01 The research
Researchers spent $14 and an AI engine cited them
Getting named in an AI answer can cost fourteen dollars, and that is the number any AI-visibility invoice is measured against.
Eight of the ten platforms tested went on to cite a concept the researchers had invented, within seven days of publishing it.
Qi Liu, Geng Hong, Xinyang Zhang, Pei Chen, Yutong Li and Min Yang published a measurement study on 8 October on arXiv, the preprint site where most AI research appears before any journal sees it. They mapped 17,211 citations across 6,356 different websites on ten AI search platforms, then tested whether an outsider could get into those citations on purpose.
The concentration came first. The top twenty websites supplied between 20.5% and 70.8% of each platform’s citations, and of twenty-two publishing sites tied to those cited domains, fifteen had low or medium barriers to both opening an account and posting. The places these engines quote most are, in many cases, places anyone can publish.
Then the team published. Eight of the ten platforms cited a concept the researchers had fabricated, within seven days. One article on a site the engines favoured carried more citation weight than over twenty matched posts on sites they did not.
The last finding is the one to read twice. The team bought a commercial search-optimization service for fourteen dollars, it produced thirteen public posts, and one platform quoted the marked content within an hour.
This is the market Forever Cited sells into, and the study describes its cheap end accurately. Any business paying anyone for AI visibility is entitled to ask which of those two things it is buying.
What to do about it
Ask whoever handles your AI visibility to name the sites your mentions are published on. Then check two things: whether you could have opened an account there yourself in ten minutes, and whether a client would recognise the name. This matters most where credibility is the product and a regulator reads the claim, which covers an accounting practice, a tax professional working under Circular 230, and an admissions consultant answering to the Federal Trade Commission. A mention on a site anyone can post to is worth very little, and it is awkward to explain when a client traces it.
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,” 8 October 2026 · arxiv.org/abs/2610.11932 (a preprint, not peer reviewed; 17,211 citation instances over 6,356 source domains across 10 AI-search platforms, plus marker-controlled publication experiments on 22 tested publishing platforms. Eleven days old at the time of writing, and carried here as context rather than as this week’s news)
02 The research
Seven AI models moved one client’s stock advice 35 points
A client who asks an AI the same question twice can get two different answers, then arrive at your office certain of one.
Asked the identical question twice, the models barely moved; change one client answer and the recommended stock share moved by up to 35 points.
Soovadeep Bakshi, Apurba Bose, Aishwarya Bose and Sandip Nandi audited seven large language models from five providers against 800 real household profiles, published on arXiv on 16 October. Between profiles only one thing changed: the client’s answer to the standard question about how much risk they are willing to take.
That single answer did most of the work. The recommended share in stocks moved 10 to 16 percentage points on average, and 24 to 35 points when the answer changed, against 0.6 to 4.2 points when the identical prompt was simply asked twice. So this is not the model being erratic. It is the model taking one self-reported number extremely seriously.
The trouble is that people do not give that answer consistently. In a national US panel, 41% of households chose a different answer two years later, and in a laboratory retest a third of people moved two or more points on a 0 to 10 scale after six months. The swing is largest for clients under 40.
One remedy in the paper is almost comically cheap. A single clarifying question cut the swing from 33.5 points to 1.9 in a scripted test where the client then corrected their own answer.
What to do about it
This reads as a wealth-advisory finding and it is really a finding about every first appointment. Your next client may walk in holding a number an assistant produced from one answer they would give differently tomorrow. A wealth advisor or a registered investment adviser has the Securities and Exchange Commission Marketing Rule as the frame here, but the practical move is the same for a CPA asked about a filing position or a real estate broker asked about a price: ask what the client was told, then ask what question they had to answer to get it.
Source arXiv, Soovadeep Bakshi, Apurba Bose, Aishwarya Bose and Sandip Nandi, “Same Person, Different Advice: Auditing LLM Financial Advice with Repeated Risk-Tolerance Responses,” 16 October 2026 · arxiv.org/abs/2610.23060 (a preprint, not peer reviewed; 7 models from 5 providers giving one-turn allocation advice on 800 real household profiles, with answer reliability taken from a national US panel and a six-month laboratory retest)
03 The research
AI answers scored well on accuracy by copying their sources
An assistant can look perfectly sourced while telling your client something the page it linked to never actually said.
The score used to train long AI answers to stay faithful counts a copied sentence as supported, so models learn to copy.
Hoang Pham, Duc Anh Vu, Duy Tung Doan, Cong-Duy T Nguyen and Anh Tuan Luu published this on arXiv on 13 October. Their target is the measure the industry leans on to keep long AI answers honest: the share of sentences in an answer that the cited sources actually support.
A sentence copied straight out of the source is supported by construction, so a model trained to raise that score learns to copy rather than to reason. The authors are blunt that the usual defences only move the problem, since a copy penalty either leaves the copying in place or trades it for shorter, emptier answers and outright refusals.
