01 The consumer data
OpenAI says ChatGPT does a fifth of Google’s searches
A second search engine the size of a fifth of Google now decides which businesses get named, and you cannot see inside it.
The figure comes from OpenAI rather than from anyone measuring it independently, and OpenAI also wrote the definition of what counts as a search.
Adam Fry, who leads the consumer product at OpenAI, told Alex Kantrowitz on the Big Technology podcast that ChatGPT’s volume of search-like queries is now roughly one-fifth the volume of queries of Google. He also said ChatGPT has gone from one billion to 1.2 billion users in the last two months.
Asked what counts as a search, the answer was broad. “Every time you ask a question and go search the web, that’s a search,” Fry said, naming questions about a person or a place, and finance or health questions where the answer has to be current.
Set that against the comparison the search industry reached for within hours. Alex Kantrowitz noted that Bing holds about 5% search share, which would put ChatGPT’s web searching ahead of Microsoft’s search engine. Glenn Gabe posted it as a question rather than as a fact.
None of it is audited. Google says its own search volume is at an all-time high, and both statements can be true at once if the total is growing. The number is OpenAI’s, the definition of a search is OpenAI’s, and no third party has published a figure to check it against.
For an independent insurance brokerage the practical reading is narrower than the headline. A fifth of Google’s volume is enough that a prospective customer may now ask ChatGPT who to call and never open a page of search results at all. Unlike a page of search results, nobody can show you afterwards what it said.
What to do about it
Ask ChatGPT the question a customer would ask. Describe your service and your area, leave your business name out of it, and write down the names that come back.
Then do it again tomorrow and compare the two lists. That is a two-minute exercise, and it tells you more than the volume figure does.
Source Search Engine Roundtable, Barry Schwartz, “OpenAI Says ChatGPT Is At 20% Of Google Search Volume,” 4 November 2026 · seroundtable.com/chatgpt-20-percent-google-42451.html (the underlying interview is Alex Kantrowitz’s Big Technology, which could not be read from this session, so the quotations are as Search Engine Roundtable reproduces them; the figures are OpenAI’s own, with no stated method or sample, and Bing’s 5% share is Kantrowitz’s figure)
02 The research
Eight of ten AI platforms cited an invented term
If a made-up term can reach an AI answer in a week, a competitor’s claim about your field can reach one too.
This measurement is four weeks old rather than this week’s news, and the number that should stop you is the price.
Six researchers led by Qi Liu published a measurement on 8 October of where AI search platforms take their citations from, and how hard it is to get into that pool. Across ten AI search platforms they analyzed 17,211 citations drawn from 6,356 different websites, and found that the top twenty websites accounted for between 20.5% and 70.8% of each platform’s citations.
Then they tested the barrier to entry. Of 22 publishing platforms tied to those cited websites, 15 had low or medium barriers to both opening an account and posting. So they published ordinary posts carrying an invented term, and watched what the AI platforms said afterwards.
Eight of the ten platforms cited the fabricated concept within seven days. One article on a platform the engines preferred had more citation impact than more than twenty matched posts on platforms they did not, so placement counted for more than volume.
The last finding is the one with a price on it. The researchers bought a fourteen dollar generative engine optimization package, meaning a paid service for getting a business mentioned in AI answers, which produced thirteen public posts, and one AI search platform cited the marked content within one hour. This is the trade Forever Cited works in, so the finding is ours to answer for: a citation that fourteen dollars can buy is not evidence of authority, and anyone selling AI visibility should be the first to say so.
The limits are real: this is a preprint, it is four weeks old, and the invented term was a concept rather than a business name. For a specialty veterinary hospital the reading is still direct. The pages an AI answer draws on about emergency animal care sit on a small number of websites, and some of those websites will take a post from anybody.
What to do about it
Find out what the handful of websites in your field actually say about you. Search your own service and area, open the three or four directories, forums or round-up pages that keep reappearing, and read your own entry on each one.
Those are the pages an AI answer about your field is most likely to be built from, and on some of them anybody can post.
Source arXiv, Qi Liu, Geng Hong, Xinyang Zhang, Pei Chen, Yutong Li and Min Yang, “From Public Posts to AI-Search Citations: Measuring the Fragility of AI Search,” 8 October 2026 · arxiv.org/abs/2610.11932 (a preprint, published four weeks before this issue and carried as context rather than as news; 17,211 citation instances across 6,356 source domains on ten AI search platforms, 22 publishing platforms tested for barriers, and marker-controlled publication experiments; the authors state no figure for how often this happens outside their own experiments)
03 The evidence
A fifth of web pages answered software with no content
Your website can report itself working to the software that reads it while handing over an empty page.
The pages here were collected by software rather than visited by customers, and the share that failed without saying so is the number to take away.
When an AI assistant answers a question, it often fetches the page at that moment. Shuyang Zhang fetched 10,307 web addresses across 1,507 websites over three days and checked what those fetches actually returned.
A web server answers every request with a status, and the status meaning success is the one everything downstream trusts. Of the pages that returned a success status to a plain fetch, 20.9% did not contain the page’s main content. What came back instead was an empty shell, a bot challenge, a block page, an error page or a fragment.
The study also asked every website for an address that does not exist. 13.1% of the websites that answered at all answered a made-up address with a success status, which means software cannot tell a real page from nothing by the status alone.
