01 The platform
Google Search Live now speaks answers in 97 languages
When a customer asks out loud and hears one answer back, your website never gets the chance to make its own case.
The upgrade landed on September 15, and the spoken answer arrives before any link does.
Google published a new audio model called Gemini 3.8 Live on September 15 and put it behind Search Live, the voice mode inside the Google app, the same day. The model automatically detects and transitions between 97 supported languages mid-conversation. Google reports it scored 97.7% on a benchmark called Big Bench Audio and took first place on Artificial Analysis’ Speech to Speech Quality Index, at 82.6.
Search Live works the way it sounds. You tap a Live icon, ask a question in your own voice, and hear a spoken answer you can interrupt and follow up on. Web links do appear on screen, but they arrive alongside an answer the person has already heard.
That ordering is the whole story for a business. A spoken answer has room for two or three names and no room for a page of options. An admissions consultant whose site explains their process beautifully is no longer competing on the quality of that page. They are competing to be one of the names said out loud.
Hold the numbers loosely. These are Google’s own benchmark results, published by Google, on tests Google selected. The capability is real and the direction is not in doubt. The precision belongs to the vendor.
What to do about it
Open the Google app, tap the Live icon, and ask the question your best customer would ask before they knew your name. Listen to who gets named and whether you are among them. It takes two minutes and it is the only version of this test that reflects what your customers actually hear.
Source
Google, “Introducing Gemini 3.8 Live and 3.8 Live Extended Thinking,” by Tom Ouyang and Malini Jaganathan, September 15, 2026 · blog.google · reported by Search Engine Land, Barry Schwartz, September 15, 2026 · searchengineland.com
02 The evidence
Researchers found reformatting wins mentions, not entries
Restructuring your page can raise how often an answer names you, but it does not reliably get you into answers that skip you.
Two researchers replayed 113 real searches to separate what formatting actually causes from what merely travels alongside it.
Sriram Selvam and Anneswa Ghosh published an audit on September 14 that asks a narrow question with a large consequence. When several pages all support the same fact, an AI assistant cites some of them and ignores the others. The pair took 129 recordings of real back-and-forth searches, found 113 pairs of pages that independently supported the same fact, and had human reviewers confirm 103 of those pairs.
Then they rewrote one page in each pair two ways, once as structured content with clear headings and lists, once as ordinary prose, and replayed the identical search. Structured rendering raised how many times a page was cited inside an answer, by half a citation per answer. Whether the page got cited at all rose 4.5 percentage points, and that second result was inconclusive by the authors’ own measure. Total citations did not rise and competitors did not lose credit.
In plain terms, formatting moved credit around inside answers the page was already part of. It did not reliably buy a way in.
The second finding deserves more attention than it will get. Out in the wild, pages sitting first were cited 42.3 percentage points more often than pages sitting fifth. When the researchers actually moved pages up and down themselves, the effect was 7.9 points, and zero in the portion of data they held back as a check. Most of that apparent advantage is not the position doing the work.
They also measured their own error rate and published it, which is rare. Re-running the same test flipped 15% of the yes-or-no citation decisions, and roughly 45% of the variation came from the model’s own randomness rather than anything about the page. A private aviation charter operator shown a flattering before-and-after from a single run is being shown something that would partly reshuffle if it were run again. This is uncomfortable for everyone selling AI visibility, Forever Cited included.
What to do about it
Before you pay for AI visibility work, ask for the before-and-after to be run more than once on different days, and ask to see both runs. A result that holds across repeats is worth acting on. One that moves is the measurement talking, not your business.
Source
arXiv, “CITECHOICE: A Causal Audit of How Document Presentation Redistributes Citation Credit in Agentic Search,” by Sriram Selvam and Anneswa Ghosh, September 14, 2026 · arxiv.org
03 The evidence, again
Retested on ten current engines, the 2023 levers moved nothing
A good deal of the AI visibility advice being sold today rests on measurements taken from engines that have since been replaced.
The only published cause-and-effect results in this field date from 2023, and re-running them found no effect left.
Elisha Bajemon and Andre-Louis Rochet published a paper on September 7 about how to check whether a cheap scoring tool actually predicts anything. They tested their method on the business of getting cited by AI engines, and one check failed in a way worth reading.
They re-measured the only published cause-and-effect results in this field, from 2023, against ten families of current AI engines. The levers those 2023 results identified moved citation on none of them. The authors’ own word for the older results is expired.
