Google's billions-of-clicks claim: why it can't be checked against your own site data

Google says its AI Search features send “billions of clicks to the web every week” and that the volume is “relatively stable.”¹ That claim cannot be checked against any individual site’s own analytics. It sits next to a separate, reproducible pattern in the traffic data: The New York Times, citing Cloudflare data, reports human traffic to finance, publishing, and retail sites down nearly 40% in less than a year, and Wikipedia counts an 8% decline in human visitors over the same period.¹ The practical test for a site owner is simple: can Google’s number be checked against your own Search Console data? On the evidence available, no.

Google says its AI Search features send billions of clicks to the web every week and calls the number stable, while Cloudflare’s own report counts more than half of all web traffic as now nonhuman.

What did Google actually claim about AI Search clicks?

Google has repeated its billions-of-clicks-per-week claim through three different voices over eleven months, and none of the three came with a baseline, a denominator, or a published methodology. Search vice president Liz Reid wrote in an August 2025 blog post that Google sends a “relatively stable” number of clicks to the web, numbering in the billions, and called reports of a traffic decline “inaccurate,” built on “flawed methodologies.”¹

Eleven months later, and just days before a Google spokesman gave a near-identical line to The New York Times, Nick Fox, a Google senior vice president, stated the same weekly-clicks figure, again with no baseline, denominator, or methodology attached.²³ Fox’s statement said AI features in Search now send billions of clicks to websites every week, on top of the billions Search sends daily overall, and no earlier version of that weekly AI-specific figure appears anywhere in Google’s own posts before mid-July 2026.²³ The spokesman’s version, in a Times article published July 20, 2026, read: “Google’s AI Search features send billions of clicks to the web every week, meeting people’s evolving preferences for how they want to find information while driving significant traffic to websites.”¹

Both statements describe a real product change. AI Mode, which Google introduced in 2025, replaces search-results lists of hyperlinks with a conversational answer written by Gemini, and Google told developers in May 2026 that people now write AI Mode queries roughly three times as long as the keyword-style queries they used to type into a search box.¹ A search box that now answers directly, in conversational paragraphs written by Gemini, is the product generating both the stable-clicks claim and the traffic complaints it is meant to rebut.

Why can’t you verify Google’s “billions of clicks” from your own data?

Search Console, the free tool Google gives every site owner, shows impressions: how often your pages appeared in search results, including AI Overviews and AI Mode. It does not show a separate count of clicks attributable to AI features the way Google’s public figure implies exists. That gap is a structural feature of Search Console, and it is why Google’s number cannot be reconciled against any individual site’s own data.²³

Nick Fox’s number illustrates the problem precisely, and the fact that Search Engine Journal and Search Engine Land had to report it themselves, rather than Google publishing a methodology alongside it, is itself the point.²³ A number with no prior baseline cannot be evaluated as “stable,” “growing,” or “shrinking,” because there is nothing published anywhere, by Google or anyone else, to compare it to.

This is not a small methodological quibble. A weekly click count is only useful as a trend indicator if it can be checked against a defined baseline period, a stated cohort, or, at minimum, a site’s own referral data. Google’s figure clears none of those bars: no baseline period, no cohort definition, and no way for Search Console to confirm or contradict it at the level of an individual property.

What does the traffic data on the other side show?

Cloudflare, the web infrastructure company that sits in front of a large share of the internet’s traffic, reports in its own July 2026 bot-traffic analysis that more than half of all web traffic is now nonhuman.⁴ The sharper cut, that human traffic to websites in finance, publishing, and retail fell nearly 40% between June 2025 and April 2026, is The New York Times’ rendering of Cloudflare data rather than a phrase in Cloudflare’s own post.¹ Cloudflare’s own framing is broader: it says some of the most heavily crawled categories have seen human traffic decline as much as 40% in under a year.⁴ Cloudflare describes the shift as a change to the web’s underlying economic model.⁴

The direction of the finding matches a second, independent measurement: the Wikimedia Foundation, which runs Wikipedia, says it has seen an 8% decline in human visitors to its site over the past year, even as bot traffic scraping the site to train AI models has increased.¹ Wikimedia’s chief product and technology officer, Selena Deckelmann, put the mission plainly: “My work is to ensure that people don’t forget about us because Google’s user interface changed.”¹ Wikimedia has started promoting its own app and social channels, and it began charging Google and other AI companies for access to its training data.¹

Regulators are already treating the pattern as real enough to act on. Britain’s Competition and Markets Authority ordered Google last month to add clear attribution links inside AI-generated responses in the UK and to let website operators opt out of having their content used in AI summaries, changes Google is rolling out globally rather than confining to Britain.¹ A regulator mandating attribution links is itself a judgment that AI answers were displacing outbound links often enough to require a remedy.

