01 The research
OpenAI’s model credited PayPal’s site, five months out of date
An assistant can cite the exact page you corrected last week and still tell your customer the old answer.
Answers written from preview text alone were out of date 6.4% of the time, against 1.3% when the model opened a page and cited it.
Owen Walters, Lihu Chen and Francesca Toni at Imperial College London assembled 225 facts whose answers had changed after the tested models finished training, and the paper opens on one trace that holds the whole finding. Asked on August 8 whether Alex Chriss was the chief executive of PayPal, the model ran one search, opened none of the 18 results, answered yes, and credited paypal.com. Enrique Lores had held the job since March 1.
That was OpenAI’s GPT-5.6 Luna, on OpenAI’s own web search at default settings. Across 1,747 questions it searched every time and opened at least one page in only 46.4% of them. Given an identical set of results, six other model setups answered with no page opened between 2.6% and 76.0% of the time.
Reading mattered more than searching. Answers written from preview text alone were out of date 6.4% of the time, against 1.3% when the model opened a page and cited it, and carried an unsupported citation at the same 6.4% against 2.3%. The link between the two is the part to keep: when an answer was out of date, 81.8% of the citations attached to the claim did not support it, against 0.9% when it was current.
The limits are in the paper. It is a preprint, one commercial model plus six smaller open ones, and the 225 facts cover drug labels, company filings, sports transfers and deaths, not professional services. The mechanism carries across: a page that gets found is not a page that gets read.
What to do about it
Ask ChatGPT, Gemini and Google’s AI answers about the three facts that changed most recently: your address, your prices, who is on staff. If one gives the old answer while crediting your corrected page, move that fact into the page title and first sentence, where preview text comes from. This bites hardest where facts change on a schedule: a software company that reprices, a cybersecurity practice whose certifications lapse, a charter operator whose fleet changes.
Source arXiv, Owen Walters, Lihu Chen and Francesca Toni (Imperial College London), “Retrieved but Not Read: Understanding Why Search-Augmented LLMs Fail to Use Available Evidence,” 15 October 2026 · arxiv.org/abs/2610.21731 (a preprint, not peer reviewed; 225 facts, one commercial model on its own live web search plus six smaller open-weight setups, 4,108 answers hand-graded by the authors)
02 The platform
Google put more links in AI answers, still no clicks
Google will show your link in more places inside its AI answers and still will not tell you whether anyone clicked it.
Google says use of its Preferred Sources feature has grown more than sevenfold this year, and that more metrics will arrive “over time”, with no date attached.
Hema Budaraju, a vice president of product management at Google, published five changes to AI Overviews and AI Mode on October 15. More links embedded in the text of an answer, and those links placed higher up more often. A new section at the top of some answers holds the one or two links that complete a specific task, with a line on why each is useful.
The other three: news carousels for timely topics reaching more countries and languages, a more prominent block of related articles with thumbnails, and clickable publisher and creator names on video results. Google’s own reason for doing it is that use of its Preferred Sources feature “has grown by more than 7x since the start of the year.”
Two things are missing, and both are the reader’s problem rather than Google’s. Those clickable publisher and creator names lead to a Google profile page first, not to the publisher’s own site, which Barry Schwartz of Search Engine Roundtable checked himself. And Google still publishes no click data for AI answers in the reports it gives site owners.
Google’s answer to Schwartz, in full: “We’re continuing to work with website owners to understand what insights will be most helpful to inform their strategies, and we’ll introduce additional metrics over time.” No date. Nick Fox of Google separately said more than 400 publications are now in a pilot program in which Google pays for content used in AI features, which covers news publishers only.
What to do about it
If anyone reports to you on visits from AI answers, ask which part is measured and which is inferred. Google does not publish the clicks, so any figure is an estimate somebody built. Bing Webmaster Tools, which is free, remains the only official report naming the questions where AI answers cite other sites in your field and not yours.
Source Google, Hema Budaraju, “More ways to connect with the web from Search,” 15 October 2026 · blog.google/…/updates-ai-mode-overviews-explore-web · Google’s statement on click data and Nick Fox’s post reported by Search Engine Roundtable, Barry Schwartz, 15 October 2026 · seroundtable.com/…/42308 (Google states both figures itself, neither independently audited; two of the five changes are tests, not finished launches)
03 The evidence
Singapore researchers tested 11 page rewrites on 2,939 answers
Rewriting a page to be named by an AI did almost nothing measurable unless the rewrite named the business it was meant to promote.
Pages the subject published itself drew 24.8% to 44.3% of the citations inside AI answers, against 5.1% of the links in Google’s own results.
Shiqi Zhang and Xiaokui Xiao at the National University of Singapore built a test of whether rewriting a web page changes who an AI answer names, the practice the industry calls generative engine optimization. They ran it on election questions, across 2,939 question-and-engine combinations, with ChatGPT and Gemini models answering from a fixed set of pages.
Start with where the citations came from, because that part is not about elections. News sites supplied 46.4% to 60.9% of citations, pages the subject published itself 24.8% to 44.3%, and open forums such as Reddit as much as 22.2%. In ordinary United States Google results the same two channels accounted for 5.1% and 4.5%.
