01 The research
Simulated recommendations stopped all visits to a fifth of restaurants
A system that learns from the customers it already sent you can quietly stop sending any, and keep doing it.
The authors built a simulation rather than measuring real cities, and the finding that should worry you is what happened when the system retrained on its own results.
Nine researchers led by Guangyuan Weng published a study of location recommendation on 30 October. They built a long-running simulation of four United States metropolitan areas from detailed movement data, with travel distance and venue capacity as limits, and pointed it at food venues. Under every algorithm they tried, and in every metropolitan area, at least a fifth of venues stopped receiving visits.
The mechanism is the part worth your time. More than half of that loss appeared only after the system retrained on the visits it had itself produced. A business that drops out of the suggestions generates no visits, so the next version of the system has even less reason to suggest it.
The second finding inverts the usual assumption. The more accurate an algorithm was by the standard test, the more it narrowed who people encountered, and the most accurate one left people 15 to 21% more segregated by income than they actually are. Accuracy here means predicting past visits well, which turns out not to be the same thing as serving a city well.
This is the mechanism Forever Cited was built around. When software is asked who to use, it answers with a few names, and a business that is not among them is not seen at all.
The authors tested one fix, and the limits are their own. Re-ranking that favors less popular venues, running on top of the existing algorithm, reduced both the disappearance and the segregation. Every figure here is simulated rather than observed, the venues are food venues, and the paper states no count of venues or of algorithms behind the proportions.
What to do about it
Ask one narrow question of whoever handles your marketing: when a customer asks an assistant for a business like yours, do you appear, and did you appear last month? A specialty or emergency veterinary hospital can check it in five minutes by asking the same question a week apart from a signed-out browser.
Write down what came back. This paper’s point is that the direction matters more than any single answer, because the direction is what the system learns from next.
Source arXiv, Guangyuan Weng, Lin Chen, Ciro Beneduce, Giovanni Mauro, Massimiliano Luca, Takahiro Yabe, Bruno Lepri, Luca Pappalardo and Esteban Moro, “Recommendation Feedback Loops Shape Urban Income Segregation and Venue Visitation,” 30 October 2026 · arxiv.org/abs/2610.41911 (a preprint; every figure is the output of the authors’ simulation rather than an observed measurement, the venues are food venues in four United States metropolitan areas, and the paper states no count of venues or of algorithms behind the “at least a fifth” and “more than half” proportions)
02 The market
Reddit's chief executive says Google stopped being a traffic driver
If one of the web's most quoted sites cannot hold its Google visits, a smaller business should not expect them back.
A competing explanation sits in the same article, and it matters: some of Reddit's lost traffic may be its own machine-translated pages being demoted, not AI answers.
Search Engine Roundtable’s Barry Schwartz reported on 2 November that Reddit’s chief executive, Steve Huffman, has confirmed Google search traffic is declining for the site. Quoting a CNBC interview, the article says Reddit’s shrinking search-referral traffic from Google went “down meaningfully in this quarter” because of the “shift to AI mode and AI overviews.”
Huffman’s framing is the line to keep. Earlier Google changes would “kind of come and go,” he said, while the company’s AI work is “a fundamental shift in how the Google product works, from being a traffic driver to whatever’s going on now.” The article also reports that the decline is driving Reddit’s plan to require people to become logged-in users.
Two cautions, both from inside the same article. Its headline says the stock dropped over 33% and its body says 36% over the past week, and the piece does not reconcile the two. And Glenn Gabe is quoted pointing at a different cause: Reddit’s machine-translated pages, tens of millions of web addresses, took a hit from Google’s May 2026 core update and June spam update.
So the honest reading is narrower than the headline. A named chief executive has said on the record that Google referrals are falling and blamed AI answers, and a named practitioner has offered a competing explanation for part of it. Both can be true at once, and neither is measured here.
What to do about it
If you have ever been recommended in a forum thread, that thread is a source an AI assistant may quote today and stop quoting tomorrow. A wealth advisor or registered investment adviser whose name appears in a few community threads should save copies of them now, because the page carrying your name is not yours and can be put behind a login.
Nothing here falls due. The point is that a mention you do not own is a mention you can lose without being told.
Source Search Engine Roundtable, Barry Schwartz, “Reddit Stock Drops Over 33% After Declining Google Search Traffic,” 2 November 2026 · seroundtable.com/reddit-stock-drops-declining-google-search-traffic-42432.html (secondary reporting; the Huffman quotations are Search Engine Roundtable quoting a CNBC interview, which could not be reached from this run, and the article’s own headline and body disagree on the size of the share-price fall, over 33% against 36%)
03 The evidence
Software picked pages to read by link text, not keywords
The words in your own navigation decide which of your pages software collects when it cannot afford to read them all.
This is a benchmark paper rather than a study of live AI assistants, and its labels were voted on by language models, which is the reason to read it carefully rather than to dismiss it.
Manuel Rey-Escobar, David Otero and Javier Parapar published a test collection for focused crawling on 29 October. A focused crawler is software that wants pages on one topic and can only afford to download a limited number, so its whole job is choosing which links to follow. The collection is built from static copies of 621 company and organization websites, about 1.1 million pages.
The result is specific and useful. Keyword matching and text classifiers did not do significantly better than simply crawling everything in order. What did better was comparing the meaning of link text against links whose target pages were already known.