Is the open web actually collapsing, or is this cyclical?

The strongest evidence for structural decline is behavioral: how long people stay inside AI Mode, and whether they click through to the web afterward at all. Nilay Patel, editor in chief of The Verge, has warned publishers for years about “Google Zero,” the moment traffic from Google collapses to nothing, and says that moment has arrived: “For publishers, Google Zero is already here.”¹ A single research newsletter’s finding points the same direction.

Growth Memo, a search and marketing newsletter, published an October study finding that in roughly 75% of AI Mode sessions, users never left for the open web.¹ That sits alongside a wider pattern The New York Times found across three researcher studies: people now spend one to nine more minutes in AI Mode than they spend on a traditional Google search.¹ Google disputes the Growth Memo number specifically. A company spokesman told the Times it was flawed, relying on a small sample and specific assigned tasks that did not reflect normal search behavior.¹ That is a fair objection, and it cuts both ways. A small-sample newsletter study is weak evidence on its own, and Google’s weekly-clicks figure fails the identical evidentiary test for the identical reason: neither publishes enough to check.

Google is also not standing still on the outbound side. Since Reid’s August 2025 post, Google has added website previews to some AI responses, a “preferred sources” setting that lets people weight familiar publishers more heavily in what AI cites, and a profile tool that helps publishers and creators promote their work inside search.¹ Whether those additions offset the dwell-time and non-return patterns above at the aggregate level is exactly the figure Google has not published.

How should a practitioner read AI-usage numbers now?

A practitioner should read any AI-usage number, whether it favors the open web’s decline or Google’s stability, by applying what I will call the checkable-methodology test: does the source publish enough methodology to let you check the number against a comparison point? In my own prior analysis, Anthropic’s June 2026 Economic Index passed that test by publishing its sampling method, its output classifier, and a survey linked to real usage logs.

I’ve made that argument in more detail elsewhere: AI usage is a legitimate demand signal, but only when its methodology is visible, the way Anthropic showed its work down to the classifier categories and the sampling cadence. Google’s number fails that same test: no cohort, no baseline period, no linkable methodology, repeated by three different Google executives across eleven months without the underlying data ever being published.¹²³

There is a specific, falsifiable fix available, and its absence is the real story. Individual sites can already see impressions in Search Console.²³ If Google published a parallel per-site figure for clicks referred specifically from AI features, any site owner could check the aggregate against their own reality. Until that exists, the responsible move is to treat the billions-of-clicks-per-week figure as a public-relations number, not a metric to plan against.

I don’t have a client Search Console breakdown of AI Mode impressions against referred clicks to publish alongside this argument, so what is above is a synthesis of public reporting, not a first-hand measurement. That is a real limit on this piece, and I am naming it rather than implying evidence I do not have.

The click is not the only unit changing shape. I’ve also covered how Amazon, Google, and OpenAI are converging on AI-generated ad copy assembled at query time from a merchant’s product feed, a related but separate shift in how the commercial layer of the web gets built. Both changes point the same direction: your data has to be legible enough for a model to use it, cite it, or act on it, whether or not you ever see the click that used to come with it.

None of this requires taking either side’s word for it. Cloudflare, Wikimedia, and the UK’s regulator are describing a real, measurable pattern in their own traffic and their own enforcement actions. Google is describing a real number from its own logs. The two are not necessarily in conflict: a stable or growing aggregate number of clicks is mathematically compatible with individual categories of sites losing traffic, if enough other categories are gaining it. What is missing is the breakdown that would tell you which is happening to sites like yours, and Google is the only party positioned to publish it.