AI Overviews Send You Traffic. Your Analytics Files 22% of It Under Direct

Over nine months, one brand tracked 51,200 clicks arriving from Google AI Overview citations. Its analytics filed 11,468 of them, 22.4%, under Direct instead of Organic. In the worst month, May 2026, the leak reached 29.3%. Nearly a third of a channel’s traffic, invisible to the report that decides its budget.

The numbers come from Alex Galinos of Elife Group, published on Search Engine Land in August 2026. One transportation brand, September 2025 through June 2026, 1,661 cited snippets under instrumentation. A single site, and the author says so plainly. But it is the first public dataset of its size to put a number on something practitioners have only suspected: AI Overview traffic exists, it is material, and standard analytics undercounts it by a fifth or more.

We have argued before that AI search broke the measurement stack and that analytics platforms cannot see agent traffic at all. This dataset moves the story forward in two ways. It quantifies the leak. And it ships a free recipe for plugging it, one any team running GA4 can adopt.

How you measure a channel Google refuses to label

Google exposes no native AI Overview traffic signal. If you want the number, you build the instrument yourself.

Galinos built it on a quirk of how AI Overviews link out. When Google cites a page in an AI Overview, the citation link often appends a text fragment to the URL: the #:~:text= suffix that scrolls the visitor to the exact sentence Google quoted. That fragment is a countable signal. So he registered a GA4 custom dimension on the URL fragment and started counting.

Nine months later: 51,200 events across 1,661 cited snippets. AI Overviews accounted for 7.53% of the brand’s organic sessions over the full period. The share moved hard: it peaked at 16 to 17% in February and March 2026, then settled to 2 to 4% in recent months. A channel that swings between 2% and 17% of organic belongs on the report as its own line item.

Two honesty notes the author includes and we repeat. The #:~:text= fragment is also used by Featured Snippets, so the signal is shared; Galinos bounded the contamination at fewer than 30 Featured Snippet positions across the tracked set. And the 7.53% figure compares events to sessions, an imperfect denominator the author labels directionally accurate rather than precise.

The 22.4% that never reaches the Organic report

Here is the part that should bother anyone who owns a traffic dashboard. Of the 51,200 tracked AI Overview events, 11,468 landed in analytics attributed to Direct, the bucket where attribution goes to die.

That is 22.4% on average. The monthly picture is worse: May 2026 hit 29.3%. In a month like that, for every ten visitors Google’s AI layer sent, analytics correctly credited seven.

Think through what a marketing team does with a Direct-inflated, Organic-deflated report. Organic looks like it is declining faster than it is. Content investment gets cut based on a channel report that is structurally wrong. Meanwhile Direct swells, and nobody investigates Direct because Direct has never been investigable. The error leans in one direction, against the channel you are trying to evaluate.

This is one instrumented site in one vertical, so 22.4% should be quoted as one brand’s measured rate rather than an industry constant. The method travels further. Nothing in the tracking recipe is specific to transportation websites. If your pages get cited in AI Overviews, some slice of that traffic may be sitting in your Direct bucket right now, and without the fragment dimension you cannot say what your rate is.

What nine months of citation data says about content

The dataset also shows which content earns citations, and the pattern is specific. AI Overviews favor structured, direct-answer content. HTML comparison tables in particular were, in the author’s words, punching well above their weight in citation share. The formats that win are the ones a model can lift cleanly: a table, a definition, a numbered answer.

Citations also decay. Snippets have lifecycles: a page gets cited heavily for a stretch, then rotates out as Google reshuffles sources. Galinos’s warning is worth adopting verbatim: don’t treat a peak month as a new baseline. A February at 17% of organic does not commit Google to a June at 17%. Any reporting you build on this channel needs to price that volatility in from the start.

The measurement layer is now your job

The uncomfortable conclusion is structural. Google operates the AI surface, decides which pages it cites, and provides no instrument for the publisher to measure any of it. The measurement layer that used to come bundled with search, however imperfect, does not exist for AI search. Whoever wants the number builds the instrument.

The encouraging half of the same conclusion: the instrument is cheap. The whole recipe is one GA4 custom dimension on a URL fragment, published openly, free to adopt. The cost of knowing your misattribution rate has collapsed to roughly zero. The cost of not knowing it is a channel report your budget decisions quietly trust and should not.

Do this now

Register a GA4 custom dimension on the #:~:text= URL fragment this week, following the method in the source article. Let it collect for 30 days, then answer two questions with your own data: what share of your organic sessions arrives from AI Overview citations, and what share of those events your analytics files under Direct. Keep the Featured Snippet caveat attached to the numbers, and do not promote any single month to a baseline. If the misattributed share looks anything like this dataset’s 22.4%, your next channel review should open with that number, because every Organic-versus-Direct conclusion in it is currently wrong by that margin.

This analysis synthesizes What 9 months of AI Overview data and 51,000+ tracked events reveal (Search Engine Land, Alex Galinos of Elife Group, August 2026; Search Engine Land is owned by Semrush, as disclosed on the page).

Victorino Group helps organizations build the measurement layer AI platforms do not provide, from instrumentation to the reports leadership can trust. Let’s talk.