01 The market
Hiding from Google’s AI answers cost one site clicks
Appearing in an AI answer may be pulling clicks toward your other listings on the page rather than taking them away.
This is one website’s test, which is exactly why the result deserves repeating rather than acting on.
Cyrus Shepard used Google’s own opt-out controls to keep one website out of AI Overviews and AI Mode, then compared that period against the weeks either side of it. Barry Schwartz, who reported the test on 30 October, said nobody else has produced data like it because everyone else is probably too afraid to opt out of the generative AI features in Google Search.
The first number is modest and it is the one people have been asking for. Shepard estimated that the AI placements had drawn about a 0.9% click rate, and that those AI appearances were roughly 8% of the site’s total appearances in search. His own verdict on 0.9% was “not too shabby”, which is a fair reading of a surface most owners assume returns nothing at all.
Then the part he called most surprising. “Most surprisingly, regular organic clicks ALSO DROPPED when AI features were disabled,” he wrote. Leaving the AI answer did not send those readers down to the ordinary listing underneath it, and on this site it appears to have cost clicks there as well.
His proposed explanation is a halo. The more places a business appears on one results page, the more likely a visitor is to click at least one of them, so earning a mention inside the AI answer may lift clicks further down the page too. Greg Sterling of Near Media said their own research has seen AI Overviews create a kind of brand awareness that turns into later searches and clicks, “not consistently but many times”.
One website is not a finding, and the brief is running this because of the question it answers rather than the certainty it offers. Schwartz called the sample super limited, and flagged the limit that matters most: the AI answer still appeared on the page throughout, so this compares being inside it against being absent from it, not a results page with an AI answer against one without. For an M&A advisory firm or a tax practice the useful move is not to copy the experiment.
What to do about it
Do not treat the opt-out controls as a traffic strategy on the strength of one website. The honest state of the evidence is that nobody knows, and this is the first number anyone has published.
If you have already opted out, go and look at whether your ordinary clicks moved in the same week. That is the comparison this test was built to make, and you may already be sitting on a second data point.
Source Search Engine Roundtable, Barry Schwartz, “Google AI Overviews & AI Mode 0.9% Click-Through Rate, Opt Out Study Says,” 30 October 2026 · seroundtable.com/google-ai-overviews-mode-ctr-42421.html (a single-site case study, described by its own author as small, with the click rate estimated by subtraction rather than measured directly)
02 The research
Ten AI assistants given the same pages chose differently
Nobody can reliably tell you which of your pages an AI will quote, and that includes everyone selling the service.
The paper is an attempt to build the predictor this industry assumes already exists, and it reports how far short the best one falls.
Elisha Bajemon and Andre-Louis Rochet posted their study on 28 October. They took 229 questions and gave ten AI assistants the identical shortlist of pages for each one, which removes the usual excuse that two engines cited different sources because they searched differently. Then they measured how much of each answer’s citations each page received.
Start with the engines against each other. Handed the same pages for the same question, the measure of how closely their choices agreed came out at 0.704, on a scale where 1 would be perfect agreement and 0 would be none. Two assistants looking at an identical shortlist do not reach the same answer about who deserves the credit.
Then the predictors. The authors tested hosted and open models alongside a smaller one they built and released themselves, which reached 0.542 to 0.549. The best scorers of any kind reached 0.575 to 0.578 on that same scale, and the differences between the leaders were not distinguishable from zero across the 229 questions.
That matters for anyone being sold a forecast. The best published method for predicting which page an AI assistant will cite sits a little over halfway between a coin toss and certainty. It is a real advance on nothing, it is openly released so others can test it, and it is nowhere near a promise.
Forever Cited works in this field, and the number cuts against the most saleable thing in it, which is a confident prediction about what an assistant will quote. We would rather publish the 0.578 than price a certainty that the published evidence does not support. For a cybersecurity firm being pitched on AI visibility, the question worth asking a supplier is what their prediction is measured against.
What to do about it
When a supplier tells you what an AI assistant will cite, ask how that was tested and against what. A number with no stated scale and no test behind it is not a measurement.
Treat this as a preprint rather than settled science. It has not been peer reviewed, and 229 questions is a modest base for a claim about ten engines.
Source arXiv, Elisha Bajemon and Andre-Louis Rochet, “Predicting What Answer Engines Cite: Open Decision Models, Rerankers and a Distilled 4B Student on Ten Engines under Controlled Retrieval,” 28 October 2026 · arxiv.org/abs/2610.38090 (a preprint, not peer reviewed; 229 benchmark questions answered by ten engines given the same candidate sources)
03 The platform
Google Gemini stopped labeling the visits it sends you
Gemini visits from its iPhone app arrive with no source attached, so your analytics will file them under Direct.
The tag appeared in early September and was gone by the end of the month, so a trend line built on it describes three weeks and nothing since.
For about three weeks, links cited inside Gemini’s answers carried a label identifying where the visit came from. Google started adding it in early September and stopped around 28 September. The analyst Natzir tracked it and published a chart showing the tagged visits climb for three weeks and then vanish overnight.
A label of this kind is the ordinary way one website tells another where a visitor came from, and it is what lets your own reports say Gemini rather than guess. John Mueller of Google confirmed on Reddit that he had seen the labels a month or so earlier. Google has not said why they stopped, and the reporting rests on one analyst’s measurements rather than an announcement.
