01 The research
ChatGPT, Gemini and Grok disagreed on the best brand
If three assistants name three different winners, one flattering answer about your business proves nothing about the next one.
The brands in this test were invented, and the finding that should concern anyone paying for this work is what happened when the researchers merely reshuffled the links.
Six researchers led by Runze Kong published a test on 31 October of generative engine optimization, which is the practice of rewriting your web pages so that AI assistants recommend you. Rather than hand the text straight to a model, they published the rewritten pages as live websites, then asked three assistants to pick the best of 80 invented brands across 18 product families. The three engines chose the same top product for only 48.8% to 66.3% of brands.
Then they changed one thing that has nothing to do with the pages. Reversing the order in which the candidate links were presented moved one method’s first-place rate on ChatGPT from 73.8% to 93.8%. A twenty-point swing came from the running order alone, and the randomized order the authors used landed between the two at 82.5%.
One of the five methods they tested deserves naming. It adds statistics to a product description, and the authors’ numerical check flagged all 80 of its rewritten descriptions, out of 86 flagged among 400 in total. A method that invents figures about your business is an advertising problem in every industry, whether the rule that binds you comes from the AICPA code, your state licensing board or the Federal Trade Commission.
This is the work Forever Cited does, so the finding sits in our own lap. A number that moves when somebody reshuffles the links is not a measurement, and anyone selling you one should say so first.
The limits are the authors’ own and they are substantial. The 80 brands are fictional and the items are products rather than professional services, the engines were specific versions tested at frozen prices, and this is a preprint rather than reviewed work.
What to do about it
Ask whoever handles your marketing one question: when an assistant is asked for a business like yours, which names come back, and does the answer change if you ask again an hour later? An enterprise software company being pitched on this kind of work should also ask what any quoted win rate was measured against, and whether the order of the candidates was held still while it was measured.
Write down both answers. The point of this paper is that a single reading is not evidence of anything.
Source arXiv, Runze Kong, Haibo Jin, Xucheng Yu, Xiaoqi Wang, Huimin Zeng and Haohan Wang, “Do GEO Gains Survive the Live Web? Evaluating In-the-Wild Visibility Gains of Generative Engine Optimization via LIVE-GEO,” 31 October 2026 · arxiv.org/abs/2611.01126 (a preprint; the 80 brands are fictional and the items are products rather than professional services, the engines tested were ChatGPT with GPT-6, Gemini 3.6 Thinking and Grok in Expert mode at frozen prices, and the numerical check is the authors’ own verifier, which agreed with reference labels on 90.7% of 194 cases)
02 The brand result
Google moved brand AI Overviews into the right column
When somebody searches your business by name, Google may now answer in its own words where your panel used to be.
Nobody at Google has confirmed this, and the three people who spotted it are describing a layout change rather than a measured effect on anyone’s visits.
Google has moved the AI Overview for many brand-related searches into the right-hand column of the results page, where it previously sat at the top or the middle of the main results. On the examples published, the AI-written summary now occupies the slot the brand knowledge panel used to hold. Chris Long, who posted the change on 2 November, read it as Google showing “AI Overviews for brand terms in the right column” instead of a knowledge panel.
Two other people saw it first. David McSweeney reported the same layout on 31 October, and Sachin Patel posted on 27 October that the knowledge panel had begun carrying an AI-generated label. Three independent sightings across six days is a pattern rather than a glitch, though it is still an observation.
The distinction worth holding is what each box is. A knowledge panel was assembled from identifiable sources about you, your own website among them. An AI Overview is written prose, and the business it describes does not write it.
Google has announced nothing. The scope word in the reporting is “many” rather than all, and no figure has been published for how this changes visits to the business being searched. Barry Schwartz, who reported it, added his own view that the right column suits the box better but that he would rather a business described itself than have software do it.
What to do about it
Search your own business by name on a desktop browser you are not signed into, and read the right-hand column. An advisory practice that handles mergers or franchise work should check that the description matches what it actually does, and note which of its own pages, if any, are cited underneath.
If the summary is wrong, the repair is on your own pages rather than in the box. The summary is assembled from whatever the engine can read about you.
Source Search Engine Roundtable, Barry Schwartz, “Google AI Overviews For Brands Moved To Right Column,” 2 November 2026 · seroundtable.com (an observed layout change reported from screenshots by Chris Long, David McSweeney and Sachin Patel; Google has confirmed nothing, the source says “many” brand-related keywords rather than all, and both the reading that AI Overviews are replacing knowledge panels and the opinion about it are the reporter’s own)
03 The evidence
AI screeners preferred applications written by certain models
Software that reads your writing can score it on which tool produced it rather than on what it actually says.
This was a simulated audit rather than real hiring, and its sharpest number is how many qualified people were filtered out for nothing but their choice of tool.
Four researchers led by Nicolò Pagan published an audit on 30 October of what happens when software writes an application and other software screens it. They paired eleven writing models with nine screening models across ten job descriptions and fifty candidates per job. The screeners produced a consistent ranking of the writing models.
