured data alone: the paper does not call it useless, it says it does not separate your own pages from each other.
Source arXiv, Mahe Chen and Nick Koudas, “What Drives AI Citations? Content, Retrieval and Model Knowledge in Generative Engine Optimization (GEO),” 10 October 2026 · arxiv.org/abs/2610.15033 (a preprint, not peer reviewed; the authors release their data and code)
02 The change to check today
OpenAI stopped tagging the links ChatGPT sends you
Visitors who arrived from ChatGPT now land in your analytics as direct traffic, so the channel looks like it collapsed overnight.
One analyst measured about 31,000 ChatGPT visits a day disappearing from roughly 500 pages while direct visits rose by about 26,000.
In June, OpenAI began attaching a tracking tag to the links inside ChatGPT answers. A tracking tag, a UTM parameter, is the marker that lets your analytics tell you where a visitor came from. Barry Schwartz of Search Engine Roundtable reported this morning that the tag is gone, and that the change appears to have started with GPT-6, released about a week ago.
Cedric Huesler of Adobe traced it to how GPT-6 opens a link, with no tag and no referrer. In his words: “There’s no tag and no referrer, so your analytics files the visit under direct.”
He compared October 9 to 13 against the prior week across about 500 pages where ChatGPT drove at least 5% of visits. ChatGPT visits fell by roughly 31,000 a day while direct visits rose by roughly 26,000 a day on the same pages, and direct traffic rose on two in three of the pages that lost ChatGPT traffic.
OpenAI has not announced any of this, and Schwartz writes that he is “not sure if it was directly intentional.” A real estate brokerage or a luxury home builder that has spent six months reporting ChatGPT as a growing source of inquiries will watch that source fall away next month for a reason that has nothing to do with the business.
What to do about it
Before you read another traffic report, look at your last two weeks. If visits referred by ChatGPT fell while direct rose by a similar amount on the same pages, you are looking at this change and not a decline. Write the date and the cause into the report, and do not move a marketing budget on the strength of the drop.
Source Search Engine Roundtable, Barry Schwartz, “OpenAI Drops UTM Parameters From ChatGPT Results (GPT-6),” 14 October 2026 · www.seroundtable.com/utm-parameters-missing-chatgpt-42299.html (the evidence is one analyst’s measurement posted publicly and relayed by Schwartz, not a statement from OpenAI)
03 The platform change
Google put a new model behind AI Mode answers
The system deciding whether your business gets named in Google’s AI answers changed on October 13, and only paying subscribers see it.
Argon ranked first on an index scoring finance, legal and tax work, and it reaches Search only for Google’s paid tiers.
Google released Gemini 4 Argon on October 13. Koray Kavukcuoglu, who runs Google DeepMind, wrote that it is available “starting today in the Gemini app and AI Mode in Search if you’re a Google AI Ultra or Pro subscriber.” AI Mode is the conversational version of Google Search, the one that answers in prose instead of listing links.
Google says Argon combines reasoning and writing “across legal, finance, science, and coding to help with high-stakes professional workflows” and that it ranked first on the Vals Index, which measures performance across finance, coding, legal, and tax work.
For an accounting practice, a tax resolution business or a wealth advisory, that means the questions your prospects ask about your field are now being answered by a different system than they were last week, and the people seeing it first are those paying Google for Ultra or Pro.
Read the announcement for what it does not contain. Every benchmark figure in it is Google’s own, the Vals ranking included, and Google did not publish how often Argon names a business correctly, which is the only measure that matters here. Nothing says how sources are chosen or shown.
What to do about it
Ask the question your best customer would ask, in AI Mode, and save the answer with today’s date on it. A model change is the one moment when a before-and-after costs nothing, and the before cannot be taken once it has passed. If nobody has a paid Google tier, note which model you were reading.
Source Google, Koray Kavukcuoglu, “Gemini 4 Argon: Built to solve your hardest problems,” 13 October 2026 · blog.google (a company announcement; all performance figures in it are Google’s own)
04 The consumer data
98 patients consulted an AI before seeing a doctor
Your next customer may arrive already holding an answer from an assistant you never saw.
In a clinic study published in The Lancet, the AI’s list of likely diagnoses matched the doctor’s final one 90% of the time.
Google and Beth Israel Deaconess Medical Center ran a study in which 98 patients consulted a research chatbot called AMIE before an urgent care visit, with supervising physicians watching the conversations in real time. Mike Schaekermann, a research lead at Google, posted it on October 8. It was published in The Lancet.
The chatbot’s differential diagnoses, its ranked list of what the problem might be, matched the doctors’ final diagnoses 90% of the time. Clinicians said the AI summaries helped them prepare for the visit in 75% of cases and influenced their approach to care in more than half.
The caveats are in the post and they are real. 98 patients at one clinic is small, the authors say larger trials are needed, and Google is reporting on a system Google built. No conversation had to be stopped on safety grounds, which is the result Google leads with.
The part that travels past medicine is simpler. The patient walked in already holding an answer. That now happens in a veterinary waiting room, at a listing appointment, on a first call with an insurance broker and in an architect’s opening consultation: the person opposite you has been briefed, and the brief did not come from you.
What to do about it
Write down the three questions a new customer asks you most often, then put the same three to an assistant and read what comes back. You are not auditing the machine or grading it. You are reading the script your next customer walked in with, which is worth knowing before they say the first line of it.
Source Google, Mike Schaekermann, “Study in The Lancet suggests AI could improve patient-physician relationships,” 8 October 2026 · blog.google (a Google post about a study led by Google and Beth Israel Deaconess researchers; 98 patients at a single clinic)
05 The money
Google will now score your advertising data setup
Google is putting a score on your data setup, and the proof offered for the gain is Google’s own two-week analysis.
The headline figure is over 20% more conversions, measured by Google, across fourteen days in June.
Google announced on October 13 that a metric called Data Strength Uplift is launching in Google Ads, showing advertisers the conversion gains it attributes to how their data is set up. A conversion is whatever action you asked the ad to produce: a form, a phone call, a booking.
The number carrying the announcement is that advertisers using Google Tag Gateway “see, on average, over 20% conversion uplift in Demand Gen campaigns.” Then read the footnote. It cites Google’s own global performance data for the fourteen days from June 3 to June 17, 2026, four months old, measured by the company making the claim.
The one named customer sits in this brief’s own audience. The Zebra, an insurance platform, is quoted at a 16% increase in conversion volume and a 19% increase in overall conversion value after doing the work. One company, self-reported, with no method shown.
The setup work the score points at is probably worth doing, and that is the distinction to hold. An uplift figure produced by the company selling the uplift is a claim, not a measurement. Treat the score as a checklist and leave the percentage where you found it, because you can check the list and you cannot check the percentage.
Whatever the setup does for your numbers, the claims in the ad stay yours: state Department of Insurance advertising rules for an insurance broker, Circular 230 for a tax professional, the Securities and Exchange Commission’s Marketing Rule for a wealth advisor, and a license number on the ad where a builder’s state requires one.
What to do about it
If you run ads, open the account and read what the new score says is missing from your data setup. That part is a list you can work through. Never move a vendor’s own uplift figure into a report for a partner, a board or a lender as though it were a measurement, because repeated without its footnote it becomes your claim rather than theirs.
Source Google, “Build your data strength with October’s Demand Gen drop,” 13 October 2026 · blog.google (all figures are Google’s own internal analysis; the footnote cites Google data, global, performance, June 3 to June 17, 2026)
