MPWhat The Shift Is
AI search is a new category, and it is moving faster than anything search has done in twenty years. The platforms change how they choose sources without announcing it. Google ships updates that reach AI Overviews and blue links at the same time. New measurement tools appear monthly, and the research about them contradicts the research from the month before. A business owner cannot follow all of that, and should not have to.
So every Monday morning we read what changed across Google AI Overviews, ChatGPT, Perplexity, Copilot and the studies around them, and we write down the handful of things that matter to a business trying to get found, understood and recommended. Plain English first. The detailed brief underneath for the marketing people. If you have ever wished somebody would just tell you what changed and what to do about it, without a webinar, this is that.
Two labels do the honest work in every issue. Confirmed means a platform documented it or the data is public. Our read means it is our interpretation, which is worth exactly what you paid for it, and we say so.
In Plain English: What Changed and What To Do
Google started a spam update on September 24
Google announced a global spam update on September 24. It covers every language, and Google says the rollout can take up to two weeks. Google did not name a specific AI content tactic as the target.
Google's AI Overviews and AI Mode draw from the same pages regular search does, so a spam update reaches AI answers too. If your visibility moves this week, up or down, it might be the update settling and not anything you or your agency did. Taking credit or taking blame right now is guessing with a straight face.
Write September 24 on your reporting calendar. Keep a copy of your numbers from before that date. If traffic drops a lot, look into it, but do not conclude you were penalized. A dip during an update is weather, not a verdict.
Some AI visibility scores measure a stand-in, not the real thing
A research firm called Foglift updated its citation studies on September 21. Buried in the methodology is an honest note: its Google AI Overview numbers come from a Gemini API set up to imitate AI Overviews, not from the Google search page a person sees. The observations were collected in early August.
A visibility score is only as good as what it looked at. If a tool checks an imitation of Google's AI, it can hand you a polished number that describes an experience no customer has. Nobody has ever been recommended by a dashboard.
Before you trust any AI visibility tool, or any agency quoting one, ask five questions: which product did you check (the real app or an API), which questions did you ask, from where, on what dates, and how many times. Then spot-check a few questions yourself in the real ChatGPT and the real Google.
AI can give the same answer and credit completely different sources
A new academic paper, accepted at a workshop at EMNLP 2026, compared AI-generated search overviews on Google and Baidu in English and Chinese. The answers were often close to identical in meaning while the sources they credited were very different.
There are three separate wins in an AI answer: it describes your business accurately, it recommends you, and it links to you. You can get one without the others. An accurate description with somebody else's link is a nice compliment and zero customers.
Score those three things separately when you check your visibility. If you have changed your name or rebranded, also check that AI connects the old name to the new one.
Perplexity's search got much faster
On September 24, Perplexity announced Fast Search for its Search API, built on a new retrieval and ranking service it calls Photon. Perplexity says 95 percent of results come back in 230 milliseconds or less.
That is a speed claim from the vendor, not evidence that Perplexity cites businesses any differently. What it does mean: web search is getting cheap and quick enough to sit inside other software, including the tools people use to research purchases without ever opening a browser. Fast wrong answers are still wrong, so make sure yours are right.
Nothing to change in your content this week. Worth watching if you build software or automate research for your own business.
Your To-Do List This Week
- Mark the Google rollout.Note September 24 in every report you run. Save the before numbers. Hold off on confident cause-and-effect stories until the update finishes.
- Audit your scorekeeper.For any AI visibility score you use or get quoted, find out exactly what it measured and validate a small sample against the real thing.
- Build a question panel.Write down the questions your buyers ask before they call, and track, separately, whether AI describes you correctly, recommends you, links to you, and whether that turns into inquiries.
The Full Brief
The detailed version, for marketing directors and anyone who has to defend a number in a meeting. Every item carries its source. Where we are interpreting, it says so.
Google's September 2026 spam update is underway
Google began a global spam update on September 24, 2026 at 9:15 AM Pacific. It applies to all languages. Google says the rollout may take up to two weeks. The announcement does not identify a particular AI content tactic as its target. Google Search Status Dashboard
Google's own documentation says AI Overviews and AI Mode depend on foundational search eligibility, meaning the same crawling, indexing and quality rules that govern ordinary results. Google Search Central: AI features
Changes during the rollout window will muddy any claim that a recent content or AEO change caused a gain or a loss. A traffic decline on its own does not establish a spam violation, and a gain on its own does not prove the work paid off.
