Researchers invented a concept. Eight of ten AI platforms cited it within seven days.

01 The evidence

Eight of ten AI platforms cited a made-up concept

Anyone can publish their way into an AI’s answer about your industry, and it cost these researchers fourteen dollars.

Across 17,211 citations on ten AI search platforms, the sources an assistant leans on turned out to be few, concentrated, and open to whoever posts.

Start with the measurement, because it is the largest of its kind so far. Qi Liu, Geng Hong, Xinyang Zhang, Pei Chen, Yutong Li and Min Yang analyzed 17,211 citation instances across ten AI search platforms, covering 6,356 different source domains, and published the work on October 8.

The first finding is concentration. The top twenty domains account for between 20.5% and 70.8% of everything a given platform cites, depending which platform you look at. A handful of websites decide what the assistant says about your field.

So the authors tested how hard it is to get into that handful. Of 22 publication platforms tied to those cited domains, 15 had low or medium barriers to posting: no credential check and no editor. They published their own material there and watched what the assistants did with it.

Eight of the ten platforms cited a concept the researchers had fabricated, within seven days. One cited their marked content within one hour of it going up. A single article on a platform the engines favor outweighed more than twenty matched posts on ones they do not, so where you publish beats how much.

Then there is the price. The team bought a commercial GEO service for $14, and it produced 13 public posts. That is the whole cost of putting a claim where an answer engine may pick it up and repeat it as established.

AI visibility is what Forever Cited sells, and this paper put a price on the cheapest possible version of it. A citation is not by itself evidence of standing, and anyone offering you one should be asked which pages they intend to publish, and where.

Source arXiv, Qi Liu, Geng Hong, Xinyang Zhang, Pei Chen, Yutong Li and Min Yang, “From Public Posts to AI-Search Citations: Measuring the Fragility of AI Search,” October 8, 2026 · arxiv.org/abs/2610.11932

02 Research & evidence

The best models reached 0.97 and still approved false citations

A near-perfect checking score still lets through the fabricated citation, which is the only one that would have hurt you.

Two researchers built 3,396 labeled claims out of New York Court of Appeals decisions and asked whether a model could tell a supported citation from an invented one.

The test is narrower and more useful than most. M. Mikail Demir and M. Abdullah Canbaz, publishing on October 7, took recent New York State Court of Appeals decisions and broke them into 3,396 individual claims, each labeled Supported, Refuted or Not Found against the authority cited beside it.

Then they asked several models to make the same call, with no examples and no training on the task. The strongest reached up to 0.97 accuracy, which sounds like a solved problem.

It is not. The authors state plainly that models still incorrectly mark unsupported claims as supported, even when the full opinion text is provided. And the failure is not spread evenly: fabricated but plausible citations caused the largest drop in performance of anything they tested.

That is the wrong error in the wrong place. A checker that misses only the convincing fakes is a checker that waves through precisely what you needed it to catch, and the reassurance arrives while the verification does not.

An attorney is the obvious reader here, but the shape travels. A wealth advisor whose compliance read leans on an AI pass, an independent insurance broker checking a policy citation, a veterinary specialist confirming a dosing source: in all three, the mistake that survives the check is the one that looked right.

Source arXiv, M. Mikail Demir and M. Abdullah Canbaz, “When Citations Mislead? A Claim-Level Benchmark for Legal Hallucination Detection,” October 7, 2026 · arxiv.org/abs/2610.10971

03 How it is actually used

Clinicians sent 127,833 queries and barely asked for a diagnosis

The professionals who already use AI every day use it for paperwork, not judgment, which is the opposite of what it is sold on.

An eight-month rollout inside one institution recorded what 6,342 clinicians actually asked an AI assistant, and the answer was documentation.

Krithik Vishwanath and a large group of co-authors analyzed 127,833 queries sent by 6,342 physicians, advanced practice providers and nurses across 35 specialties to an institutional assistant over an eight-month rollout. The paper went up on October 8.

The breakdown is the finding. Documentation and administration accounted for 36.2% of use and knowledge retrieval for 28.9%, nearly two thirds between them. Diagnosis was 3.7%.

Then the authors turned the same measure on the tests everybody quotes. Across 58 public benchmarks drawn from major evaluation suites and frontier model reports, the median benchmark contained no documentation requests at all and shared 31% of the task mix of real use.

So the published scores describe a job nobody is doing. One more number is worth keeping: more than a third of the queries could not be answered well as posed, which is a fact about how people ask rather than about the model.

Read it across to your own work. The people in your field who have already folded AI into their day are using it to write things down and look things up. A franchise advisor drafting a disclosure summary, an M&A advisor assembling comparable terms, a practice manager turning notes into a letter: that is where this lands, and it is not the replacement story.

Source arXiv, Krithik Vishwanath and co-authors, “Clinician use of language models diverges from how the models are evaluated,” October 8, 2026 · arxiv.org/abs/2610.11069

04 Search

Google’s September spam update took two weeks, not two days

If your inquiries fell on one of five specific dates, you now know what did it and that it will not come back quickly.

The rollout ran from September 24 to October 8, and Google says recovery from this kind of update can take many months.

Barry Schwartz reported on October 8 that the update started on Thursday, September 24 at around 12:00 pm ET and completed on October 8 at 4:37 am ET. The comparison he draws is the useful part: this one took two weeks to roll out whereas the others took only two days.

Site owners reported impact on September 25 through 27, September 30, and October 4 through 6. If your calls, inquiries or form fills dropped on one of those and you have been hunting for a reason, start there.

What it targets is narrow. It penalizes techniques that are against Google’s spam policies, so it is not a verdict on good work that happens to rank badly. Google’s instruction to anyone affected is to read those policies against what is actually on the site.

The clock is the hard part. Google says it can take many months to recover. A two-week rollout also means a drop may have arrived in stages rather than overnight, which is how a real penalty gets written off as a slow month.

Source Search Engine Roundtable, Barry Schwartz, “Google September 2026 Spam Update Has Finished Rolling Out,” October 8, 2026 · seroundtable.com/google-september-2026-spam-update-done-42235.html

05 What changes next

Anthropic’s revised usage policy takes effect November 12

Paying somebody to manufacture mentions of your business now breaks the stated rules of a company whose model answers questions about you.

The update consolidates the rules on deceptive campaigns, naming fake accounts and fabricated news sites used to amplify content.

Anthropic published a usage policy update on October 8, and the updated policy takes effect on November 12. Most of it governs the people building on the models rather than the people typing into them.

One section matters here. “Do Not Engage in Deceptive Campaigns or Artificial Activity” covers deception of any kind, commercial as much as political, and names two practices in particular: hiding who is behind a message, and amplifying content through fake accounts. The company cites networks of fake accounts and fabricated news sites as what it has in view.

Anthropic is straightforward that this was already prohibited and the section consolidates existing rules rather than inventing new ones. The requirements for high-risk uses are also unchanged: a qualified person must be able to review and change the model’s recommendations, and people affected have to be told AI was used. If you put a model anywhere near advice a client acts on, that pair is the one to read in full.

Now set it beside the first story. Researchers showed that posting on low-barrier sites moves what AI search cites, a commercial service does it for pocket money, and a model provider has written a rule naming the pattern. The two documents describe the same practice from opposite ends, which is the clearest signal yet that bought citations are a liability rather than an asset.

Source Anthropic, “2026 Usage Policy update,” October 8, 2026 · anthropic.com/news/2026-usage-policy-update