01 Research & evidence
An audit of 100 AI startups found three overstatements
The next honest AI claim your business makes will be read against every inflated one your buyer has already met.
The audit names the exact shapes an inflated AI claim takes, and one of them is the performance number with no method beside it.
A Pranav, Luke Hartmann and Anne Lauscher audited 100 startups in the German market against a codebook of seven dimensions, and interviewed 63 experts on how the claims land. The paper went up on October 8 carrying a note that it is accepted at AIES 2026, so it has been through peer review.
The startups overstated what their systems could do in three specific forms: claims with no evidence anyone could inspect, performance numbers with no methodology, and claims with no acknowledged limits. The capabilities claimed rarely matched the underlying system.
The consumer half is the part that costs you money. People who had met an exaggerated AI claim before withdrew trust from later AI marketing, so the price is paid by whoever makes the next honest claim. If you run a software company or an IT and cybersecurity practice, you are both the buyer and the seller of claims like these.
The audit also flagged 38% of the sample for moderate or serious ethical concern, in areas including surveillance and automated decisions in regulated fields. The authors want startup AI claims brought under disclosure requirements, and offer their codebook to regulators and consumer-protection bodies.
AI visibility is what Forever Cited sells, and a performance number with no method attached is this industry’s house style, ours included. Ask us, or anyone else, what was counted and how, before believing a figure.
What to do about it
Pull the AI claims your own marketing currently makes, and write the method under each one: what was counted, over what period, and by whom.
Any claim you cannot finish that sentence for is one to cut. It is the exact category this audit found buyers have already learned to discount.
Source arXiv, A Pranav, Luke Hartmann and Anne Lauscher, “Auditing AI-Washing in German Startups: A Mixed-Methods Study of Marketing Claims and Perceptions,” October 8, 2026 · arxiv.org/abs/2610.11788
02 The evidence
Six AI models hit 95% on rules, 37% on claims
Work that is a chain of small correct steps is exactly where an AI’s good score stops predicting a right answer.
An insurance benchmark separated getting each rule right from getting the whole set right, and the gap between the two is almost 59 points.
Linqi Zhang and six co-authors built InsClaimBench from real claim materials and structured policy rules: 3,780 cases in 375 case families across auto, property and health insurance, holding 86,656 individual rule judgments. Six models were run against it, and the paper published on October 7.
On the individual rules the models did well, with accuracy reaching 95.48%. On assembling the complete correct set of rules for a case, the best score was 36.90%.
The gap carries through to the decision that matters. Payout-decision accuracy ranges from 74.23% to 80.19%, and getting the decision and the amount right together drops to between 47.54% and 73.15%.
The authors then changed facts in a case to see whether the models updated properly. Correct local judgments still produced incorrect payouts, and correct payouts still concealed intermediate errors. A right answer was not evidence of right working.
Independent insurance brokers are the obvious reader here, and the shape travels to any work built as a chain. An accounting firm’s return, a tax practice’s filing position, an architecture and engineering submittal: each is many small correct steps that have to compose, and composition is where this measurement says the score falls apart.
What to do about it
Take one piece of work an AI tool touched this month and check the middle of it, not the answer.
A right answer is the condition under which a wrong step in the middle is hardest to find, which is this paper’s finding stated as a working habit.
Source arXiv, Linqi Zhang, Chong Qi, Yan Cheng, Wanqing Cao, Yu Liu, Chenwei Lin and Xian Xu, “InsClaimBench: Benchmarking Insurance Claim Adjudication Across the Decision Chain,” October 7, 2026 · arxiv.org/abs/2610.09671
03 The counterweight
A study of 712,930 submissions found AI use at 1%
The AI panic that would have justified throwing out your market research does not show up in the data.
The first count taken from people’s own chat histories rather than from guesswork, and it points the opposite way to the fear.
Neil Sehgal and four co-authors surveyed 2,500 workers on a high quality online research platform, then did what nobody had done: they linked donated ChatGPT histories to those workers’ own submission records. The result covers 712,930 submissions across more than 127,000 studies.
Self-report and observation disagreed sharply. One in eight surveyed workers said they had ever used AI on a study and 68% of the observed workers had, yet assistance appeared in only 1% of submissions, with no detectable increase over more than 3 years.
It is also concentrated rather than general. 5% of workers accounted for 64% of assisted submissions, so the behavior belongs to a small group and not to the population the data describes.
The uncomfortable finding is about enforcement rather than prevalence. Workers were virtually always paid, even when study instructions prohibited AI use, at approval rates similar to unassisted work, and half of assisted submissions were closed-answer questions, the multiple-choice kind, which sit outside what the platform’s detector examines.
