01 The evidence
Cornell and McMaster researchers found AI explanations barely moved judgment
Your customer is barely checking the AI’s answer about your field, and showing them its reasoning moves that very little.
One hundred and ten people were given an AI assistant on work they had no training in, and they overruled it on fewer than one answer in twenty.
The setup matters more than the subject. Eun Jeong Kang of Cornell University, with Peter Duan and Swati Mishra of McMaster University, gave 110 people a medical annotation task they had no training for, put an AI assistant beside them, and deliberately withheld any feedback about whether each decision had been right.
That is the condition most of your customers are in. Across the two AI conditions, participants accepted the assistant’s suggestion 75.0% of the time, modified it 20.2% of the time, and rejected it 4.8% of the time.
Half the group also saw the assistant’s own explanation, its confidence score and a dictionary they could look terms up in. Those people accepted slightly less, 72.0% against 79.0%, and the authors did find a small improvement in how closely their choices tracked whether the assistant was actually right. On the study’s main measure of well-judged trust, though, the explanation made no credible difference at all.
Two things drifted over the session. They went along with the assistant more as they worked through the tasks and overruled it less, so the longer they used it the more they deferred to it. They had rated their own competence at 4.33 out of 5 before starting, and 4.03 by the end.
Nothing here is about medicine. It is about what happens when somebody with no training in your subject is handed a confident answer and no way to find out whether it was right. A tax professional reading a client’s forwarded AI answer, an accounting practice asked how a new business should be taxed, and a managed IT provider asked why a backup failed are all on the other side of that same gap.
What to do about it
Ask an AI assistant three questions a customer would ask about your field, and write down which answers you would have had to overrule. That number is what your prospects are not catching.
Then put the correction where they will hit it. A page on your own site that says plainly what the wrong answer gets wrong is the only version of this a customer can check without calling you.
Source arXiv, Eun Jeong Kang, Peter Duan and Swati Mishra, “Novice Reliance Calibration in AI-Assisted Decision Making: The Role of Explanations and Self-Assessment,” October 6, 2026 · arxiv.org/abs/2610.07800
02 Research & evidence
Claude Sonnet asserted a source link 26 times in 28
An AI naming your business as a source does not mean the page it pointed at backs up what it said.
Researchers deleted the record that tied a citation number to its source, then watched whether two AI readers would admit they could no longer tell.
The test is small and deliberately airtight. Taehyeon Yun and five co-authors took 64 saved runs of a banking AI agent, each with one known right answer, and removed the table that tied each citation number to the record it stood for.
Without that table the honest answer is that you cannot tell which record the citation refers to. Claude Sonnet found the right place in the records every time. Then, in 26 of the 28 cases that needed the missing link, it asserted the link anyway.
Here is the part that matters for anyone who checks AI output by reading it. In 22 cases the answer still matched the correct reference anyway. The model was right while being unable to show why, and nothing in the output told the two apart.
Giving the model the link table back improved things for both AI readers. Renaming the identifiers on their own fixed nothing. A plain mechanical comparison, with no AI in it, either resolved the question correctly or declined to answer, every time.
The setting is narrow and the authors claim nothing about web search. The distinction travels anyway. A franchise advisory practice or a real estate brokerage that turns up in an AI answer is being cited by a system that may not be able to show which page supports the sentence your name is attached to.
What to do about it
Pick one claim an AI assistant makes about your business and open the page it cites. Check that the sentence it attributed is actually on that page.
Where it is not, the fix is a page of your own that states the fact plainly, so the next answer has something that does support it.
Source arXiv, Taehyeon Yun, Dongho Kim, Geonwoo Kim, Juyoung Seo, Minseok Hur and Moohong Min, “Correct Answers, Unsupported Findings: Evidence Binding in Forensic Reconstruction of LLM Agent Logs,” October 7, 2026 · arxiv.org/abs/2610.09581
03 What did not happen
Two researchers found Reddit question-asking never fell after ChatGPT
The claim that AI took away demand for human expertise does not survive a measurement with a control group.
The study compares 26 communities where an AI can answer against 90 where it cannot, and finds the drop everybody cites is mostly a trend that predates ChatGPT.
Published on September 11 and not widely picked up, so treat it as context rather than news. Hazem Ibrahim and Yasir Zaki asked whether people stopped asking other people for advice online once an AI could answer instead.
Earlier studies reported declines of 8% to 25%. This one tracked 26 Reddit communities built around asking for information against 90 similar-sized hobby communities over the same six months before and after ChatGPT’s launch, and it rules out any decline larger than 3.4%.
Steady numbers could still hide a switch to AI-written posts, so the authors scored 274,411 posts and 223,775 comments with AI-text detectors. AI-written posts rose 2 to 3 percentage points more in the information communities than in the hobby ones, short of the 5.1 points that would have been needed to mask even the smallest decline previously claimed. The replies people received showed no such rise at all.
