01 The market data
Microsoft counted almost 2 million AI agents visiting sites
Software, not people, is now a measurable share of website traffic, and almost nobody is checking what it does once it arrives.
Microsoft published the figure about its own product and named nothing for it to be a share of, so the direction is worth more than the total.
Microsoft Clarity, the company’s free website analytics product, added a feature it calls Agentic Analytics. It reports which AI agents visit a site and what they do once they are on it. Announcing it on 4 November, Microsoft said “AI agents nearly doubled in three weeks to almost 2 million”.
The announcement states no denominator, no sample and no method. Two million agents across how many sites, counted how, is not published anywhere in it. The three-week doubling is the part worth keeping, because a direction is harder to get wrong than a total.
The same announcement expanded Brand Agents to local businesses, which Microsoft describes as helping them capture leads and book appointments across Microsoft Copilot, Bing and MSN. That makes two platforms in one week moving the booking inside the assistant.
For an information technology or cybersecurity firm this is operational before it is commercial. Agent traffic behaves nothing like human traffic: it arrives with no referring page, ignores your navigation, and opens pages no person has looked at in months. Forever Cited measures AI visibility for a living, which is the reason a vendor’s own number gets reported here with its gap named rather than repeated as a fact.
What to do about it
Open your website analytics and look for visits with no referring page that load your pages in an order no person would choose.
If your analytics cannot separate software from people, close that gap before you buy anything priced in agents.
Source Search Engine Roundtable, Barry Schwartz, “Microsoft Clarity With New Agentic Analytics & Ask Clarity,” 4 November 2026 · seroundtable.com/microsoft-clarity-gains-agentic-analytics-and-ask-clarity-42449.html (the figure is Microsoft’s own, quoted from its announcement; no sample, denominator or counting method is stated, and Microsoft’s own post could not be read from this session)
02 The research
Researchers hid one buyer requirement and success fell to 28%
An AI agent recommends from what your pages already say, because it will not stop to ask the customer what they actually need.
The benchmark ran on books, so the subject is not your field, but the failure it isolates is.
Bharath Sivaram Narasimhan and Karthik R. Narasimhan published a new benchmark on 3 November. It runs 450 tasks against a catalogue of 20,000 books drawn from Wikidata, with four AI agents and two search interfaces, and it scores recommendations against the catalogue itself rather than asking another model to judge them.
The variable is when the buyer says what they want. With every requirement stated up front, average success is 97.5%. With one stated and the rest emerging only if the agent asks, it falls to 75.6%. With some never stated at all, it falls to 28.3%.
The cause is not reasoning. The agents mostly did not ask. Told explicitly to ask, every agent recovered to at least 93.1% on the middle condition, which the authors report as the main source of those failures.
Read that against the booking news above. For an architecture and engineering practice that only takes historic restoration work, the requirement that decides the job is the one a client does not know to say out loud. If your pages do not state it, nothing in that conversation will.
What to do about it
Write one page that says who you are for and who you are not for, in the words a client would use.
Not a list of services. The specifics an agent can match against: the kind of building, the kind of case, the kind of patient.
Source arXiv, Bharath Sivaram Narasimhan and Karthik R. Narasimhan, “tau-Rec-Books: A Disclosure-Controlled Benchmark for Requirement Elicitation in Agentic Recommender Systems,” submitted 3 November 2026 · arxiv.org/abs/2611.04459 (a preprint, not peer reviewed; the buyer is simulated by rules rather than by a model, and the catalogue is books, not professional services)
03 The platform
Google AI Mode now books home services through partners
A customer who finds you inside an AI answer may now book on a lead platform instead of ever reaching your office.
Google announced it through an employee’s post rather than a blog entry, and it runs in the United States only.
Robby Stein of Google posted on 4 November that AI Mode now carries direct booking links for home services, naming Angi Home, Thumbtack, Housecall and GetJobber as partners, with ServiceTitan listed as coming soon. It is available in the United States only.
This follows the pattern Google set when it first put booking links in AI Mode. It said then that the feature was “expanding soon to local service appointments and event tickets”, and restaurant reservations went through OpenTable, Resy and Tock on the same model.
The booking does not happen on your website. It happens on a partner platform, and several of those platforms sell leads. A customer Google’s AI sends you can arrive as a paid lead rather than as a direct enquiry.
