The week the numbers came in

Last week Google changed three defaults and called none of them an update. This week it announced one update, and it was about spam.

That is the only event of 21–27 September that a volatility chart will show. Everything else that arrived this week was measurement — and, unusually, not from rank trackers, which are having a bad month of their own. A consumer panel, a publisher analytics network, an SEO platform's referral data, a randomised field experiment and Microsoft's internal numbers, produced in a copyright case, all published within seven days. They were counting different things in different ways. They point the same way.

The one update Google did announce

At 09:15 Pacific time on Thursday 24 September Google posted the September 2026 spam update to its Search Status Dashboard. The notice is two sentences long.

The movement came a day later. Barry Schwartz's timeline has it starting on Friday 25 September, growing on Saturday and easing on Sunday, with fourteen trackers — Semrush, Mozcast, Sistrix, AccuRanker, DataForSEO and others — all showing heat. Glenn Gabe, who reads more hit sites than most, put it plainly on Monday.

Gabe's early examples lean towards AI-generated and programmatic sites — Grokipedia, which had already fallen, fell further. A week earlier Google had also cleaned up about twelve thousand public Gemini Notebook pages that black-hat SEOs were using to push coupon codes, peptides and porn into the results: parasite SEO hosted on Google's own domain.

Two cautions before reading this update off a chart. It has more than a week still to run. And the charts themselves are compromised: on 26 September SISTRIX added a new line to the incident it has kept open since August.

Six measurements, one direction

The rest of the week belongs to people counting. What makes this batch worth reading together is that no two of them used the same method, so they cannot share the same blind spot.

Taken one at a time each has an obvious objection. A panel is a panel. Chartbeat's network skews to news, and news was always going to be hit first. BrightEdge counts its own data set. Microsoft's figures were produced by the defence in a copyright suit. Taken together, the objections stop mattering, because the thing they measure moved the same way under every method.

The experiment is the one to read twice, because it is the only one that can speak about cause. Researchers at Penn and Northeastern installed a browser extension for 1,444 people and randomly assigned them for a week to ordinary Google with AI features hidden, to Google as it is, or to forced AI Mode. Removing the AI features raised clicks to websites with no detectable loss in how people rated their searches. Forcing AI Mode did the opposite on both counts — and sent more of them to Bing, DuckDuckGo and Yahoo.

Microsoft's numbers are the other one worth dwelling on, because nobody would publish them voluntarily. In the publisher copyright litigation, internal Bing data showed what Copilot answers did to clicks on the plaintiffs' own domains: New York Times properties fell 87–93%, Daily News domains 83–91%, Ziff Davis 51–94%. The same filing puts the share of ChatGPT users who visit no external site during a search — a zero-click search in the plainest sense — at 87.78%, against 26.91% for Google.

The paradox in the middle: the share of questions is not the share of traffic

Two of this week's releases seem to contradict each other, and the contradiction is the most useful thing in either of them.

Comscore says ChatGPT is losing its grip on AI questions: its share of AI prompt volume in the US fell from 70% in January to 50% in June, while Gemini went from 17% to 30% and Claude from 2% to 11%.

BrightEdge, two days later, says ChatGPT has never been more dominant: it produced 95.1% of all AI referral traffic in its data for August, and its referrals doubled between January and August. Gemini accounted for 2.4%, Perplexity 1.1% and Claude 0.8%.

Both can be true, because they count different things: one counts questions asked, the other counts people who left the answer for a website. The two samples are not the same population, so the ratio between them is not a precise figure. But the shape is hard to miss. Gemini takes almost a third of the questions and sends a fortieth of the traffic. The assistant whose share of conversations is falling is the one that still sends people out; the ones growing fastest keep them.

That is last week's story again, from the outside. When the answer, the next answer and the checkout all live on the same page, a growing assistant grows on its own page.

Chartbeat's number closes the loop from the publishers' side. Across its network, AI referrals grew eightfold in two years and still make up 0.01% of traffic — and in 2026 they are falling for ChatGPT, Claude, Perplexity, Copilot and Gemini alike. Google Search's share of all pageviews fell from about 9% in July 2024 to 5% in July 2026. What grew instead is the traffic nobody can attribute: dark social went from 7.1% to 11.3% of pageviews, direct from 13.5% to 15.9%.

More links in the answer, and fewer that leave

One number this week moved the other way, and it deserves a careful reading rather than a cheer. Peec AI's tracking shows that the share of AI Overviews with external links inside the answer text went from almost zero to over 25% within a fortnight, with the jump around 11 September.

