Your AI Traffic Is Hiding in Direct

Gemini began adding UTM tags to its outbound links in the last days of September 2026. Google has not documented the change. Site owners spotted it in their own reports.

That is good news. It is also a reminder of how much AI traffic you still cannot see.

Where AI visits go in your analytics

An analytics tool needs a clue to name a source. The clue is a referrer or a UTM tag. With neither, the visit drops into “Direct.” Direct is the bucket for “we don’t know.”

Most AI traffic has landed there. The referrer often breaks between the AI app and your site. One practitioner reported that only about 9% of visits from the Gemini mobile app carried a referrer. Android assistant visits carried none. Desktop web visits fared better.

A UTM tag lives inside the link itself. It survives the in-app browser. It shows up in raw server logs. So the new tags will move some Gemini visits out of Direct and into a named channel.

Some, not all. ChatGPT adds UTM tags only to links from answers grounded in live web results. An answer drawn from training data carries no tag. Google has not said when Gemini adds its tag. Expect gaps.

Agents that look like people

A second group of visitors is harder still. Personal AI agents now shop and browse for users.

One consumer agent app passed 2.5 million downloads in its first two weeks after a September 8 launch. It browses from home internet connections. It presents as an ordinary visitor. It does not name itself. It does not cite the pages it reads in its answers.

You cannot filter these visits out. You can only spot their shadow. One research team tracked unverified bot requests that appeared just before these agent sessions. They compared the 20 days before launch with the 20 days after.

Region	Change in unverified agent requests
US and Canada (launch markets)	+1,313%
Rest of world	+827%
Ratio	1.6x

Chart: change in unverified agent-linked requests, 20 days before versus 20 days after a September 8 launch. Data from a study of 280 retail sites.

The sites in launch markets saw the bigger jump. That gap is a rough measure of agent traffic. It is not a count.

What you can measure today

Here is the state of play for each source.

Source	What reaches your analytics	Where it lands
Gemini, desktop	Referrer, now a UTM tag	Named channel
Gemini, mobile app	UTM tag on some links	Named channel or Direct
ChatGPT, web-grounded answer	UTM tag	Named channel
ChatGPT, answer from training data	Nothing	No visit at all
Personal shopping agents	Looks like a human visit	Direct or organic

Three steps follow from this table.

First, build a channel group for AI sources. Match on known referrers and on UTM source values. Check it each month. The values will change.

Second, read your server logs. Logs keep UTM tags that some scripts drop. Logs also show bot requests your analytics tool hides.

Third, accept the blind spot. The largest share of AI influence never makes a visit. A user reads the answer and acts. Your site gets no session. Your brand still got judged.

Some of that influence does reach your site, just under the wrong name. A user who sees your brand in an AI answer often types your URL or searches for your name. The visit lands in Direct or in branded organic search, and the credit goes to a channel that did not earn it. Watch for direct and branded search traffic that rises with no campaign behind it. That pattern is often AI at work.

The gap cuts both ways. You may be winning in AI answers and not know it, or losing and not know it. Most of your competitors share the same blind spot, so the teams that start measuring now get an early read on the channel.

Measure the answer, not only the click

Clicks tell you who arrived. They do not tell you what the AI said before the user decided. If ChatGPT gives a wrong price, the user may never come. If Gemini names a rival, you see nothing.

So track two things. Track the visits you can attribute. And track the answers themselves. Ask each AI tool the questions your buyers ask. Score each answer against your real facts. Fix what is wrong at the source.

Beket AI does the second part. It sends standard queries to ChatGPT, Gemini, and Perplexity. It scores each answer against your business data. It reports the errors, the gaps, and the conflicts. Where a fix can be automated, Beket applies it. Start with a Beket AI audit and see what AI tells your customers before they ever click.