01 The money
Google will write ad copy from your reviews
Your advertisement can carry praise you never wrote, assembled from reviews on sites you have never visited.
The compliance for a sentence about your business still sits with you when Google’s AI composed it.
Google published a help document describing review summaries on Search ads, reported on 26 October by Barry Schwartz at Search Engine Roundtable. Google’s wording is that they are AI-powered summaries of online reviews that appear directly as part of your Search ads. The reviews are aggregated from across the web, not only from the advertiser’s own Google listing.
The scope is set out plainly. It runs on Search campaigns, in the United States, Australia, the United Kingdom and India, in English, at all positions on Google.com and on all supported devices. Every advertiser category is covered except sensitive ones, with pharmaceuticals given as the example, and the summaries are described as positive only.
A testimonial rule does not stop applying because a machine wrote the sentence. For a wealth advisor or a registered investment adviser, the Securities and Exchange Commission’s marketing rule governs how a client testimonial may be used in advertising, including the disclosures that have to travel with it. A real estate brokerage answers to state license law and the Fair Housing Act for the claims its advertising makes, whoever composed them.
One thing the reporting does not settle is the part that matters most to a regulated advertiser. It does not say whether an advertiser can switch the summaries off, and that is the question worth asking before the words appear rather than after.
What to do about it
Search your own business name on Google and read your own ads as a customer sees them. Then ask your Google representative two questions: whether review summaries are active on your account, and whether they can be turned off. If you work under advertising rules, get both answers in writing before the next campaign runs.
Source Google Ads Help, “Review summaries on Search,” reported by Search Engine Roundtable (Barry Schwartz), 26 October 2026 · seroundtable.com/google-ads-ai-review-summaries-doc-42370.html
02 The consumer data
Google AI answers appear on 62% to 82.5% of searches
The figure a marketing vendor quotes you for AI’s reach depends entirely on which of three unaudited studies they picked.
Each number describes a different set of searches, which is why the gap between them runs to twenty points.
Barry Schwartz brought the three together on 26 October. The headline figure is that AI Overviews now appear on over 80% of Google searches, up from 50% in June of last year and about 15% the year before that. It is attributed to data from Semrush, SE Ranking and seoClarity collectively, without each number being assigned to one source.
The three do not agree. SE Ranking put AI Overviews on 82.5% of United States brand searches, which is the narrowest and best defined of the three, because a brand search is somebody typing a company name. seoClarity reported over 62% of queries, ranging from 55% to 84% by industry, so those two describe different populations and the 80% figure describes a third.
The brand-search number is the one that matters most here. If 82.5% is right, an AI answer is what greets most people who look your business up by name. For a real estate brokerage or a luxury home builder, where a referred buyer very often searches the company before calling it, that is the screen a referred buyer now sees first.
Not one of the three published a sample size or a method in the reporting, and the three disagree by twenty points. Forever Cited sells AI visibility measurement, so this is our own number as much as anyone’s, and a share that cannot be traced to a stated method is not a measurement.
What to do about it
When anyone quotes you a percentage for how much of search AI now answers, ask two questions: which searches it counted, and how many. A figure for brand searches and a figure for all queries are not the same claim, and today they are twenty points apart.
Source Search Engine Roundtable, Barry Schwartz, “Over 80% Of Google Searches Have AI Overviews & Growing,” 26 October 2026 (figures attributed to Semrush, SE Ranking and seoClarity; no sample size stated for any) · seroundtable.com/over-80-of-google-searches-have-ai-overviews-42372.html
03 The research
Confident AI falsehoods cut test scores from 7.4 to 2.9
A customer who hears a confident wrong answer about your business will keep it, and stop being able to tell.
Confidence is what broke: people could no longer tell their right answers from their wrong ones.
A paper posted on 22 October, “Who Is Sycophantic to Whom?”, by Deepthika S, Pawni Yadav and B. Sankar, ran a within-subjects experiment. Thirty-two participants learned up to five science topics from assistants whose briefings carried a rising share of plausible falsehoods, each delivered with full confidence.
Across 104 quizzes and 1,040 answers, scores fell from 7.4 to 2.9 out of 10. The planted option was chosen on 60% of falsified items against 11% of correctly taught ones, an odds ratio of 12.5. Participants held the wrong answers confidently, and the authors report that confidence stopped separating right answers from wrong ones.
