51,200 Events and 22.4% Misfiled as Direct: A Nine-Month AI Overview Tracking Report — and the Three Holes Its Own Author Discloses

On August 18, 2026, Search Engine Land published something rare in the GEO space — not a vendor report, but nine months of first-party tracking done by one practitioner. In "What 9 months of AI Overview data and 51,000+ tracked events reveal," Alex Galinos disclosed that between September 2025 and June 2026 he recorded 51,200 AI Overview events across 1,661 cited snippets for a brand in the transportation industry. The number that stops you is 22.4% — the share of that AI Overview traffic that GA4 filed under Direct instead of Organic Search. That is 11,468 events quietly deleted from the organic ledger.

To understand why anyone would build this themselves, look at what the platform provides. Almost nothing. Google only launched Search generative AI performance reports in Search Console on June 3, 2026, and they surface impressions only — no clicks, no CTR, no query strings — initially limited to a subset of UK site owners. Meanwhile third-party estimates of how AI Overviews even pick sources contradict each other: Ahrefs, analyzing 863,000 SERPs and 4 million AI Overview URLs, found only 37.9% of AIO citations also rank in the top 10 for the same query, down from roughly 76% in July 2025. When the platform withholds numbers and the tools disagree, brands are pushed into measuring for themselves.

That vacuum is exactly why the GEO tooling category has absorbed so much capital in eighteen months. Profound closed a $96 million Series C at a $1 billion valuation on February 24, 2026, taking total funding past $155 million. London-based geoSurge raised a $12 million seed round in July, billed as the largest seed financing in the GEO category to date. Every one of those valuations rests on a single premise: AI visibility is measurable. The most honest thing about this August 18 report is that the person who actually ran the measurement for nine months wrote three holes in that premise into his own article.

For small and mid-sized businesses and freelance studios, the practical value here is unusually high: the entire tracking setup requires no SaaS purchase — one GA4 custom dimension gets you started. Equally important is knowing how far this signal can be pushed. Below: the full numbers, what three reader types should do today, the three methodological holes, and a zero-subscription build.

Event Details and the Full Numbers

The method is simple — simple enough that most teams never considered it. When a user clicks a cited snippet inside an AI Overview, Google sometimes appends a #:~:text= text fragment to the destination URL. That is the WICG Scroll To Text Fragment specification, which tells the browser to scroll to and highlight the matching passage. Galinos built a GA4 custom dimension that fires whenever a session lands carrying that fragment.

• Dataset size: September 2025 to June 2026 — 51,200 events across 1,661 cited snippets.
• Extreme concentration: the best-performing snippet drove 2,276 events, while the average across all 1,661 was just 31. A small minority of pages does nearly all the work.
• Attribution gap: an average of 22.4% of AI Overview events landed in Direct — 11,468 events. Worst month was May 2026 at 29.3%; best was April 2026 at 16.8%.
• Share of organic: across the full period, AI Overviews accounted for 7.53% of organic sessions — but it peaked at 16–17% in February–March 2026, then fell to roughly 2–4% by the time of writing.
• Content types cited: transfer-time and pricing content dominated, along with comparison tables written as actual HTML tables. Generic destination guides underperformed badly.

Be precise about the logic behind 22.4%, because it is the figure most likely to be misused. AI Overviews exist only inside Google Search, so those sessions should belong to Organic Search. When GA4 files them as Direct, the referrer was usually dropped during redirection — and Google's own definition of the Direct channel explicitly includes traffic from untracked sources. This is not Analytics breaking. It is Analytics following rules written before AI Overviews existed.

Immediate Actions for Three Reader Types

Brand owners and SMB operators

• Ask your marketer or agency one question today: how much has our Direct share grown over the past twelve months? If it grew meaningfully and you ran no brand advertising, that is not brand strength. That is attribution leaking.
• Do not reset KPIs off this dataset. 7.53% is a nine-month average for one transportation brand, not an industry benchmark. Treat it as a reason to spend two days measuring, not as a target.
• Change the budget question from "where do we rank in AI" to "do we own a measurement pipeline we can read ourselves." The first buys somebody else's dashboard. The second buys an asset.

