01 The market data
Google’s AI answers cut Wikipedia’s search visits 5.45%
The page an AI answer is built from does not get the visit; the answer keeps it.
Nothing on Wikipedia’s pages changed, which means the whole loss came from what Google placed above them.
Two researchers, Mehrzad Khosravi and Hema Yoganarasimhan, posted a new version of their study on 25 October. Where Google switched AI Overviews on by default, English Wikipedia’s visits from search fell 5.45%.
The comparison is what makes the finding hold up. Google turned the feature on in different countries at different times, so the authors lined up the same Wikipedia articles in English, German and French and watched only the English ones lose referrals. Measured against the German versions the fall was 5.45%; against the French versions, 4.82%.
Wikipedia is the page an AI answer is most likely to be built from, and it lost visits anyway. A physician’s explainer on a procedure, or a surgical practice’s page on recovery time, sits in exactly the same position with none of Wikipedia’s standing to protect it.
Every agency in this field, Forever Cited included, sells being named in AI answers as the win. This study is the clearest case yet that being named and being visited are two different things. One limit is worth stating plainly: Wikipedia sells nothing, so what it lost was attention rather than customers, and the study measures referrals, not revenue.
What to do about it
Ask whoever reports on your website for visits from Google search last month against the same month a year ago. If that number fell while your position in the results held, this is what you are looking at. The page worth your attention is the one a customer still needs after reading a summary of it.
Source arXiv, Mehrzad Khosravi and Hema Yoganarasimhan, “Impact of AI Search Summaries on Website Traffic: Evidence from Google AI Overviews and Wikipedia,” revised 25 October 2026 (first posted 5 February 2026; the 25 October revision is the current version) · arxiv.org/abs/2602.18455
02 The research
Pages that were true but one-sided moved four AI models
Nobody has to lie about your field to move the answer, only to publish one side of it.
A machine summarizing a contested question in your field follows the weight of what has been published, not the merits.
Five researchers published a study on 25 October on what they call stance poisoning. Every page they added to the pool the AI system read was factually accurate. The only thing they controlled was how many of those pages argued the same side of a contested question.
At the point where 18.9% of the documents on a topic leaned one way, all four of the commercial AI models tested shifted in that direction. The four came from four different vendors. Across four levels of concentration the shift grew each time, which is the pattern that makes a result hard to dismiss as noise.
This matters most where your customers ask something with two defensible answers. An adviser who brokers the sale of businesses, asked whether now is a bad time to sell, or a franchise consultant asked whether a franchise beats starting out independently, is being summarized by a system whose position follows the balance of what has been published rather than who is right.
The authors also built a repair, and it is the more interesting half. Sitting between the step that gathers documents and the step that writes the answer, forcing a fixed mix of two supporting, two opposing and one neutral document cut the lean substantially. That is a fix for whoever runs the AI system, not for you, which is the part worth noticing.
What to do about it
You cannot balance the internet. You can make sure the defensible case for your own position exists in full sentences on a page you own. Where your field has a genuine two-sided argument, a page that sets out both sides and says which one you take is more use to a machine writing a summary than a page that only asserts.
Source arXiv, Puyuan Yang, Shaoxuan Dong, Liujie Wu, Xiaochun Cheng and Bo Li, “Factual Yet One-Sided: Stance Poisoning and Evidence Balancing in Retrieval-Augmented Generation,” 25 October 2026 · arxiv.org/abs/2610.34294
03 The research
AI models found the right website 37.5% of the time
When a customer asks an assistant for you by name, the page it has seen most often wins, not the best one.
The gap between looking a page up and recalling it from memory is 22 percentage points, and your customer never learns which one answered.
Six researchers built a system that indexes 201 million web pages and tested it on real traffic from people searching for one specific site. The paper went up on 26 October.
Asked to name the right website outright, from memory, the leading models managed at most 37.5%. A system that goes and looks the page up reached 59.6%, and an older method built on what people had clicked before reached 46.5%. Answering from memory also cost at least forty times as much per question.
