AI referral traffic, the count of sessions that analytics can attribute to an answer engine, was always a borrowed metric. It depended on engines voluntarily tagging their outbound links, and in the last five weeks Google quietly stopped doing it.
The sequence is short and worth getting exactly right. In early September, Gemini began appending utm_source=gemini to the URLs it cited in answers. John Mueller confirmed on Reddit that he was seeing the parameters. Analytics teams built dashboards on them. Around September 28, they disappeared overnight. Natzir charted the rise and the cliff and posted it on X, Xyborg put the drop on the same date, and Barry Schwartz wrote the change up on October 30.
Three weeks of clean data, then nothing. No announcement went out when the tagging started and none went out when it stopped, because the implementation was never documented in the first place. A measurement surface that arrives without notice and leaves without notice is not a measurement surface. It is a coincidence you were charting.
On the same day, Cyrus Shepard published the other half of the problem: a small case study that used Google's generative AI opt-out controls to pull one site out of AI Overviews and AI Mode, then watched what happened to the numbers. His estimate for the click-through rate on AI impressions was about 0.9 percent.
What Happened To Gemini AI Referral Traffic?
Gemini AI referral traffic did not fall in volume. What fell was the ability to identify it, because Google removed the utm_source=gemini parameter from cited links around September 28 after roughly three weeks of appending it, leaving most of those visits indistinguishable from direct traffic.
The distinction matters more than it sounds. Nothing suggests fewer people clicked through from Gemini answers. The visits with a gemini.google.com referrer stayed flat across the whole period, which is the detail that tells you this was a tagging change and not a traffic change. The clicks are still arriving. They are arriving anonymously.
The iOS behaviour is the part that should change how you build reports. Natzir tested the app against the web version and found the app sends no referrer, so every click from it is filed as Direct. If your audience skews mobile, and in most consumer and prosumer categories it does, the majority of your genuine Gemini traffic has been sitting in the Direct bucket this whole time, next to bookmark visits and pasted links. The UTM parameters briefly papered over that. Their removal just put the problem back where it always was.
There is a reasonable objection to all of this, and it deserves airing: UTM tagging is a courtesy, not an obligation, and engines are under no duty to help marketers attribute anything. That is true. It is also exactly the argument for not building a reporting program on top of it. OpenAI has expanded ChatGPT's analytics parameters, Google has now added and removed its own, and every one of those decisions can reverse next quarter. We made the same argument when rank tracking data turned out to be less solid than the dashboards implied: if a vendor or an engine can change your metric overnight without telling you, it is not a metric you can run a program on.
What Does A 0.9 Percent Click Through Rate Really Tell You?
A 0.9 percent click-through rate on AI impressions tells you that answer-engine visibility is an awareness channel with a narrow click path, not a traffic channel. Shepard reached the figure by subtracting known AI impressions from his site's overall impressions and clicks during an opt-out window.
Treat the number as directional and the method as the real finding. Shepard called it a small case study. Barry Schwartz stressed the same point in his write-up. One site, one window, one vertical, with the arithmetic done by subtraction rather than by a reported AI clicks column, because no such column exists. Anyone presenting 0.9 percent as an industry benchmark is doing the thing we complained about when cold email benchmarks started citing each other in a circle.
What the figure does do is settle an argument about magnitude. If AI impressions were about 8 percent of total impressions on the tested site and they convert at roughly a hundredth of the rate of a traditional result, then no amount of reporting rigour turns AI referral clicks into a line item worth a quarterly review. The value is upstream of the click, in whether the engine names you at all.
