01 The market data
Microsoft researchers found half of no-click searches succeeded
A drop in website visits from search is not by itself proof that fewer people wanted what you sell.
If a search with no click is usually a success, the number this whole industry quotes most often has been measuring the wrong thing.
Click status alone cannot tell you whether a search worked. A click can lead somewhere useless, and no click can mean the answer was already on the page. The team built a scoring method around that problem, training judges to label whole sessions and then scoring those labels from how people actually behaved.
On 6,904 scored searches, the split was stark. 53.2% of the searches where nobody clicked had succeeded anyway, and 6.5% of the searches that did get a click had failed. In the authors’ words, “interaction status is not an adequate outcome”: whether somebody clicked does not tell you whether the search worked.
Forever Cited sells AI visibility, so a finding that no-click searches usually succeed is a finding against the most profitable alarm in this industry, and it is reported here for that reason. The paper does not name the search engine it studied, which is a real limit on how far it travels.
For a charter operator, the practical version is this. Somebody asks what a light jet from Van Nuys to Aspen runs, reads the answer, and books a week later by phone. The search that started it never appeared as a visit to anybody’s website.
What to do about it
Before you read a fall in search visits as lost business, check whether your phone, your inbox and your booking form moved the same way.
If those held steady while visits fell, the visits were never the thing you were selling. If they fell together, you have a real problem and now you have evidence for it.
Source arXiv 2611.01582, Minha Hwang, Shavak Agrawal, David Chen, Carto Wong, Xin Du, Tiannan Wang and Georg Buscher of Microsoft with Zheng Chen of Simplex AI, “Measuring User Success in Randomized Search Experiments,” submitted 1 November 2026 · arxiv.org/abs/2611.01582 (6,904 scored queries; a preprint, not peer reviewed; the abstract does not name the search engine studied, and the affiliations are taken from the arXiv listing rather than from any company statement, so nothing here is a Microsoft announcement)
02 The evidence
Twelve AI models invented more sources when asked to cite
If you ask an AI to include citations, you are asking it for the part it is most likely to invent.
The models tested were open research versions rather than ChatGPT, Claude or Gemini, and the author says the cause is not proven.
A single researcher ran four experiments on 900 fixed questions drawn from US federal case law, scientific literature and medical literature. The method was filed publicly before the experiments ran, so it could not be adjusted afterwards to suit the result. The models were twelve freely downloadable versions from three model families, each tested at four stages of the extra training that turns a raw model into a helpful assistant.
Two findings matter. That extra training raised invented citations across the board, from 4.8% to 56.7%. And the request itself was the trigger: asking for citations raised fabrication sharply in the trained models and not in the raw ones, a gap of 47.8 percentage points across the twelve.
A control ruled out the obvious alternative. Matching answer length separated the citation request from a simple demand for more words, so the effect belongs to citations rather than verbosity. A scorer built only on surface citation features preferred the tuned output 86% of the time, and rewarded sheer quantity of sources.
For an attorney this is not abstract, because federal case law was one of the three test domains. The author’s own verdict on the cause is “possibly, but not proven”, and these were research builds rather than the assistants a practice actually buys.
What to do about it
If a draft comes back with citations attached, check every one before it leaves your office.
The study’s finding is that the citation is the least trustworthy part of the answer rather than the most. Treat a list of sources as a list of things to verify.
Source arXiv 2611.01313, Zain Naboulsi, “Are Fake Citations a Form of Reward Hacking? A Pre-Registered Study of Post-Training and Citation Cues,” submitted 31 October 2026 · arxiv.org/abs/2611.01313 (900 fixed questions across three domains, twelve open-weights checkpoints refit on fifteen; pre-registered, with code and data published; a preprint, not peer reviewed; the models are open research checkpoints and not commercial assistants, and the author states the reward-hacking explanation is not proven)
03 The platform change
Google moved the Message button next to Call on Maps
A customer who cannot reach you by phone can now message you from your Maps listing, and it lands wherever you pointed it.
Google says the chat option used to be buried deep in the profile, and the setting that decides where a message lands is yours.
Google publishes a monthly Small Business Bulletin. In the latest one, Lisa Landsman of Google wrote that messaging has moved to the front of the Maps listing on both Android and iOS: “customers on both Android and iOS can now find a prominent Message button right next to ‘Call’ on your Google Maps listing”.
Her stated reason is that the option used to be buried. “Previously, finding chat on Maps required scrolling deep down into your profile,” she wrote, adding that it now appears whether a customer opens the profile, the expanded header or the bottom action bar.