Annotate September 24 in client reporting. Preserve pre-rollout baselines. Investigate material losses. Defer confident attribution until the update is marked complete.
New citation research exposes a dashboard problem
Foglift updated its Q3 2026 citation studies on September 21. They report large differences in which sources each AI engine cites, and a Q2-to-Q3 citation turnover of roughly 77 to 87 percent depending on the engine. Its Google AI Overview measurements use what it calls a grounded Gemini API call shape rather than the consumer Google Search page, and the underlying observations were collected August 1 through 10. Foglift AI Search Industry Reports
Two lessons. First, a visibility tool can hand you a clean score that describes a proxy experience; treating that score as real customer exposure overstates what was measured. Second, if the set of cited sources turns over by three quarters in a single quarter, a one-time screenshot of an AI answer is not a market position. It is a photograph of a moving car.
Before adopting a tool or quoting its findings, require disclosure of the exact product surface (consumer app or API), the prompts, the geography, the collection dates, and the repeat sampling. Validate a small sample against the experience a prospect actually sees.
Similar answers can give visibility to different sources
An academic study submitted September 21 and accepted at a workshop at EMNLP 2026 audited AI-generated overviews on Google and Baidu, using English queries and their Chinese translations. It found the answers were semantically similar across platform and language settings (cosine similarities between 0.70 and 0.81) while the source domains given visibility differed substantially. arXiv: Auditing Source Exposure in Baidu and Google AI Search
The study measures source exposure, not US buyer behavior or conversions, so do not stretch it further than that. What it does establish is that accuracy, recommendation and credit are separate outcomes. The strategic question for any business is whether AI describes its positioning correctly, recommends it, and gives it a visible link. Those need separate scores.
Score factual accuracy, brand recommendation and citation separately. For a client that has recently rebranded, test whether answers correctly connect the former and current identities, then record whether they recommend and link to the current business.
Perplexity's faster retrieval deserves watching
On September 24, Perplexity announced Fast Search in its Search API, powered by a new Rust-based retrieval and ranking service called Photon. Perplexity claims 95 percent of results return in 230 milliseconds or less. Perplexity API Platform Forum
That is a vendor performance claim, not evidence of improved brand visibility or a changed citation policy. Faster, cheaper retrieval makes live web search more practical inside automated buying and research workflows, which is where a growing share of decisions will get made without a person ever seeing a results page.
Worth a controlled test if you are building such workflows. It does not yet justify changing anyone's content strategy.
Questions People Ask
What is the Google September 2026 spam update?
A global update to how Google detects and demotes spam, announced on September 24, 2026. It covers all languages and Google says the rollout can take up to two weeks. Google did not name a specific tactic it targets. Because Google's AI Overviews and AI Mode draw from the same index and quality rules as regular search, the update can affect AI answers as well as blue links.
My traffic dropped this week. Was I hit by the spam update?
Not necessarily. Rankings and AI citations move around while an update rolls out, and some of that movement reverses when it finishes. Note the date, keep your pre-update numbers, investigate large losses, and wait until Google marks the update complete before drawing a conclusion. A dip alone does not mean a spam violation.
Are AI visibility tools accurate?
Some are careful and some are not, and the difference is in the methodology. Ask which product surface the tool checked (the real consumer app or an API that imitates it), which questions it asked, from where, on which dates, and how many times. A score built on an imitation of Google's AI does not describe what your customers see. Spot-check a few questions yourself in the real apps.
What is the difference between an AI mention, a recommendation and a citation?
A mention is your business appearing in the answer. A recommendation is the AI suggesting you as the choice. A citation is a visible link to your website as a source. New research shows AI can give nearly the same answer while crediting different sources, so you can be described accurately and still get no link and no customer. Track all three separately.
Does Perplexity's Fast Search change how I get recommended?
Not today. It is a speed improvement to Perplexity's search API, aimed at developers who build search into their own software. The signals that get a business cited have not changed. It is worth watching because faster search inside other apps means more buying decisions happen without anyone opening a browser.
What is The Shift?
The Shift is Agency AEO's weekly brief on AI search news. Every Monday morning we review what changed across Google AI Overviews, ChatGPT, Perplexity, Copilot and the research around them, explain in plain English what it means for a business, and separate what is confirmed from what is our read. It exists because AI search is a new category that changes weekly.