Read that as a statement about your own site. The crawler is judging the words you put in a link, not the page behind it, and it is judging them by meaning rather than by whether they contain the right term. A link reading “Learn more” tells it nothing, while a link reading “Fixed-fee business tax returns” tells it a great deal.
The caveats are the authors’ own. Every page label came from a majority vote among three large language models rather than from a person, and the task measured is finding product pages, not answering a customer’s question. This is evidence about how crawling software chooses, not proof about any particular assistant.
What to do about it
Open your own site and read only the link text, ignoring the pages. An accounting firm whose menu says Services, About and Contact has told crawling software almost nothing. The same firm with Business tax returns, Audit support and Bookkeeping has told it what the practice actually does.
This costs nothing and needs nobody. Rewrite the three or four links that describe what you sell, in the words a customer would use.
Source arXiv, Manuel Rey-Escobar, David Otero and Javier Parapar, “ProdCrawl: A Replayable Test Collection for Focused Crawling,” 29 October 2026 · arxiv.org/abs/2610.40855 (a preprint describing a benchmark; the page labels come from a majority vote among three large language models rather than from people, and the task measured is finding product pages on company websites, not answering a customer’s question in an AI assistant)
04 The operations
Google asked for a tax ID before switching on messages
Turning on the message button can now put your profile into a review that takes three to four working days.
This is one named user's screenshot rather than a Google announcement, which is exactly how much weight to give it and exactly why it is worth two minutes of preparation.
Search Engine Roundtable reported on 28 October that Google is asking some businesses for an Employer Identification Number and a link to their privacy policy when they update parts of a Google Business Profile. The evidence is one reader, Syed Amir, who said he hit it while setting up the message button and posted a screenshot.
The screenshot is the useful part. It reads that a Google partner will review your details before you can receive messages, and that the review takes about three to four working days. So the cost of not having a privacy policy ready is not a refusal, it is a delay.
Treat the scope with care. Google has announced nothing, the report covers some businesses rather than all, and a single screenshot is a single screenshot. The article’s own lead ties the prompt to updating a profile while the quoted reader ties it to the message button, and the piece reconciles neither.
The preparation is cheap enough to do on a report this thin. An Employer Identification Number is already on file for any business that has one, and a privacy policy page is a requirement of running a contact form in most of the United States anyway.
What to do about it
If you plan to turn on messaging, have two things to hand before you start: your Employer Identification Number and a live address for your privacy policy. A luxury home builder who begins the setup without the second one will sit in a three to four day review for no reason.
If you already take inquiries through a form on your own site, you have a privacy policy somewhere. Find its address now, while it costs you nothing.
Source Search Engine Roundtable, Barry Schwartz, “Google Business Profiles Asking Some Businesses For Employer Identification Number (EIN),” 28 October 2026 · seroundtable.com/google-business-profiles-ein-request-42382.html (a single reader’s report with a screenshot, quoted in the article as Syed Amir’s own words; Google has made no announcement, and the article’s lead attributes the prompt to updating parts of a profile while the quoted reader attributes it to setting up the message button)
05 The money
Google confirmed ads sit among unpaid results in AI Mode
Inside an AI answer, the only thing separating a paid listing from an earned one is the word Sponsored.
A Google spokesperson said this on the record, which makes it the firmest item in today's issue and the one that says most about where AI answers are heading.
Search Engine Roundtable reported on 29 October that Google is mixing paid product listings in among the unpaid ones inside AI Mode, in the sideways-scrolling row of products the trade calls a carousel. The headers above it read “Sponsored” and “All products.” A Google spokesperson confirmed it to the publication: “Yes, this format is part of our ongoing experiments with ads in AI Mode.”
The spokesperson’s second sentence defines the surface. “Shopping Ads can appear alongside organic visual search results within carousels for commercially relevant queries,” Google said, using its own word for the unpaid results, and “All sponsored listings are labeled as Sponsored.” The test was spotted by Sachin Patel, who posted screenshots.
This is not new behavior, which is rather the point. Google confirmed mixing ads with organic results back in October 2023 and calls it dynamic ad placement. What is new is that the mixing now happens inside the answer rather than above a list of links.
Be precise about who this touches today. These are product carousels on commercially relevant queries, so a reader selling a service is not in this surface yet. Enterprise and specialized software is the readership it reaches first, because software is a product people compare in exactly that way.
The reason to watch it anyway is the direction. An answer that carries paid placement has a reason to be shorter, and a shorter answer names fewer businesses.
What to do about it
Nothing to do this week unless you sell a product. If you do, open AI Mode on your own category and count how many of the names in the carousel carry the Sponsored label.
If you sell a service, the useful move is simply to know this exists before it arrives in your category, so that the first time a paid name sits above yours it is not a surprise.
And if you do buy a placement there, your own advertising rules travel with you. A paid listing inside an AI answer is an advertisement, so the same disclosures apply as in any other channel: state bar rules for attorneys, the Securities and Exchange Commission marketing rule for wealth advisors and registered investment advisers, and the Federal Trade Commission guidance on endorsements and AI claims for software and cybersecurity firms.
Source Search Engine Roundtable, Barry Schwartz, “Google AI Mode Tests Blending All Products With Sponsored Products,” 29 October 2026 · seroundtable.com/google-ai-mode-all-sponsored-products-42402.html (the quoted sentences are a Google spokesperson’s words given to the publication, read from the article’s raw page this morning; the test was spotted by Sachin Patel, the article carries no figure for how many products, queries or users are affected, and an embedded post dated 1 November comes from the account SERP Alert rather than from Google)