Their alternative credits only sentences the model wrote itself, that say more than the question did, and that the sources support. Across four models and five bodies of documents, it cut the error rate of self-written sentences by 38% and raised the count of supported ones by 46% at the median, judged by evaluators held back from the training.
What that leaves a business owner with is a warning about what a citation means. The link under an AI answer is evidence that a page was fetched, not that the page says what the answer says.
What to do about it
Read the AI answers that name your business against the page they link to, and not just for whether you appear. The gap between the two is your exposure, and it is widest wherever a number or a condition is involved: a luxury home builder whose quoted all-in cost has moved since the page was written, an independent insurance broker whose coverage language has been paraphrased into something tidier, a specialty veterinary hospital whose after-hours cover is summarised wrongly. Correcting your own page is the only lever you hold.
Source arXiv, Hoang Pham, Duc Anh Vu, Duy Tung Doan, Cong-Duy T Nguyen and Anh Tuan Luu, “Copying Is Not Grounding: Training Faithful Long-Form RAG with Authored-Sentence Credit,” 13 October 2026 · arxiv.org/abs/2610.18442 (a preprint, not peer reviewed; four models across two in-domain and three held-out bodies of documents. The 38% and 46% figures are median improvements produced by the authors’ own method, not a measure of how often AI answers mislead)
04 Your ads
Google is removing featured professionals from Local Services Ads
If your Local Services Ads sell on one named professional, that person stops appearing on them this month.
The notice says the change starts in October without naming a date, and the help document it links to did not describe it when a reporter checked.
Google Ads put a message in Local Services Ads accounts, spotted by advertiser Anthony Higman and reported by Barry Schwartz at Search Engine Roundtable on 19 October. The notice reads: “Starting in October, featured professionals will no longer show on your ads. Headshot is optional and won’t affect ad serving.”
Local Services Ads are the ads that sit above the ordinary results carrying a profile, reviews and, until now, a named professional you could nominate. The feature that let you put a specific person forward appears to have been removed, and the headshot is now optional and no longer affects whether the ad serves at all.
Two cautions before anyone acts on it. Schwartz notes the linked help document did not describe the change when he checked, and the notice gives a month rather than a date. This is one advertiser’s screenshot plus a named reporter, which is reporting rather than documentation.
Google’s notice says the rest of the ad holds: photos, reviews and the other elements do not change.
What to do about it
If a person’s name and face are doing the selling in your Local Services Ads, assume that stops this month and look hard at what is left, which is the reviews, the photos and your response time. These ads run heavily in legal services and the home trades, so law firms and contractors feel it first, along with any practice that bought the slot specifically to feature a named principal. Where your own regulator requires an identified licensee in advertising, confirm that requirement is still carried by what remains on the ad rather than by the headshot that is going away.
Source Search Engine Roundtable, Barry Schwartz, “Google Ads To No Longer Show Featured Professionals, Headshots Are Optional,” 19 October 2026 · seroundtable.com/google-ads-featured-professionals-headshots-42325.html (based on an in-account notice screenshotted by advertiser Anthony Higman; the Google help document the notice links to did not describe the change at the time of reporting, and no specific date in October is given)
05 Money
OpenAI is testing new ad formats on your existing creative
Your advertisement can change shape inside ChatGPT this week, using artwork you uploaded for a different one.
OpenAI says the tests use existing campaigns and creative with no action required from advertisers, which also means no approval from them.
OpenAI announced a set of ChatGPT Ads changes late on Friday, reported by Search Engine Roundtable on 19 October from the email OpenAI sent advertisers. The format list covers wider cards, larger images on desktop and more prominent full-width images on mobile, with a multi-image format said to be coming.
The sentence worth noticing is OpenAI’s own. The tests “use existing campaigns and creative, with no action required from advertisers.” Convenient, and it also means artwork composed for one shape can be re-presented in another without anyone at your business approving the result.
The rest is less dramatic and more useful. A redesigned conversions page now puts event volume, the share of campaigns optimising for conversions and any active warnings in one view. Advertisers with live campaigns now get a weekly performance email, and six new measurement partners were added: Northbeam, DV Rockerbox and INCRMNTAL, plus Kochava, Tenjin and Airbridge for mobile.
For a regulated advertiser the format change is the part that matters. A required disclosure sitting in the corner of an image can be lost to a crop, and the advertisement still runs.
What to do about it
Open your ChatGPT Ads creative and look at every image at a wider crop and at full width on a phone. If a disclosure, a licence number or a disclaimer lives inside the image rather than in the ad text, move it into the text now. This is not housekeeping for everyone: a luxury home builder in California has to carry a licence number in advertising, a tax professional has Circular 230 to satisfy, and an admissions consultant has the Federal Trade Commission. A crop is not a defence.
Source Search Engine Roundtable, Barry Schwartz, “ChatGPT Ads New Ad Formats, Updated Conversion Page, Performance Emails & More,” 19 October 2026 · seroundtable.com/openai-chatgpt-ads-updates-42321.html (reported from the email OpenAI sent to advertisers on Friday 16 October; OpenAI gives no figure for how many accounts are in the format tests)
