01 The market data
Apple says Siri answers 2.5 billion requests a day
A launch-day figure with no definition and no method, describing an answer surface that sends your website no visitors at all.
Nine minutes into Apple’s September 9 keynote, Lilian Rincon, the company’s vice president for AI product marketing, gave the one number of the event that matters for anyone who sells expertise. “Our users make over 2.5 billion requests every single day,” she said of Siri, before adding that “our new version of Siri goes further than ever before.” Search Engine Roundtable’s Barry Schwartz pulled the quote from the video the next morning.
Read it as a claim, not a measurement. Apple did not say what counts as a request, so a timer, a text message and a question about a divorce lawyer in Denver all sit inside the same figure. There is no breakdown by kind, no country, and no comparison. Schwartz assumes Google still handles more, and Apple offered nothing to check that against. It is a stage number from the company that sells the product.
The scale is still the point. A surface taking billions of spoken requests a day is a place where a business is either named or it is not, and a spoken answer sends no visitor, so none of it appears in your analytics. It is a phone line into your market that never rings at your desk. We noted on August 14 that Apple had been approaching publishers about paying for content in the rebuilt Siri; the keynote said the new Siri goes further and said nothing a business could verify about where its answers come from.
For a firm, the useful version of the number is an order of magnitude. Voice assistants are now a discovery channel of the same rough size as the ones you measure, and the only way to know what one says about your practice is to ask it.
What to do about it
On an iPhone, ask Siri the three questions a prospective client asks about your practice area and your city, and write down which firms it names and where it says the answer came from. Do it once a month. It takes five minutes and it is the only measurement of this channel you will get.
Add “voice assistant” to the “how did you hear about us” list on your intake form. The number will be small and it will be the first real data you have.
Source
Search Engine Roundtable, “Apple Says Users Make 2.5 Billion Siri Requests Daily,” by Barry Schwartz, September 10, 2026, quoting Lilian Rincon of Apple at the September 9 Apple Event keynote (9:13 mark) · seroundtable.com
02 The evidence
The machine that explains your AI invisibility was right 39% of the time
A Georgia Tech prototype that diagnoses why AI assistants skip a seller names the correct cause in 39 of 100 trials, and the paper is unusually honest about what that means.
Every vendor in this field now sells a diagnosis along with the score: you are not being named because of this, so fix this. Spandan Ghose Chowdhury of Georgia Tech’s College of Computing built that machine and, unusually, published how often it is right. In a paper posted to arXiv on Thursday and accepted for the Decision Sciences Institute’s 2026 conference, the diagnostic agent named the right cause in 39 of 100 trials, with a 95% confidence interval of 30.0% to 48.8%.
The system measures what the paper calls Agentic Share-of-Search: how often a retailer’s products appear in answers from ChatGPT, Perplexity and Gemini, weighted by how high up they appear. A second agent then compares the seller’s products with competitors’ on 14 signals, such as review counts, a questions-and-answers section, price and structured data, and recommends the fix. The test removed one signal at random and asked the agent to find it. Chance would have scored 7.1%, so 39% is a 5.5 times improvement, and 63.9% when the missing lever was one of the three strongest predictors, which the paper notes is the same 39 successes counted a different way.
Then the caveats, all of them the author’s. The study covers one category, running shoes, with 23 products, 7 retailers and 34 questions, and most of the answers were synthetic stand-ins for the real platforms. The one live test ran ten questions against Gemini, thirty calls in all, and only 70% of the answers named a specific product at all; the rest hedged in prose. Live Gemini also credited marketplaces over makers, handing Zappos 39.2% of the visible share for shoes it does not make. The assistant reads the shelf label, not the maker’s mark.
That last finding is the one for a firm. The marketplaces of legal services are the directories, and an assistant that credits Zappos with a Nike shoe will credit Avvo with your expertise. The paper’s own verdict on its diagnoses: they “warrant human review before action; they do not yet establish the agent as a validated proxy for managerial judgment.” Every claim in it is an association, not a cause, and it says so. A vendor’s explanation of why you are invisible is a correlation, and the most careful purpose-built attempt at one was right under half the time.
What to do about it
When anyone tells you why the assistants skip your firm, ask two questions: how do you know, and how often has that diagnosis been right before. A vendor with a good answer to the second question is rare and worth keeping.
Ask whether the answers name your firm or name a directory profile of your firm. The fix is different in each case, and the second is more common than the first.
Source
arXiv, “Agentic Share-of-Search: A Multi-Agent AI System for Competitive Decision-Making in LLM-Mediated E-Commerce,” Spandan Ghose Chowdhury (Georgia Institute of Technology), submitted September 10, 2026, accepted for the 2026 Decision Sciences Institute Annual Conference (100-trial synthetic ablation; live pilot 10 queries, 30 calls, Gemini only) · arxiv.org
03 The structure
Google is shrinking the sources and writing the next question
Two tests spotted the same day: a thinner citations panel beside AI Overviews, and AI Mode generating its own follow-up questions with a skip button.
Search Engine Roundtable reported two Google experiments on Monday, and together they describe where the conversation is going. The first is cosmetic and consequential. Google is testing a noticeably thinner version of the panel on the right of an AI Overview that lists the sources the answer drew on. Khushal Bherwani spotted it, Barry Schwartz reproduced it, and Schwartz’s reading was blunt: it “may be less noticeable to searchers, maybe leading to fewer clicks.”
The second changes who asks the questions. In AI Mode, Google’s chat-style search, Sachin Patel found “paginated follow-up questions and responses, with a skip button to skip them.” Once an answer arrives, “AI Mode generated a new query within the same chat based on the response.” The searcher can skip it. The searcher did not write it.