Marketing and SEO practitioners

• Build the text-fragment custom dimension in GA4 this week. It costs nothing, and the data is not retroactive — whatever you do not capture today is gone permanently.
• Build a monthly Direct-to-Organic ratio chart annotated with Google update dates. The August 2026 spam update finished rolling out on August 21, so movement in that window must be attributed separately, not blended into one narrative.
• Check whether your comparison content uses real <table> markup — structured tables were the biggest overperformer here. The same week, Search Engine Land also warned that JavaScript-only links can render pages invisible to AI search.
• Write "snippets have lifecycles" into your content calendar. Some peak then decay; others emerge months after publication and keep climbing. Refresh cadence is a variable, not a one-time task.

Developers and agencies

• This is directly productizable: deliver an "AI exposure attribution pipeline" — GA4 custom dimension, server-side event collection, persistence into the client's own database. Concrete acceptance criteria, zero ranking promises.
• Parse the text fragment in backend logic rather than a front-end script. Ad blockers, consent platforms and privacy settings all eat front-end events. Server-side receipt is the durable copy.
• Turn the citation prerequisites in Google's official AI features documentation into a delivery checklist: server-side rendering, real links, structured data, clearly delimited question-and-answer passages. Engineering problems, not folklore.

AI Visibility and Attribution Tooling Comparison

The important thing about this table is that the last row does not answer the same question as the four above it. The first four measure exposure and mentions. The last measures actual arrival behaviour. Most brands buying the first four believe they are buying the fifth.

What Nobody Will Tell You

• The signal is not exclusive to AI Overviews. The author states it plainly: #:~:text= is also used by Featured Snippets and People Also Ask. His defence is that an Ahrefs check showed fewer than 30 Featured Snippet appearances — but that uses one tool's sampling to calibrate another dataset's noise. It bounds the problem; it does not remove it.
• The numerator and denominator are different units. This deserves the loudest amplification, and again the author disclosed it himself: the GA4 custom dimension is event-scoped, not session-scoped. So "7.53% of organic sessions" is in fact events divided by sessions, and one visitor can fire multiple events. The direction may be right; the absolute value cannot sit alongside any session-based metric.
• n = 1, in an unusually favourable vertical. The sample is one transport and transfer brand, whose queries are natively "how long from A to B, and how much" — precisely the shape AI Overviews cite directly. Ecommerce, B2B services and healthcare have entirely different query structures. There is no basis for porting 7.53% across.
• Part of that 22.4% may not be an error at all. URLs carrying text fragments can be copied, pasted into messaging apps, and bookmarked. Those genuinely direct visits get counted here as "misattributed AI Overview traffic," and the dataset offers no way to separate them.
• The incentive structure is worth labelling. Search Engine Land is owned by Semrush, and the article page carries inline promotions for Semrush AI visibility tooling. The author, alongside his brand-side role, co-founded an SEO agency. None of that implies fabricated data. It does mean "you should start tracking immediately" is a conclusion convenient for every party involved.

The Zero-Subscription SMB Alternative

This is a rare case where you genuinely do not need to buy a tool. One engineer can complete the following four steps in two to three working days.

• Step 1 (half a day): create an event-scoped custom dimension in GA4 capturing the #:~:text= fragment from the landing URL. Fragments are never sent to the server by default — read location.hash on the client and dispatch it deliberately.
• Step 2 (one day): mirror the same event to your own backend (a single Laravel API endpoint suffices), storing the raw URL, parsed fragment text, landing page, timestamp and the channel GA4 assigned. The front end gets blocked; the backend copy is your insurance.
• Step 3 (half a day): produce two charts — monthly event counts ranked by snippet, and a monthly trend of the share filed as Direct. The first shows which pages feed you; the second shows how much your reporting misses.
• Step 4 (ongoing): reconcile backend against GA4 monthly. The gap is itself a health metric; sudden widening usually means a consent platform, cache layer or CDN configuration changed.
• Cost: two to three developer-days, then nothing beyond existing hosting. No monthly fee.

Run this self-check before you start:

• ☐ My key comparison content uses real <table> markup, not images or div layouts
• ☐ Navigation and in-body links are real <a href> elements, not JavaScript handlers
• ☐ Prices, times and specifications sit in distinct passages rather than buried mid-article
• ☐ The GA4 custom dimension exists, and I know it is event-scoped rather than session-scoped
• ☐ I hold a backend copy independent of GA4, and my monthly report annotates algorithm update dates

Frequently Asked Questions

I have no engineering resource. Can the GA4 interface alone do this?