In the most adverse pairing, about two of every twenty-five qualified candidates were filtered out purely because of which model wrote their text. The ranking held across the screeners, and models running on local hardware were marked down regardless of their license.
The mechanism is the useful part. Screeners rewarded or penalized particular habits of style, and they often did so unanimously, so the penalty followed the writing rather than any hidden signature of the tool. Screeners also favored text produced by their own model.
The limits matter more than usual here. Every writer and every screener in this study was a model, and the candidates were generated rather than real people, so it measures a mechanism and not any real employer’s decisions. The authors offer it as an auditing method to be run again as models change.
What to do about it
If software screens anything you receive, or drafts anything you send, keep a named person in the loop and record that they reviewed it. An admissions consultant should be careful with this one: the study looked at job applications rather than college ones, and nobody may promise an admission on the strength of a finding like this in any case.
The practical version is short. Do not hand a final decision, or a final draft, to a model on its own.
Source arXiv, Nicolò Pagan, Michael Andreas Ebling, Stefania Ionescu and Anikó Hannák, “AI evaluators favor job applications written by certain AI models, independent of candidate quality,” 30 October 2026 · arxiv.org/abs/2610.41866 (a preprint; eleven writer models, nine screening models, ten job descriptions and fifty candidates per job, all of them generated rather than real applicants, and the figure of two in twenty-five is the single most adverse pairing rather than an average)
04 The money
Google says its automated campaigns may filter brand searches
Searches for your own business name may be routed away from the campaign you assumed was catching them.
This is a quiet edit to a Google help document rather than an announcement, and it changes where the budget for your own name needs to sit.
Google added a paragraph to the help document describing how Performance Max, its most automated ad product, interacts with the other campaigns in the same account. The new text says those campaigns “may also filter out navigational or brand-specific search queries to prioritize serving them through dedicated Search campaigns.”
The reason Google gives is control. Its own wording is that this “ensures brand-related queries are handled by campaigns offering more precise control”, and the document tells advertisers to make sure a dedicated brand campaign has enough budget and eligible ads to catch that traffic.
The change was spotted by Hana Kobzová, who noted the paragraph is absent from the archived version of the page. Google published no announcement, and the document carries no date, no share of traffic and no account-level detail, so there is no way to tell from it how much moves, or when.
The consequence is narrow and real. If an automated campaign was quietly covering searches for your own business name and it stops, that traffic does not disappear: it goes to whichever campaign is eligible to serve it, or to nobody.
What to do about it
Open your ad account and check two things: whether a campaign dedicated to your own business name exists, and whether it still has budget at the end of a normal day. An architecture or engineering practice that pays for its own name should also confirm the ads in that campaign are approved and running, because Google’s own note makes eligibility the deciding factor.
Then apply your own advertising rules to it exactly as you would to any other ad. A state licensing board does not care which campaign type served the click.
Source Search Engine Roundtable, Barry Schwartz, “Google Ads PMax Campaigns May Filter Out Navigational & Brand Queries,” 3 November 2026 · seroundtable.com (the quoted sentences are from Google’s own updated help document, which the article reproduces; the change was spotted by Hana Kobzová and posted on PPC News Feed, Google made no announcement, and the document states no date, no share of traffic and no rollout detail)
05 The self-preference
Google AI Mode linked only Google Flights in one test
When the engine sells the thing a customer is asking about, every link in the answer can belong to the engine.
This is one reporter’s weekend of searches rather than a study, and it is the most concrete example yet of an engine answering with its own product.
Barry Schwartz spent a weekend asking Google’s AI Mode for flight information. He reports that 100% of the time, the only links inside the answers were to Google Flights, across several destinations both inside the United States and international.
Natzir reported the same thing on X on 30 October, writing “Ask AI Mode for flights, get Google Flights.” Schwartz says he holds dozens of screenshots. He states no count of searches, so that 100% is a complete run of an informal test rather than a measured rate.
The shape of this matters more than the example. A results page lists other people’s websites, while an AI answer lists whatever the engine chooses to link, and here the engine had a competing product of its own in the same category. Nothing in the answer tells a customer that every link belongs to one company.
Schwartz adds his own aside that European regulators will take an interest in this. That is his speculation rather than a filing, and no investigation is being reported here. What is reported is a pattern two people observed in the same week.
What to do about it
Ask an assistant the question your customers actually ask, then look at where the links point rather than at what the text says. A private charter operator can run this one today, because flights are the exact category in the example.
If every link in the answer belongs to the platform, the useful work is being the business named in the sentences. That is a different job from ranking tenth on a results page.
Source Search Engine Roundtable, Barry Schwartz, “Google AI Mode Biased: It Only Shows Links To Google Flights,” 3 November 2026 · seroundtable.com (the reporter’s own informal test across one weekend; he states “100% of the time” and “dozens” of screenshots but no count of searches, the same pattern was posted by Natzir on 30 October, and the reference to European regulators is the reporter’s speculation rather than any announced action)