| SIGNAL | CAN AN ENGINE SWITCH IT OFF? | MOBILE COVERAGE | WHAT IT IS HONESTLY GOOD FOR |
|---|---|---|---|
| UTM parameters on cited links | Yes, and Gemini just did, without notice | Partial, and nil inside in-app WebViews | Short-term sanity checks while the tagging happens to exist |
| Referrer header in analytics | Partly, since apps can suppress it | Poor, because in-app browsers strip it | Desktop trend lines, read as a floor and never as a total |
| Server log user agents | No, the request has to identify itself to be served | Good, including the GeminiiOS case | The only durable count of answer-engine clicks you own |
| AI impression share in Search Console | No, while Google keeps reporting it | Full | Sizing how much of your visibility is now answer-shaped |
| Citation share across a fixed question set | No, you run the sampling yourself | Full | The actual performance metric for answer-engine work |
| Branded search and direct volume | No | Full | Catching the assisted demand that never shows as a referral |
Read the first and third rows together. The signal that is easiest to put in a dashboard is the one an engine can revoke overnight, and the signal nobody wants to maintain is the one that cannot be taken away. That trade is the whole reason AI visibility reporting is in the state it is in, and it is why most SEO reporting software still cannot show you answer-engine clicks honestly: the vendors are reselling the same revocable parameters you are.
“If a number can vanish from your dashboard because a product manager at an engine shipped a change nobody announced, it was never your number. Measure the things that require your own effort to collect, because those are the only ones you control.”
Why Opting Out Is Not A Clean Measurement Test
Opting out of AI features to measure their value looks like a controlled experiment and is not one, because removing a site from AI Overviews and AI Mode changes how the rest of the result page performs. Shepard's own data is the clearest evidence of that.
The headline result from his test was not the 0.9 percent. It was that organic clicks also fell when AI features were disabled, which he called the most surprising outcome and attributed to what he labelled a Search SERP Halo Effect: appearing in more places on a result page makes a user more likely to click something of yours anywhere on it. Average position told a matching story. Sitewide it improved, which he put down to new posts and brand search, while on the URLs with the most AI visibility position dropped during the opt-out and recovered when visibility came back.
Greg Sterling of Near Media added the longer-horizon version of the same point, noting that AI Overviews can produce a kind of brand lift or awareness that surfaces later as a search and a click. If that is right, then a subtraction test measures the immediate loss and misses the deferred gain entirely, which biases the result against staying in.
There is a second, quieter reason to distrust subtraction tests right now. Impressions themselves have been unreliable this autumn, as anyone who lived through the reporting discrepancy we covered when impressions climbed while clicks sat still can attest. A test whose arithmetic depends on an impressions denominator inherits every problem that denominator has.
How To Rebuild AI Referral Traffic Reporting
Rebuild AI referral traffic reporting around three sources an engine cannot switch off: server log user agents for clicks, a fixed question set sampled on a schedule for citation share, and branded search plus direct volume for the demand that never arrives as a referral at all.
Where answer-engine measurement effort should concentrate once referral tagging is treated as unreliable, our recommended allocation rather than measured data
The last row is deliberately small, and it is the row most teams spend most of their time on. Five percent is the right amount of effort for a signal that disappeared overnight in September without an announcement. The two highlighted rows are the ones that require real work and produce numbers that are yours.
What Should Go In The Board Deck Instead?
A board deck should report answer-engine performance as share of voice on the questions that matter, backed by a click count from server logs and a note on assisted demand, with referral sessions removed from the headline entirely. Nobody should be defending a budget with a 0.9 percent channel.
This is a reframe, not a retreat. The work that earns a citation is the same work it always was, and it pays off whether or not a click follows, which is the argument we made in detail about where engines actually source the pages they cite. What changes is the scoreboard. A program measured on AI referral sessions will look like it is failing in a quarter where it is winning, because the engines got better at answering and worse at tagging in the same six weeks.
Expect the tagging to come back, probably more than once, and expect it to leave again. Treat each return as a bonus data source with a short shelf life rather than as the foundation of the report. The teams that will handle the next reversal calmly are the ones who built their reporting on logs, sampling and demand signals before they had to.
See where you are cited today
A free snapshot audit of your rankings and AI citations before we ever talk.
Tyler leads work at the intersection of SEO and generative engines at Something Inc., helping B2B brands get ranked and cited across every major AI engine.