01 The market data
Google Ads data shows short searches collapsing, 42% to 24%
If your website and your ads only answer three-word questions, you are missing the sentences most customers now type.
The queries that convert have moved from two words to four, and copy written for the old shape answers the wrong question.
Jason Tabeling pulled Google Ads data through August 2026 and compared it against January 2025. One and two word searches fell from 42% of impressions to 24%. Three and four word searches rose from 33% to 48%.
The conversion numbers moved further. One and two word queries produced 62% of conversions in January 2025 and 52% in August 2026. Three and four word queries went from 20% to 46%. Five and six word queries tripled, from 3% to 9%, and queries of seven words or more quadrupled from 1% to 4%.
Take the caveat seriously. Search Engine Land does not say how many accounts this covers or which industries they sit in, so this is one practitioner’s aggregate rather than a published study. The direction matches every other measurement of AI-era search behavior, and the size of the move is not independently verifiable.
What it describes is people talking to search the way they talk to a person. An emergency veterinary hospital used to be found by someone typing “emergency vet”. Now it is found by someone typing what their dog swallowed and when. Those two searches want completely different pages.
What to do about it
Open the search terms report in Google Ads, or the queries report in Search Console, and sort by length. Read the longest twenty. Those are the sentences your customers are actually typing, and most of them will be questions no page on your site answers directly. Write one page for the three most common, in the words the searcher used.
Source
Search Engine Land, “Google Ads data shows query length shift post-AI Mode,” by Jason Tabeling, September 11, 2026 · searchengineland.com
02 The evidence
Six AI engines kept naming new businesses after fifteen asks
A one-time check of whether AI mentions your business is not evidence of anything, in either direction.
The same fifty buying questions, put to six engines fifteen times each, produced a list that kept growing.
Dmitrij Żatuchin published a preprint on September 4 that does the slow, expensive version of a question this industry usually answers with one screenshot. Fifty buying questions, six AI engines, fifteen runs each, 300 question-engine cells and 1,470 organizations adjudicated by hand. The question was whether a large language model, the technology behind ChatGPT and Gemini, eventually runs out of businesses to recommend.
Mostly it does not. The five engines answering without a web search were still adding businesses nobody had seen in earlier runs at the fifteenth ask, in 86% to 92% of cells, with typical lists of 15 to 31 organizations. The one engine that searched the web closed its list far faster, with a median of eight organizations.
The line that matters to anyone who has ever checked is this. A single run shows 62% to 77% of the businesses that five runs would show. And across engines the median question draws 38 organizations, of which about 15 appear on exactly one engine. So being named by ChatGPT and missing from Gemini is the normal result, not a diagnosis.
The caveats are real. This is a preprint rather than peer-reviewed work, it covers buying questions in general rather than any one industry, and at ten days old it is not this week’s news. It is here because it is the most rigorous thing published this month on whether an AI visibility check means anything. An AI visibility number is what every agency in this field now sells, Forever Cited included, and this paper says the single-look version of that number is worth 62% to 77% of itself.
What to do about it
Pick the three questions a customer would actually type to find a business like yours. Ask each one on ChatGPT, Gemini and Perplexity, three times each, and write down every business named. That is about nine minutes per question, and it is the difference between a measurement and a screenshot. Repeat it monthly, and treat any single result as partial.
Source
arXiv preprint 2609.05059, “Repeated Queries Exhaust an LLM’s Brand Recommendations but Not Its Sources,” by Dmitrij Żatuchin, submitted September 4, 2026 · arxiv.org
03 The platform change
Google’s assistant is calling businesses and logging the answers
Whoever answers your phone is now editing what Google knows about you, and nobody on your staff was told.
The automated call your front desk took last month may already be part of how Google matches you to customers.
Google Business Profiles gained a tab called Collected Info this weekend, which Barry Schwartz found under Edit profile. It lists information Google gathered from phone calls, text messages and WhatsApp conversations with your verified business number, with the source, the date, and the language it was given in.
Google’s own wording is plain. “This info was collected in phone or chat conversations with your business. It’s used to match your business with customers looking for services like yours.” So an automated call that a receptionist or an answering service handled has been feeding the record that decides when you appear.
You can delete entries. Google’s documentation says “Deleting collected information removes it from Google’s records but does not change other details already on your Business Profile,” which means a wrong detail that already reached the public side of your profile has to be fixed separately. The tab is available only in selected regions, languages and business categories, so it may not have arrived yet.
