01 The evidence
Claude quoted 98% of its claims and proved 37%
A quote under an AI answer looks like proof, and most of the time it does not check out.
Twelve AI models were graded not on whether they cited something, but on whether the quote actually held the claim up.
Four researchers published a study on September 14 that tested something most citation research skips. They evaluated twelve AI models on 222 clinical questions, drawn from four medical practice guidelines including the 2026 American Diabetes Association Standards of Care. The models were asked to quote the exact words that backed each statement they made.
Attaching a quote turned out to be the easy part. Most models produced a genuine word-for-word quote for more than 90% of their claims. The quotes were not invented.
Then the researchers checked whether each quote actually substantiated the full claim it sat under. Claude Opus 5 quoted a real source for 98.0% of its claims and fully substantiated 37.1% of them.
The best performer, GPT-5.4, reached 75.6%. The weakest, GPT-4.1-mini, reached 13.2%.
The spread is the finding worth carrying. A reader looking at two answers, both with quotes attached, has no way to tell a 76% from a 13%. For a specialty veterinary hospital checking a treatment protocol, or a wealth advisor asking a model to pull the relevant passage out of a prospectus, the quote is doing the work of reassurance that the checking never did.
The authors state their own limits plainly: the questions are synthetic and every answer came from a single prompt with no variations tested. That is a small study describing itself accurately, which is more than most numbers in this field manage.
What to do about it
Next time an AI answer hands you a quote, open the source and read the sentence on either side of it. The question is not whether the quote is real, it is whether it covers the whole claim. That check takes a minute and it is the one nobody runs.
Source
arXiv, Jiashuo Zhang, Yuling Chen, Yvonne Commodore-Mensah and Michael Oberst, “Verifiable by Construction: Claim-Level Evaluation of Verbatim Citation in Clinical Question Answering,” September 14, 2026 · arxiv.org
02 The evidence
Researchers found invented citations in 2.3% of 2026 papers
Invented sources are now surviving peer review, so a citation in a published paper is no longer proof on its own.
The count covers citations confirmed as fabricated in published papers, not ones anyone merely suspected.
A second study, published September 15 by Paul Denny and seven colleagues, went looking for fabricated references inside work that had already been accepted and published. It examined 24,751 computing education papers and checked them against an American Computing Machinery corpus of more than 723,000 papers and 15 million references.
The trend is short but steep. Verified fabricated references rose from 3 in 2025 to 17 in 2026. At one conference, fabricated references appeared in 2.3% of the 2026 proceedings papers.
Seventeen is a small number and the authors do not dress it up. The reason it matters is where those references now live. A fabricated citation that clears peer review stops being a mistake and becomes part of the permanent record, available to be cited again by the next paper and read by the next AI model that crawls it.
That is the mechanism worth understanding if your business publishes anything with sources attached. A tax professional citing a revenue ruling, or an admissions consultant citing published outcomes data, is relying on a chain where every link used to be checked by a person. The checking did not stop, but the volume of plausible-looking references it has to catch went up.
What to do about it
If anything published under your name carries citations, click through to two of them at random before it goes out. A fabricated reference is usually obvious the moment you try to open it, and almost invisible if you never do.
Source
arXiv, Paul Denny and seven colleagues, “Testing Our Foundations: Citation Trends, Errors, and Emerging Hallucinations in the Computing Education Literature,” September 15, 2026 · arxiv.org
03 The platform change
OpenAI started testing ads that answer follow-up questions
A competitor can now buy a conversation with your customer inside the assistant they trusted for a recommendation.
The sponsored conversation is labeled and sits apart from the assistant's own answer, which is the part worth understanding.
OpenAI began testing a new advertising format called Sponsored Agents on September 16. A person who clicks an ad inside ChatGPT can now start a conversation with an agent the advertiser paid for, rather than being sent straight to a website.
OpenAI describes what happens next in plain terms: the user "can explain what matters to them, ask follow-up questions, and follow a link to the business's website" when they are ready. The company is explicit that "the conversation with a Sponsored Agent is distinct from ChatGPT's independent answers and separate from the original conversation that the user started."