Put the two together and the shape is clear. The list of where the answer came from gets smaller, and the conversation gets longer under Google’s steering. On Thursday we reported that 97% of the questions in Google’s People Also Ask box are now answered by AI; this is the same move inside the chat. Google has become a tour guide who talks more, points less, and chooses the next room. The second question your prospective client asks is now, in part, Google’s choice, and the page that answers it is the one that gets cited.
Both are tests. Google has issued no statement on either, published no click data, and runs many experiments that never ship. Nothing here is confirmed policy. It is a direction, visible in the product, and consistent with everything else Google has shipped this year.
What to do about it
Search the one question that brings you the most clients and count how many sources are visible in the AI Overview without clicking anything. Write the number down with the date, so you can tell later whether the panel got thinner for you.
Then read your own key pages and check whether each one answers the obvious next question: what it costs, and what happens first. That is the question Google is starting to ask on the prospect’s behalf, and the page that already answers it is the one in the running to be cited.
Source
Search Engine Roundtable, “Google Tests Thinner AI Overview Citations Panel,” by Barry Schwartz, September 8, 2026 · seroundtable.com · Search Engine Roundtable, “Google AI Mode Tests Paginated Follow Up Responses With Skip Button,” by Barry Schwartz, September 8, 2026 · seroundtable.com
04 The one to check today
A stranger can propose a new name for your firm, and you get four days
Google Maps is emailing business owners to approve or reject user-suggested edits within four days, in the same week Search Console lost a month of your indexing history.
The email reads, in part: “A Google Maps user noticed that some of the information on your Business Profile might be out of date and suggested an update.” Then the sentence that matters: “There is a suggestion to update the title for this business. Please share your feedback within 4 days.” Two buttons follow, correct or incorrect. Lluc B. Penycate shared it, Search Engine Roundtable published it on Monday, and Barry Schwartz said the format was new to him although Google regularly sends similar notices.
“Title” is Google’s word for the business name. The name on your Maps listing is the name Google’s AI answers read when they describe your firm, and for a law firm it is also regulated text under the advertising rules of most state bars. Anyone using Google Maps can propose a change to it, and the notice gives you four days to object. The article does not say what happens on day five. It is a stranger filing an amendment to your letterhead, and the clerk gives you until Friday.
The second notice came Friday, about the other record you rely on. Google Search Console’s page indexing report is missing a large stretch of June 2026, the report that shows which of your pages Google could read and when. Google’s John Mueller explained: “This is likely from the time in June where the data was delayed - there just isn’t any data for that period, the page indexing report wasn’t updated.” Google does not backfill indexing data. The gap is permanent.
Neither is a crisis. Together they are one lesson. Google’s record of your firm is edited by other people and kept on Google’s terms, and the only version you control is the one you keep yourself.
Do this today
Search your inbox and your junk folder for “Does this look right to you” and for any message from Google Business Profile, then sign in to the profile and look for pending suggested edits. Answer them. Firms on Microsoft 365 should check quarantine as well, since these notices are exactly the kind of mail that gets held.
Then open Search Console and export the indexing and performance reports to a folder you own. Do it monthly. If Google loses a month again, you will still have it.
Source
Search Engine Roundtable, “Google Business Profiles Email: Does This Look Right To You,” by Barry Schwartz, September 8, 2026 · seroundtable.com · Search Engine Roundtable, “Google Search Console Indexing Report Missing Old June Data,” by Barry Schwartz, September 11, 2026, quoting John Mueller of Google · seroundtable.com
05 The money
Google Ads now writes the explanation of your own spend
Gemini-built dashboards appearing in some accounts turn a typed question into a chart and a paragraph explaining the “why,” written by the company that sold the clicks.
Search Engine Land reported on Monday that Google Ads AI Dashboards have started appearing in advertiser accounts. They “let advertisers create visual performance reports with simple text prompts,” and “each report also includes a real-time AI summary that explains the ‘why’ behind the data.” Paid-search specialist Thomas Eccel spotted them live and posted screenshots; Search Engine Roundtable quoted Google’s own description the same morning: “an insights tool for data visualization, analysis, and export built with Gemini capabilities.”
Until now, someone had to build the report. You or your agency picked the metrics, the date range and the chart, and then a person wrote the sentence that said what it meant. Now you type the question and Gemini builds the chart and writes the sentence. As Search Engine Land put it, the dashboards “shift much of that work to Gemini, letting you describe what you want to analyze while Google builds the report.”
Follow the incentive rather than the feature. The company that runs the auction and sells the placement now also writes the paragraph explaining why your cost per lead went up. Nobody has shown the summaries to be wrong, and this is not an accusation. It is a description of who holds the pen. The landlord now writes the tenant’s annual review of the building. For a firm that pays an agency, the report you receive next quarter may be this paragraph with a logo on top.
The caveats are ordinary. The rollout is to some accounts, not all, and no one has independently tested whether the “why” is accurate. And the ads themselves are unchanged: an ad for legal services is attorney advertising in every state that regulates it, whoever writes the report about it.
What to do about it
Keep one number Google cannot compute: cost per signed client, by month, from your own intake records. It is the only figure in the whole chain that lives outside the platform, and it is the one that decides whether the spend was worth it.
Ask whoever sends your ads report whether the paragraph explaining the results was written by a person or by the dashboard. Require the person’s version, with the dashboard’s attached if they like.
Source
Search Engine Land, “Google Ads AI Dashboards start appearing in advertiser accounts,” by Anu Adegbola, September 8, 2026 · searchengineland.com · Search Engine Roundtable, “Google Ads AI Dashboards Live For Some Advertisers,” by Barry Schwartz, September 8, 2026 · seroundtable.com