A rewrite that did not name who it was meant to promote almost never increased how often that name appeared. Of the four kinds tried, reworking a page’s organization, or arguing a position more strongly, beat adding quotations and statistics or changing the wording. In Japan, restructuring lifted mentions of the strongest and weakest parties by 4.28 and 4.12 percentage points, a 14.4% improvement on the weakest party’s own level.
Two limits, then the uncomfortable reading. The test used election pages, older and smaller models, and live web search switched off, so nothing here measures a professional services page in a live engine. And who was asking moved exposure by up to 11.6 percentage points, more than rewriting a single page did, a finding every agency selling page rewrites for AI visibility, Forever Cited included, has to answer for.
What to do about it
If you are paying for AI visibility work, ask whether the rewritten pages actually name your business, your city and your service, because an unnamed rewrite did nothing here. Then ask what is being done with the pages you already control and the public threads where your field gets argued out, admissions consulting and software most of all.
Source arXiv, Shiqi Zhang and Xiaokui Xiao (National University of Singapore), “VoteGEO-Bench: Measuring the Impact of Generative Engine Optimization on Electoral Visibility,” 15 October 2026 · arxiv.org/abs/2610.21330 (a preprint, not peer reviewed; 2,289 United States and 650 Japanese question-and-engine combinations from the 2024 presidential and 2026 Japanese lower-house elections, answered from fixed page sets with live web search off, using gpt-4o-mini, gpt-5-mini and gemini-2.5-flash)
04 The measurement
Outside tools found only a quarter of ChatGPT’s citations
Any report telling you which pages ChatGPT cites is rebuilt from the outside, and most of what it cites is not visible there.
Google Search results and OpenAI’s own developer interface, combined, recovered about one quarter of the web addresses ChatGPT actually cited.
A paper led by Yunxuan Jiang, with Shiqi Zhang and Xiaokui Xiao among its authors, the same two who wrote the study above, watched ChatGPT’s web mode from the inside. Four settings were covered, from ChatGPT 5.2 to 5.5, in the quick mode and the slower thinking mode. The authors recorded how it rewrote the question, which pages it pulled in, and which of those it cited.
Then they tried to rebuild those citations from outside, using Google Search results and OpenAI’s own developer interface. Together, the two recovered only about one quarter of the web addresses ChatGPT had actually cited.
The rebuild worked better at a coarser level. Matching at the level of the website, or of the business named, rather than the exact page, lined up considerably better, and a model combining the outside signals scored above 0.74 on a standard accuracy measure where 0.5 is a coin flip, within the set of pages those tools could see at all.
The authors’ conclusion is a warning about measurement rather than about pages. The quick settings usually search once while the slower mode searches repeatedly, and the questions ChatGPT writes for itself are not the question the customer typed. Being pulled into the shortlist and being cited are two different things, and a tactic measured on one version of ChatGPT has to be measured again when the version changes.
What to do about it
If a vendor shows you a list of the pages ChatGPT cites for your questions, ask how it was built. Where the answer is Google results or a direct data connection rather than the chat window, this study puts the recoverable share at about a quarter. Worth reading, and nobody selling it can make it the whole picture.
Source arXiv, Yunxuan Jiang, Shiqi Zhang, Anqi Xing, Xiangyu Chang and Xiaokui Xiao, “Understanding ChatGPT Citations: Evidence from Production Web Search Across Model Generations,” 11 October 2026 · arxiv.org/abs/2610.15382 (a preprint, not peer reviewed; an observational study of four ChatGPT web settings, no question count stated on the abstract page, two of its authors also wrote the study above)
05 The ad controls
Google Ads now lets you mark which answers qualify
You can now tell Google which answers on your ad’s own form mean a real prospect, instead of paying the same for all of them.
Up to 12 answers per question, each one marked as qualifying or not, inside the form Google opens underneath your ad.
This one is small, it shipped on October 16, and it is the only item today that hands the reader a control rather than a measurement. Google Ads lead forms, the short contact form that opens inside a Google ad so the person never reaches your website, now let you mark which answers to your own custom question count as a qualified prospect.
You can set up to 12 answers per question and check off which of them qualify. The wording inside the product reads: “You can now indicate answers that is a lead to your business to get higher quality leads.” The path is Assets, then Lead Form, then Create Lead Form, then Custom Questions.
The useful part is what you put in the question. An engineering practice that only bids projects above a certain value, a charter operator that flies one region, a specialty veterinary hospital that takes referrals from certain counties: that filter can now sit in the form rather than in a phone call. Hana Kobzová spotted the change on PPC News Feed, from a post by Dafne Nazar.
Two cautions. Google has published no documentation on how these marks feed its bidding, so treat it as a sorting aid until it says otherwise. And a qualifying question is advertising copy, so the answers you offer have to match what you are licensed and able to do, under whichever body regulates your advertising: a state veterinary or medical board, a Department of Insurance, a state bar.
What to do about it
If you run Google ads with a lead form attached, open one custom question this week and write the answers so that one of them, and only one, describes the work you actually want. Then check the Leads page, because a qualifying mark is only useful if somebody reads the result.
Source Search Engine Roundtable, Barry Schwartz, “Google Ads Lead Form Adds Lead Scoring For Lead Qualification,” 16 October 2026 · seroundtable.com/google-ads-lead-form-scoring-42310.html (reporting a change spotted by Hana Kobzová on PPC News Feed via Dafne Nazar; Google has published no announcement, and the quoted wording comes from the product screen in a screenshot)