The knock-on was measured rather than asserted. Treating success as content overstated how many claims could be checked against a live page by 12.2 percentage points, and raised the apparent rate of citations drifting away from their sources from 5.8% to 21.9%.
For a wealth advisory practice this is a question for whoever runs the website rather than a marketing question. A bot challenge in front of your service pages reads to an AI assistant as a page that loaded and said nothing, and your own analytics will show no problem at all.
What to do about it
Ask whoever runs your website one question: when software that is not Google asks for our main service page, what comes back? If the answer is a challenge screen, a block page or an empty shell, the AI assistants fetching that page at question time are getting the same thing.
The study names three cheap checks that caught most of it: how much text came back, whether a made-up address gets the same answer, and whether the page quietly redirects to the home page.
Source arXiv, Shuyang Zhang, “When 200 OK Is Not an Answer: Silent Fetch Failures in Agentic Web Content Acquisition,” 31 October 2026 · arxiv.org/abs/2611.00884 (a preprint; 10,307 web addresses across 1,507 sites in five groups, plus the 4,795 addresses cited by expert-verified claims in ExpertQA, labeled automatically with 300 items reviewed and 60 of those checked by a person; on the held-out test split the reviewed sample lowers the 20.9% estimate to 20.0%)
04 The deadline
Google will pass full visitor addresses to its ad tags
From January your website can hand Google’s advertising tags the full internet address of everyone who visits.
The date is fixed and the change is to a default, so a setting nobody revisits will decide what leaves your site.
Google has emailed advertisers about a change to server-side Google Tag Manager, the plumbing that sits between a website and the advertising and analytics services it reports to. Google wrote: “Starting January 11, 2027, server-side Google Tag Manager (SGTM) will allow tags to process standard network signals like IP address to facilitate improvements to Ads measurement and personalization.”
Today that plumbing can strip the visitor’s internet address across the whole container at once. From 11 January 2027 the stripping moves from the container to each tag, and the tags will not behave alike: tags sending data to Google Ads, Display & Video 360, Search Ads 360 and Campaign Manager 360 will accept the full address, while Google Analytics tags carry on stripping it.
Third-party tags each decide for themselves. Google’s own email says the stripping can still be applied first, and that “transformations can redact these signals before any tag accesses the data”, so this is a change to what happens when nobody touches it rather than an instruction to send anything.
An internet address is personal data under European data-protection law in many circumstances, and receiving one does not by itself identify anybody by name. The exposure is not even across industries: for a tax practice, the people reading a page about unpaid tax debt are identifiable as a group by the page they chose.
None of this is automatic for a business that does not run server-side tagging, which is most of them. The question is whether yours does.
What to do about it, before 11 January 2027
Ask whoever handles your website and your advertising two questions. Do we use server-side Google Tag Manager? And if we do, is there a transformation that strips visitor addresses before any tag sees them?
If the first answer is yes and the second is no, the default changes on that date without anybody doing anything. Whatever rule governs advertising in your profession applies to this the same way it applies to every other channel you use.
Source Search Engine Roundtable, Barry Schwartz, “Server Side Google Tag Manager To See Full IP Addresses,” 4 November 2026 · seroundtable.com/google-tag-manager-full-ip-address-42453.html (the quoted text is Google’s own email to advertisers as reproduced by Search Engine Roundtable from a copy posted publicly by Govind Singh; Google’s own help documentation could not be read from this session, so the email text is what was verified, and Google has published no figure for how many websites this reaches)
05 The platform
Google will switch Search off for K-12 student accounts
A whole population of users will lose Google Search by default, and only a school administrator can turn it back on.
Google has named no date for this, and the announcement is about who holds the controls rather than any measured effect on anybody’s visits.
Chris Phillips, who runs education at Google, published a post on 3 November about AI in schools. The sentence that matters is about defaults: “Access to these products will be entirely off for K-12 student accounts until school administrators explicitly turn them on.” The products named are Gemini, Gemini Notebook, Gemini in Classroom and Google Search.
Google Search being on that list is the part the search industry noticed. A K-12 student account will reach Google Search only if an administrator switches it on, and Google says it will “share more specifics with schools in the coming months ahead of the changes rolling out” rather than naming a date.
Google posted a clarification the same day, saying the announcement gives schools more choice rather than less and that each service can be turned on individually.
Alongside it, Google is building what it calls Guided Learning into Search for student accounts, which answers a homework question with steps instead of the answer. That starts as a pilot in selected United States schools later this month.
For an admissions consultant the consequence is concrete and easy to miss. A high-school student researching programs on a school account may have no Google Search at all, and the families who do reach you will have arrived by some other route. Nobody has published a figure for how many student accounts this covers.
What to do about it
If students or their families are the people who find you, do not file this as a story about schools.
Ask the two or three schools or counselors who refer work to you what their district has decided, and whether their students can reach Google Search on a school account at all. The answer varies by district, and Google has not said when it changes.
Source Google, Chris Phillips, “Designing responsible AI for learning,” 3 November 2026 · blog.google/products-and-platforms/products/education/ai-learning-era-updates-2026 · reported by Search Engine Roundtable, Barry Schwartz, 3 November 2026 · seroundtable.com/google-search-off-k-12-schools-42447.html (Google names no date for the controls change and publishes no figure for the number of accounts affected; the Guided Learning pilot in Search is stated as select United States schools later this month)
