01 The platform
One analyst sees AI Overview citations vanishing, Google stays silent
If your search impressions fell this week, the cause may be a change nobody at Google has confirmed.
The reduction is reported by one analyst, unreplicated by another and unacknowledged by Google, which is what makes it hard to act on.
Barry Schwartz reported on 29 October that Google may be testing the removal of the citations card from AI Overviews. That card is the small panel of clickable source links that sits beside an AI answer, and on a desktop search it is the only part of the answer that names a business at all.
The change was spotted by Brodie Clark, who tracks search result layouts and posted side-by-side screenshots. In his words, “Google has made recent changes to AI overviews that are brutal for websites.” He said Google had previously featured citations for two or three domains in visible cards alongside the result, and that he is “now seeing a drastic reduction in their appearance”.
Clark went further and told readers what to check. If impressions and site visibility have declined in recent days, he wrote, this is likely a contributing factor.
Then the honest part. Schwartz could not reproduce it. He noted that the citation card does load in later on his own searches, so it may be a loading bug rather than a deliberate test, and he asked Rajan Patel at Google directly which it was. He had no answer by the time he published. Clark replied the following day that he was seeing it “consistently across a lot of test queries”.
So this is one observer reporting a change, a second unable to confirm it, and the company that would know declining to say. It is in the brief anyway, because the thing being reported is the credit line itself. A tax practice or an architecture firm whose inquiries arrive through search cannot tell the difference between a quiet week and a layout change that removed its name, and right now neither can anyone else.
What to do about it
Do not rebuild anything on the strength of an unconfirmed test. Instead write down where your search impressions stood this week, so that if the change is confirmed you can date it.
Schwartz also noted Google has tested putting the citation card below the AI answer rather than beside it. A card that moved is not a card that went, and the two look identical in a traffic report.
Source Search Engine Roundtable, Barry Schwartz, “Is Google Testing Dropping Citations Cards From AI Overviews?,” 29 October 2026 · seroundtable.com/citations-cards-google-ai-overviews-42413.html
02 The market
Google search ads reached $82 billion in one quarter
People are asking more buying questions than ever, and the company answering them keeps the visit.
The revenue is not the news. The sentence about query growth underneath it is.
Google reported third-quarter results on 28 October. Search advertising revenue rose 10.5% to $82 billion, up from about $74 billion a year earlier, and overall revenue rose 25% to $127.6 billion.
The line worth reading twice came from Sundar Pichai. “Our AI investments are driving overall Search query growth, including commercial queries,” he said. “In fact, our year-over-year query growth rate is the highest it’s been in the last five years.”
Read that from the other side of the screen. People are not asking fewer buying questions now that a machine answers them. They are asking more, and the fastest growth in five years is happening inside a surface that answers in place.
Microsoft reported the same week. Its search and advertising revenue excluding traffic acquisition costs rose 12%, two points better than the prior quarter, on company revenue of $92.3 billion. The smaller engine is growing its advertising faster than the larger one.
Pichai also said the Gemini app has reached one billion users, which he called the fastest-growing product in Google’s history. Google’s stock fell 7% in after-hours trading on the report, which tells you about spending rather than about demand.
For an insurance broker or a veterinary hospital the practical reading is narrow and firm. The question volume is there and it is rising; what has changed is how often answering it requires visiting you.
Source Search Engine Roundtable, Barry Schwartz, “Google Ad Revenue Up 10.5% With Overall Revenue Up 25%,” 28 October 2026 · seroundtable.com/google-earnings-q326-42403.html
Source Search Engine Roundtable, Barry Schwartz, “Bing Ads Revenue Up 12% With Overall Revenue Up 19%,” 28 October 2026 · seroundtable.com/microsoft-earnings-q127-42404.html
03 The research
Researchers found page features explain 2.8% of citation differences
The page checklist most agencies sell moves almost nothing; being a name the model already knows moves more.
The study separates what your page does from what the engine’s own search step does, and the gap is wide.
Mahe Chen and Nick Koudas posted their study on 10 October. They gathered two 90-day collections of answers to commercial questions in two contrasting markets: 158,099 answers carrying 468,037 citations, then asked which pages got cited and why.
One content finding survives. Comparing pages on the same website, pages whose title and paragraphs match the questions people actually ask are cited more. The rest of the usual list did not survive: FAQ blocks, structured data, summary boxes and page speed appeared to work only when comparing one website against another, which is the comparison that confounds everything else about the two businesses.
Then the number that reframes the field. Page properties explained 2.8% of the variation between pages on the same site. The engine’s own retrieval step explained 32%.