Where the message goes is a setting you own. Google says the chats route to where the business already works, including WhatsApp and SMS, and its own advice is to confirm the preferred channel is linked in the Google Business Profile.
For a real estate agent the value is the case Google names: the lead that arrives while you cannot pick up. Somebody looking at a listing on a Saturday asks whether a showing is still open, and the question survives instead of going to the next agent.
What to do about it
Open your Google Business Profile today and confirm which messaging channel is linked.
A message routed to an app nobody on your team watches is a lead you will never see arrive, and the button is now sitting next to your phone number.
Source Search Engine Roundtable, Barry Schwartz, “Google Expands Business Messaging To Google Maps On Android and iOS,” 6 November 2026 · seroundtable.com/google-maps-messaging-android-ios-42472.html (quoting Lisa Landsman of Google in the Google Business Help Forums, via the monthly Google Small Business Bulletin; the forum post itself is unreachable from this session, so the quotes are as the publisher reproduces them, and Google names no regions, no business categories and no date)
04 The document risk
Four in ten AI-written research memos carried a false claim
A claim-level accuracy near 98% still put a false statement in roughly four of every ten finished documents.
The memos were investment research checked against official filings, and the gap between a good claim rate and a bad document rate is the finding.
Four researchers had four AI agents write 1,440 investment research memos, covering 30 companies over 12 quarters. They then broke the memos into 77,272 individual financial claims and checked each one against the same bundle of official filings.
The claim-level result looks excellent. Only 1.17% to 1.77% of the individual claims were contradicted by the filings. The document-level result does not. Between 36.94% and 41.11% of the memos contained at least one contradicted claim, which is roughly four in ten.
Both numbers describe the same output. A long document has many chances to be wrong, so a small per-sentence error rate still lands a false statement in most batches of ten. The authors put it plainly: a low claim-level error rate does not necessarily mean a memo is reliable. Their error groups were time, unit, context and concept.
For an advisor assembling deal memos or diligence summaries, the relevant number is the document rate, and it is the one nobody publishes. The study covers investment research for 30 companies, so it is not evidence about tax, legal or clinical writing.
What to do about it
If software drafts anything a client will rely on, ask your vendor the document question rather than the sentence question.
Not how accurate each claim is, but how often a finished document contains at least one wrong one. Nobody volunteers that number, and it is the one that reaches your client.
Source arXiv 2611.04375, Xukang Gong, Chenri Luo, Binyan Yu and Xiao-Yang Liu, “LLM-Generated Investment Research Memos Need Validation,” submitted 3 November 2026 · arxiv.org/abs/2611.04375 (1,440 memos from four agents, 30 companies, 12 quarters, decomposed into 77,272 claims; a preprint, not peer reviewed; the abstract names only one of the four agents and offers no human-written comparison)
05 The measurement
OpenAI is building a ChatGPT dashboard for businesses
The first official count of how often ChatGPT names your business may arrive without telling you why the number moved.
Nothing here is announced, the screenshot came from one analyst, and the early version appears to need a storefront many practices do not have.
OpenAI appears to be building a dashboard that tells a business how often ChatGPT mentions it. The analyst Juozas Kaziukenas found a banner offering to “Claim your business in ChatGPT” and posted a screenshot. By his account it would count how often a business appears in ChatGPT answers, and how many people click through to its website.
He describes those two counts as tracked daily and drawn as a chart, inside a larger Business Hub. He says he first saw early signs in August, and that it is limited to businesses that have verified ownership, giving a connected Shopify store as the example.
The limits are the interesting part. He reports that it shows no prompt data, no user data and no competitive ranking, so it would tell a business that mentions rose or fell without telling it why. It also looks separate from ChatGPT advertising.
None of this is announced. One person saw one banner, OpenAI has said nothing, and the ownership check described would exclude an admissions consultant with nothing to connect. A count of ChatGPT mentions published by the platform itself is close to what firms in this field, Forever Cited included, are paid to measure, which is reason to read the claim carefully rather than to repeat it.
What to do about it
Nothing to do yet, and that is the point worth holding.
If you are told that a ChatGPT visibility number is proprietary, note that the platform itself may be about to publish a basic version of it for free, and that even then it will not say why the number moved.
Source Search Engine Roundtable, Barry Schwartz, “OpenAI Creating ChatGPT Tools For Business For Analytics & Citations,” 6 November 2026 · seroundtable.com/chatgpt-tools-for-business-42476.html (relaying a LinkedIn post and screenshot by Juozas Kaziukenas; OpenAI has announced nothing and is quoted nowhere, the LinkedIn post is unreachable from this session, and there is no launch date, no stated availability for service businesses and no independent confirmation)