The explanation for the disagreement is the useful part. The two kinds of community were already drifting apart before ChatGPT existed, at rates comparable to every published estimate, which the authors show by running their whole method at 66 earlier dates when no ChatGPT effect could exist. Their own largest number, an 18% fall in posts to low-stakes curiosity communities, matches that community’s pre-existing trend.
This cuts against the premise the AI visibility field sells on, Forever Cited included, that people have moved their questions to the assistants and the old places are emptying out. The measurement says they have not, and a number repeated without a control group is worth less than the people repeating it would like.
What to do about it
Before you accept that your customers have moved, look at your own numbers. A charter operator and a real estate brokerage both have a phone log and a contact-form count going back three years, and those are better evidence about their own buyers than any industry figure.
Treat any percentage about AI replacing something as unproven until you can see what it was compared against.
Source arXiv, Hazem Ibrahim and Yasir Zaki, “Informational Help-Seeking on Reddit Did Not Decline After ChatGPT,” September 11, 2026 · arxiv.org/abs/2609.12447
04 Your ad bill
Google will charge for calls that rang past 20 seconds
If your ads buy phone calls, you now pay for the ones nobody picked up.
Google rewrote the page explaining which Local Services Ads leads are billable, and the new version charges for unanswered calls and drops the published list of what qualifies for a credit.
Local Services Ads are the ones that sit above the search results with a Google badge, and they are billed per lead rather than per click. Google rewrote the page that explains which leads count on October 8, and the rewrite is not cosmetic.
The main change is that an unanswered call can now be charged if the caller stays connected for more than 20 seconds during business hours. Google also spelled out when that clock starts for calls that reach a menu or a voicemail.
Three other changes matter. A lead can be re-evaluated for up to 30 days and credited later, so your bill is not final when you first see it. Follow-up calls and messages from the same customer are generally not charged again within 15 days, and a lead now shows one of four statuses: charged, not charged, in review, or credited.
The removal is the one to notice. The old page carried a long list of situations that would not qualify for a credit, and that list is largely gone, replaced by Google’s automated lead-quality assessment. You are now disputing a charge against a judgment you cannot read.
This is not everybody’s problem, and it is expensive for the businesses it is. Local Services Ads run in the categories where the phone call is the sale, so an attorney’s office, a real estate brokerage and a luxury home builder are the ones who should open that page today and check what happens to their calls after hours.
What to do about it
Call your own Local Services Ads number from a phone that is not in your contacts, let it ring 25 seconds, and hang up without speaking. Then check whether that shows up as a charged lead.
Two settings decide most of this: what your after-hours routing does, and whether a caller reaches a menu before a person. Check those before you dispute anything.
Source Search Engine Roundtable, Barry Schwartz, “Google Local Service Ads Expands How Leads Work Help Document,” October 8, 2026 · seroundtable.com/google-local-service-ads-how-leads-work-42259.html
05 The money
Google opened lift testing to advertisers with 1,000 conversions
The test that shows whether your ads caused anything is now self-serve, and most expertise-led businesses still cannot reach the minimum.
Conversion Lift no longer needs a Google representative to set up, but it does need 1,000 recorded conversions and a $5,000 campaign budget.
Conversion Lift is Google’s own experiment. It withholds your ads from a randomly chosen slice of people and compares what that group does against everybody else, which makes it the one thing Google offers that answers whether the ads caused the result rather than whether the result happened near the ads.
Until this week you needed a Google Ads representative to set one up for Search and Performance Max campaigns. Hana Kobzova spotted the rewritten help page, and that requirement is gone.
The eligibility is now published, and it is the story. You need at least 1,000 observed conversions, and conversions carrying supplementary data do not count toward that minimum. You also need a campaign budget of at least $5,000.
Work out what 1,000 conversions means for a business that sells a handful of engagements a year. A specialized software company or a managed IT practice counting form fills and demo bookings may clear it. A practice whose conversion is a signed engagement will not come close.
That is not a complaint about the threshold, which exists because a smaller test cannot separate a real effect from noise. It is the reason the honest answer to whether the ads worked is usually that nobody has measured it, and that is worth knowing before anyone quotes you a figure implying otherwise.
What to do about it
Count your recorded conversions over the last 90 days and multiply by four. If that is under 1,000, nobody can run this test on your account, and any report implying someone did is worth a question.
Where the number is too small, the usable substitute is a held-back period. Stop one campaign for a defined stretch, keep everything else steady, and write down what changed.
Source Search Engine Roundtable, Barry Schwartz, citing Hana Kobzova at PPC News Feed, “Google Ads Conversion Lift Measurement Now Self-Service,” October 8, 2026 · seroundtable.com/google-ads-conversion-lift-measurement-self-service-42250.html