If you commission survey research, which admissions consultants and wealth advisors do more than most, the honest reading is that the data is probably sound and the controls are not. That is a better question for your provider than whether it bans AI.
What to do about it
If a survey figure is load-bearing in your marketing, ask the provider two things: what share of responses its detector can examine, and what happens to one it flags.
For the platform in this paper the answers were about half, and paid anyway.
Source arXiv, Neil K. R. Sehgal, Manuel Tonneau, Dunigan Folk, Lyle Ungar and Sharath Chandra Guntuku, “AI-Assisted Submissions in Online Research Are Rare and Highly Concentrated but Routinely Approved,” October 7, 2026 · arxiv.org/abs/2610.09279
04 Operational change
Google penalties for scaled content are hitting established sites
A cheap content program on a long-standing site is now drawing a penalty applied by a person at Google.
The penalties name thin content and scaled content, and the sites collecting them are established businesses rather than fly-by-night operations.
Barry Schwartz reported on October 8 that Google may be issuing more manual action penalties than it had over the weeks before. A manual action is a person at Google deciding to demote or remove a site rather than an algorithm, and Search Console is the only place one is disclosed to you.
Two named consultants describe the same rise. Marie Haynes, who has consulted on these penalties for two decades, went from essentially no requests to several per week and named thin content, pure spam, and scaled content, that last one meaning many pages produced quickly to a template. Her note on who receives them is the line to read twice: most of these sites are legitimate long standing businesses.
Daniel Foley Carter put a cause to it. In his account businesses cheaped out on SEO or cut spend because AI could do it, the cheaper end of the industry turned to scaled content for short-term gains, and the penalties are the bill arriving. He says a manual action can cost a site 90% or more of its traffic overnight, which is his characterization of cases he has worked and not a measured figure.
Then the report does something worth noticing in an issue about unchecked numbers. Schwartz asked ChatGPT and Gemini to read the forums and chart whether complaints had spiked; both said yes, more complaints but not a large spike, and the two charts they drew disagreed with each other.
Google has not confirmed any increase, and one consultant’s caseload is not a measurement of the web. The category is what is solid: a luxury home builder with forty near-identical location pages, or an architecture and engineering firm running a service page per city, is holding the exact pattern both consultants name.
What to do about it
Open Google Search Console and look under Security and Manual actions. That page is the only place a penalty is disclosed to you, and reading it takes a minute.
Then count your near-duplicate pages. Where one page would serve and eight now sit, that is this week’s cleanup.
Source Search Engine Roundtable, Barry Schwartz, “Google Search Issuing More Google Manual Action Penalties?,” October 8, 2026 · seroundtable.com/google-search-manual-action-spike-42258.html
05 Platform change
Google opened a faster route into Search for forum content
The pages where your customers discuss you can now reach Google faster than your own website can.
Approved platforms get to push content straight to Google, while an ordinary business site still waits to be crawled.
Google announced its UGC Fresh Data Program on October 8. In Google’s words it is an initiative designed to help platforms display fresh, high quality, and authentic user-generated content (UGC) across various features on Search.
Mechanically, approved platforms can use a dedicated data ingestion pipeline to send content and interaction signals to Google instead of waiting to be crawled. Google says the program is independent of the Google Indexing API and traditional organic crawling for general web content.
You cannot apply unless you are that kind of platform. Eligibility requires primarily hosting user-generated content on dedicated pages with stable URLs, at high volume with a significant user base, public pages rather than anything behind a login or paywall, active moderation, and the ability to submit regular updates of engagement counters within 72 hours of creation.
The consequence is about relative speed, not a door opening for you. A forum thread about your practice, an industry community, a review-and-discussion site: those can be current in Search quickly, while the page you control waits its turn to be crawled.
Google is explicit that the program doesn’t guarantee that content will appear in Search, and says applicants hear back within 6-8 weeks. For admissions consultants, whose market runs on discussion forums, and for private aviation and charter operators talked about in enthusiast communities, the useful move is knowing what those threads currently say about you.
What to do about it
Search the two or three communities your buyers actually read for your own name, and read what is there.
Those pages are now on a faster path into Search than the page you control, so they are worth knowing before a customer quotes one back to you.
Source Search Engine Roundtable, Barry Schwartz, “Google Reveals The UGC Fresh Data Program,” October 8, 2026 · seroundtable.com/google-ugc-fresh-data-program-42265.html