The line in the paper that should concern an owner is the one about what holds the result back. Accuracy is limited by how much a page has been seen, not by how capable the model is. A newer enterprise software company with a better product and a thin record online loses to an older one with a weaker product and a thick one.
Read the three figures together and the practical version is short. When a machine looks your site up it is roughly right. When it answers from memory it is mostly wrong, and nobody lets you choose which of the two happens.
What to do about it
Ask an assistant who your business is, by name, and then ask it for your website address. Note whether the answer is right and whether it cites a page or simply asserts. Then count how many pages on your own site name your business, your location and what you do in full sentences, because that is the record the looking-up step reads.
Source arXiv, Guangwei Xu, Wenkai Zhang, Yawen Liu, Dongxu Zhang, Xijie Xu and Kuan Zou, “Octen-WebGR: Scaling Generative Retrieval to 200 Million Web Pages,” 26 October 2026 · arxiv.org/abs/2610.35256
04 Your listing
Google now tells owners which pages to keep unindexed
The pages where your clients sign in and share files were never meant to be found, and Google now says so.
Most owners have never looked at how many of their own pages Google is holding, and some were only ever meant for existing clients.
On 26 October Google updated its guidance on preventing user-generated spam and added a section headed “Only index content meant for Search”.
The new wording explicitly covers tools that let users publish shared pages, generated content, or web apps. For pages built for sharing, collaboration or utility rather than public discovery, Google now recommends keeping them out of its index altogether, using a no-index instruction or a crawling restriction.
Read that as a list of your own pages. A client sign-in portal, a shared document folder, a quote calculator, a generated PDF, a booking confirmation: an accounting practice running client document exchange, or an IT and cybersecurity company with a customer-facing tool, probably has dozens.
The cost here is not embarrassment. Pages with no public purpose sit in the index alongside the ones you actually want found, and clearing them out is a short job for whoever maintains the site rather than a project.
What to do about it
Ask whoever maintains your website for a list of every page Google has indexed, and read it once. Mark everything that exists for clients who are already yours rather than for people trying to find you. Google’s updated guidance is the reason to take those out of the index, and it is now written down where your developer can read it too.
Source Search Engine Roundtable, Barry Schwartz, “Google Warns Tools On User-Generated Spam Guidance & More,” 26 October 2026, reporting Google’s updated “Prevent user-generated spam on your site and platform” help document · seroundtable.com/google-user-generated-spam-guidance-42378.html
05 What to stop watching
Google says stop watching the page indexing report
If someone sends you a weekly screenshot of that report, Google’s own advocate says the number moves last.
If you pay somebody to watch your website, this changes which screen they should be watching.
John Mueller of Google was asked about the page indexing report running slow in Search Console, which is Google’s free report on how its search engine sees a website. His answer, posted on Reddit on 27 October: “it can feel like it’s important to focus on it, but honestly, save your energy”.
His reasoning was specific. “Indexing doesn’t change quickly, and when it does, you see that reflected in the performance reports first”. The report telling you whether Google holds your pages moves after the report telling you whether anybody found them.
That is worth money to a reader who pays for reporting. If a weekly update is built around a count of indexed pages, it is built on the slower number. The question that pays is whether searches are reaching you, and the answer sits in a different report inside the same free tool.
Mueller added the reason Google does not keep announcing how fresh each report is: “this is why we put dates in there”. So before anybody acts on a figure from one of these reports, read the date printed on the data.
What to do about it
Open Search Console and go to the performance report rather than the indexing one. Compare the searches that reached your site last month with the same month last year. If your reporting leads on indexed-page counts instead, ask for it to lead on this.
Source Search Engine Roundtable, Barry Schwartz, “Google Downplays The Importance Of Fresh Index Reporting,” 27 October 2026, quoting John Mueller of Google on Reddit · seroundtable.com/google-fresh-index-reporting-42381.html
