01 The evidence
ChatGPT named 13% of the pages behind its answers
Your website can shape an AI answer without ever appearing in it, which means the customer never learns your name.
The study counted citations in 171,264 conversations people actually had, not in a laboratory test set.
Researchers at the Max Planck Institute for Software Systems published a study on September 16 that followed what four AI assistants actually do when they search the web. The dataset covers 171,264 conversations from 613 people, recorded as they happened rather than staged for a test. It tracked every step, from the question typed to the searches the assistant ran, to the pages it pulled up, to the pages it finally named.
The gap between reading and naming is the finding. ChatGPT’s searches returned about 43 pages for every question asked, and it named 13% of them.
Claude named 20% and Grok named under 2%. Across the four assistants the range ran from 1.7% to 34%.
Those uncredited pages are not idle. Between 14% and 53% of the claims in the answers traced back to pages the assistant had read but never cited. For an architecture practice or an independent insurance brokerage, that is the difference between informing the answer and being recommended in it.
One number cuts the other way and is worth keeping. Invented citations were rare, at 1% to 2%, so the links these assistants do show are usually real pages. The problem is not fabrication, it is omission.
What to do about it
Ask the assistant your customers use a question you would want to win, then read the list of sources it names rather than the answer. Open two of those sources and check whether your business appears anywhere on them. When your own site is not the page being cited, being named on the pages that are cited is the practical route in.
Source
arXiv, Mahsa Amani and eleven colleagues, Max Planck Institute for Software Systems, “Characterizing Web Search by Conversational LLM Agents: From Search Decisions and Strategies to Results and Responses,” September 16, 2026 · arxiv.org
02 The evidence
OpenAI, Google and four rivals lost accuracy on buried facts
An AI reading your long documents is likeliest to be wrong exactly where the detail is hardest to find.
The same questions were asked twice, once with the supporting fact easy to reach and once with it buried deeper in the file.
A second paper, published on September 14 by Luis M. Sanchez, tested a narrower thing carefully. It ran six flagship models from OpenAI, Anthropic, Google, Meta, Alibaba and Zhipu over the same 41 questions, first with the supporting fact easy to reach and then with it buried deeper in the document.
Accuracy fell in every case. Google’s Gemini went from 85.2% correct to 67.5%, and Anthropic’s flagship went from 89.4% to 77.9%. The cost of getting a right answer rose as much as sevenfold, because the models made far more tool calls hunting for the buried fact.
The confidence is the part worth carrying into a meeting. Of 167 wrong answers across the six models, 120 were stated with a confidence of 80 or higher out of 100. Nothing in the answer itself signals which of the two conditions produced it.
This matters most where the work is long documents. An accounting practice reading a disclosure schedule, or an adviser running due diligence on a business sale, is asking a machine to find the one clause that was written not to stand out. The study is small at 41 questions and its author says so plainly, which is the right way to publish a result like this.
What to do about it
Before you rely on an AI summary of a long document, pick two facts you already know are inside it and check that the summary got both right. If the facts you picked were easy to find, you have tested the easy case only, which is the case this study says the models already pass.
Source
arXiv, Luis M. Sanchez, “Clean Scores, Buried Evidence, and Confident Wrong: A Receipt-Based Audit of Frontier Agentic QA,” September 14, 2026 · arxiv.org
03 The money
Google now shows retailers how often AI names them
Retailers can now see how often AI names them, and no equivalent report exists for a business that sells expertise.
The report compares a brand against its competitors inside AI Mode and AI Overviews, and reaching it requires a product feed.
Google announced on September 16 that AI performance insights in Merchant Center is now live in Australia, Canada, India, New Zealand and the United States. In Google’s own words it lets a business compare “their share of voice with other brands across surfaces like AI Mode and AI Overviews.”
Share of voice means how often you are named when an AI answers a question in your category, measured against named competitors. It is the number this whole industry now sells, Forever Cited included, and Google has just begun giving a version of it away. That version needs a Merchant Center product feed, so it reaches shops and not practices.
The same post carried a figure worth noting. Google said that in testing with lululemon, product details the brand wrote itself were used in 50% of relevant AI Mode recommendations. That is one brand in one test, reported by the company selling the placement, and it should be read that way.
For a real estate brokerage, a luxury home builder or an IT and cybersecurity firm, the practical read is blunt. The surface that decides whether you get named is now measured for the people who sell products and unmeasured for the people who sell expertise.
What to do about it
If you sell products, open Merchant Center this week and look at AI performance insights. If you sell a service, the same question has to be asked by hand. Pick five questions a customer would put to an assistant, ask each one, and count how often your name comes back.
Source
Google, Ashish Gupta, VP and GM, Merchant Shopping, “Boost your holiday sales with these agentic commerce updates,” September 16, 2026 · blog.google
04 What to do this week
Google cut its Search profile threshold to 10,000 followers
If anyone at your business has 10,000 followers, Google will now give them a profile inside its search results.
The bar has fallen from 100,000 to 35,000 to 10,000 in under two months.
Google announced on September 16 that the follower requirement for a Search profile is now 10,000 across YouTube, Instagram, X or TikTok. Ibrahim Badr, a product manager on Search, wrote that more United States publishers and creators are now eligible.
The bar started at 100,000 and passed through 35,000 last month, which is a fast retreat for an eligibility rule. A Search profile is a Google-hosted page that gathers a person’s or a brand’s published work and can appear in search results under their name.
Google also now allows up to ten Search profiles to be managed from a single login, which is aimed at organizations running several brands. Article listings on the profile were given larger thumbnails and longer headlines.
The threshold is the story for readers here. An admissions consultant, a luxury home builder or an insurance broker with a working Instagram or YouTube audience may already clear 10,000 without ever having counted it as a search asset.
What to do about it
Check the follower count on your best-performing account today. If it is above 10,000 and you are in the United States, claiming the profile costs one login and puts a page you control under your own name in Google’s results.
Source
Google, Ibrahim Badr, Product Manager, Search, “3 new ways we’re improving Search profiles for publishers,” September 16, 2026 · blog.google · Search Engine Roundtable, Barry Schwartz, on the threshold history
05 Local
Google let directory sites feed local listings into search
Directory sites gained a new supported route into the local results your business is trying to win.
The change arrived as a documentation edit, with no announcement, no launch date and no figures attached.
Google updated the documentation for two of its search units on September 18. In Google’s own wording, the “aggregator unit and supplier unit now support local business queries,” and the documentation now points readers to the Local Point of Interest Feed.
An aggregator unit is a block in search results fed by a site that collects many businesses, which is a directory in plain terms. Until this change those units handled products and suppliers, not local businesses.
Google published no numbers, no launch date and no announcement post. It was a documentation edit, spotted by Barry Schwartz at Search Engine Roundtable, and it is worth reporting as that rather than dressing it up as a launch.
The direction is the point, and it matches what directories already do to you. A structured feed from a site listing hundreds of businesses now has a supported path into local results, while your own site competes as a single business. For a real estate brokerage or a specialty practice, the directory entry may once again be the thing that gets seen.
What to do about it
Search the two or three phrases a local customer would use to find you, and write down which directory sites appear above your own. Those are the entries to get right first, because Google has just given that kind of site another supported way in.
Source
Search Engine Roundtable, Barry Schwartz, “Google Supports Local Business Queries In Aggregator & Supplier Search Units,” September 18, 2026 · seroundtable.com
