Around September 28, 2026, Gemini quietly stopped adding utm_source=gemini to the links it cites. Teams that had just wired that tag into their dashboards lost their cleanest signal for AI referral traffic overnight. The fix is to measure AI traffic by referrer and server logs, not campaign tags.
The short version
- Gemini tagged cited links with utm_source=gemini in early September 2026, then dropped the tag around September 28. Any dashboard built on that tag is now undercounting.
- Desktop Gemini clicks still carry a gemini.google.com referrer and stay recoverable. App clicks often send no referrer and fall into Google Analytics 4's Direct bucket.
- Google Search Console's Generative AI performance report shows impressions in AI Overviews and AI Mode, not clicks or queries.
- Measure AI traffic as a blend of referrer data, server logs, and Search Console impressions, not one tidy number.
What actually changed with Gemini
Gemini added tracking tags for about three weeks, then took them away, and Google never documented either move. In early September 2026, Gemini started appending a UTM (Urchin Tracking Module) parameter, utm_source=gemini, to the outbound links in its answers. Around September 28, the tags vanished from one day to the next, per analyst data from Natzir surfaced by Search Engine Roundtable. Google announced neither the start nor the stop.
Here is why that matters. If you stood up a channel group or a Looker Studio view keyed on utm_source=gemini in mid-September, it captured a three-week window and then flatlined. On a chart that looks exactly like a traffic collapse. It is not. The visits are still coming. The label you were counting them by is gone.
Why so much AI traffic hides in the Direct bucket
The real problem is not UTMs, it is how AI assistants hand off a click. Google Analytics 4 (GA4) assigns a session to Referral when the medium is referral, app, or link, and to Direct when the source is (direct) and the medium is (none) or (not set), per Google's default channel definitions. A UTM tag or a referring hostname is what fills in source and medium in the first place.
Gemini's desktop web experience still passes gemini.google.com as the referrer, so those visits can be rebuilt into a Referral channel even without the tag. The Gemini iOS app is the hole. It opens cited links in its own in-app WebView, which sends no referrer, so GA4 has nothing to attribute and files the session under Direct. The only durable trace of those visits is the GeminiiOS string in your server request logs. This is the pattern across assistants: a link opened in a normal browser tab passes a referrer, a link opened inside a native app usually does not.
What Search Console's AI report does and does not give you
Search Console now reports your AI visibility, but it is impressions, not clicks, and it lives on Google's side of the fence. The Generative AI performance report, which Google began rolling out in 2026, shows how often your links appeared in generative AI features on Search, broken out for AI Overviews and AI Mode, and it includes multimodal searches like Google Lens and Circle to Search.
Google defines an impression here as a link to your site being shown in a generative AI feature. What the report does not surface is click data or query data, and it excludes experiments in Search Labs. So Search Console tells you that you are being cited and roughly how often. It does not tell you who clicked through or what they did next. That half of the story only exists in your own analytics and logs, which is exactly the half the Gemini change just degraded.
A measurement stack that survives the next change
Stop chasing one perfect number and build three overlapping views you control. Each one covers a gap in the others, so no single vendor decision can blind you again.
Define a durable AI channel group on referrer, not UTMs. In GA4, build a custom channel group that matches on referring hostname: gemini.google.com, chatgpt.com, perplexity.ai, copilot.microsoft.com, and the rest. Referrer-based rules keep working when a vendor drops its tags, which UTM-based rules do not. Keep a UTM capture as a bonus layer on top, so you catch the extra detail when an assistant does tag links, without depending on it.
Pull the app traffic out of Direct with server logs. Filter your request logs by AI user agents such as GeminiiOS, and you can size the app-driven slice that GA4 will never attribute. It will not be perfect, but it turns an invisible chunk of Direct into a number you can defend in a board deck.
Reconcile against Search Console impressions. Line your referrer and log data up against the Generative AI report. If impressions are flat but your referral traffic fell, you have an attribution problem, not a visibility one. If impressions themselves dropped, that is a content and citation problem, and a different fix entirely.
When I rebuilt this for a B2B SaaS client this fall, the UTM-only dashboard showed AI referrals falling off a cliff in early October. The server logs told the real story: the GeminiiOS slice had grown, not shrunk, and the entire drop was the tag going away. We moved the channel group onto referrer rules and reconciled it against Search Console impressions, and the decline disappeared. The traffic had been there the whole time, filed under Direct.
What this does not cover
This is a measurement post, not a visibility one. It does not cover how to earn more AI citations in the first place, which is a content, structured-data, and authority problem worth its own playbook. It does not cover paid placement inside AI surfaces, since text link ads in AI Mode were still in testing as of late October 2026. And it assumes GA4. If you run a server-side or warehouse-first analytics setup, the referrer and user-agent logic still holds, but the implementation will look different.
FAQ
Did my AI traffic actually drop, or did the tracking break?
Check whether the decline lines up with late September 2026, when Gemini stopped tagging links. If your fall-off starts there and your Direct channel rose by a similar amount over the same window, you are looking at an attribution gap, not a real loss. Server logs filtered by AI user agents will confirm which one it is.
Should I still set up UTM tracking for AI assistants?
Capture UTMs when they appear, but do not build your reporting on them. Vendors add and remove tags without notice, as Gemini just demonstrated. Referrer-based channel rules and server logs are the stable foundation. UTMs are a bonus layer, not the base.
Can Google Search Console show me AI clicks?
No. The Generative AI performance report shows impressions in AI Overviews and AI Mode, not clicks or queries. For click-level data you need your own analytics and server logs, and even those will miss app traffic that arrives with no referrer.
Why does Gemini app traffic show up as Direct in GA4?
The Gemini iOS app opens cited links in an in-app WebView that sends no referrer and no UTM. With nothing to read, GA4 defaults the session to Direct, because its Direct channel is defined as a (direct) source with a (none) or (not set) medium. The only trace left is the GeminiiOS user agent in your logs.
Get a read you can trust
If you want a clear picture of how much of your pipeline is actually coming from AI search, I will rebuild your GA4 channel grouping on referrer rules, reconcile it against your server logs and Search Console, and hand you a dashboard that holds up the next time a vendor changes the rules. Book a 30-minute AI traffic audit read-out.