How to measure a client's traffic from AI search.

When somebody clicks a link inside an AI assistant, the browser may or may not pass on where they came from. When it does not, analytics has nothing to attribute the visit to and files it as Direct, alongside people who typed the address in.

So the first thing to understand is that any AI traffic figure is a floor rather than a total. You are measuring the share that arrived with a referrer attached. Treat it as a trend line, not a count.

Google Analytics added a native AI Assistant channel group in July 2026, which is the fastest way to see some of this traffic. It does not cover every assistant, so on its own it undercounts.

The working setup is both: the native channel, plus a custom channel group of your own that matches the hosts the assistants actually send from. Build the custom group by looking at real referrer reports rather than by copying a list from somewhere else, because the hosts change and the reports show what is genuinely arriving.

For an agency, that means one shared list, built from what your clients' referrer reports actually show, applied to every client's property. Then check each client's own report now and then for hosts the shared list misses.

Three things worth knowing, and one worth ignoring.

• Which pages get cited

  Assistants tend to send people to explanatory content rather than to product pages. Look at which of the client's pages actually receive this traffic, because that tells you what they are being cited for.

• Whether it converts differently

  Somebody who arrived after an assistant summarised you has already had part of the conversation. Compare their behaviour with search traffic rather than assuming it is the same visitor.

• The direction of travel

  Month on month share matters more than any single figure, because the undercount is roughly consistent while the platforms are not.

• Ignore the market share reports

  Published figures for which assistant sends the most traffic move constantly and are not about the client's site. Their own referrer report is the only one that describes their audience.

Being cited by an assistant and being visited are different events, and only one of them shows up in analytics. Somebody can read a summary of your page, get what they needed and never click. That is not a tracking gap you can close.

Which means the honest version of this metric is narrow: it tells you that assistants are sending people, roughly how many, and to which pages. It does not tell you how often the client was cited, and no analytics setup will. Say so in the client's report, because a client who reads the AI traffic figure as a count of citations will expect it to track something it never can.

Each client's analytics lands in a warehouse you own beside their ad platforms, so a question that spans them can be answered at all. Channel groupings you define in SQL are yours, and they do not have to be rebuilt inside a reporting tool for every client each time you change them.

History matters here more than usual. Assistants and their referrer behaviour change month to month, and a warehouse that adds to its tables rather than replacing them lets you see a change in the data instead of losing the comparison.