But that is the same fortnight in which some inline links started opening AI Mode rather than a page, and Search Engine Journal pointed out what that does to the books. Google's rule is that "clicking a link to an external page in the AI Overview counts as a click". A link that starts a new Google query is a query refinement, which is neither a click nor an impression. And the generative AI report in Search Console shows impressions only. So a site can now be linked more often inside the answer and have no report that tells it which of those links led anywhere.

On 24 September Search Console did add a new filter: Web multimodal, for queries made with an image — Lens, Circle to Search, screenshots. Useful, and telling. Google found a way to separate a search made with a camera; it has not yet separated a click into AI Mode from a click to your site.

In advertising, the report arrives after the move

Google expects the September migration of Search campaigns to AI Max — those with automatically created assets or campaign-level broad match — to finish by 30 September. On 23 September it announced the report that shows what AI Max does.

Its availability: "We'll share additional details and availability later this year." So the forced move completes this week and the instrument for judging it comes later. The same post brought the AI Brief closed beta to seven more languages, among them French and Spanish.

The reporting changed on the API side too, on 21 September, and not in one direction.

And on the other side of the market, the move continues. On 23 September OpenAI took ChatGPT Ads to seven more markets — Indonesia, Malaysia, the Philippines, Singapore, Thailand, Vietnam and Taiwan — with self-serve from the first day, bringing the total to 63 countries. Ads still show only to Free and Go users.

A regulator writes the assistant into the choice

On 23 September the UK's Competition and Markets Authority strengthened its proposal for a search choice screen on Android and Chrome: shown at first use and then once a year, and — this is the new part — open to AI assistants that meet technical and security criteria. The consultation closes on 9 October.

It is a small procedural step with a large premise inside it: a regulator is treating an AI assistant as a search service, in the same list as Google. The Comscore numbers are the reason. So is the last line of the week, from Cloudflare's founders' letter on 27 September: they had forecast that automated traffic would pass human traffic in the second half of 2027, and "the rise of agents and AI crawlers pulled that date forward to May of 2026."

What to count when the numbers disagree

The week's lesson is not that one study is right. It is that each one answers a different question, and a report that uses one to answer another will mislead.

What to do this week

• Read the spam update off your own data. It may run to around 8 October. Compare landing pages before and after 24 September in your analytics; treat any tracker's weekend line as a lead, not a verdict.
• Look at your programmatic and AI-assisted pages first. That is where the early hits cluster. If they are thin, the fix is the pages, not the links.
• Split AI referrals by assistant. One line called "AI" is a ChatGPT line with noise on it.
• Measure presence in answers separately from traffic. The assistants gaining share of questions are the ones that send the fewest visits, so traffic alone will under-report exactly where the growth is.
• Check your PMax dashboards. From 21 September one row can become seven. A "Cross-network" filter now returns nothing.
• Write down which campaigns moved to AI Max, and read their search terms weekly until Google's unified report actually arrives.

Last week Google moved where the click leads. This week everybody else counted what that moved — with a panel, a network, a court and an experiment — and got the same answer. The only instruments that did not come back clearer were the ones that read Google's results page from outside.

Sources

Every claim above is dated and linked. Where a number comes from a measurement rather than an announcement, whoever measured it is named.

• Google Search Status Dashboard: September 2026 spam update — Google, 24 September 2026
• Google September 2026 Spam Update Weekend Impact — Search Engine Roundtable, 28 September 2026
• Gemini Notebook Pages Spamming Google Search — Search Engine Roundtable, 22 September 2026
• SISTRIX for Google: status — SISTRIX, entry of 26 September 2026
• Comscore's latest AI Intelligence report — Comscore, 22 September 2026
• BrightEdge: ChatGPT referral traffic more than doubles in 2026 — BrightEdge, 24 September 2026
• Google Search decline for publishers, Chartbeat data — Search Engine Journal, 24 September 2026
• Microsoft measured Copilot cutting publisher clicks by up to 94% — PPC Land, on filings in MDL 25-md-3143, 21 September 2026
• Research shows Google AI Mode sends fewer clicks — Search Engine Journal, 21 September 2026; paper: arXiv 2608.18352
• Google AI Overviews: more external links — Search Engine Roundtable, 23 September 2026
• AI Overviews have more links, but not all reach the web — Search Engine Journal, 25 September 2026
• Search Console multimodal search type filter — Search Engine Roundtable, 24 September 2026
• New AI Max language and reporting features — Google Ads & Commerce Blog, 23 September 2026
• AI Max unified reporting and the September migration — Digital Applied, 23 September 2026
• Google Ads reporting changes, 21 September 2026 — Supermetrics
• OpenAI gains 7 Asian markets for ChatGPT Ads — PPC Land, 23 September 2026
• CMA strengthens proposals allowing people choice over their search service — UK Competition and Markets Authority, 23 September 2026
• Cloudflare's 2026 annual founders' letter — Cloudflare, 27 September 2026