It gets you about 70% of the way. A GA4 custom dimension plus an exploration report reveals the volume of text-fragment traffic and its landing-page distribution. Engineering is needed for the backend backup in step two and the reconciliation in step four. To simply confirm whether the volume exists on your site, do step one — half a day.

Is 22.4% landing in Direct a Google Analytics bug I should fix?

It is not a bug, and it cannot be fixed. GA4's definition of Direct explicitly includes traffic from untracked sources, so when a referrer is lost during redirection, filing it as Direct is the rule working as written. Your move is not to correct it but to quantify it — know how much your organic reporting understates, and disclose that margin when reporting.

This data comes from a transportation brand. Can ecommerce or B2B apply it directly?

Apply the method, not the numbers. Transport queries are natively "how long and how much" — the exact shape AI Overviews cite directly — so 7.53% is likely at the high end. Measure your own site with the same method and make budget decisions from your own figures, not somebody else's percentage.

Search Console already has generative AI reports. Why build my own?

They answer different questions. The reports Google launched in June 2026 provide impressions only — no clicks, no CTR, no query strings — and were initially limited to site owners in certain regions. You learn how often you were seen, not how many people arrived or what they did afterwards. Text-fragment tracking covers the second half. Run both.

AI Overview share fell from a 16–17% peak to 2–4%. Is AI traffic disappearing?

The data does not support that. This is one site's relative share, and the decline could stem from that site's own ranking movements, seasonal intent shifts, changes in how often Google triggers AI Overviews, or all three. The safer reading: AI Overview prominence in the SERP is unstable, and brands treating a peak month as a new baseline will miscalibrate their models.

My Take

The mainstream reaction treats 22.4% as a leak to be patched. I think the real value of this story sits entirely elsewhere: it is the clearest evidence yet that the AI visibility industry is selling a measurement with no stated margin of error.

Notice the difference between this report and most vendor white papers. It puts three fatal limitations inside its own text — the signal is not exclusive, the sample is n=1, the numerator and denominator are different units. First-party data honest enough to undermine itself is more trustworthy than tool reports citing million-scale samples while disclosing nothing about method. That exposes the category's problem: while Profound commands a $1 billion valuation and geoSurge takes the category's largest seed round, none of their public material discusses error sources with anything like this candour.

So my forecast diverges from consensus. Most people expect GEO tooling to keep consolidating through 2027. I expect the first wave of "AI visibility data disputes" within twelve months — a client running two tools in parallel, finding the same brand in the same month scores more than three times differently, and cancelling both. Once those disputes go public, pricing power shifts from "whose dashboard looks best" to "who publishes their methodology and error bars." The first vendors culled will be those unwilling to disclose even their sampling frequency.

For a Laravel and Flutter studio like ScriptWalker, the opportunity is concrete: sell attribution pipeline engineering, not AI visibility scores. Three auditable deliverables — GA4 custom dimensions with server-side event collection, raw data persisted in the client's own database, and a monthly reconciliation report annotated with error sources and algorithm update dates. It promises no rankings. It guarantees that the next time a platform changes the rules without notice, the client holds their own historical data and knows how wrong it might be. In a market where nobody publishes the denominator, that is the differentiator.

Sources

• First-party: Search Engine Land — What 9 months of AI Overview data and 51,000+ tracked events reveal (2026-08-18)
• First-party: Google Search Central Blog — Introducing Search Generative AI performance reports in Search Console (2026-06-03)
• First-party: Google Search Central — AI features and your website (official documentation)
• First-party: Google Analytics Help — Default channel group (Direct definition)
• First-party: WICG — URL Fragment Text Directives (Scroll To Text Fragment specification)
• Third-party: Ahrefs — Update: 38% of AI Overview Citations Pull From The Top 10 (863K SERP study)
• Third-party: Fortune — Profound raises $96 million at a $1B valuation (2026-02-24)
• Third-party: citybiz — geoSurge Raises $12 Million Seed Round to Expand AI Visibility Platform
• Third-party: Search Engine Land — Google August 2026 spam update done rolling out (2026-08-21)

Want a Measurement Pipeline You Can Actually Read?

ScriptWalker builds with Laravel and Flutter. We do not sell AI visibility scores. We deliver an auditable attribution pipeline: GA4 custom dimensions with server-side event collection, raw data persisted in your own database, and a monthly reconciliation report that states its error sources.

• Email: [email protected]
• Phone: 0916-224-047
• LINE: @ufv9089p