Google's traffic grew 10.72% while its clicks to sites fell, and GA4 files the gap as direct

Google's US traffic grew 10.72% over the past year while the clicks it hands to everyone else kept shrinking, and the analytics tool most software companies still run their reporting on has quietly stopped being able to explain the difference. Two pieces published on 4 September — a traffic teardown in Search Engine Land and a research round-up in Search Engine Journal — arrive at the same uncomfortable place from opposite directions. Your audience did not leave. Your referral did.

The audience held, the queries dipped, the clicks fell

Search Engine Land's Greg Jarboe pulled Semrush Traffic Analytics figures for July 2026 against the same month a year earlier. Google still ranks first in the US at 25.3 billion monthly visits, up 10.72%. YouTube is second at 10.3 billion, up 36.6%. ChatGPT.com now sits ninth with 1.09 billion monthly visits, up 48.38%, ahead of both DuckDuckGo and Bing — and Bing is down 50.43% year over year. These are one vendor's panel estimates rather than census figures, and Search Engine Land runs Semrush-sponsored programming, so trust the direction more than the decimal places.

Search Engine Journal's Matt G. Southern spent the same week reconciling three research findings that look contradictory and are not. Similarweb puts the share of ChatGPT users who also appear in Google's audience at 95% in September 2025, still 95% through May 2026, even as visits to generative AI platforms grew 70% year over year. A Bocconi University working paper, using Comscore US desktop clickstream data from October 2024 to July 2025, found households cut traditional search queries by 3.14 per week after gaining ChatGPT Search access — about 9.4% off a 33.51 baseline, widening to 17.0% after twenty weeks. And a field experiment posted to arXiv on 18 August, which routed 1,100 US Chrome users through AI Mode for a week, measured an 18.8 percentage point drop in click-through to external sites.

All three hold at once because they count different things. Overlap counts people, and a single Google search in a month makes you a Google user. Bocconi counts queries per household. The arXiv experiment counts clicks. Southern's point is that a marketer watching referrals fall while the trade press insists nobody left Google is standing in the gap between three honest measurements.

The category detail matters more to software marketers than the headline numbers. Bocconi found search referrals to developer sites down 15.1%, reference sites down 26.5% and academic sites down 32.8%, while marketplaces and entertainment barely moved. Documentation, comparison content and explainers — the exact assets B2B software teams have published for a decade — sit in the categories losing most. Both papers are preprints, and the AI Mode study forced full adoption rather than the gradual opt-in most people experience. Read them as direction, not settlement.

The bigger problem is what replaced the referral

Jarboe's most useful contribution is not the traffic table. It is a GA4 account he checked for a publisher he advises, comparing 1 January to 26 August 2026 against the same window in 2025. Sessions up more than 105%. Engaged sessions up almost 111%. Key events up around 80.5%. On that reading the site is winning. Then the same report shows direct traffic up more than 157%, now over 78% of all traffic, against roughly 62.5% a year earlier.

"'Direct' isn't a traffic source. It's the label GA4 slaps on everything it can't attribute."

That is Jarboe, writing in Search Engine Land, and it is the sentence to take into your next reporting meeting. Links surfaced inside chat interfaces routinely lose referrer data, so they land in the bucket GA4 uses for everything it cannot identify. The native AI Assistant channel Google added on 13 May 2026 is not the rescue: it accounted for 0.47% of that publisher's traffic since launch, and it excludes AI Overviews and AI Mode entirely. It is one site, so treat the percentages as illustration rather than benchmark — but the mechanism applies to every site you own.

Google's own framing sits awkwardly alongside all of this. Sundar Pichai told I/O in May that AI Overviews had passed 2.5 billion monthly active users and AI Mode a billion in its first year, and summarised the company's position as "When people use our AI-powered features in Search, they use Search more." That may well be true. It says nothing about whether they leave.

What to change in your reporting

For a SaaS marketer this is a budget-defence problem before it is an SEO problem. The organic line in your dashboard is systematically understated and the direct line is a container of unrelated things, which means any channel comparison built on GA4 defaults is quietly arguing to cut the channel that still works.

• Add a self-reported attribution field to trial and demo forms. It survives referrer stripping, and it is cheap.
• Stop treating direct as one line. Segment it by landing page and new-versus-returning — direct traffic landing on a deep comparison page is not someone typing your domain.
• Track Search Console impressions and branded query volume as leading indicators, and expect clicks per impression to fall without your rankings moving.
• Re-baseline your targets, not your team. A 20% referral decline against flat impressions is a market change, not underperformance.