The mechanism is plain once stated. Every page ChatGPT cites is one its own search retrieved, and it retrieves against searches it rewrites for itself on every single answer, searches the authors found “rarely coincide with Google’s or Bing’s results”. Content still matters, but it matters by matching those rewritten queries in meaning rather than by carrying a particular block of markup.
The model’s memory does the rest. The authors found 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 brand names reordered its rankings of products whose stated facts were identical. Their summary is six words long: pages are cited when searched, brands are also named when remembered. The pattern held again in a market of small business-software vendors.
The paper opens by naming a growing industry of generative engine optimization, which is the business of getting a company named in AI answers. Forever Cited is in that industry, and this finding cuts against the easiest thing anyone in it can sell, which is a checklist of page changes. We would rather publish the 2.8% than sell around it.
What to do about it
Stop reading your Google ranking as a proxy for whether an assistant will cite you. A companion study published a day later found that Google Search and OpenAI’s own interface, combined, recover only about a quarter of the URLs ChatGPT actually cites.
Write the way the question is asked. Of everything the authors tested on a page, matching real questions in the title and the opening paragraphs was the one thing that held up within a single website.
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
Source arXiv, Jiang, Zhang, Xing, Chang and Xiao, “Understanding ChatGPT Citations: Evidence from Production Web Search Across Model Generations,” 11 October 2026 · arxiv.org/abs/2610.15382
04 The mechanics
Google says blocking an AI crawler is not blocking training
The file that decides whether AI can read your site at all is one almost nobody has tested.
Two settings that look like one thing are different things, and mixing them up leaves content already taken.
Gary Illyes and Chery Prommawin released the second video in Google’s Non-Canonical Cabinet series on 21 October, titled “What can you actually put in a robots.txt?” A robots.txt file is a plain text file at the root of a website that tells automated visitors which pages they may read.
The distinction they drew is the one worth carrying. Blocking an AI crawler from reaching a page is a different act from telling a company not to train on content it has already obtained. The first is a door; the second is a permission, and the file only operates the door.
Their second warning is about false confidence. Do not assume a rule is working because it reads sensibly to a human. Google’s advice was to validate the syntax and test the actual URL patterns against real addresses.
Encoding is the quiet failure. Rules covering addresses with non-English characters or emoji need percent-encoding, and the team noted that inconsistencies there can undermine a rule entirely, particularly across different crawler implementations. A rule that works for one company’s crawler may do nothing against another’s.
The file should follow RFC 9309, the published standard for this format, and the rules apply to search crawlers, AI training crawlers and other well-behaved crawlers alike. The phrase well-behaved is carrying weight: the file asks, it does not enforce.
For a cybersecurity firm or an architecture practice this is a decision rather than a chore. Every page you close to an AI crawler is a page that cannot be cited back to you, and this issue’s research says being retrieved is most of what citation is.
Source Search Engine Roundtable, Barry Schwartz, “Google On robots.txt Rules For AI Crawlers, Emojis, International URLs & More,” 21 October 2026 · seroundtable.com/google-video-robots-txt-rules-42355.html
05 The money
Google is putting ads inside AI Mode follow-up questions
Your competitor can now buy the second question a customer asks, not just the first.
Ten more countries now, three more by the end of the year, and follow-up turns in the United States.
Google announced on 28 October that it is expanding ads inside AI Mode to Australia, Canada, India, Kenya, Malaysia, New Zealand, Nigeria, Pakistan, the Philippines and Singapore, with Brazil, Mexico and the United Kingdom following in the fourth quarter.
The structural change is the other half. “To allow ads in follow-up queries in AI Mode, we are launching multiturn ads in the U.S.,” Google said. A spokesperson put it plainly: multiturn ads appear on responses to follow-up questions as people go back and forth with AI Mode.
Until now a paid placement sat against the opening question. The second, third and fourth turns of a conversation are now inventory too, which is where a customer stops asking what something is and starts asking who should do it.
Google also offered a number on how the placements perform: “click-through rates are already similar between sponsored and organic product list items within AI Mode”. Read carefully, that is a claim about product list items rather than about text links, and Google did not give a figure for either.
Everything will be labeled “Sponsored”, which is the right disclosure and also the whole point: it is a label a customer sees after they have already been given the answer.
A charter operator or a wealth advisory firm should take one thing from this. The conversation that used to end at your website now continues inside the answer, and the later turns can be bought by somebody else.
Source Search Engine Roundtable, Barry Schwartz, “Google Expands Ads In AI Mode To More Countries & Multiturn Ads In U.S.,” 28 October 2026 · seroundtable.com/google-expands-ads-ai-mode-42412.html