For a luxury home builder the exposure is narrower than for a plumber, because the partner list is weighted towards trades and jobs rather than towards projects. The question worth settling now is whether your business already appears on any of those four platforms, and on what terms.
What to do about it
Search the four partner names and check whether your business already has a profile on any of them, whether or not you created it.
If it does, read what it says and what it charges. That is the page Google’s AI may now hand your customer.
Source Search Engine Roundtable, Barry Schwartz, “Google AI Mode Integrates With Home Services Booking,” 5 November 2026 · seroundtable.com/google-ai-mode-home-services-booking-42459.html · the booking-partner pattern and the quoted expansion line are from Google’s own post, blog.google/products-and-platforms/products/search/ai-mode-agentic-personalized, which predates this change (Google published no blog post for this one; the announcement is a post on X by a named Google employee, and no figure for reach or partner terms is stated)
04 The deadline
Google Ads now labels AI-made ad assets in California
If a tool generated or retouched your ad images, people in California may now see a label on the advertisement saying so.
The change is live now rather than on a future date, and the label can appear on the ad itself rather than in a policy page.
Google updated its advertising documentation on 5 November to add California, South Korea and Vietnam to the places where AI use in an ad must be disclosed. The European Union, India and New York were already on that list.
The mechanism has two parts. An advertiser designates an asset as AI created or edited, and in some places Google then puts a visible overlay on the ad. Google’s documentation says “In some geographies, AI labels will also show on the ad itself.”
No compliance deadline is stated, which in practice means it reaches campaigns that are running now. The Merchant Center product specification names the same six places as requiring a disclosure or a label.
For a California surgical practice this sits on top of rules you already answer to. Your state medical board and the Federal Trade Commission already govern what a before-and-after image may imply, and a label beside that image invites a question you should be ready to answer plainly. The same holds for a tax practice under Circular 230 and an insurance brokerage under its state Department of Insurance: a new channel does not come with new advertising rules, it comes with your existing ones.
What to do about it
Open your Google Ads account and find which assets are marked as AI created or edited.
If you advertise into California and cannot say which images were generated or retouched, settle that this week, before a label appears beside one of them.
Source Search Engine Roundtable, Barry Schwartz, “Google Ads Expands AI Content Disclosure Label To More Regions,” 5 November 2026 · seroundtable.com/google-ads-expands-ai-content-label-disclosure-42461.html (Google’s own Help documentation could not be read from this session, so the quoted line is as Search Engine Roundtable reproduces it; Google states no effective date and no compliance deadline)
05 The evidence
No AI model passed 45% on expert-grade buying advice
An assistant can be reliably correct and still not give the advice a specialist would, especially where the decision is still open.
The standard was built from replies that led to completed orders, which is harder to clear than simply being right.
Twelve researchers led by Yuqin Dai published CARTBench on 4 November, a benchmark for shopping advice on complex and ambiguous requests. It holds 1,603 product records across 17 retail domains and 1,328 criteria written by experts, and its rubrics were distilled from excellent replies picked by hand out of orders that actually completed.
The scoring separates what a reply is required to do from what a good reply does. Three of ten models cleared 90% on the required behaviour. None cleared 45% on the higher standard. The authors call the difference an understanding-to-execution gap.
Where the models failed together was product evidence and unresolved needs, which the paper ties to changing requirements and decisive unknowns. That is the same fault the book benchmark found from the other side: the hard part is not knowing things, it is settling what the buyer has not settled.
For an advisor handling a franchise or merger decision, that distinction is the useful one. A client arriving with an AI-assembled shortlist holds something correct and probably not something expert, and the difference sits exactly where their decision was still open.
What to do about it
Ask the question your least certain client arrives with, in an AI assistant, and read the answer as that client would.
Note what it left unresolved. Then make sure your own pages resolve it.
Source arXiv, Yuqin Dai and eleven co-authors, “CARTBench: Rubric-Based Evaluation of Shopping Advice for Complex and Ambiguous Requests with Timelines,” submitted 4 November 2026 · arxiv.org/abs/2611.05862 (a preprint, not peer reviewed; the domains are retail rather than professional services, and the abstract does not define how the two scoring tiers are calculated)