For a surgical practice, an emergency veterinary hospital or an insurance brokerage, this is a staffing question rather than a marketing one. The person who takes calls at four in the afternoon now has editorial control over your Google record, and has never been briefed on it. We wrote on September 13 about a stranger being able to propose a new name for your listing, which at least comes with a four-day warning email. This one is Google itself, and it arrives with no warning at all.
What to do about it
Open your Google Business Profile, choose Edit profile, and look for Collected info. Read every line. Delete anything wrong, and check your public profile separately for the same error, because deleting one does not fix the other. Then tell whoever answers your phone that an automated caller from Google is not a nuisance call, and that what they say goes on the record.
Source
Search Engine Land, “Google Business Profiles rolls out Collected Info,” by Barry Schwartz, September 14, 2026 · searchengineland.com
04 The one to check today
Microsoft now requires AI-made ads to keep their watermarks
Your agency’s AI-generated images are your legal problem, in every market where the ad actually runs.
Microsoft has put responsibility for synthetic ad creative on the advertiser, not on whoever produced the file.
Microsoft Advertising published rules on Monday for ads built with AI-generated or AI-altered images and video. Advertisers have to preserve the watermarks, the metadata and the provenance information that identify content as AI-made rather than stripping them out, which many creative tools do by default on export.
Disclosure has to be visible and adjacent. Microsoft asks advertisers to make it clear and place it close to the relevant content, and suggests embedding the disclosure into the image or video asset itself rather than relying on a caption. Ads can be rejected outright for prohibited deepfakes, for impersonating a person or an organization, or for using someone’s likeness or voice without authorization.
The sentence to read twice is about who carries the risk. “Advertisers are responsible for complying with applicable laws wherever their campaigns run, including requirements around disclosures, permissions and the use of a person’s likeness or voice.” Microsoft did not publish an enforcement date, so treat the rules as live.
A luxury home builder running an AI-rendered interior that was never built, or a medical practice using a synthetic patient image, is now the responsible party twice over: once under Microsoft’s rules, and once under the advertising code of whatever board licenses them. Your state contractor licensing rules and your state medical board’s advertising rules did not change this week. What changed is that the platform now expects you to have thought about them already.
What to do about it
Ask whoever makes your ad creative one question in writing: is any of this generated or altered by AI, and are the watermarks and metadata still intact in the files sent to Microsoft. If the answer is yes to the first and no to the second, those ads are out of policy now. Keep the reply on file, because you are the party the rules name.
Source
Search Engine Land, “Microsoft Advertising sets rules for AI-generated ads,” by Anu Adegbola, September 14, 2026 · searchengineland.com
05 The money
Google started paying publishers whose pages feed AI answers
Being the source of an answer now carries a price, and for almost every business that price is still nothing.
A pilot called AI Contribution pays a short invited list when their content shapes what Gemini and AI Overviews say.
Google confirmed on Monday that it is testing payments to publishers through Google Search Console. The program is called AI Contribution, and it pays when a publisher’s content “contributes significantly to the creation of AI-generated responses” in Gemini, AI Overviews and AI Mode.
This is not a licensing deal with a fixed fee. Payment is tied to the value Google assigns to a piece of content each time it shapes an answer, tracked in a reporting screen inside Search Console, and publishers can leave whenever they want. Google described it to Digiday as an early-stage learning pilot to test how best to reward high-quality content.
The invitation list is the story. At least dozens of publishers have been approached, according to the reporting that Search Engine Land and Search Engine Roundtable both cite, and interest has been stronger among small and mid-sized publishers than large ones. Some participants have said the payouts so far are small next to advertising revenue.
Read it as a direction rather than an offer. For the first time, Google is attaching money to being the source of an answer instead of the destination of a click. An architecture practice or an admissions consultant publishing genuinely useful explainers is producing exactly the kind of content this pilot was built to pay for, and is not on the list. The price of being cited is being set right now, by the party that does the citing.
What to do about it
You are almost certainly not in this pilot and cannot apply for it. The useful move is to find out what you are already giving away. In Search Console, look for the pages that collect impressions and almost no clicks. Those are the pages being read and not visited, and they are the asset Google has just started putting a price on.
Source
Search Engine Land, “Google tests paying publishers for using its content in AI Mode, AI Overviews and Gemini,” by Barry Schwartz, September 14, 2026 · searchengineland.com · Search Engine Roundtable, “Google AI Contribution Pilot,” by Barry Schwartz, September 14, 2026 · seroundtable.com