Follow that one step further. A large share of what is sold as AI visibility technique traces back to that 2023 work. When the engines changed, the evidence was not re-run. It was repeated.
They found something else worth knowing. A score calculated without knowing what the customer asked barely tracks whether you actually get cited: the relationship they measured was 0.11 on a scale where 1.0 is a perfect match, which is close to none at all. Their reading is that such scores work as content quality filters rather than as predictions of citation.
The paper is unusually honest about itself. The authors disclose a bug in their own first ranking evaluation and a failed confidence check, and correct both in the open. An IT and cybersecurity firm paying a monthly retainer for AI optimization is entitled to ask which of those tactics has been tested on an engine that exists now. This paper is nine days old rather than this week’s news, so read it as context rather than a fresh event.
What to do about it
Ask whoever handles your marketing a single question: which of the things we are doing has been tested on an AI engine that exists today, and when was that test run? A confident answer with a date is a good sign. A confident answer without one is the pattern this paper describes.
Source
arXiv, “Scoring Without the Engine: Validating a Deterministic, Manipulation-Resistant Content Score for Generative Engine Optimization,” by Elisha Bajemon and Andre-Louis Rochet, September 7, 2026 · arxiv.org
04 The money
Google finishes moving two ad settings to AI this month
If you run Google Ads, a setting you chose is being replaced automatically before October, and eligible accounts cannot opt out.
Two older settings are replaced automatically before October, and a third change Google once scheduled sooner has moved to 2027.
Google is auto-upgrading two older Search advertising settings, automatically created assets and campaign-level broad match, into a newer product called AI Max. The upgrades began in September and Google expects all eligible campaigns to be finished by the end of the month. Eligible advertisers cannot opt out, though they can move early by choice.
Google’s own figure for AI Max is 7% more conversions or conversion value at a similar cost per conversion, when the full feature set is switched on. That is Google measuring its own product, which is worth holding in mind when you set expectations for the quarter.
There is a third change that is easy to confuse with these two, and reading Google’s own post settles it. Google says the Dynamic Search Ads sunset and auto-upgrade was extended and will begin in February 2027, not this month. What ends in September is the ability to create new ones.
The distinction matters to anyone budgeting. An accounting practice running a Dynamic Search Ads campaign through to filing season does not have to rebuild it in the next two weeks. The two settings being auto-upgraded now are the other ones. Google published this post in April and updated it in June, and the September deadline it sets falls inside this month.
What to do about it
Open Google Ads, look for an upgrade notice on your Search campaigns, and write down your current cost per enquiry before the end of September. Whatever AI Max does next, you will only be able to tell if you recorded where you started. Five minutes now is the difference between a comparison and a guess.
Source
Google, “Google’s Dynamic Search Ads are upgrading to AI Max,” by Brandon Ervin, Director of Product Management, Google Ads, published April 15, 2026 and updated June 11, 2026 · blog.google
05 The rules
OpenAI stopped rival image and audio tools advertising in ChatGPT
The company writing the answer also sells the advertising beside it and competes in some of the same categories.
Adobe ran in the first ChatGPT advertising pilot, and a named Adobe executive confirmed its campaigns would no longer be approved.
OpenAI has updated its advertising policies to prevent campaigns promoting standalone image and audio generation products from running in ChatGPT. Video generation tools can still advertise. The change was first reported by The Information and picked up by Search Engine Land on September 10.
Adobe is among the advertisers affected, having taken part in OpenAI’s first advertising pilot for products including Acrobat Studio and its Firefly image generator. Doug Wyatt, Adobe’s senior director for Americas media, said OpenAI informed advertising partners that campaigns for standalone image and voice generation products would no longer be approved. OpenAI has not made a public statement in the reporting.
The structural point matters more than the specific products. The same company writes the answer, sells the advertisement beside it, and sells competing products in some of those categories. Where those roles meet, it decides.
For most readers this is not a problem to solve this week. It is a reason to be careful about how much of your visibility you rent from one place. A luxury home builder whose enquiries arrive through a single platform is exposed to a policy page they do not control and are not consulted on. The reporting here is solid but single-sourced, and OpenAI has not confirmed it directly.
What to do about it
Write down where your last twenty enquiries actually came from. If more than half trace back to one platform, that concentration is worth a conversation, whatever the platform is currently doing. The point is not to leave it. The point is to know the number.
Source
Search Engine Land, “OpenAI reportedly blocks rival AI tools from advertising in ChatGPT,” by Anu Adegbola, September 10, 2026, reporting The Information · searchengineland.com