This brief reported on 17 October that models shift their answers when a user pushes back. This paper measures the other direction, the person deferring to the confident machine, which is why its authors argue that trustworthiness belongs to the human and the assistant together rather than to the model alone.
The sample is 32 people and the subject matter was science, not who to hire, so this is a mechanism and not a measurement of your market. The mechanism is still the one that matters. An admissions consultant whose fee or record is described wrongly by an assistant is not meeting a skeptical parent, but a confident one holding the wrong figure.
What to do about it
Ask an assistant the three questions a customer asks before hiring you, and read the answers for facts rather than for tone. Anything wrong is worth correcting at the source it came from, because on this evidence a confident wrong answer does not get discounted by the person who hears it.
Source arXiv, Deepthika S, Pawni Yadav and B. Sankar, “Who Is Sycophantic to Whom? Confident LLMs, Human Deference, and the Collapse of Trust Calibration,” 22 October 2026 (a preprint, not peer reviewed; 32 participants, 104 quizzes, 1,040 answers) · arxiv.org/abs/2610.31579
04 The research
The best AI agent answered 39% of web research tasks
When an assistant goes out to research your business on your own website, the leading systems get it wrong more often than right.
A system that cannot finish the research step still produces an answer about your business.
“Deep Search in Real-World Web Environments” was posted on 19 October by 25 authors led by Xiaomin Yu. Its benchmark, OpenDS-World, is 200 tasks that require finding and combining information from live websites rather than from a frozen copy of them. Twelve frontier models were evaluated under one shared tool environment and execution protocol.
The best-performing model reached 39.0% accuracy, and no model in the set answered a majority of the tasks correctly. The paper also reports a figure from 1,000 independent human attempts, but the abstract does not define what those attempts covered, so it is not repeated here.
This is the step an assistant performs when a customer asks about you. For a luxury home builder whose service area, warranty terms and finished projects each sit on a different page, the finding is that the failure happens while the facts are being gathered, not in the writing that follows.
The tasks that defeat these systems are the ones needing two or more facts joined from different places. A page that answers one whole question by itself is the format these systems complete most reliably, which is a decision about layout rather than technology.
What to do about it
Pick the question a customer asks most often before buying from you, and check whether any single page on your site answers it completely. If the answer is spread across three pages, it is a question an assistant is likely to get wrong, and one page is the fix.
Source arXiv, Xiaomin Yu and 24 co-authors, “Deep Search in Real-World Web Environments,” 19 October 2026 (a preprint, not peer reviewed; 200 tasks, 12 frontier models) · arxiv.org/abs/2610.30181
05 The deadline
Google added a one-click test for the February ad change
You can now test the ad system you will be moved to in February against your current one, before the move.
February’s upgrade arrives whether or not an advertiser has measured what it changes.
Search Engine Roundtable reported the change on 26 October. Google’s own FAQ says that if an account is eligible, a one-click experiment recommendation appears on the Recommendations page, and that selecting it sets up a 50/50 split test comparing your base Search campaign against an AI Max treatment arm without requiring manual configuration.
The date behind it is not new, and this brief has carried it before. Google’s announcement, updated on 11 June, says the Dynamic Search Ads sunset and automatic upgrade will begin in February 2027, and that campaigns containing dynamic ad groups will transition to standard ad groups. What is new today is the one-click way to see what that does to your own results first.
Google’s wording in the same announcement is that transitioning now is how an advertiser keeps full control over your campaign setup. The advertiser who tests early picks the settings, and the advertiser who waits inherits the migration’s defaults. Attorneys, real estate brokerages and independent insurance brokers are the heaviest Search advertisers among the industries this brief serves.
Two limits are worth knowing before starting. The report does not say how long the experiment runs, and refinement is available at ad-group level inside the treatment arm, so a test can be narrowed to the campaigns that matter rather than run across everything.
What to do about it
Open the Recommendations page in Google Ads and look for the one-click experiment. If it is there, run it on one campaign you understand well, and let it gather enough data to be worth reading before February. The result is the only thing that will tell you what the migration costs or saves you.
Source Search Engine Roundtable, Barry Schwartz, “Google Ads 1-Click AI Max Experiment Recommendation,” 26 October 2026, with Google’s own upgrade announcement (Brandon Ervin, updated 11 June 2026) · seroundtable.com/google-ads-1-click-ai-max-experiment-42324.html · blog.google/products/ads-commerce/dsa-upgrade-to-ai-max-2026/
