The week Google closed the exit

The first fortnight of September was spent arguing about whether there had been an update. The seven days that followed mattered more — and almost none of it was about ranking.

Between 16 and 23 September 2026 Google changed three defaults. Links inside AI Overviews began leading to AI Mode rather than to a page on a site. Merchant Center switched on payment inside AI Mode and Gemini for Shopify stores, without being asked. And the tools the industry uses to measure both started getting noticeably less data out of Google after 13 September. None of the three was called an update, and none of them will appear on a volatility chart.

Three locks in seven days

The three are worth reading together, because separately each looks like a small news item and together they describe one route: query, answer, next answer, payment — all of it on Google's page. Below is that route in two states, before and after, with the measurement strip underneath that went darker this week.

Lock one: a link that does not lead to a site

On 21–22 September Gagan Ghotra spotted a new kind of anchor link inside AI Overviews. They look like the familiar links under an answer, but instead of the source page they open AI Mode with a follow-up question already filled in.

This is a test, not an announcement, and Google did not comment on it. But it shows the direction of travel precisely: a click on an answer is spent on the next answer.

To see why Google would do this, look at how small AI Mode still is. On Similarweb clickstream data as read by Rand Fishkin, 0.34% of all searches reached AI Mode between January and April 2026. For comparison: AI Overviews appear on more than 20% of searches, and 68.01% of all Google searches end without a click at all — ten years ago that figure was around 45%.

So the product the links now lead into has almost no traffic of its own, while the surface leading into it is enormous. That is not a feature maturing slowly. That is a transplant.

Lock two: a purchase that never leaves the answer

On Friday 19 September Google widened the rollout of native checkout, and on 22 September merchants began receiving emails telling them it was already on. The mechanism: a Shopify store is matched to its Merchant Center account, and eligible products become buyable directly inside AI Mode and Gemini — through Google Pay and the Universal Commerce Protocol, which launched back in February 2026.

Two words in that email carry the whole thing. The first is was: the merchant was told after it was switched on, not asked before. The second is native: the payment does not happen on the merchant's site. Google stresses that merchants "manage and retain full ownership of your customer relationships, data, and sales", which is most likely legally exact — but the page the transaction happens on is still not yours.

It can be turned off, and not at Google: in the Shopify admin, under Sales channels → Agentic. That switch is worth visiting deliberately, in either direction — for some stores a presence in somebody else's checkout is new sales, and for others it is a lost upsell, a discount code that does not apply and an empty remarketing list. But the default has already been decided for you.

Lock three: the lights went out on measurement

From around 13 September Google became markedly better at blocking automated access to its results — both AI company scrapers and third-party rank trackers. The effect is uneven, and the unevenness is what makes this week's data dangerous: two reports built on two tools now disagree for reasons that have nothing to do with the site.

Glenn Gabe flagged the worse version of this problem separately: tools were showing recoveries that the sites' own analytics did not confirm. So it is not simply less data — in places it is wrong data that looks perfectly normal.

Lay John Mueller's 13 September admission on top of that — that Search Console tracks an AI Overview as a block rather than the position of an individual link — and the picture is symmetrical. Google's own analytics does not show the AI layer by construction. Third-party tools this week show less of it, and in places show it wrong. And the user's route was rearranged in that very week.

It is the worst possible moment to draw conclusions from a rankings report — and the best one to stop treating position as a complete description of visibility.

Why the top 10 is no longer the ticket

That position describes less and less is not an impression of the week; it is a measured thing. Ahrefs looked at 863K SERPs and 4M AI Overview links: the share of citations coming from pages in the organic top 10 fell from 76% to 38%. The rest splits almost evenly between positions 11–100 and pages outside the first hundred altogether.

The reason, on Ahrefs' reading, is mechanical: AI Overviews lean less on the results for the query itself and more on the results for the fan-out sub-queries. A page that turns up often across those small SERPs gets cited, even if it sits on page four for the head term.