The rest of the signal is intact, with one exception that matters. Visits from Gemini on the web still identify themselves, and Natzir said that count stayed flat throughout. Visits from the iPhone app do not: it opens links inside its own window and sends nothing, so in Google Analytics those arrivals land in Direct, the bucket for people who appear to have typed your address in.
The practical effect is a quiet understatement. If you are counting how much traffic AI assistants send you, the answer has been wrong since the end of September and wrong in the direction that makes the number look smaller. For a software company reporting a growth channel to a board, that is a reporting problem before it is a marketing one.
What to do about it
Look at your Direct traffic for the last two months and ask whether it grew without an obvious cause. That is where these visits now sit.
The one remaining trace is in your raw server logs rather than your analytics dashboard, where the Gemini iPhone app still identifies itself by name. Anyone who maintains your site can pull that.
Source Search Engine Roundtable, Barry Schwartz, “Google Gemini Stopped Attaching UTM Parameters To Cited URLs,” 30 October 2026 · seroundtable.com/gemini-utm-parameters-42399.html (observed and charted by the analyst Natzir; not announced or confirmed by Google)
04 The mechanics
Google is testing an AI agent that answers customer messages
An AI agent will answer your customers out of whatever your Google listing says, so stale hours become wrong answers.
The pilot is free, and the thing to check before switching it on is not the agent but your own listing.
Google is rolling out AI-powered messaging on Business Profiles, the listing that appears in Search and Maps. Customers message your business from their phone’s own texting app, and an agent answers the routine questions instantly. It was spotted by Claudia Tomina, who described the agent fielding questions like what are today’s specials and whether Saturday can be booked.
The handover is the sensible part of the design. When the agent cannot answer something it passes the conversation to you, and you can take over at any point, with every message kept in the Business Profile. An email Google sent to some owners, shared by Mike Blumenthal, offers early access to the pilot at no cost and says the agent will redirect complex questions.
Two limits are worth knowing before anyone turns it on. The agent only works for customers whose phone and network support RCS, the newer texting standard, and everyone else’s messages arrive as ordinary texts straight to your phone at standard rates. Schwartz could not see the feature on his own account, so this is a pilot rather than a general release.
The real exposure is your own data. Tomina’s advice was to make sure your hours and services are current first, because the agent can only answer as well as your profile does. A surgical practice with last year’s hours, or a luxury home builder whose service area was never filled in, now has those gaps answering customers in a text thread rather than sitting unread on a listing page.
What to do about it
Open your Google Business Profile and read it as a customer would: hours, holiday hours, services, service area and the questions people actually ask. Fix what is stale before you let anything answer on your behalf.
Look for the red dot on Messages in your profile, which is how the rollout is appearing. If it is not there yet, the profile work is the part you can do today regardless.
Source Search Engine Roundtable, Barry Schwartz, “Google Business Profiles AI-Powered Messaging Expands,” 29 October 2026 · seroundtable.com/google-business-profiles-ai-powered-messaging-42419.html (a pilot; reported from owner screenshots and Google’s own invitation email, not a general release)
05 The money
Google says AI Mode questions earn less than searches
Google is widening ads into a surface that currently pays it less, which tells you where it expects the volume to go.
Google said so on the same earnings call that reported a record advertising quarter, which is what makes it worth reading.
On Google’s third-quarter earnings call, Philipp Schindler addressed something the company had not said plainly before. Each AI Mode question currently earns Google less than a traditional search does, and he paired that with the expectation that it will improve over time. He also said many AI Mode questions are upper funnel, meaning people researching early rather than deciding what to buy.
Schindler was not retreating. “To be clear, we feel very good about our trajectory here,” he said, adding that the ad experiments “have shown promising results” and that Google is now scaling ads inside both AI Overviews and AI Mode to more users and more questions.
Google’s own announcement two days earlier supplies the mechanism. Writing on 28 October, Shashi Thakur, a vice president in Google Search Ads, said that “In AI Mode, queries are, on average, three times longer than traditional searches”. A longer, more conversational question is usually an earlier one, and an earlier question is worth less to an advertiser at the moment it is asked.
Put the two together and the direction is legible. Google is expanding paid placement into the surface that pays it least per question, which is a bet on volume rather than on price. For a franchise advisory practice or a tax practice, the early research question is exactly the one you want to be the answer to, and it is now the question somebody else can buy.
What to do about it
If you advertise, stop judging AI Mode placements by the standards of a classic search ad. Google has now said itself that the questions arrive earlier, which means the value shows up later and somewhere harder to attribute.
Whatever the channel, your own advertising rules follow you into it. Comparative claims, results and required disclaimers apply inside an AI answer exactly as they do on your website, and the governing rules are your own profession’s rather than the platform’s.
Source Search Engine Roundtable, Barry Schwartz, “Google Says Ads In AI Mode Earning Less Than Traditional Search Ads,” 29 October 2026 · seroundtable.com/google-ai-mode-ads-earn-less-42417.html (reporting Philipp Schindler on Google’s third-quarter earnings call; Google gave no figure for the gap)
Source Google, Shashi Thakur, VP and GM, Google Search Ads, “Ads in AI Mode expand to new markets,” 28 October 2026 · blog.google/products/ads-commerce/ads-in-ai-mode-global/