That separation is the whole design, and it is worth taking seriously rather than dismissing. The assistant's own recommendation is still unpaid. What changed is that a business can now buy the follow-up questions, which is the part of a sale where objections get handled.
The test is limited to select advertisers in the United States. OpenAI shipped a set of less visible changes alongside it, including attribution windows of 7, 14 or 30 days and integrations with Shopify and HubSpot. For a software company selling a subscription, or a charter operator quoting a route, those are the pieces that make the format buyable rather than experimental.
Every advertising rule your business already follows applies here unchanged. A conversational ad is still an advertisement, and the claims inside it are subject to the same professional standards as a printed one.
What to do about it
Ask ChatGPT the question a customer would ask before hiring someone in your field, and note whether an ad appears beside the answer. Knowing whether your category is being sold against yet is worth more right now than deciding what to do about it.
Source
Search Engine Roundtable, Barry Schwartz, “OpenAI Testing Sponsored Agents For ChatGPT Ads,” September 16, 2026 · seroundtable.com
04 The operational change
Google Search Console lost a day of crawl data
If your search reporting shows a dip around September 15, it is Google's gap and not your website's problem.
The missing day is on every Search Console account, so a report that looks like a crawl problem is not one.
Google's Search Console is missing crawl statistics for September 15. The gap appears on every account, which is the detail that matters. It is not a signal about any one website.
Search Console is the free tool that reports how often Google's crawler visits your pages, and a drop in it is normally worth investigating. Google has published no statement and no restoration timeline. Search Engine Roundtable notes the same thing has happened repeatedly, listing November 2021, February 2022, May 2022 and October 2025, and that the data is normally restored afterward.
The cost is not the missing day. It is the diagnosis somebody makes from it. A medical practice whose agency reports a sudden crawl drop in mid-September is about to spend a meeting on a problem that belongs to Google, and possibly money fixing a site that is working.
Any report covering the week of September 15 is built on an incomplete week, and should be read that way until the data returns. A one-day hole is not a trend, however it renders on a chart.
What to do about it
If a search report lands this month showing a mid-September drop in crawling, check the date before acting. A dip that starts and ends on September 15 is this gap. A dip that continues past it is worth a real look.
Source
Search Engine Roundtable, Barry Schwartz, “Google Search Console Crawl Stats Missing A Day Of Data: September 15th,” September 20, 2026 · seroundtable.com
05 The money
Google started paying publishers whose pages feed AI answers
Content that feeds AI answers now has a price, and the businesses being paid for it are not yours.
Google’s payments are small and the formula behind them is undisclosed, which is the part publishers are asking about.
Google is running a limited test that pays publishers when their pages are used in AI Overviews, AI Mode and Gemini. Pages accrue earnings when they "contribute significantly to Google's AI-generated responses," and participants see a monthly earnings figure inside Search Console.
Google calls it "an early-stage learning pilot to test how best to reward high-quality content." One participant described it to Digiday more bluntly, as "an early test of a marketplace for inference data." That phrase is the useful one. It says out loud that being read by a machine is becoming a thing with a price.
The pilot is invitation-only and at least dozens of publishers have been approached, with the program reported as more appealing to small and mid-sized publishers than to large ones. The calculation behind the monthly figure is not disclosed.
That is the complaint participants keep returning to. One publisher executive put it to Digiday this way: "Do I wish they were more transparent? Definitely."
Read it for the precedent rather than the payment. A software company whose documentation answers questions all day, or an admissions consultant whose explainers get quoted back by an assistant, is supplying the same raw material. There is no route into this pilot for a business that sells expertise rather than publishing, and no equivalent number telling you how often your pages are used.
What to do about it
Nothing to sign up for here. Note which of your pages answer a real question completely, because those are the pages this pilot has just put a value on, whether or not anyone is paying you for them.
Source
Search Engine Roundtable, Barry Schwartz, “Google AI Contribution Pilot - Paying Publishers For Use Of Content In AI,” September 14, 2026 · seroundtable.com
