When someone researches a purchase through an AI assistant and then visits your site directly, that journey appears in analytics as direct traffic with no context. The research happened, the recommendation happened, and none of it is visible in the reporting that decides your budget.
The practical response is prompt tracking. Assemble the thirty to fifty questions a real buyer in your category would ask an assistant, run them on a fixed schedule across the major systems, and record whether you are mentioned, in what position, and with what characterisation. It is manual, it is imperfect, and it is far better than nothing.
What that data reveals is usually uncomfortable. Most brands appear in a small fraction of relevant prompts, and when they do appear the description is often drawn from a review site or a directory rather than from their own pages. Both problems are addressable.
The traffic you cannot attribute is still traffic. Measuring it badly beats assuming it does not exist.
The second, softer signal is a rise in direct and branded search traffic that does not correspond to any campaign. When assistant citations increase, branded search reliably follows a few weeks later, because people hear a name and then search for it.
Treat these as directional rather than precise. The goal is not attribution to the decimal place; it is knowing whether the channel is growing and whether your work is affecting it.