A second Ahrefs study — almost 17 million citations across ChatGPT, Perplexity, Gemini, Copilot and AI Overviews — adds a second criterion, freshness. The average cited page is 1,064 days old against 1,432 days in ordinary organic results, or 25.7% fresher. The spread between systems is wide: ChatGPT cites material an average of 458 days newer than organic, while AI Overviews cite material 16 days older. The study's author warns against over-reading it: the average cited page is still 2.9 years old, and restamping a publication date without changing the content does not work.

And in advertising, the same thing inside the account

September is a month of forced migration in Google Ads. Through 1–30 September, campaigns with automatically created assets (ACA) or the campaign-level broad match setting are upgraded to AI Max. Dynamic Search Ads campaigns were meant to move at the same time, but Google pushed them to February 2027 after advertiser feedback. If you run ordinary RSAs with keywords and no ACA, nothing changes this month.

Google's own estimate of the gain is careful and specific: campaigns running the full AI Max feature set see on average "7% more conversions or conversion value at a similar CPA/ROAS" — but the comparison is not against the old campaign, it is against an AI Max campaign with only search term matching enabled. That is a narrower claim than the one usually repeated.

More interesting than the figure is the history of the controls. Every lever in AI Max arrived after a wave of objections, and the sequence reads like a transcript of the market's complaints.

And in parallel, a move from the other side. Over August, OpenAI took ChatGPT Ads to 31 European markets and opened self-service in Ads Manager; ads are shown only to Free and Go users, and the buying models have moved past CPM and CPC — the company writes that it "expanded beyond CPM and CPC bidding to support conversion optimization". By its own count, "tens of thousands of marketers" have advertised on ChatGPT since February.

Both platforms are making the same move from opposite directions: paid placement moves inside the answer, and the level at which an advertiser steers it rises from the keyword to the intent. The keyword disappears as a unit of control in advertising by the same motion that position disappears as a unit in organic.

What to measure when position measures the wrong thing

Strip the news out of this week and what is left is a short table of metric substitutions. It needs no new tooling — only honesty about what the old metric cannot see.

What to do this week

• Check the switch in Shopify. Sales channels → Agentic. The decision can go either way, but it should be yours rather than a default left standing.
• Do not report tracker data from 13–20 September without a caveat. Collection fell unevenly, and a "recovery" in one tool this week proves nothing.
• Check the tracker against your own analytics. Where they disagree, the one holding data from your server is right.
• Find out which campaigns moved to AI Max in September. Not whether, but which, and what is on by default inside them. The brand toggle and negative keywords are the control you have left.
• Split native checkout orders out in reporting. Otherwise a fall in on-site conversion will look like a fall in demand.
• Measure presence in answers separately from positions. After this week that is no longer a methodological preference; it is the only way to see the leg of the route that is left.

In seven days Google confirmed no ranking change at all — and moved where a click leads, where a purchase completes and how much data reaches the people counting. An update moves sites around inside the list. This week moved the list.

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 AI Overview Links Pushing You To AI Mode Over Web Pages — Search Engine Roundtable, 22 September 2026
• Google May Be More Successful In Blocking Scrapers & Tracking Tools — Search Engine Roundtable, 18 September 2026
• Google Merchant Center Auto Enables Native Checkout — Search Engine Roundtable, 22 September 2026
• Gemini Notebook Pages Spamming Google Search — Search Engine Roundtable, 22 September 2026
• Google's Dynamic Search Ads are upgrading to AI Max — Google Ads & Commerce Blog
• Google Ads gets branded search controls inside AI Max — PPC Land, 1 June 2026
• ChatGPT Ads expands across Europe — OpenAI, updated 31 August 2026
• In 2026, Less than One Third of Google Searches Still Send a Click — Rand Fishkin, SparkToro, on Similarweb clickstream data, 9 June 2026
• Update: 38% of AI Overview Citations Pull From The Top 10 — Ahrefs, 863K SERPs, May 2026
• New Study: AI Assistants Prefer to Cite "Fresher" Content — Ahrefs, 16.975M citations
• Google Admits Search Console Reporting For AI Search Is Inadequate — Search Engine Journal, 13 September 2026